Merge branch 'mh/object-code-cleanup'
[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 "string-list.h"
11 #include "delta.h"
12 #include "pack.h"
13 #include "blob.h"
14 #include "commit.h"
15 #include "run-command.h"
16 #include "tag.h"
17 #include "tree.h"
18 #include "tree-walk.h"
19 #include "refs.h"
20 #include "pack-revindex.h"
21 #include "sha1-lookup.h"
22 #include "bulk-checkin.h"
23 #include "streaming.h"
24 #include "dir.h"
25
26 #ifndef O_NOATIME
27 #if defined(__linux__) && (defined(__i386__) || defined(__PPC__))
28 #define O_NOATIME 01000000
29 #else
30 #define O_NOATIME 0
31 #endif
32 #endif
33
34 #define SZ_FMT PRIuMAX
35 static inline uintmax_t sz_fmt(size_t s) { return s; }
36
37 const unsigned char null_sha1[20];
38
39 static const char *no_log_pack_access = "no_log_pack_access";
40 static const char *log_pack_access;
41
42 /*
43  * This is meant to hold a *small* number of objects that you would
44  * want read_sha1_file() to be able to return, but yet you do not want
45  * to write them into the object store (e.g. a browse-only
46  * application).
47  */
48 static struct cached_object {
49         unsigned char sha1[20];
50         enum object_type type;
51         void *buf;
52         unsigned long size;
53 } *cached_objects;
54 static int cached_object_nr, cached_object_alloc;
55
56 static struct cached_object empty_tree = {
57         EMPTY_TREE_SHA1_BIN_LITERAL,
58         OBJ_TREE,
59         "",
60         0
61 };
62
63 /*
64  * A pointer to the last packed_git in which an object was found.
65  * When an object is sought, we look in this packfile first, because
66  * objects that are looked up at similar times are often in the same
67  * packfile as one another.
68  */
69 static struct packed_git *last_found_pack;
70
71 static struct cached_object *find_cached_object(const unsigned char *sha1)
72 {
73         int i;
74         struct cached_object *co = cached_objects;
75
76         for (i = 0; i < cached_object_nr; i++, co++) {
77                 if (!hashcmp(co->sha1, sha1))
78                         return co;
79         }
80         if (!hashcmp(sha1, empty_tree.sha1))
81                 return &empty_tree;
82         return NULL;
83 }
84
85 int mkdir_in_gitdir(const char *path)
86 {
87         if (mkdir(path, 0777)) {
88                 int saved_errno = errno;
89                 struct stat st;
90                 struct strbuf sb = STRBUF_INIT;
91
92                 if (errno != EEXIST)
93                         return -1;
94                 /*
95                  * Are we looking at a path in a symlinked worktree
96                  * whose original repository does not yet have it?
97                  * e.g. .git/rr-cache pointing at its original
98                  * repository in which the user hasn't performed any
99                  * conflict resolution yet?
100                  */
101                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
102                     strbuf_readlink(&sb, path, st.st_size) ||
103                     !is_absolute_path(sb.buf) ||
104                     mkdir(sb.buf, 0777)) {
105                         strbuf_release(&sb);
106                         errno = saved_errno;
107                         return -1;
108                 }
109                 strbuf_release(&sb);
110         }
111         return adjust_shared_perm(path);
112 }
113
114 enum scld_error safe_create_leading_directories(char *path)
115 {
116         char *next_component = path + offset_1st_component(path);
117         enum scld_error ret = SCLD_OK;
118
119         while (ret == SCLD_OK && next_component) {
120                 struct stat st;
121                 char *slash = next_component, slash_character;
122
123                 while (*slash && !is_dir_sep(*slash))
124                         slash++;
125
126                 if (!*slash)
127                         break;
128
129                 next_component = slash + 1;
130                 while (is_dir_sep(*next_component))
131                         next_component++;
132                 if (!*next_component)
133                         break;
134
135                 slash_character = *slash;
136                 *slash = '\0';
137                 if (!stat(path, &st)) {
138                         /* path exists */
139                         if (!S_ISDIR(st.st_mode))
140                                 ret = SCLD_EXISTS;
141                 } else if (mkdir(path, 0777)) {
142                         if (errno == EEXIST &&
143                             !stat(path, &st) && S_ISDIR(st.st_mode))
144                                 ; /* somebody created it since we checked */
145                         else if (errno == ENOENT)
146                                 /*
147                                  * Either mkdir() failed because
148                                  * somebody just pruned the containing
149                                  * directory, or stat() failed because
150                                  * the file that was in our way was
151                                  * just removed.  Either way, inform
152                                  * the caller that it might be worth
153                                  * trying again:
154                                  */
155                                 ret = SCLD_VANISHED;
156                         else
157                                 ret = SCLD_FAILED;
158                 } else if (adjust_shared_perm(path)) {
159                         ret = SCLD_PERMS;
160                 }
161                 *slash = slash_character;
162         }
163         return ret;
164 }
165
166 enum scld_error safe_create_leading_directories_const(const char *path)
167 {
168         /* path points to cache entries, so xstrdup before messing with it */
169         char *buf = xstrdup(path);
170         enum scld_error result = safe_create_leading_directories(buf);
171         free(buf);
172         return result;
173 }
174
175 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
176 {
177         int i;
178         for (i = 0; i < 20; i++) {
179                 static char hex[] = "0123456789abcdef";
180                 unsigned int val = sha1[i];
181                 char *pos = pathbuf + i*2 + (i > 0);
182                 *pos++ = hex[val >> 4];
183                 *pos = hex[val & 0xf];
184         }
185 }
186
187 const char *sha1_file_name(const unsigned char *sha1)
188 {
189         static char buf[PATH_MAX];
190         const char *objdir;
191         int len;
192
193         objdir = get_object_directory();
194         len = strlen(objdir);
195
196         /* '/' + sha1(2) + '/' + sha1(38) + '\0' */
197         if (len + 43 > PATH_MAX)
198                 die("insanely long object directory %s", objdir);
199         memcpy(buf, objdir, len);
200         buf[len] = '/';
201         buf[len+3] = '/';
202         buf[len+42] = '\0';
203         fill_sha1_path(buf + len + 1, sha1);
204         return buf;
205 }
206
207 /*
208  * Return the name of the pack or index file with the specified sha1
209  * in its filename.  *base and *name are scratch space that must be
210  * provided by the caller.  which should be "pack" or "idx".
211  */
212 static char *sha1_get_pack_name(const unsigned char *sha1,
213                                 char **name, char **base, const char *which)
214 {
215         static const char hex[] = "0123456789abcdef";
216         char *buf;
217         int i;
218
219         if (!*base) {
220                 const char *sha1_file_directory = get_object_directory();
221                 int len = strlen(sha1_file_directory);
222                 *base = xmalloc(len + 60);
223                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
224                         sha1_file_directory, which);
225                 *name = *base + len + 11;
226         }
227
228         buf = *name;
229
230         for (i = 0; i < 20; i++) {
231                 unsigned int val = *sha1++;
232                 *buf++ = hex[val >> 4];
233                 *buf++ = hex[val & 0xf];
234         }
235
236         return *base;
237 }
238
239 char *sha1_pack_name(const unsigned char *sha1)
240 {
241         static char *name, *base;
242
243         return sha1_get_pack_name(sha1, &name, &base, "pack");
244 }
245
246 char *sha1_pack_index_name(const unsigned char *sha1)
247 {
248         static char *name, *base;
249
250         return sha1_get_pack_name(sha1, &name, &base, "idx");
251 }
252
253 struct alternate_object_database *alt_odb_list;
254 static struct alternate_object_database **alt_odb_tail;
255
256 /*
257  * Prepare alternate object database registry.
258  *
259  * The variable alt_odb_list points at the list of struct
260  * alternate_object_database.  The elements on this list come from
261  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
262  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
263  * whose contents is similar to that environment variable but can be
264  * LF separated.  Its base points at a statically allocated buffer that
265  * contains "/the/directory/corresponding/to/.git/objects/...", while
266  * its name points just after the slash at the end of ".git/objects/"
267  * in the example above, and has enough space to hold 40-byte hex
268  * SHA1, an extra slash for the first level indirection, and the
269  * terminating NUL.
270  */
271 static int link_alt_odb_entry(const char *entry, const char *relative_base, int depth)
272 {
273         const char *objdir = get_object_directory();
274         struct alternate_object_database *ent;
275         struct alternate_object_database *alt;
276         int pfxlen, entlen;
277         struct strbuf pathbuf = STRBUF_INIT;
278
279         if (!is_absolute_path(entry) && relative_base) {
280                 strbuf_addstr(&pathbuf, real_path(relative_base));
281                 strbuf_addch(&pathbuf, '/');
282         }
283         strbuf_addstr(&pathbuf, entry);
284
285         normalize_path_copy(pathbuf.buf, pathbuf.buf);
286
287         pfxlen = strlen(pathbuf.buf);
288
289         /*
290          * The trailing slash after the directory name is given by
291          * this function at the end. Remove duplicates.
292          */
293         while (pfxlen && pathbuf.buf[pfxlen-1] == '/')
294                 pfxlen -= 1;
295
296         entlen = pfxlen + 43; /* '/' + 2 hex + '/' + 38 hex + NUL */
297         ent = xmalloc(sizeof(*ent) + entlen);
298         memcpy(ent->base, pathbuf.buf, pfxlen);
299         strbuf_release(&pathbuf);
300
301         ent->name = ent->base + pfxlen + 1;
302         ent->base[pfxlen + 3] = '/';
303         ent->base[pfxlen] = ent->base[entlen-1] = 0;
304
305         /* Detect cases where alternate disappeared */
306         if (!is_directory(ent->base)) {
307                 error("object directory %s does not exist; "
308                       "check .git/objects/info/alternates.",
309                       ent->base);
310                 free(ent);
311                 return -1;
312         }
313
314         /* Prevent the common mistake of listing the same
315          * thing twice, or object directory itself.
316          */
317         for (alt = alt_odb_list; alt; alt = alt->next) {
318                 if (!memcmp(ent->base, alt->base, pfxlen)) {
319                         free(ent);
320                         return -1;
321                 }
322         }
323         if (!strcmp(ent->base, objdir)) {
324                 free(ent);
325                 return -1;
326         }
327
328         /* add the alternate entry */
329         *alt_odb_tail = ent;
330         alt_odb_tail = &(ent->next);
331         ent->next = NULL;
332
333         /* recursively add alternates */
334         read_info_alternates(ent->base, depth + 1);
335
336         ent->base[pfxlen] = '/';
337
338         return 0;
339 }
340
341 static void link_alt_odb_entries(const char *alt, int len, int sep,
342                                  const char *relative_base, int depth)
343 {
344         struct string_list entries = STRING_LIST_INIT_NODUP;
345         char *alt_copy;
346         int i;
347
348         if (depth > 5) {
349                 error("%s: ignoring alternate object stores, nesting too deep.",
350                                 relative_base);
351                 return;
352         }
353
354         alt_copy = xmemdupz(alt, len);
355         string_list_split_in_place(&entries, alt_copy, sep, -1);
356         for (i = 0; i < entries.nr; i++) {
357                 const char *entry = entries.items[i].string;
358                 if (entry[0] == '\0' || entry[0] == '#')
359                         continue;
360                 if (!is_absolute_path(entry) && depth) {
361                         error("%s: ignoring relative alternate object store %s",
362                                         relative_base, entry);
363                 } else {
364                         link_alt_odb_entry(entry, relative_base, depth);
365                 }
366         }
367         string_list_clear(&entries, 0);
368         free(alt_copy);
369 }
370
371 void read_info_alternates(const char * relative_base, int depth)
372 {
373         char *map;
374         size_t mapsz;
375         struct stat st;
376         const char alt_file_name[] = "info/alternates";
377         /* Given that relative_base is no longer than PATH_MAX,
378            ensure that "path" has enough space to append "/", the
379            file name, "info/alternates", and a trailing NUL.  */
380         char path[PATH_MAX + 1 + sizeof alt_file_name];
381         int fd;
382
383         sprintf(path, "%s/%s", relative_base, alt_file_name);
384         fd = git_open_noatime(path);
385         if (fd < 0)
386                 return;
387         if (fstat(fd, &st) || (st.st_size == 0)) {
388                 close(fd);
389                 return;
390         }
391         mapsz = xsize_t(st.st_size);
392         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
393         close(fd);
394
395         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
396
397         munmap(map, mapsz);
398 }
399
400 void add_to_alternates_file(const char *reference)
401 {
402         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
403         int fd = hold_lock_file_for_append(lock, git_path("objects/info/alternates"), LOCK_DIE_ON_ERROR);
404         char *alt = mkpath("%s\n", reference);
405         write_or_die(fd, alt, strlen(alt));
406         if (commit_lock_file(lock))
407                 die("could not close alternates file");
408         if (alt_odb_tail)
409                 link_alt_odb_entries(alt, strlen(alt), '\n', NULL, 0);
410 }
411
412 void foreach_alt_odb(alt_odb_fn fn, void *cb)
413 {
414         struct alternate_object_database *ent;
415
416         prepare_alt_odb();
417         for (ent = alt_odb_list; ent; ent = ent->next)
418                 if (fn(ent, cb))
419                         return;
420 }
421
422 void prepare_alt_odb(void)
423 {
424         const char *alt;
425
426         if (alt_odb_tail)
427                 return;
428
429         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
430         if (!alt) alt = "";
431
432         alt_odb_tail = &alt_odb_list;
433         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
434
435         read_info_alternates(get_object_directory(), 0);
436 }
437
438 static int has_loose_object_local(const unsigned char *sha1)
439 {
440         return !access(sha1_file_name(sha1), F_OK);
441 }
442
443 int has_loose_object_nonlocal(const unsigned char *sha1)
444 {
445         struct alternate_object_database *alt;
446         prepare_alt_odb();
447         for (alt = alt_odb_list; alt; alt = alt->next) {
448                 fill_sha1_path(alt->name, sha1);
449                 if (!access(alt->base, F_OK))
450                         return 1;
451         }
452         return 0;
453 }
454
455 static int has_loose_object(const unsigned char *sha1)
456 {
457         return has_loose_object_local(sha1) ||
458                has_loose_object_nonlocal(sha1);
459 }
460
461 static unsigned int pack_used_ctr;
462 static unsigned int pack_mmap_calls;
463 static unsigned int peak_pack_open_windows;
464 static unsigned int pack_open_windows;
465 static unsigned int pack_open_fds;
466 static unsigned int pack_max_fds;
467 static size_t peak_pack_mapped;
468 static size_t pack_mapped;
469 struct packed_git *packed_git;
470
471 void pack_report(void)
472 {
473         fprintf(stderr,
474                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
475                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
476                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
477                 sz_fmt(getpagesize()),
478                 sz_fmt(packed_git_window_size),
479                 sz_fmt(packed_git_limit));
480         fprintf(stderr,
481                 "pack_report: pack_used_ctr            = %10u\n"
482                 "pack_report: pack_mmap_calls          = %10u\n"
483                 "pack_report: pack_open_windows        = %10u / %10u\n"
484                 "pack_report: pack_mapped              = "
485                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
486                 pack_used_ctr,
487                 pack_mmap_calls,
488                 pack_open_windows, peak_pack_open_windows,
489                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
490 }
491
492 /*
493  * Open and mmap the index file at path, perform a couple of
494  * consistency checks, then record its information to p.  Return 0 on
495  * success.
