]> git.scripts.mit.edu Git - git.git/blob - sha1_file.c
howto: add article on recovering a corrupted object
[git.git] / sha1_file.c
1 /*
2  * GIT - The information manager from hell
3  *
4  * Copyright (C) Linus Torvalds, 2005
5  *
6  * This handles basic git sha1 object files - packing, unpacking,
7  * creation etc.
8  */
9 #include "cache.h"
10 #include "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 static struct packed_git *last_found_pack;
64
65 static struct cached_object *find_cached_object(const unsigned char *sha1)
66 {
67         int i;
68         struct cached_object *co = cached_objects;
69
70         for (i = 0; i < cached_object_nr; i++, co++) {
71                 if (!hashcmp(co->sha1, sha1))
72                         return co;
73         }
74         if (!hashcmp(sha1, empty_tree.sha1))
75                 return &empty_tree;
76         return NULL;
77 }
78
79 int mkdir_in_gitdir(const char *path)
80 {
81         if (mkdir(path, 0777)) {
82                 int saved_errno = errno;
83                 struct stat st;
84                 struct strbuf sb = STRBUF_INIT;
85
86                 if (errno != EEXIST)
87                         return -1;
88                 /*
89                  * Are we looking at a path in a symlinked worktree
90                  * whose original repository does not yet have it?
91                  * e.g. .git/rr-cache pointing at its original
92                  * repository in which the user hasn't performed any
93                  * conflict resolution yet?
94                  */
95                 if (lstat(path, &st) || !S_ISLNK(st.st_mode) ||
96                     strbuf_readlink(&sb, path, st.st_size) ||
97                     !is_absolute_path(sb.buf) ||
98                     mkdir(sb.buf, 0777)) {
99                         strbuf_release(&sb);
100                         errno = saved_errno;
101                         return -1;
102                 }
103                 strbuf_release(&sb);
104         }
105         return adjust_shared_perm(path);
106 }
107
108 int safe_create_leading_directories(char *path)
109 {
110         char *pos = path + offset_1st_component(path);
111         struct stat st;
112
113         while (pos) {
114                 pos = strchr(pos, '/');
115                 if (!pos)
116                         break;
117                 while (*++pos == '/')
118                         ;
119                 if (!*pos)
120                         break;
121                 *--pos = '\0';
122                 if (!stat(path, &st)) {
123                         /* path exists */
124                         if (!S_ISDIR(st.st_mode)) {
125                                 *pos = '/';
126                                 return -3;
127                         }
128                 }
129                 else if (mkdir(path, 0777)) {
130                         if (errno == EEXIST &&
131                             !stat(path, &st) && S_ISDIR(st.st_mode)) {
132                                 ; /* somebody created it since we checked */
133                         } else {
134                                 *pos = '/';
135                                 return -1;
136                         }
137                 }
138                 else if (adjust_shared_perm(path)) {
139                         *pos = '/';
140                         return -2;
141                 }
142                 *pos++ = '/';
143         }
144         return 0;
145 }
146
147 int safe_create_leading_directories_const(const char *path)
148 {
149         /* path points to cache entries, so xstrdup before messing with it */
150         char *buf = xstrdup(path);
151         int result = safe_create_leading_directories(buf);
152         free(buf);
153         return result;
154 }
155
156 static void fill_sha1_path(char *pathbuf, const unsigned char *sha1)
157 {
158         int i;
159         for (i = 0; i < 20; i++) {
160                 static char hex[] = "0123456789abcdef";
161                 unsigned int val = sha1[i];
162                 char *pos = pathbuf + i*2 + (i > 0);
163                 *pos++ = hex[val >> 4];
164                 *pos = hex[val & 0xf];
165         }
166 }
167
168 /*
169  * NOTE! This returns a statically allocated buffer, so you have to be
170  * careful about using it. Do an "xstrdup()" if you need to save the
171  * filename.
172  *
173  * Also note that this returns the location for creating.  Reading
174  * SHA1 file can happen from any alternate directory listed in the
175  * DB_ENVIRONMENT environment variable if it is not found in
176  * the primary object database.
177  */
178 char *sha1_file_name(const unsigned char *sha1)
179 {
180         static char buf[PATH_MAX];
181         const char *objdir;
182         int len;
183
184         objdir = get_object_directory();
185         len = strlen(objdir);
186
187         /* '/' + sha1(2) + '/' + sha1(38) + '\0' */
188         if (len + 43 > PATH_MAX)
189                 die("insanely long object directory %s", objdir);
190         memcpy(buf, objdir, len);
191         buf[len] = '/';
192         buf[len+3] = '/';
193         buf[len+42] = '\0';
194         fill_sha1_path(buf + len + 1, sha1);
195         return buf;
196 }
197
198 static char *sha1_get_pack_name(const unsigned char *sha1,
199                                 char **name, char **base, const char *which)
200 {
201         static const char hex[] = "0123456789abcdef";
202         char *buf;
203         int i;
204
205         if (!*base) {
206                 const char *sha1_file_directory = get_object_directory();
207                 int len = strlen(sha1_file_directory);
208                 *base = xmalloc(len + 60);
209                 sprintf(*base, "%s/pack/pack-1234567890123456789012345678901234567890.%s",
210                         sha1_file_directory, which);
211                 *name = *base + len + 11;
212         }
213
214         buf = *name;
215
216         for (i = 0; i < 20; i++) {
217                 unsigned int val = *sha1++;
218                 *buf++ = hex[val >> 4];
219                 *buf++ = hex[val & 0xf];
220         }
221
222         return *base;
223 }
224
225 char *sha1_pack_name(const unsigned char *sha1)
226 {
227         static char *name, *base;
228
229         return sha1_get_pack_name(sha1, &name, &base, "pack");
230 }
231
232 char *sha1_pack_index_name(const unsigned char *sha1)
233 {
234         static char *name, *base;
235
236         return sha1_get_pack_name(sha1, &name, &base, "idx");
237 }
238
239 struct alternate_object_database *alt_odb_list;
240 static struct alternate_object_database **alt_odb_tail;
241
242 static int git_open_noatime(const char *name);
243
244 /*
245  * Prepare alternate object database registry.