496  */
497 static int check_packed_git_idx(const char *path, struct packed_git *p)
498 {
499         void *idx_map;
500         struct pack_idx_header *hdr;
501         size_t idx_size;
502         uint32_t version, nr, i, *index;
503         int fd = git_open_noatime(path);
504         struct stat st;
505
506         if (fd < 0)
507                 return -1;
508         if (fstat(fd, &st)) {
509                 close(fd);
510                 return -1;
511         }
512         idx_size = xsize_t(st.st_size);
513         if (idx_size < 4 * 256 + 20 + 20) {
514                 close(fd);
515                 return error("index file %s is too small", path);
516         }
517         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
518         close(fd);
519
520         hdr = idx_map;
521         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
522                 version = ntohl(hdr->idx_version);
523                 if (version < 2 || version > 2) {
524                         munmap(idx_map, idx_size);
525                         return error("index file %s is version %"PRIu32
526                                      " and is not supported by this binary"
527                                      " (try upgrading GIT to a newer version)",
528                                      path, version);
529                 }
530         } else
531                 version = 1;
532
533         nr = 0;
534         index = idx_map;
535         if (version > 1)
536                 index += 2;  /* skip index header */
537         for (i = 0; i < 256; i++) {
538                 uint32_t n = ntohl(index[i]);
539                 if (n < nr) {
540                         munmap(idx_map, idx_size);
541                         return error("non-monotonic index %s", path);
542                 }
543                 nr = n;
544         }
545
546         if (version == 1) {
547                 /*
548                  * Total size:
549                  *  - 256 index entries 4 bytes each
550                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
551                  *  - 20-byte SHA1 of the packfile
552                  *  - 20-byte SHA1 file checksum
553                  */
554                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
555                         munmap(idx_map, idx_size);
556                         return error("wrong index v1 file size in %s", path);
557                 }
558         } else if (version == 2) {
559                 /*
560                  * Minimum size:
561                  *  - 8 bytes of header
562                  *  - 256 index entries 4 bytes each
563                  *  - 20-byte sha1 entry * nr
564                  *  - 4-byte crc entry * nr
565                  *  - 4-byte offset entry * nr
566                  *  - 20-byte SHA1 of the packfile
567                  *  - 20-byte SHA1 file checksum
568                  * And after the 4-byte offset table might be a
569                  * variable sized table containing 8-byte entries
570                  * for offsets larger than 2^31.
571                  */
572                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
573                 unsigned long max_size = min_size;
574                 if (nr)
575                         max_size += (nr - 1)*8;
576                 if (idx_size < min_size || idx_size > max_size) {
577                         munmap(idx_map, idx_size);
578                         return error("wrong index v2 file size in %s", path);
579                 }
580                 if (idx_size != min_size &&
581                     /*
582                      * make sure we can deal with large pack offsets.
583                      * 31-bit signed offset won't be enough, neither
584                      * 32-bit unsigned one will be.
585                      */
586                     (sizeof(off_t) <= 4)) {
587                         munmap(idx_map, idx_size);
588                         return error("pack too large for current definition of off_t in %s", path);
589                 }
590         }
591
592         p->index_version = version;
593         p->index_data = idx_map;
594         p->index_size = idx_size;
595         p->num_objects = nr;
596         return 0;
597 }
598
599 int open_pack_index(struct packed_git *p)
600 {
601         char *idx_name;
602         int ret;
603
604         if (p->index_data)
605                 return 0;
606
607         idx_name = xstrdup(p->pack_name);
608         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
609         ret = check_packed_git_idx(idx_name, p);
610         free(idx_name);
611         return ret;
612 }
613
614 static void scan_windows(struct packed_git *p,
615         struct packed_git **lru_p,
616         struct pack_window **lru_w,
617         struct pack_window **lru_l)
618 {
619         struct pack_window *w, *w_l;
620
621         for (w_l = NULL, w = p->windows; w; w = w->next) {
622                 if (!w->inuse_cnt) {
623                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
624                                 *lru_p = p;
625                                 *lru_w = w;
626                                 *lru_l = w_l;
627                         }
628                 }
629                 w_l = w;
630         }
631 }
632
633 static int unuse_one_window(struct packed_git *current)
634 {
635         struct packed_git *p, *lru_p = NULL;
636         struct pack_window *lru_w = NULL, *lru_l = NULL;
637
638         if (current)
639                 scan_windows(current, &lru_p, &lru_w, &lru_l);
640         for (p = packed_git; p; p = p->next)
641                 scan_windows(p, &lru_p, &lru_w, &lru_l);
642         if (lru_p) {
643                 munmap(lru_w->base, lru_w->len);
644                 pack_mapped -= lru_w->len;
645                 if (lru_l)
646                         lru_l->next = lru_w->next;
647                 else
648                         lru_p->windows = lru_w->next;
649                 free(lru_w);
650                 pack_open_windows--;
651                 return 1;
652         }
653         return 0;
654 }
655
656 void release_pack_memory(size_t need)
657 {
658         size_t cur = pack_mapped;
659         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
660                 ; /* nothing */
661 }
662
663 void *xmmap(void *start, size_t length,
664         int prot, int flags, int fd, off_t offset)
665 {
666         void *ret = mmap(start, length, prot, flags, fd, offset);
667         if (ret == MAP_FAILED) {
668                 if (!length)
669                         return NULL;
670                 release_pack_memory(length);
671                 ret = mmap(start, length, prot, flags, fd, offset);
672                 if (ret == MAP_FAILED)
673                         die_errno("Out of memory? mmap failed");
674         }
675         return ret;
676 }
677
678 void close_pack_windows(struct packed_git *p)
679 {
680         while (p->windows) {
681                 struct pack_window *w = p->windows;
682
683                 if (w->inuse_cnt)
684                         die("pack '%s' still has open windows to it",
685                             p->pack_name);
686                 munmap(w->base, w->len);
687                 pack_mapped -= w->len;
688                 pack_open_windows--;
689                 p->windows = w->next;
690                 free(w);
691         }
692 }
693
694 /*
695  * The LRU pack is the one with the oldest MRU window, preferring packs
696  * with no used windows, or the oldest mtime if it has no windows allocated.
697  */
698 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
699 {
700         struct pack_window *w, *this_mru_w;
701         int has_windows_inuse = 0;
702
703         /*
704          * Reject this pack if it has windows and the previously selected
705          * one does not.  If this pack does not have windows, reject
706          * it if the pack file is newer than the previously selected one.
707          */
708         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
709                 return;
710
711         for (w = this_mru_w = p->windows; w; w = w->next) {
712                 /*
713                  * Reject this pack if any of its windows are in use,
714                  * but the previously selected pack did not have any
715                  * inuse windows.  Otherwise, record that this pack
716                  * has windows in use.
717                  */
718                 if (w->inuse_cnt) {
719                         if (*accept_windows_inuse)
720                                 has_windows_inuse = 1;
721                         else
722                                 return;
723                 }
724
725                 if (w->last_used > this_mru_w->last_used)
726                         this_mru_w = w;
727
728                 /*
729                  * Reject this pack if it has windows that have been
730                  * used more recently than the previously selected pack.
731                  * If the previously selected pack had windows inuse and
732                  * we have not encountered a window in this pack that is
733                  * inuse, skip this check since we prefer a pack with no
734                  * inuse windows to one that has inuse windows.
735                  */
736                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
737                     this_mru_w->last_used > (*mru_w)->last_used)
738                         return;
739         }
740
741         /*
742          * Select this pack.
743          */
744         *mru_w = this_mru_w;
745         *lru_p = p;
746         *accept_windows_inuse = has_windows_inuse;
747 }
748
749 static int close_one_pack(void)
750 {
751         struct packed_git *p, *lru_p = NULL;
752         struct pack_window *mru_w = NULL;
753         int accept_windows_inuse = 1;
754
755         for (p = packed_git; p; p = p->next) {
756                 if (p->pack_fd == -1)
757                         continue;
758                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
759         }
760
761         if (lru_p) {
762                 close(lru_p->pack_fd);
763                 pack_open_fds--;
764                 lru_p->pack_fd = -1;
765                 return 1;
766         }
767
768         return 0;
769 }
770
771 void unuse_pack(struct pack_window **w_cursor)
772 {
773         struct pack_window *w = *w_cursor;
774         if (w) {
775                 w->inuse_cnt--;
776                 *w_cursor = NULL;
777         }
778 }
779
780 void close_pack_index(struct packed_git *p)
781 {
782         if (p->index_data) {
783                 munmap((void *)p->index_data, p->index_size);
784                 p->index_data = NULL;
785         }
786 }
787
788 /*
789  * This is used by git-repack in case a newly created pack happens to
790  * contain the same set of objects as an existing one.  In that case
791  * the resulting file might be different even if its name would be the
792  * same.  It is best to close any reference to the old pack before it is
793  * replaced on disk.  Of course no index pointers nor windows for given pack
794  * must subsist at this point.  If ever objects from this pack are requested
795  * again, the new version of the pack will be reinitialized through
796  * reprepare_packed_git().
797  */
798 void free_pack_by_name(const char *pack_name)
799 {
800         struct packed_git *p, **pp = &packed_git;
801
802         while (*pp) {
803                 p = *pp;
804                 if (strcmp(pack_name, p->pack_name) == 0) {
805                         clear_delta_base_cache();
806                         close_pack_windows(p);
807                         if (p->pack_fd != -1) {
808                                 close(p->pack_fd);
809                                 pack_open_fds--;
810                         }
811                         close_pack_index(p);
812                         free(p->bad_object_sha1);
813                         *pp = p->next;
814                         if (last_found_pack == p)
815                                 last_found_pack = NULL;
816                         free(p);
817                         return;
818                 }
819                 pp = &p->next;
820         }
821 }
822
823 static unsigned int get_max_fd_limit(void)
824 {
825 #ifdef RLIMIT_NOFILE
826         {
827                 struct rlimit lim;
828
829                 if (!getrlimit(RLIMIT_NOFILE, &lim))
830                         return lim.rlim_cur;
831         }
832 #endif
833
834 #ifdef _SC_OPEN_MAX
835         {
836                 long open_max = sysconf(_SC_OPEN_MAX);
837                 if (0 < open_max)
838                         return open_max;
839                 /*
840                  * Otherwise, we got -1 for one of the two
841                  * reasons:
842                  *
843                  * (1) sysconf() did not understand _SC_OPEN_MAX
844                  *     and signaled an error with -1; or
845                  * (2) sysconf() said there is no limit.
846                  *
847                  * We _could_ clear errno before calling sysconf() to
848                  * tell these two cases apart and return a huge number
849                  * in the latter case to let the caller cap it to a
850                  * value that is not so selfish, but letting the
851                  * fallback OPEN_MAX codepath take care of these cases
852                  * is a lot simpler.
853                  */
854         }
855 #endif
856
857 #ifdef OPEN_MAX
858         return OPEN_MAX;
859 #else
860         return 1; /* see the caller ;-) */
861 #endif
862 }
863
864 /*
865  * Do not call this directly as this leaks p->pack_fd on error return;
866  * call open_packed_git() instead.
867  */
868 static int open_packed_git_1(struct packed_git *p)
869 {
870         struct stat st;
871         struct pack_header hdr;
872         unsigned char sha1[20];
873         unsigned char *idx_sha1;
874         long fd_flag;
875
876         if (!p->index_data && open_pack_index(p))
877                 return error("packfile %s index unavailable", p->pack_name);
878
879         if (!pack_max_fds) {
880                 unsigned int max_fds = get_max_fd_limit();
881
882                 /* Save 3 for stdin/stdout/stderr, 22 for work */
883                 if (25 < max_fds)
884                         pack_max_fds = max_fds - 25;
885                 else
886                         pack_max_fds = 1;
887         }
888
889         while (pack_max_fds <= pack_open_fds && close_one_pack())
890                 ; /* nothing */
891
892         p->pack_fd = git_open_noatime(p->pack_name);
893         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
894                 return -1;
895         pack_open_fds++;
896
897         /* If we created the struct before we had the pack we lack size. */
898         if (!p->pack_size) {
899                 if (!S_ISREG(st.st_mode))
900                         return error("packfile %s not a regular file", p->pack_name);
901                 p->pack_size = st.st_size;
902         } else if (p->pack_size != st.st_size)
903                 return error("packfile %s size changed", p->pack_name);
904
905         /* We leave these file descriptors open with sliding mmap;
906          * there is no point keeping them open across exec(), though.
907          */
908         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
909         if (fd_flag < 0)
910                 return error("cannot determine file descriptor flags");
911         fd_flag |= FD_CLOEXEC;
912         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
913                 return error("cannot set FD_CLOEXEC");
914
915         /* Verify we recognize this pack file format. */
916         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
917                 return error("file %s is far too short to be a packfile", p->pack_name);
918         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
919                 return error("file %s is not a GIT packfile", p->pack_name);
920         if (!pack_version_ok(hdr.hdr_version))
921                 return error("packfile %s is version %"PRIu32" and not"
922                         " supported (try upgrading GIT to a newer version)",
923                         p->pack_name, ntohl(hdr.hdr_version));
924
925         /* Verify the pack matches its index. */
926         if (p->num_objects != ntohl(hdr.hdr_entries))
927                 return error("packfile %s claims to have %"PRIu32" objects"
928                              " while index indicates %"PRIu32" objects",
929                              p->pack_name, ntohl(hdr.hdr_entries),
930                              p->num_objects);
931         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
932                 return error("end of packfile %s is unavailable", p->pack_name);
933         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
934                 return error("packfile %s signature is unavailable", p->pack_name);
935         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
936         if (hashcmp(sha1, idx_sha1))
937                 return error("packfile %s does not match index", p->pack_name);
938         return 0;
939 }
940
941 static int open_packed_git(struct packed_git *p)
942 {
943         if (!open_packed_git_1(p))
944                 return 0;
945         if (p->pack_fd != -1) {
946                 close(p->pack_fd);
947                 pack_open_fds--;
948                 p->pack_fd = -1;
949         }
950         return -1;
951 }
952
953 static int in_window(struct pack_window *win, off_t offset)
954 {
955         /* We must promise at least 20 bytes (one hash) after the
956          * offset is available from this window, otherwise the offset
957          * is not actually in this window and a different window (which
958          * has that one hash excess) must be used.  This is to support
959          * the object header and delta base parsing routines below.
960          */
961         off_t win_off = win->offset;
962         return win_off <= offset
963                 && (offset + 20) <= (win_off + win->len);
964 }
965
966 unsigned char *use_pack(struct packed_git *p,
967                 struct pack_window **w_cursor,
968                 off_t offset,
969                 unsigned long *left)
970 {
971         struct pack_window *win = *w_cursor;
972
973         /* Since packfiles end in a hash of their content and it's
974          * pointless to ask for an offset into the middle of that
975          * hash, and the in_window function above wouldn't match
976          * don't allow an offset too close to the end of the file.
977          */
978         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
979                 die("packfile %s cannot be accessed", p->pack_name);
980         if (offset > (p->pack_size - 20))
981                 die("offset beyond end of packfile (truncated pack?)");
982
983         if (!win || !in_window(win, offset)) {
984                 if (win)
985                         win->inuse_cnt--;
986                 for (win = p->windows; win; win = win->next) {
987                         if (in_window(win, offset))
988                                 break;
989                 }
990                 if (!win) {
991                         size_t window_align = packed_git_window_size / 2;
992                         off_t len;
993
994                         if (p->pack_fd == -1 && open_packed_git(p))
995                                 die("packfile %s cannot be accessed", p->pack_name);
996
997                         win = xcalloc(1, sizeof(*win));
998                         win->offset = (offset / window_align) * window_align;
999                         len = p->pack_size - win->offset;
1000                         if (len > packed_git_window_size)
1001                                 len = packed_git_window_size;
1002                         win->len = (size_t)len;
1003                         pack_mapped += win->len;
1004                         while (packed_git_limit < pack_mapped
1005                                 && unuse_one_window(p))
1006                                 ; /* nothing */
1007                         win->base = xmmap(NULL, win->len,
1008                                 PROT_READ, MAP_PRIVATE,
1009                                 p->pack_fd, win->offset);
1010                         if (win->base == MAP_FAILED)
1011                                 die("packfile %s cannot be mapped: %s",
1012                                         p->pack_name,
1013                                         strerror(errno));
1014                         if (!win->offset && win->len == p->pack_size
1015                                 && !p->do_not_close) {
1016                                 close(p->pack_fd);
1017                                 pack_open_fds--;
1018                                 p->pack_fd = -1;
1019                         }
1020                         pack_mmap_calls++;
1021                         pack_open_windows++;
1022                         if (pack_mapped > peak_pack_mapped)
1023                                 peak_pack_mapped = pack_mapped;
1024                         if (pack_open_windows > peak_pack_open_windows)
1025                                 peak_pack_open_windows = pack_open_windows;
1026                         win->next = p->windows;
1027                         p->windows = win;
1028                 }
1029         }
1030         if (win != *w_cursor) {
1031                 win->last_used = pack_used_ctr++;
1032                 win->inuse_cnt++;
1033                 *w_cursor = win;
1034         }
1035         offset -= win->offset;
1036         if (left)
1037                 *left = win->len - xsize_t(offset);
1038         return win->base + offset;
1039 }
1040
1041 static struct packed_git *alloc_packed_git(int extra)
1042 {
1043         struct packed_git *p = xmalloc(sizeof(*p) + extra);
1044         memset(p, 0, sizeof(*p));
1045         p->pack_fd = -1;
1046         return p;
1047 }
1048
1049 static void try_to_free_pack_memory(size_t size)
1050 {
1051         release_pack_memory(size);
1052 }
1053
1054 struct packed_git *add_packed_git(const char *path, int path_len, int local)
1055 {
1056         static int have_set_try_to_free_routine;
1057         struct stat st;
1058         struct packed_git *p = alloc_packed_git(path_len + 2);
1059
1060         if (!have_set_try_to_free_routine) {
1061                 have_set_try_to_free_routine = 1;
1062                 set_try_to_free_routine(try_to_free_pack_memory);
1063         }
1064
1065         /*
1066          * Make sure a corresponding .pack file exists and that
1067          * the index looks sane.
1068          */
1069         path_len -= strlen(".idx");
1070         if (path_len < 1) {
1071                 free(p);
1072                 return NULL;
1073         }
1074         memcpy(p->pack_name, path, path_len);
1075
1076         strcpy(p->pack_name + path_len, ".keep");
1077         if (!access(p->pack_name, F_OK))
1078                 p->pack_keep = 1;
1079
1080         strcpy(p->pack_name + path_len, ".pack");
1081         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1082                 free(p);
1083                 return NULL;
1084         }
1085
1086         /* ok, it looks sane as far as we can check without
1087          * actually mapping the pack file.
1088          */
1089         p->pack_size = st.st_size;
1090         p->pack_local = local;
1091         p->mtime = st.st_mtime;
1092         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1093                 hashclr(p->sha1);
1094         return p;
1095 }
1096
1097 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1098 {
1099         const char *path = sha1_pack_name(sha1);
1100         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
1101
1102         strcpy(p->pack_name, path);
1103         hashcpy(p->sha1, sha1);
1104         if (check_packed_git_idx(idx_path, p)) {
1105                 free(p);
1106                 return NULL;
1107         }
1108
1109         return p;
1110 }
1111
1112 void install_packed_git(struct packed_git *pack)
1113 {
1114         if (pack->pack_fd != -1)
1115                 pack_open_fds++;
1116
1117         pack->next = packed_git;
1118         packed_git = pack;
1119 }
1120
1121 void (*report_garbage)(const char *desc, const char *path);
1122
1123 static void report_helper(const struct string_list *list,
1124                           int seen_bits, int first, int last)
1125 {
1126         const char *msg;
1127         switch (seen_bits) {
1128         case 0:
1129                 msg = "no corresponding .idx nor .pack";
1130                 break;
1131         case 1:
1132                 msg = "no corresponding .idx";
1133                 break;
1134         case 2:
1135                 msg = "no corresponding .pack";
1136                 break;
1137         default:
1138                 return;
1139         }
1140         for (; first < last; first++)
1141                 report_garbage(msg, list->items[first].string);
1142 }
1143
1144 static void report_pack_garbage(struct string_list *list)
1145 {
1146         int i, baselen = -1, first = 0, seen_bits = 0;
1147
1148         if (!report_garbage)
1149                 return;
1150
1151         sort_string_list(list);
1152
1153         for (i = 0; i < list->nr; i++) {
1154                 const char *path = list->items[i].string;
1155                 if (baselen != -1 &&
1156                     strncmp(path, list->items[first].string, baselen)) {
1157                         report_helper(list, seen_bits, first, i);
1158                         baselen = -1;
1159                         seen_bits = 0;
1160                 }
1161                 if (baselen == -1) {
1162                         const char *dot = strrchr(path, '.');
1163                         if (!dot) {
1164                                 report_garbage("garbage found", path);
1165                                 continue;
1166                         }
1167                         baselen = dot - path + 1;
1168                         first = i;
1169                 }
1170                 if (!strcmp(path + baselen, "pack"))
1171                         seen_bits |= 1;
1172                 else if (!strcmp(path + baselen, "idx"))
1173                         seen_bits |= 2;
1174         }
1175         report_helper(list, seen_bits, first, list->nr);
1176 }
1177
1178 static void prepare_packed_git_one(char *objdir, int local)
1179 {
1180         /* Ensure that this buffer is large enough so that we can
1181            append "/pack/" without clobbering the stack even if
1182            strlen(objdir) were PATH_MAX.  */
1183         char path[PATH_MAX + 1 + 4 + 1 + 1];
1184         int len;
1185         DIR *dir;
1186         struct dirent *de;
1187         struct string_list garbage = STRING_LIST_INIT_DUP;
1188
1189         sprintf(path, "%s/pack", objdir);
1190         len = strlen(path);
1191         dir = opendir(path);
1192         if (!dir) {
1193                 if (errno != ENOENT)
1194                         error("unable to open object pack directory: %s: %s",
1195                               path, strerror(errno));
1196                 return;
1197         }
1198         path[len++] = '/';
1199         while ((de = readdir(dir)) != NULL) {
1200                 int namelen = strlen(de->d_name);
1201                 struct packed_git *p;
1202
1203                 if (len + namelen + 1 > sizeof(path)) {
1204                         if (report_garbage) {
1205                                 struct strbuf sb = STRBUF_INIT;
1206                                 strbuf_addf(&sb, "%.*s/%s", len - 1, path, de->d_name);
1207                                 report_garbage("path too long", sb.buf);
1208                                 strbuf_release(&sb);
1209                         }
1210                         continue;
1211                 }
1212
1213                 if (is_dot_or_dotdot(de->d_name))
1214                         continue;
1215
1216                 strcpy(path + len, de->d_name);
1217
1218                 if (has_extension(de->d_name, ".idx")) {
1219                         /* Don't reopen a pack we already have. */
1220                         for (p = packed_git; p; p = p->next) {
1221                                 if (!memcmp(path, p->pack_name, len + namelen - 4))
1222                                         break;
1223                         }
1224                         if (p == NULL &&
1225                             /*
1226                              * See if it really is a valid .idx file with
1227                              * corresponding .pack file that we can map.
1228                              */
1229                             (p = add_packed_git(path, len + namelen, local)) != NULL)
1230                                 install_packed_git(p);
1231                 }
1232
1233                 if (!report_garbage)
1234                         continue;
1235
1236                 if (has_extension(de->d_name, ".idx") ||
1237                     has_extension(de->d_name, ".pack") ||
1238                     has_extension(de->d_name, ".bitmap") ||
1239                     has_extension(de->d_name, ".keep"))
1240                         string_list_append(&garbage, path);
1241                 else
1242                         report_garbage("garbage found", path);
1243         }
1244         closedir(dir);
1245         report_pack_garbage(&garbage);
1246         string_list_clear(&garbage, 0);
1247 }
1248
1249 static int sort_pack(const void *a_, const void *b_)
1250 {
1251         struct packed_git *a = *((struct packed_git **)a_);
1252         struct packed_git *b = *((struct packed_git **)b_);
1253         int st;
1254
1255         /*
1256          * Local packs tend to contain objects specific to our
1257          * variant of the project than remote ones.  In addition,
1258          * remote ones could be on a network mounted filesystem.
1259          * Favor local ones for these reasons.
1260          */
1261         st = a->pack_local - b->pack_local;
1262         if (st)
1263                 return -st;
1264
1265         /*
1266          * Younger packs tend to contain more recent objects,
1267          * and more recent objects tend to get accessed more
1268          * often.
1269          */
1270         if (a->mtime < b->mtime)
1271                 return 1;
1272         else if (a->mtime == b->mtime)
1273                 return 0;
1274         return -1;
1275 }
1276
1277 static void rearrange_packed_git(void)
1278 {
1279         struct packed_git **ary, *p;
1280         int i, n;
1281
1282         for (n = 0, p = packed_git; p; p = p->next)
1283                 n++;
1284         if (n < 2)
1285                 return;
1286
1287         /* prepare an array of packed_git for easier sorting */
1288         ary = xcalloc(n, sizeof(struct packed_git *));
1289         for (n = 0, p = packed_git; p; p = p->next)
1290                 ary[n++] = p;
1291
1292         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
1293
1294         /* link them back again */
1295         for (i = 0; i < n - 1; i++)
1296                 ary[i]->next = ary[i + 1];
1297         ary[n - 1]->next = NULL;
1298         packed_git = ary[0];
1299
1300         free(ary);
1301 }
1302
1303 static int prepare_packed_git_run_once = 0;
1304 void prepare_packed_git(void)
1305 {
1306         struct alternate_object_database *alt;
1307
1308         if (prepare_packed_git_run_once)
1309                 return;
1310         prepare_packed_git_one(get_object_directory(), 1);
1311         prepare_alt_odb();
1312         for (alt = alt_odb_list; alt; alt = alt->next) {
1313                 alt->name[-1] = 0;
1314                 prepare_packed_git_one(alt->base, 0);
1315                 alt->name[-1] = '/';
1316         }
1317         rearrange_packed_git();
1318         prepare_packed_git_run_once = 1;
1319 }
1320
1321 void reprepare_packed_git(void)
1322 {
1323         prepare_packed_git_run_once = 0;
1324         prepare_packed_git();
1325 }
1326
1327 static void mark_bad_packed_object(struct packed_git *p,
1328                                    const unsigned char *sha1)
1329 {
1330         unsigned i;
1331         for (i = 0; i < p->num_bad_objects; i++)
1332                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1333                         return;
1334         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1335         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1336         p->num_bad_objects++;
1337 }
1338
1339 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1340 {
1341         struct packed_git *p;
1342         unsigned i;
1343
1344         for (p = packed_git; p; p = p->next)
1345                 for (i = 0; i < p->num_bad_objects; i++)
1346                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1347                                 return p;
1348         return NULL;
1349 }
1350
1351 /*
1352  * With an in-core object data in "map", rehash it to make sure the
1353  * object name actually matches "sha1" to detect object corruption.
1354  * With "map" == NULL, try reading the object named with "sha1" using
1355  * the streaming interface and rehash it to do the same.
1356  */
1357 int check_sha1_signature(const unsigned char *sha1, void *map,
1358                          unsigned long size, const char *type)
1359 {
1360         unsigned char real_sha1[20];
1361         enum object_type obj_type;
1362         struct git_istream *st;
1363         git_SHA_CTX c;
1364         char hdr[32];
1365         int hdrlen;
1366
1367         if (map) {
1368                 hash_sha1_file(map, size, type, real_sha1);
1369                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1370         }
1371
1372         st = open_istream(sha1, &obj_type, &size, NULL);
1373         if (!st)
1374                 return -1;
1375
1376         /* Generate the header */
1377         hdrlen = sprintf(hdr, "%s %lu", typename(obj_type), size) + 1;
1378
1379         /* Sha1.. */
1380         git_SHA1_Init(&c);
1381         git_SHA1_Update(&c, hdr, hdrlen);
1382         for (;;) {
1383                 char buf[1024 * 16];
1384                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1385
1386                 if (readlen < 0) {
1387                         close_istream(st);
1388                         return -1;
1389                 }
1390                 if (!readlen)
1391                         break;
1392                 git_SHA1_Update(&c, buf, readlen);
1393         }
1394         git_SHA1_Final(real_sha1, &c);
1395         close_istream(st);
1396         return hashcmp(sha1, real_sha1) ? -1 : 0;
1397 }
1398
1399 int git_open_noatime(const char *name)
1400 {
1401         static int sha1_file_open_flag = O_NOATIME;
1402
1403         for (;;) {
1404                 int fd = open(name, O_RDONLY | sha1_file_open_flag);
1405                 if (fd >= 0)
1406                         return fd;
1407
1408                 /* Might the failure be due to O_NOATIME? */
1409                 if (errno != ENOENT && sha1_file_open_flag) {
1410                         sha1_file_open_flag = 0;
1411                         continue;
1412                 }
1413
1414                 return -1;
1415         }
1416 }
1417
1418 static int stat_sha1_file(const unsigned char *sha1, struct stat *st)
1419 {
1420         struct alternate_object_database *alt;
1421
1422         if (!lstat(sha1_file_name(sha1), st))
1423                 return 0;
1424
1425         prepare_alt_odb();
1426         errno = ENOENT;
1427         for (alt = alt_odb_list; alt; alt = alt->next) {
1428                 fill_sha1_path(alt->name, sha1);
1429                 if (!lstat(alt->base, st))
1430                         return 0;
1431         }
1432
1433         return -1;
1434 }
1435
1436 static int open_sha1_file(const unsigned char *sha1)
1437 {
1438         int fd;
1439         struct alternate_object_database *alt;
1440
1441         fd = git_open_noatime(sha1_file_name(sha1));
1442         if (fd >= 0)
1443                 return fd;
1444
1445         prepare_alt_odb();
1446         errno = ENOENT;
1447         for (alt = alt_odb_list; alt; alt = alt->next) {
1448                 fill_sha1_path(alt->name, sha1);
1449                 fd = git_open_noatime(alt->base);
1450                 if (fd >= 0)
1451                         return fd;
1452         }
1453         return -1;
1454 }
1455
1456 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1457 {
1458         void *map;
1459         int fd;
1460
1461         fd = open_sha1_file(sha1);
1462         map = NULL;
1463         if (fd >= 0) {
1464                 struct stat st;
1465
1466                 if (!fstat(fd, &st)) {
1467                         *size = xsize_t(st.st_size);
1468                         if (!*size) {
1469                                 /* mmap() is forbidden on empty files */
1470                                 error("object file %s is empty", sha1_file_name(sha1));
1471                                 return NULL;
1472                         }
1473                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1474                 }
1475                 close(fd);
1476         }
1477         return map;
1478 }
1479
1480 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1481                 unsigned long len, enum object_type *type, unsigned long *sizep)
1482 {
1483         unsigned shift;
1484         unsigned long size, c;
1485         unsigned long used = 0;
1486
1487         c = buf[used++];
1488         *type = (c >> 4) & 7;
1489         size = c & 15;
1490         shift = 4;
1491         while (c & 0x80) {
1492                 if (len <= used || bitsizeof(long) <= shift) {
1493                         error("bad object header");
1494                         size = used = 0;
1495                         break;
1496                 }
1497                 c = buf[used++];
1498                 size += (c & 0x7f) << shift;
1499                 shift += 7;
1500         }
1501         *sizep = size;
1502         return used;
1503 }
1504
1505 int unpack_sha1_header(git_zstream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1506 {
1507         /* Get the data stream */
1508         memset(stream, 0, sizeof(*stream));
1509         stream->next_in = map;
1510         stream->avail_in = mapsize;
1511         stream->next_out = buffer;
1512         stream->avail_out = bufsiz;
1513
1514         git_inflate_init(stream);
1515         return git_inflate(stream, 0);
1516 }
1517
1518 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1519 {
1520         int bytes = strlen(buffer) + 1;
1521         unsigned char *buf = xmallocz(size);
1522         unsigned long n;
1523         int status = Z_OK;
1524
1525         n = stream->total_out - bytes;
1526         if (n > size)
1527                 n = size;
1528         memcpy(buf, (char *) buffer + bytes, n);
1529         bytes = n;
1530         if (bytes <= size) {
1531                 /*
1532                  * The above condition must be (bytes <= size), not
1533                  * (bytes < size).  In other words, even though we
1534                  * expect no more output and set avail_out to zero,
1535                  * the input zlib stream may have bytes that express
1536                  * "this concludes the stream", and we *do* want to
1537                  * eat that input.
1538                  *
1539                  * Otherwise we would not be able to test that we
1540                  * consumed all the input to reach the expected size;
1541                  * we also want to check that zlib tells us that all
1542                  * went well with status == Z_STREAM_END at the end.
1543                  */
1544                 stream->next_out = buf + bytes;
1545                 stream->avail_out = size - bytes;
1546                 while (status == Z_OK)
1547                         status = git_inflate(stream, Z_FINISH);
1548         }
1549         if (status == Z_STREAM_END && !stream->avail_in) {
1550                 git_inflate_end(stream);
1551                 return buf;
1552         }
1553
1554         if (status < 0)
1555                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1556         else if (stream->avail_in)
1557                 error("garbage at end of loose object '%s'",
1558                       sha1_to_hex(sha1));
1559         free(buf);
1560         return NULL;
1561 }
1562
1563 /*
1564  * We used to just use "sscanf()", but that's actually way
1565  * too permissive for what we want to check. So do an anal
1566  * object header parse by hand.
1567  */
1568 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1569 {
1570         char type[10];
1571         int i;
1572         unsigned long size;
1573
1574         /*
1575          * The type can be at most ten bytes (including the
1576          * terminating '\0' that we add), and is followed by
1577          * a space.
1578          */
1579         i = 0;
1580         for (;;) {
1581                 char c = *hdr++;
1582                 if (c == ' ')
1583                         break;
1584                 type[i++] = c;
1585                 if (i >= sizeof(type))
1586                         return -1;
1587         }
1588         type[i] = 0;
1589
1590         /*
1591          * The length must follow immediately, and be in canonical
1592          * decimal format (ie "010" is not valid).
1593          */
1594         size = *hdr++ - '0';
1595         if (size > 9)
1596                 return -1;
1597         if (size) {
1598                 for (;;) {
1599                         unsigned long c = *hdr - '0';
1600                         if (c > 9)
1601                                 break;
1602                         hdr++;
1603                         size = size * 10 + c;
1604                 }
1605         }
1606         *sizep = size;
1607
1608         /*
1609          * The length must be followed by a zero byte
1610          */
1611         return *hdr ? -1 : type_from_string(type);
1612 }
1613
1614 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1615 {
1616         int ret;
1617         git_zstream stream;
1618         char hdr[8192];
1619
1620         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1621         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1622                 return NULL;
1623
1624         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1625 }
1626
1627 unsigned long get_size_from_delta(struct packed_git *p,
1628                                   struct pack_window **w_curs,
1629                                   off_t curpos)
1630 {
1631         const unsigned char *data;
1632         unsigned char delta_head[20], *in;
1633         git_zstream stream;
1634         int st;
1635
1636         memset(&stream, 0, sizeof(stream));
1637         stream.next_out = delta_head;
1638         stream.avail_out = sizeof(delta_head);
1639
1640         git_inflate_init(&stream);
1641         do {
1642                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1643                 stream.next_in = in;
1644                 st = git_inflate(&stream, Z_FINISH);
1645                 curpos += stream.next_in - in;
1646         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1647                  stream.total_out < sizeof(delta_head));
1648         git_inflate_end(&stream);
1649         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1650                 error("delta data unpack-initial failed");
1651                 return 0;
1652         }
1653
1654         /* Examine the initial part of the delta to figure out
1655          * the result size.
1656          */
1657         data = delta_head;
1658
1659         /* ignore base size */
1660         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1661
1662         /* Read the result size */
1663         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1664 }
1665
1666 static off_t get_delta_base(struct packed_git *p,
1667                                     struct pack_window **w_curs,
1668                                     off_t *curpos,
1669                                     enum object_type type,
1670                                     off_t delta_obj_offset)
1671 {
1672         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1673         off_t base_offset;
1674
1675         /* use_pack() assured us we have [base_info, base_info + 20)
1676          * as a range that we can look at without walking off the
1677          * end of the mapped window.  Its actually the hash size
1678          * that is assured.  An OFS_DELTA longer than the hash size
1679          * is stupid, as then a REF_DELTA would be smaller to store.
1680          */
1681         if (type == OBJ_OFS_DELTA) {
1682                 unsigned used = 0;
1683                 unsigned char c = base_info[used++];
1684                 base_offset = c & 127;
1685                 while (c & 128) {
1686                         base_offset += 1;
1687                         if (!base_offset || MSB(base_offset, 7))
1688                                 return 0;  /* overflow */
1689                         c = base_info[used++];
1690                         base_offset = (base_offset << 7) + (c & 127);
1691                 }
1692                 base_offset = delta_obj_offset - base_offset;
1693                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1694                         return 0;  /* out of bound */
1695                 *curpos += used;
1696         } else if (type == OBJ_REF_DELTA) {
1697                 /* The base entry _must_ be in the same pack */
1698                 base_offset = find_pack_entry_one(base_info, p);
1699                 *curpos += 20;
1700         } else
1701                 die("I am totally screwed");
1702         return base_offset;
1703 }
1704
1705 /*
1706  * Like get_delta_base above, but we return the sha1 instead of the pack
1707  * offset. This means it is cheaper for REF deltas (we do not have to do
1708  * the final object lookup), but more expensive for OFS deltas (we
1709  * have to load the revidx to convert the offset back into a sha1).
1710  */
1711 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1712                                                 struct pack_window **w_curs,
1713                                                 off_t curpos,
1714                                                 enum object_type type,
1715                                                 off_t delta_obj_offset)
1716 {
1717         if (type == OBJ_REF_DELTA) {
1718                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1719                 return base;
1720         } else if (type == OBJ_OFS_DELTA) {
1721                 struct revindex_entry *revidx;
1722                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1723                                                    type, delta_obj_offset);
1724
1725                 if (!base_offset)
1726                         return NULL;
1727
1728                 revidx = find_pack_revindex(p, base_offset);
1729                 if (!revidx)
1730                         return NULL;
1731
1732                 return nth_packed_object_sha1(p, revidx->nr);
1733         } else
1734                 return NULL;
1735 }
1736
1737 int unpack_object_header(struct packed_git *p,
1738                          struct pack_window **w_curs,
1739                          off_t *curpos,
1740                          unsigned long *sizep)
1741 {
1742         unsigned char *base;
1743         unsigned long left;
1744         unsigned long used;
1745         enum object_type type;
1746
1747         /* use_pack() assures us we have [base, base + 20) available
1748          * as a range that we can look at.  (Its actually the hash
1749          * size that is assured.)  With our object header encoding
1750          * the maximum deflated object size is 2^137, which is just
1751          * insane, so we know won't exceed what we have been given.
1752          */
1753         base = use_pack(p, w_curs, *curpos, &left);
1754         used = unpack_object_header_buffer(base, left, &type, sizep);
1755         if (!used) {
1756                 type = OBJ_BAD;
1757         } else
1758                 *curpos += used;
1759
1760         return type;
1761 }
1762
1763 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
1764 {
1765         int type;
1766         struct revindex_entry *revidx;
1767         const unsigned char *sha1;
1768         revidx = find_pack_revindex(p, obj_offset);
1769         if (!revidx)
1770                 return OBJ_BAD;
1771         sha1 = nth_packed_object_sha1(p, revidx->nr);
1772         mark_bad_packed_object(p, sha1);
1773         type = sha1_object_info(sha1, NULL);
1774         if (type <= OBJ_NONE)
1775                 return OBJ_BAD;
1776         return type;
1777 }
1778
1779 #define POI_STACK_PREALLOC 64
1780
1781 static enum object_type packed_to_object_type(struct packed_git *p,
1782                                               off_t obj_offset,
1783                                               enum object_type type,
1784                                               struct pack_window **w_curs,
1785                                               off_t curpos)
1786 {
1787         off_t small_poi_stack[POI_STACK_PREALLOC];
1788         off_t *poi_stack = small_poi_stack;
1789         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1790
1791         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1792                 off_t base_offset;
1793                 unsigned long size;
1794                 /* Push the object we're going to leave behind */
1795                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1796                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1797                         poi_stack = xmalloc(sizeof(off_t)*poi_stack_alloc);
1798                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1799                 } else {
1800                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1801                 }
1802                 poi_stack[poi_stack_nr++] = obj_offset;
1803                 /* If parsing the base offset fails, just unwind */
1804                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1805                 if (!base_offset)
1806                         goto unwind;
1807                 curpos = obj_offset = base_offset;
1808                 type = unpack_object_header(p, w_curs, &curpos, &size);
1809                 if (type <= OBJ_NONE) {
1810                         /* If getting the base itself fails, we first
1811                          * retry the base, otherwise unwind */
1812                         type = retry_bad_packed_offset(p, base_offset);
1813                         if (type > OBJ_NONE)
1814                                 goto out;
1815                         goto unwind;
1816                 }
1817         }
1818
1819         switch (type) {
1820         case OBJ_BAD:
1821         case OBJ_COMMIT:
1822         case OBJ_TREE:
1823         case OBJ_BLOB:
1824         case OBJ_TAG:
1825                 break;
1826         default:
1827                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1828                       type, (uintmax_t)obj_offset, p->pack_name);
1829                 type = OBJ_BAD;
1830         }
1831
1832 out:
1833         if (poi_stack != small_poi_stack)
1834                 free(poi_stack);
1835         return type;
1836
1837 unwind:
1838         while (poi_stack_nr) {
1839                 obj_offset = poi_stack[--poi_stack_nr];
1840                 type = retry_bad_packed_offset(p, obj_offset);
1841                 if (type > OBJ_NONE)
1842                         goto out;
1843         }
1844         type = OBJ_BAD;
1845         goto out;
1846 }
1847
1848 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1849                               struct object_info *oi)
1850 {
1851         struct pack_window *w_curs = NULL;
1852         unsigned long size;
1853         off_t curpos = obj_offset;
1854         enum object_type type;
1855
1856         /*
1857          * We always get the representation type, but only convert it to
1858          * a "real" type later if the caller is interested.
1859          */
1860         type = unpack_object_header(p, &w_curs, &curpos, &size);
1861
1862         if (oi->sizep) {
1863                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1864                         off_t tmp_pos = curpos;
1865                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1866                                                            type, obj_offset);
1867                         if (!base_offset) {
1868                                 type = OBJ_BAD;
1869                                 goto out;
1870                         }
1871                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1872                         if (*oi->sizep == 0) {
1873                                 type = OBJ_BAD;
1874                                 goto out;
1875                         }
1876                 } else {
1877                         *oi->sizep = size;
1878                 }
1879         }
1880
1881         if (oi->disk_sizep) {
1882                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1883                 *oi->disk_sizep = revidx[1].offset - obj_offset;
1884         }
1885
1886         if (oi->typep) {
1887                 *oi->typep = packed_to_object_type(p, obj_offset, type, &w_curs, curpos);
1888                 if (*oi->typep < 0) {
1889                         type = OBJ_BAD;
1890                         goto out;
1891                 }
1892         }
1893
1894         if (oi->delta_base_sha1) {
1895                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1896                         const unsigned char *base;
1897
1898                         base = get_delta_base_sha1(p, &w_curs, curpos,
1899                                                    type, obj_offset);
1900                         if (!base) {
1901                                 type = OBJ_BAD;
1902                                 goto out;
1903                         }
1904
1905                         hashcpy(oi->delta_base_sha1, base);
1906                 } else
1907                         hashclr(oi->delta_base_sha1);
1908         }
1909
1910 out:
1911         unuse_pack(&w_curs);
1912         return type;
1913 }
1914
1915 static void *unpack_compressed_entry(struct packed_git *p,
1916                                     struct pack_window **w_curs,
1917                                     off_t curpos,
1918                                     unsigned long size)
1919 {
1920         int st;
1921         git_zstream stream;
1922         unsigned char *buffer, *in;
1923
1924         buffer = xmallocz(size);
1925         memset(&stream, 0, sizeof(stream));
1926         stream.next_out = buffer;
1927         stream.avail_out = size + 1;
1928
1929         git_inflate_init(&stream);
1930         do {
1931                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1932                 stream.next_in = in;
1933                 st = git_inflate(&stream, Z_FINISH);
1934                 if (!stream.avail_out)
1935                         break; /* the payload is larger than it should be */
1936                 curpos += stream.next_in - in;
1937         } while (st == Z_OK || st == Z_BUF_ERROR);
1938         git_inflate_end(&stream);
1939         if ((st != Z_STREAM_END) || stream.total_out != size) {
1940                 free(buffer);
1941                 return NULL;
1942         }
1943
1944         return buffer;
1945 }
1946
1947 #define MAX_DELTA_CACHE (256)
1948
1949 static size_t delta_base_cached;
1950
1951 static struct delta_base_cache_lru_list {
1952         struct delta_base_cache_lru_list *prev;
1953         struct delta_base_cache_lru_list *next;
1954 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
1955
1956 static struct delta_base_cache_entry {
1957         struct delta_base_cache_lru_list lru;
1958         void *data;
1959         struct packed_git *p;
1960         off_t base_offset;
1961         unsigned long size;
1962         enum object_type type;
1963 } delta_base_cache[MAX_DELTA_CACHE];
1964
1965 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
1966 {
1967         unsigned long hash;
1968
1969         hash = (unsigned long)p + (unsigned long)base_offset;
1970         hash += (hash >> 8) + (hash >> 16);
1971         return hash % MAX_DELTA_CACHE;
1972 }
1973
1974 static struct delta_base_cache_entry *
1975 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
1976 {
1977         unsigned long hash = pack_entry_hash(p, base_offset);
1978         return delta_base_cache + hash;
1979 }
1980
1981 static int eq_delta_base_cache_entry(struct delta_base_cache_entry *ent,
1982                                      struct packed_git *p, off_t base_offset)
1983 {
1984         return (ent->data && ent->p == p && ent->base_offset == base_offset);
1985 }
1986
1987 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
1988 {
1989         struct delta_base_cache_entry *ent;
1990         ent = get_delta_base_cache_entry(p, base_offset);
1991         return eq_delta_base_cache_entry(ent, p, base_offset);
1992 }
1993
1994 static void clear_delta_base_cache_entry(struct delta_base_cache_entry *ent)
1995 {
1996         ent->data = NULL;
1997         ent->lru.next->prev = ent->lru.prev;
1998         ent->lru.prev->next = ent->lru.next;
1999         delta_base_cached -= ent->size;
2000 }
2001
2002 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
2003         unsigned long *base_size, enum object_type *type, int keep_cache)
2004 {
2005         struct delta_base_cache_entry *ent;
2006         void *ret;
2007
2008         ent = get_delta_base_cache_entry(p, base_offset);
2009
2010         if (!eq_delta_base_cache_entry(ent, p, base_offset))
2011                 return unpack_entry(p, base_offset, type, base_size);
2012
2013         ret = ent->data;
2014
2015         if (!keep_cache)
2016                 clear_delta_base_cache_entry(ent);
2017         else
2018                 ret = xmemdupz(ent->data, ent->size);
2019         *type = ent->type;
2020         *base_size = ent->size;
2021         return ret;
2022 }
2023
2024 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2025 {
2026         if (ent->data) {
2027                 free(ent->data);
2028                 ent->data = NULL;
2029                 ent->lru.next->prev = ent->lru.prev;
2030                 ent->lru.prev->next = ent->lru.next;
2031                 delta_base_cached -= ent->size;
2032         }
2033 }
2034
2035 void clear_delta_base_cache(void)
2036 {
2037         unsigned long p;
2038         for (p = 0; p < MAX_DELTA_CACHE; p++)
2039                 release_delta_base_cache(&delta_base_cache[p]);
2040 }
2041
2042 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2043         void *base, unsigned long base_size, enum object_type type)
2044 {
2045         unsigned long hash = pack_entry_hash(p, base_offset);
2046         struct delta_base_cache_entry *ent = delta_base_cache + hash;
2047         struct delta_base_cache_lru_list *lru;
2048
2049         release_delta_base_cache(ent);
2050         delta_base_cached += base_size;
2051
2052         for (lru = delta_base_cache_lru.next;
2053              delta_base_cached > delta_base_cache_limit
2054              && lru != &delta_base_cache_lru;
2055              lru = lru->next) {
2056                 struct delta_base_cache_entry *f = (void *)lru;
2057                 if (f->type == OBJ_BLOB)
2058                         release_delta_base_cache(f);
2059         }
2060         for (lru = delta_base_cache_lru.next;
2061              delta_base_cached > delta_base_cache_limit
2062              && lru != &delta_base_cache_lru;
2063              lru = lru->next) {
2064                 struct delta_base_cache_entry *f = (void *)lru;
2065                 release_delta_base_cache(f);
2066         }
2067
2068         ent->p = p;
2069         ent->base_offset = base_offset;
2070         ent->type = type;
2071         ent->data = base;
2072         ent->size = base_size;
2073         ent->lru.next = &delta_base_cache_lru;
2074         ent->lru.prev = delta_base_cache_lru.prev;
2075         delta_base_cache_lru.prev->next = &ent->lru;
2076         delta_base_cache_lru.prev = &ent->lru;
2077 }
2078
2079 static void *read_object(const unsigned char *sha1, enum object_type *type,
2080                          unsigned long *size);
2081
2082 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2083 {
2084         static FILE *log_file;
2085
2086         if (!log_pack_access)
2087                 log_pack_access = getenv("GIT_TRACE_PACK_ACCESS");
2088         if (!log_pack_access)
2089                 log_pack_access = no_log_pack_access;
2090         if (log_pack_access == no_log_pack_access)
2091                 return;
2092
2093         if (!log_file) {
2094                 log_file = fopen(log_pack_access, "w");
2095                 if (!log_file) {
2096                         error("cannot open pack access log '%s' for writing: %s",
2097                               log_pack_access, strerror(errno));
2098                         log_pack_access = no_log_pack_access;
2099                         return;
2100                 }
2101         }
2102         fprintf(log_file, "%s %"PRIuMAX"\n",
2103                 p->pack_name, (uintmax_t)obj_offset);
2104         fflush(log_file);
2105 }
2106
2107 int do_check_packed_object_crc;
2108
2109 #define UNPACK_ENTRY_STACK_PREALLOC 64
2110 struct unpack_entry_stack_ent {
2111         off_t obj_offset;
2112         off_t curpos;
2113         unsigned long size;
2114 };
2115
2116 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2117                    enum object_type *final_type, unsigned long *final_size)
2118 {
2119         struct pack_window *w_curs = NULL;
2120         off_t curpos = obj_offset;
2121         void *data = NULL;
2122         unsigned long size;
2123         enum object_type type;
2124         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2125         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2126         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2127         int base_from_cache = 0;
2128
2129         if (log_pack_access != no_log_pack_access)
2130                 write_pack_access_log(p, obj_offset);
2131
2132         /* PHASE 1: drill down to the innermost base object */
2133         for (;;) {
2134                 off_t base_offset;
2135                 int i;
2136                 struct delta_base_cache_entry *ent;
2137
2138                 ent = get_delta_base_cache_entry(p, curpos);
2139                 if (eq_delta_base_cache_entry(ent, p, curpos)) {
2140                         type = ent->type;
2141                         data = ent->data;
2142                         size = ent->size;
2143                         clear_delta_base_cache_entry(ent);
2144                         base_from_cache = 1;
2145                         break;
2146                 }
2147
2148                 if (do_check_packed_object_crc && p->index_version > 1) {
2149                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2150                         unsigned long len = revidx[1].offset - obj_offset;
2151                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2152                                 const unsigned char *sha1 =
2153                                         nth_packed_object_sha1(p, revidx->nr);
2154                                 error("bad packed object CRC for %s",
2155                                       sha1_to_hex(sha1));
2156                                 mark_bad_packed_object(p, sha1);
2157                                 unuse_pack(&w_curs);
2158                                 return NULL;
2159                         }
2160                 }
2161
2162                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2163                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2164                         break;
2165
2166                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2167                 if (!base_offset) {
2168                         error("failed to validate delta base reference "
2169                               "at offset %"PRIuMAX" from %s",
2170                               (uintmax_t)curpos, p->pack_name);
2171                         /* bail to phase 2, in hopes of recovery */
2172                         data = NULL;
2173                         break;
2174                 }
2175
2176                 /* push object, proceed to base */
2177                 if (delta_stack_nr >= delta_stack_alloc
2178                     && delta_stack == small_delta_stack) {
2179                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2180                         delta_stack = xmalloc(sizeof(*delta_stack)*delta_stack_alloc);
2181                         memcpy(delta_stack, small_delta_stack,
2182                                sizeof(*delta_stack)*delta_stack_nr);
2183                 } else {
2184                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2185                 }
2186                 i = delta_stack_nr++;
2187                 delta_stack[i].obj_offset = obj_offset;
2188                 delta_stack[i].curpos = curpos;
2189                 delta_stack[i].size = size;
2190
2191                 curpos = obj_offset = base_offset;
2192         }
2193
2194         /* PHASE 2: handle the base */
2195         switch (type) {
2196         case OBJ_OFS_DELTA:
2197         case OBJ_REF_DELTA:
2198                 if (data)
2199                         die("BUG in unpack_entry: left loop at a valid delta");
2200                 break;
2201         case OBJ_COMMIT:
2202         case OBJ_TREE:
2203         case OBJ_BLOB:
2204         case OBJ_TAG:
2205                 if (!base_from_cache)
2206                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2207                 break;
2208         default:
2209                 data = NULL;
2210                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2211                       type, (uintmax_t)obj_offset, p->pack_name);
2212         }
2213
2214         /* PHASE 3: apply deltas in order */
2215
2216         /* invariants:
2217          *   'data' holds the base data, or NULL if there was corruption
2218          */
2219         while (delta_stack_nr) {
2220                 void *delta_data;
2221                 void *base = data;
2222                 unsigned long delta_size, base_size = size;
2223                 int i;
2224
2225                 data = NULL;
2226
2227                 if (base)
2228                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2229
2230                 if (!base) {
2231                         /*
2232                          * We're probably in deep shit, but let's try to fetch
2233                          * the required base anyway from another pack or loose.
2234                          * This is costly but should happen only in the presence
2235                          * of a corrupted pack, and is better than failing outright.
2236                          */
2237                         struct revindex_entry *revidx;
2238                         const unsigned char *base_sha1;
2239                         revidx = find_pack_revindex(p, obj_offset);
2240                         if (revidx) {
2241                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2242                                 error("failed to read delta base object %s"
2243                                       " at offset %"PRIuMAX" from %s",
2244                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2245                                       p->pack_name);
2246                                 mark_bad_packed_object(p, base_sha1);
2247                                 base = read_object(base_sha1, &type, &base_size);
2248                         }
2249                 }
2250
2251                 i = --delta_stack_nr;
2252                 obj_offset = delta_stack[i].obj_offset;
2253                 curpos = delta_stack[i].curpos;
2254                 delta_size = delta_stack[i].size;
2255
2256                 if (!base)
2257                         continue;
2258
2259                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2260
2261                 if (!delta_data) {
2262                         error("failed to unpack compressed delta "
2263                               "at offset %"PRIuMAX" from %s",
2264                               (uintmax_t)curpos, p->pack_name);
2265                         data = NULL;
2266                         continue;
2267                 }
2268
2269                 data = patch_delta(base, base_size,
2270                                    delta_data, delta_size,
2271                                    &size);
2272
2273                 /*
2274                  * We could not apply the delta; warn the user, but keep going.
2275                  * Our failure will be noticed either in the next iteration of
2276                  * the loop, or if this is the final delta, in the caller when
2277                  * we return NULL. Those code paths will take care of making
2278                  * a more explicit warning and retrying with another copy of
2279                  * the object.
2280                  */
2281                 if (!data)
2282                         error("failed to apply delta");
2283
2284                 free(delta_data);
2285         }
2286
2287         *final_type = type;
2288         *final_size = size;
2289
2290         unuse_pack(&w_curs);
2291         return data;
2292 }
2293
2294 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2295                                             uint32_t n)
2296 {
2297         const unsigned char *index = p->index_data;
2298         if (!index) {
2299                 if (open_pack_index(p))
2300                         return NULL;
2301                 index = p->index_data;
2302         }
2303         if (n >= p->num_objects)
2304                 return NULL;
2305         index += 4 * 256;
2306         if (p->index_version == 1) {
2307                 return index + 24 * n + 4;
2308         } else {
2309                 index += 8;
2310                 return index + 20 * n;
2311         }
2312 }
2313
2314 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2315 {
2316         const unsigned char *index = p->index_data;
2317         index += 4 * 256;
2318         if (p->index_version == 1) {
2319                 return ntohl(*((uint32_t *)(index + 24 * n)));
2320         } else {
2321                 uint32_t off;
2322                 index += 8 + p->num_objects * (20 + 4);
2323                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2324                 if (!(off & 0x80000000))
2325                         return off;
2326                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2327                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2328                                    ntohl(*((uint32_t *)(index + 4)));
2329         }
2330 }
2331
2332 off_t find_pack_entry_one(const unsigned char *sha1,
2333                                   struct packed_git *p)
2334 {
2335         const uint32_t *level1_ofs = p->index_data;
2336         const unsigned char *index = p->index_data;
2337         unsigned hi, lo, stride;
2338         static int use_lookup = -1;
2339         static int debug_lookup = -1;
2340
2341         if (debug_lookup < 0)
2342                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2343
2344         if (!index) {
2345                 if (open_pack_index(p))
2346                         return 0;
2347                 level1_ofs = p->index_data;
2348                 index = p->index_data;
2349         }
2350         if (p->index_version > 1) {
2351                 level1_ofs += 2;
2352                 index += 8;
2353         }
2354         index += 4 * 256;
2355         hi = ntohl(level1_ofs[*sha1]);
2356         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2357         if (p->index_version > 1) {
2358                 stride = 20;
2359         } else {
2360                 stride = 24;
2361                 index += 4;
2362         }
2363
2364         if (debug_lookup)
2365                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2366                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2367
2368         if (use_lookup < 0)
2369                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2370         if (use_lookup) {
2371                 int pos = sha1_entry_pos(index, stride, 0,
2372                                          lo, hi, p->num_objects, sha1);
2373                 if (pos < 0)
2374                         return 0;
2375                 return nth_packed_object_offset(p, pos);
2376         }
2377
2378         do {
2379                 unsigned mi = (lo + hi) / 2;
2380                 int cmp = hashcmp(index + mi * stride, sha1);
2381
2382                 if (debug_lookup)
2383                         printf("lo %u hi %u rg %u mi %u\n",
2384                                lo, hi, hi - lo, mi);
2385                 if (!cmp)
2386                         return nth_packed_object_offset(p, mi);
2387                 if (cmp > 0)
2388                         hi = mi;
2389                 else
2390                         lo = mi+1;
2391         } while (lo < hi);
2392         return 0;
2393 }
2394
2395 int is_pack_valid(struct packed_git *p)
2396 {
2397         /* An already open pack is known to be valid. */
2398         if (p->pack_fd != -1)
2399                 return 1;
2400
2401         /* If the pack has one window completely covering the
2402          * file size, the pack is known to be valid even if
2403          * the descriptor is not currently open.
2404          */
2405         if (p->windows) {
2406                 struct pack_window *w = p->windows;
2407
2408                 if (!w->offset && w->len == p->pack_size)
2409                         return 1;
2410         }
2411
2412         /* Force the pack to open to prove its valid. */
2413         return !open_packed_git(p);
2414 }
2415
2416 static int fill_pack_entry(const unsigned char *sha1,
2417                            struct pack_entry *e,
2418                            struct packed_git *p)
2419 {
2420         off_t offset;
2421
2422         if (p->num_bad_objects) {
2423                 unsigned i;
2424                 for (i = 0; i < p->num_bad_objects; i++)
2425                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2426                                 return 0;
2427         }
2428
2429         offset = find_pack_entry_one(sha1, p);
2430         if (!offset)
2431                 return 0;
2432
2433         /*
2434          * We are about to tell the caller where they can locate the
2435          * requested object.  We better make sure the packfile is
2436          * still here and can be accessed before supplying that
2437          * answer, as it may have been deleted since the index was
2438          * loaded!
2439          */
2440         if (!is_pack_valid(p)) {
2441                 warning("packfile %s cannot be accessed", p->pack_name);
2442                 return 0;
2443         }
2444         e->offset = offset;
2445         e->p = p;
2446         hashcpy(e->sha1, sha1);
2447         return 1;
2448 }
2449
2450 /*
2451  * Iff a pack file contains the object named by sha1, return true and
2452  * store its location to e.
2453  */
2454 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2455 {
2456         struct packed_git *p;
2457
2458         prepare_packed_git();
2459         if (!packed_git)
2460                 return 0;
2461
2462         if (last_found_pack && fill_pack_entry(sha1, e, last_found_pack))
2463                 return 1;
2464
2465         for (p = packed_git; p; p = p->next) {
2466                 if (p == last_found_pack)
2467                         continue; /* we already checked this one */
2468
2469                 if (fill_pack_entry(sha1, e, p)) {
2470                         last_found_pack = p;
2471                         return 1;
2472                 }
2473         }
2474         return 0;
2475 }
2476
2477 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2478                                   struct packed_git *packs)
2479 {
2480         struct packed_git *p;
2481
2482         for (p = packs; p; p = p->next) {
2483                 if (find_pack_entry_one(sha1, p))
2484                         return p;
2485         }
2486         return NULL;
2487
2488 }
2489
2490 static int sha1_loose_object_info(const unsigned char *sha1,
2491                                   struct object_info *oi)
2492 {
2493         int status;
2494         unsigned long mapsize, size;
2495         void *map;
2496         git_zstream stream;
2497         char hdr[32];
2498
2499         if (oi->delta_base_sha1)
2500                 hashclr(oi->delta_base_sha1);
2501
2502         /*
2503          * If we don't care about type or size, then we don't
2504          * need to look inside the object at all. Note that we
2505          * do not optimize out the stat call, even if the
2506          * caller doesn't care about the disk-size, since our
2507          * return value implicitly indicates whether the
2508          * object even exists.
2509          */
2510         if (!oi->typep && !oi->sizep) {
2511                 struct stat st;
2512                 if (stat_sha1_file(sha1, &st) < 0)
2513                         return -1;
2514                 if (oi->disk_sizep)
2515                         *oi->disk_sizep = st.st_size;
2516                 return 0;
2517         }
2518
2519         map = map_sha1_file(sha1, &mapsize);
2520         if (!map)
2521                 return -1;
2522         if (oi->disk_sizep)
2523                 *oi->disk_sizep = mapsize;
2524         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2525                 status = error("unable to unpack %s header",
2526                                sha1_to_hex(sha1));
2527         else if ((status = parse_sha1_header(hdr, &size)) < 0)
2528                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2529         else if (oi->sizep)
2530                 *oi->sizep = size;
2531         git_inflate_end(&stream);
2532         munmap(map, mapsize);
2533         if (oi->typep)
2534                 *oi->typep = status;
2535         return 0;
2536 }
2537
2538 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi, unsigned flags)
2539 {
2540         struct cached_object *co;
2541         struct pack_entry e;
2542         int rtype;
2543         const unsigned char *real = lookup_replace_object_extended(sha1, flags);
2544
2545         co = find_cached_object(real);
2546         if (co) {
2547                 if (oi->typep)
2548                         *(oi->typep) = co->type;
2549                 if (oi->sizep)
2550                         *(oi->sizep) = co->size;
2551                 if (oi->disk_sizep)
2552                         *(oi->disk_sizep) = 0;
2553                 if (oi->delta_base_sha1)
2554                         hashclr(oi->delta_base_sha1);
2555                 oi->whence = OI_CACHED;
2556                 return 0;
2557         }
2558
2559         if (!find_pack_entry(real, &e)) {
2560                 /* Most likely it's a loose object. */
2561                 if (!sha1_loose_object_info(real, oi)) {
2562                         oi->whence = OI_LOOSE;
2563                         return 0;
2564                 }
2565
2566                 /* Not a loose object; someone else may have just packed it. */
2567                 reprepare_packed_git();
2568                 if (!find_pack_entry(real, &e))
2569                         return -1;
2570         }
2571
2572         rtype = packed_object_info(e.p, e.offset, oi);
2573         if (rtype < 0) {
2574                 mark_bad_packed_object(e.p, real);
2575                 return sha1_object_info_extended(real, oi, 0);
2576         } else if (in_delta_base_cache(e.p, e.offset)) {
2577                 oi->whence = OI_DBCACHED;
2578         } else {
2579                 oi->whence = OI_PACKED;
2580                 oi->u.packed.offset = e.offset;
2581                 oi->u.packed.pack = e.p;
2582                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
2583                                          rtype == OBJ_OFS_DELTA);
2584         }
2585
2586         return 0;
2587 }
2588
2589 /* returns enum object_type or negative */
2590 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2591 {
2592         enum object_type type;
2593         struct object_info oi = {NULL};
2594
2595         oi.typep = &type;
2596         oi.sizep = sizep;
2597         if (sha1_object_info_extended(sha1, &oi, LOOKUP_REPLACE_OBJECT) < 0)
2598                 return -1;
2599         return type;
2600 }
2601
2602 static void *read_packed_sha1(const unsigned char *sha1,
2603                               enum object_type *type, unsigned long *size)
2604 {
2605         struct pack_entry e;
2606         void *data;
2607
2608         if (!find_pack_entry(sha1, &e))
2609                 return NULL;
2610         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
2611         if (!data) {
2612                 /*
2613                  * We're probably in deep shit, but let's try to fetch
2614                  * the required object anyway from another pack or loose.
2615                  * This should happen only in the presence of a corrupted
2616                  * pack, and is better than failing outright.
2617                  */
2618                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2619                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2620                 mark_bad_packed_object(e.p, sha1);
2621                 data = read_object(sha1, type, size);
2622         }
2623         return data;
2624 }
2625
2626 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2627                       unsigned char *sha1)
2628 {
2629         struct cached_object *co;
2630
2631         hash_sha1_file(buf, len, typename(type), sha1);
2632         if (has_sha1_file(sha1) || find_cached_object(sha1))
2633                 return 0;
2634         if (cached_object_alloc <= cached_object_nr) {
2635                 cached_object_alloc = alloc_nr(cached_object_alloc);
2636                 cached_objects = xrealloc(cached_objects,
2637                                           sizeof(*cached_objects) *
2638                                           cached_object_alloc);
2639         }
2640         co = &cached_objects[cached_object_nr++];
2641         co->size = len;
2642         co->type = type;
2643         co->buf = xmalloc(len);
2644         memcpy(co->buf, buf, len);
2645         hashcpy(co->sha1, sha1);
2646         return 0;
2647 }
2648
2649 static void *read_object(const unsigned char *sha1, enum object_type *type,
2650                          unsigned long *size)
2651 {
2652         unsigned long mapsize;
2653         void *map, *buf;
2654         struct cached_object *co;
2655
2656         co = find_cached_object(sha1);
2657         if (co) {
2658                 *type = co->type;
2659                 *size = co->size;
2660                 return xmemdupz(co->buf, co->size);
2661         }
2662
2663         buf = read_packed_sha1(sha1, type, size);
2664         if (buf)
2665                 return buf;
2666         map = map_sha1_file(sha1, &mapsize);
2667         if (map) {
2668                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2669                 munmap(map, mapsize);
2670                 return buf;
2671         }
2672         reprepare_packed_git();
2673         return read_packed_sha1(sha1, type, size);
2674 }
2675
2676 /*
2677  * This function dies on corrupt objects; the callers who want to
2678  * deal with them should arrange to call read_object() and give error
2679  * messages themselves.
2680  */
2681 void *read_sha1_file_extended(const unsigned char *sha1,
2682                               enum object_type *type,
2683                               unsigned long *size,
2684                               unsigned flag)
2685 {
2686         void *data;
2687         const struct packed_git *p;
2688         const unsigned char *repl = lookup_replace_object_extended(sha1, flag);
2689
2690         errno = 0;
2691         data = read_object(repl, type, size);
2692         if (data)
2693                 return data;
2694
2695         if (errno && errno != ENOENT)
2696                 die_errno("failed to read object %s", sha1_to_hex(sha1));
2697
2698         /* die if we replaced an object with one that does not exist */
2699         if (repl != sha1)
2700                 die("replacement %s not found for %s",
2701                     sha1_to_hex(repl), sha1_to_hex(sha1));
2702
2703         if (has_loose_object(repl)) {
2704                 const char *path = sha1_file_name(sha1);
2705
2706                 die("loose object %s (stored in %s) is corrupt",
2707                     sha1_to_hex(repl), path);
2708         }
2709
2710         if ((p = has_packed_and_bad(repl)) != NULL)
2711                 die("packed object %s (stored in %s) is corrupt",
2712                     sha1_to_hex(repl), p->pack_name);
2713
2714         return NULL;
2715 }
2716
2717 void *read_object_with_reference(const unsigned char *sha1,
2718                                  const char *required_type_name,
2719                                  unsigned long *size,
2720                                  unsigned char *actual_sha1_return)
2721 {
2722         enum object_type type, required_type;
2723         void *buffer;
2724         unsigned long isize;
2725         unsigned char actual_sha1[20];
2726
2727         required_type = type_from_string(required_type_name);
2728         hashcpy(actual_sha1, sha1);
2729         while (1) {
2730                 int ref_length = -1;
2731                 const char *ref_type = NULL;
2732
2733                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2734                 if (!buffer)
2735                         return NULL;
2736                 if (type == required_type) {
2737                         *size = isize;
2738                         if (actual_sha1_return)
2739                                 hashcpy(actual_sha1_return, actual_sha1);
2740                         return buffer;
2741                 }
2742                 /* Handle references */
2743                 else if (type == OBJ_COMMIT)
2744                         ref_type = "tree ";
2745                 else if (type == OBJ_TAG)
2746                         ref_type = "object ";
2747                 else {
2748                         free(buffer);
2749                         return NULL;
2750                 }
2751                 ref_length = strlen(ref_type);
2752
2753                 if (ref_length + 40 > isize ||
2754                     memcmp(buffer, ref_type, ref_length) ||
2755                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2756                         free(buffer);
2757                         return NULL;
2758                 }
2759                 free(buffer);
2760                 /* Now we have the ID of the referred-to object in
2761                  * actual_sha1.  Check again. */
2762         }
2763 }
2764
2765 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2766                                     const char *type, unsigned char *sha1,
2767                                     char *hdr, int *hdrlen)
2768 {
2769         git_SHA_CTX c;
2770
2771         /* Generate the header */
2772         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2773
2774         /* Sha1.. */
2775         git_SHA1_Init(&c);
2776         git_SHA1_Update(&c, hdr, *hdrlen);
2777         git_SHA1_Update(&c, buf, len);
2778         git_SHA1_Final(sha1, &c);
2779 }
2780
2781 /*
2782  * Move the just written object into its final resting place.
2783  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2784  * "moving" is only a part of what it does, when no patch between
2785  * master to pu changes the call sites of this function.
2786  */
2787 int move_temp_to_file(const char *tmpfile, const char *filename)
2788 {
2789         int ret = 0;
2790
2791         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2792                 goto try_rename;
2793         else if (link(tmpfile, filename))
2794                 ret = errno;
2795
2796         /*
2797          * Coda hack - coda doesn't like cross-directory links,
2798          * so we fall back to a rename, which will mean that it
2799          * won't be able to check collisions, but that's not a
2800          * big deal.
2801          *
2802          * The same holds for FAT formatted media.
2803          *
2804          * When this succeeds, we just return.  We have nothing
2805          * left to unlink.
2806          */
2807         if (ret && ret != EEXIST) {
2808         try_rename:
2809                 if (!rename(tmpfile, filename))
2810                         goto out;
2811                 ret = errno;
2812         }
2813         unlink_or_warn(tmpfile);
2814         if (ret) {
2815                 if (ret != EEXIST) {
2816                         return error("unable to write sha1 filename %s: %s", filename, strerror(ret));
2817                 }
2818                 /* FIXME!!! Collision check here ? */
2819         }
2820
2821 out:
2822         if (adjust_shared_perm(filename))
2823                 return error("unable to set permission to '%s'", filename);
2824         return 0;
2825 }
2826
2827 static int write_buffer(int fd, const void *buf, size_t len)
2828 {
2829         if (write_in_full(fd, buf, len) < 0)
2830                 return error("file write error (%s)", strerror(errno));
2831         return 0;
2832 }
2833
2834 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2835                    unsigned char *sha1)
2836 {
2837         char hdr[32];
2838         int hdrlen;
2839         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2840         return 0;
2841 }
2842
2843 /* Finalize a file on disk, and close it. */
2844 static void close_sha1_file(int fd)
2845 {
2846         if (fsync_object_files)
2847                 fsync_or_die(fd, "sha1 file");
2848         if (close(fd) != 0)
2849                 die_errno("error when closing sha1 file");
2850 }
2851
2852 /* Size of directory component, including the ending '/' */
2853 static inline int directory_size(const char *filename)
2854 {
2855         const char *s = strrchr(filename, '/');
2856         if (!s)
2857                 return 0;
2858         return s - filename + 1;
2859 }
2860
2861 /*
2862  * This creates a temporary file in the same directory as the final
2863  * 'filename'
2864  *
2865  * We want to avoid cross-directory filename renames, because those
2866  * can have problems on various filesystems (FAT, NFS, Coda).
2867  */
2868 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2869 {
2870         int fd, dirlen = directory_size(filename);
2871
2872         if (dirlen + 20 > bufsiz) {
2873                 errno = ENAMETOOLONG;
2874                 return -1;
2875         }
2876         memcpy(buffer, filename, dirlen);
2877         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2878         fd = git_mkstemp_mode(buffer, 0444);
2879         if (fd < 0 && dirlen && errno == ENOENT) {
2880                 /* Make sure the directory exists */
2881                 memcpy(buffer, filename, dirlen);
2882                 buffer[dirlen-1] = 0;
2883                 if (mkdir(buffer, 0777) && errno != EEXIST)
2884                         return -1;
2885                 if (adjust_shared_perm(buffer))
2886                         return -1;
2887
2888                 /* Try again */
2889                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2890                 fd = git_mkstemp_mode(buffer, 0444);
2891         }
2892         return fd;
2893 }
2894
2895 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2896                               const void *buf, unsigned long len, time_t mtime)
2897 {
2898         int fd, ret;
2899         unsigned char compressed[4096];
2900         git_zstream stream;
2901         git_SHA_CTX c;
2902         unsigned char parano_sha1[20];
2903         static char tmp_file[PATH_MAX];
2904         const char *filename = sha1_file_name(sha1);
2905
2906         fd = create_tmpfile(tmp_file, sizeof(tmp_file), filename);
2907         if (fd < 0) {
2908                 if (errno == EACCES)
2909                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
2910                 else
2911                         return error("unable to create temporary file: %s", strerror(errno));
2912         }
2913
2914         /* Set it up */
2915         memset(&stream, 0, sizeof(stream));
2916         git_deflate_init(&stream, zlib_compression_level);
2917         stream.next_out = compressed;
2918         stream.avail_out = sizeof(compressed);
2919         git_SHA1_Init(&c);
2920
2921         /* First header.. */
2922         stream.next_in = (unsigned char *)hdr;
2923         stream.avail_in = hdrlen;
2924         while (git_deflate(&stream, 0) == Z_OK)
2925                 ; /* nothing */
2926         git_SHA1_Update(&c, hdr, hdrlen);
2927
2928         /* Then the data itself.. */
2929         stream.next_in = (void *)buf;
2930         stream.avail_in = len;
2931         do {
2932                 unsigned char *in0 = stream.next_in;
2933                 ret = git_deflate(&stream, Z_FINISH);
2934                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2935                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2936                         die("unable to write sha1 file");
2937                 stream.next_out = compressed;
2938                 stream.avail_out = sizeof(compressed);
2939         } while (ret == Z_OK);
2940
2941         if (ret != Z_STREAM_END)
2942                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2943         ret = git_deflate_end_gently(&stream);
2944         if (ret != Z_OK)
2945                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2946         git_SHA1_Final(parano_sha1, &c);
2947         if (hashcmp(sha1, parano_sha1) != 0)
2948                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2949
2950         close_sha1_file(fd);
2951
2952         if (mtime) {
2953                 struct utimbuf utb;
2954                 utb.actime = mtime;
2955                 utb.modtime = mtime;
2956                 if (utime(tmp_file, &utb) < 0)
2957                         warning("failed utime() on %s: %s",
2958                                 tmp_file, strerror(errno));
2959         }
2960
2961         return move_temp_to_file(tmp_file, filename);
2962 }
2963
2964 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
2965 {
2966         unsigned char sha1[20];
2967         char hdr[32];
2968         int hdrlen;
2969
2970         /* Normally if we have it in the pack then we do not bother writing
2971          * it out into .git/objects/??/?{38} file.
2972          */
2973         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2974         if (returnsha1)
2975                 hashcpy(returnsha1, sha1);
2976         if (has_sha1_file(sha1))
2977                 return 0;
2978         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
2979 }
2980
2981 int force_object_loose(const unsigned char *sha1, time_t mtime)
2982 {
2983         void *buf;
2984         unsigned long len;
2985         enum object_type type;
2986         char hdr[32];
2987         int hdrlen;
2988         int ret;
2989
2990         if (has_loose_object(sha1))
2991                 return 0;
2992         buf = read_packed_sha1(sha1, &type, &len);
2993         if (!buf)
2994                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
2995         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
2996         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
2997         free(buf);
2998
2999         return ret;
3000 }
3001
3002 int has_pack_index(const unsigned char *sha1)
3003 {
3004         struct stat st;
3005         if (stat(sha1_pack_index_name(sha1), &st))
3006                 return 0;
3007         return 1;
3008 }
3009
3010 int has_sha1_pack(const unsigned char *sha1)
3011 {
3012         struct pack_entry e;
3013         return find_pack_entry(sha1, &e);
3014 }
3015
3016 int has_sha1_file(const unsigned char *sha1)
3017 {
3018         struct pack_entry e;
3019
3020         if (find_pack_entry(sha1, &e))
3021                 return 1;
3022         if (has_loose_object(sha1))
3023                 return 1;
3024         reprepare_packed_git();
3025         return find_pack_entry(sha1, &e);
3026 }
3027
3028 static void check_tree(const void *buf, size_t size)
3029 {
3030         struct tree_desc desc;
3031         struct name_entry entry;
3032
3033         init_tree_desc(&desc, buf, size);
3034         while (tree_entry(&desc, &entry))
3035                 /* do nothing
3036                  * tree_entry() will die() on malformed entries */
3037                 ;
3038 }
3039
3040 static void check_commit(const void *buf, size_t size)
3041 {
3042         struct commit c;
3043         memset(&c, 0, sizeof(c));
3044         if (parse_commit_buffer(&c, buf, size))
3045                 die("corrupt commit");
3046 }
3047
3048 static void check_tag(const void *buf, size_t size)
3049 {
3050         struct tag t;
3051         memset(&t, 0, sizeof(t));
3052         if (parse_tag_buffer(&t, buf, size))
3053                 die("corrupt tag");
3054 }
3055
3056 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3057                      enum object_type type,
3058                      const char *path, unsigned flags)
3059 {
3060         int ret, re_allocated = 0;
3061         int write_object = flags & HASH_WRITE_OBJECT;
3062
3063         if (!type)
3064                 type = OBJ_BLOB;
3065
3066         /*
3067          * Convert blobs to git internal format
3068          */
3069         if ((type == OBJ_BLOB) && path) {
3070                 struct strbuf nbuf = STRBUF_INIT;
3071                 if (convert_to_git(path, buf, size, &nbuf,
3072                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3073                         buf = strbuf_detach(&nbuf, &size);
3074                         re_allocated = 1;
3075                 }
3076         }
3077         if (flags & HASH_FORMAT_CHECK) {
3078                 if (type == OBJ_TREE)
3079                         check_tree(buf, size);
3080                 if (type == OBJ_COMMIT)
3081                         check_commit(buf, size);
3082                 if (type == OBJ_TAG)
3083                         check_tag(buf, size);
3084         }
3085
3086         if (write_object)
3087                 ret = write_sha1_file(buf, size, typename(type), sha1);
3088         else
3089                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3090         if (re_allocated)
3091                 free(buf);
3092         return ret;
3093 }
3094
3095 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3096                       const char *path, unsigned flags)
3097 {
3098         struct strbuf sbuf = STRBUF_INIT;
3099         int ret;
3100
3101         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3102                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3103         else
3104                 ret = -1;
3105         strbuf_release(&sbuf);
3106         return ret;
3107 }
3108
3109 #define SMALL_FILE_SIZE (32*1024)
3110
3111 static int index_core(unsigned char *sha1, int fd, size_t size,
3112                       enum object_type type, const char *path,
3113                       unsigned flags)
3114 {
3115         int ret;
3116
3117         if (!size) {
3118                 ret = index_mem(sha1, NULL, size, type, path, flags);
3119         } else if (size <= SMALL_FILE_SIZE) {
3120                 char *buf = xmalloc(size);
3121                 if (size == read_in_full(fd, buf, size))
3122                         ret = index_mem(sha1, buf, size, type, path, flags);
3123                 else
3124                         ret = error("short read %s", strerror(errno));
3125                 free(buf);
3126         } else {
3127                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3128                 ret = index_mem(sha1, buf, size, type, path, flags);
3129                 munmap(buf, size);
3130         }
3131         return ret;
3132 }
3133
3134 /*
3135  * This creates one packfile per large blob unless bulk-checkin
3136  * machinery is "plugged".
3137  *
3138  * This also bypasses the usual "convert-to-git" dance, and that is on
3139  * purpose. We could write a streaming version of the converting
3140  * functions and insert that before feeding the data to fast-import
3141  * (or equivalent in-core API described above). However, that is
3142  * somewhat complicated, as we do not know the size of the filter
3143  * result, which we need to know beforehand when writing a git object.
3144  * Since the primary motivation for trying to stream from the working
3145  * tree file and to avoid mmaping it in core is to deal with large
3146  * binary blobs, they generally do not want to get any conversion, and
3147  * callers should avoid this code path when filters are requested.
3148  */
3149 static int index_stream(unsigned char *sha1, int fd, size_t size,
3150                         enum object_type type, const char *path,
3151                         unsigned flags)
3152 {
3153         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3154 }
3155
3156 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3157              enum object_type type, const char *path, unsigned flags)
3158 {
3159         int ret;
3160         size_t size = xsize_t(st->st_size);
3161
3162         if (!S_ISREG(st->st_mode))
3163                 ret = index_pipe(sha1, fd, type, path, flags);
3164         else if (size <= big_file_threshold || type != OBJ_BLOB ||
3165                  (path && would_convert_to_git(path, NULL, 0, 0)))
3166                 ret = index_core(sha1, fd, size, type, path, flags);
3167         else
3168                 ret = index_stream(sha1, fd, size, type, path, flags);
3169         close(fd);
3170         return ret;
3171 }
3172
3173 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3174 {
3175         int fd;
3176         struct strbuf sb = STRBUF_INIT;
3177
3178         switch (st->st_mode & S_IFMT) {
3179         case S_IFREG:
3180                 fd = open(path, O_RDONLY);
3181                 if (fd < 0)
3182                         return error("open(\"%s\"): %s", path,
3183                                      strerror(errno));
3184                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3185                         return error("%s: failed to insert into database",
3186                                      path);
3187                 break;
3188         case S_IFLNK:
3189                 if (strbuf_readlink(&sb, path, st->st_size)) {
3190                         char *errstr = strerror(errno);
3191                         return error("readlink(\"%s\"): %s", path,
3192                                      errstr);
3193                 }
3194                 if (!(flags & HASH_WRITE_OBJECT))
3195                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3196                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3197                         return error("%s: failed to insert into database",
3198                                      path);
3199                 strbuf_release(&sb);
3200                 break;
3201         case S_IFDIR:
3202                 return resolve_gitlink_ref(path, "HEAD", sha1);
3203         default:
3204                 return error("%s: unsupported file type", path);
3205         }
3206         return 0;
3207 }
3208
3209 int read_pack_header(int fd, struct pack_header *header)
3210 {
3211         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3212                 /* "eof before pack header was fully read" */
3213                 return PH_ERROR_EOF;
3214
3215         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3216                 /* "protocol error (pack signature mismatch detected)" */
3217                 return PH_ERROR_PACK_SIGNATURE;
3218         if (!pack_version_ok(header->hdr_version))
3219                 /* "protocol error (pack version unsupported)" */
3220                 return PH_ERROR_PROTOCOL;
3221         return 0;
3222 }
3223
3224 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3225 {
3226         enum object_type type = sha1_object_info(sha1, NULL);
3227         if (type < 0)
3228                 die("%s is not a valid object", sha1_to_hex(sha1));
3229         if (type != expect)
3230                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3231                     typename(expect));
3232 }