246  *
247  * The variable alt_odb_list points at the list of struct
248  * alternate_object_database.  The elements on this list come from
249  * non-empty elements from colon separated ALTERNATE_DB_ENVIRONMENT
250  * environment variable, and $GIT_OBJECT_DIRECTORY/info/alternates,
251  * whose contents is similar to that environment variable but can be
252  * LF separated.  Its base points at a statically allocated buffer that
253  * contains "/the/directory/corresponding/to/.git/objects/...", while
254  * its name points just after the slash at the end of ".git/objects/"
255  * in the example above, and has enough space to hold 40-byte hex
256  * SHA1, an extra slash for the first level indirection, and the
257  * terminating NUL.
258  */
259 static int link_alt_odb_entry(const char *entry, const char *relative_base, int depth)
260 {
261         const char *objdir = get_object_directory();
262         struct alternate_object_database *ent;
263         struct alternate_object_database *alt;
264         int pfxlen, entlen;
265         struct strbuf pathbuf = STRBUF_INIT;
266
267         if (!is_absolute_path(entry) && relative_base) {
268                 strbuf_addstr(&pathbuf, real_path(relative_base));
269                 strbuf_addch(&pathbuf, '/');
270         }
271         strbuf_addstr(&pathbuf, entry);
272
273         normalize_path_copy(pathbuf.buf, pathbuf.buf);
274
275         pfxlen = strlen(pathbuf.buf);
276
277         /*
278          * The trailing slash after the directory name is given by
279          * this function at the end. Remove duplicates.
280          */
281         while (pfxlen && pathbuf.buf[pfxlen-1] == '/')
282                 pfxlen -= 1;
283
284         entlen = pfxlen + 43; /* '/' + 2 hex + '/' + 38 hex + NUL */
285         ent = xmalloc(sizeof(*ent) + entlen);
286         memcpy(ent->base, pathbuf.buf, pfxlen);
287         strbuf_release(&pathbuf);
288
289         ent->name = ent->base + pfxlen + 1;
290         ent->base[pfxlen + 3] = '/';
291         ent->base[pfxlen] = ent->base[entlen-1] = 0;
292
293         /* Detect cases where alternate disappeared */
294         if (!is_directory(ent->base)) {
295                 error("object directory %s does not exist; "
296                       "check .git/objects/info/alternates.",
297                       ent->base);
298                 free(ent);
299                 return -1;
300         }
301
302         /* Prevent the common mistake of listing the same
303          * thing twice, or object directory itself.
304          */
305         for (alt = alt_odb_list; alt; alt = alt->next) {
306                 if (!memcmp(ent->base, alt->base, pfxlen)) {
307                         free(ent);
308                         return -1;
309                 }
310         }
311         if (!strcmp(ent->base, objdir)) {
312                 free(ent);
313                 return -1;
314         }
315
316         /* add the alternate entry */
317         *alt_odb_tail = ent;
318         alt_odb_tail = &(ent->next);
319         ent->next = NULL;
320
321         /* recursively add alternates */
322         read_info_alternates(ent->base, depth + 1);
323
324         ent->base[pfxlen] = '/';
325
326         return 0;
327 }
328
329 static void link_alt_odb_entries(const char *alt, int len, int sep,
330                                  const char *relative_base, int depth)
331 {
332         struct string_list entries = STRING_LIST_INIT_NODUP;
333         char *alt_copy;
334         int i;
335
336         if (depth > 5) {
337                 error("%s: ignoring alternate object stores, nesting too deep.",
338                                 relative_base);
339                 return;
340         }
341
342         alt_copy = xmemdupz(alt, len);
343         string_list_split_in_place(&entries, alt_copy, sep, -1);
344         for (i = 0; i < entries.nr; i++) {
345                 const char *entry = entries.items[i].string;
346                 if (entry[0] == '\0' || entry[0] == '#')
347                         continue;
348                 if (!is_absolute_path(entry) && depth) {
349                         error("%s: ignoring relative alternate object store %s",
350                                         relative_base, entry);
351                 } else {
352                         link_alt_odb_entry(entry, relative_base, depth);
353                 }
354         }
355         string_list_clear(&entries, 0);
356         free(alt_copy);
357 }
358
359 void read_info_alternates(const char * relative_base, int depth)
360 {
361         char *map;
362         size_t mapsz;
363         struct stat st;
364         const char alt_file_name[] = "info/alternates";
365         /* Given that relative_base is no longer than PATH_MAX,
366            ensure that "path" has enough space to append "/", the
367            file name, "info/alternates", and a trailing NUL.  */
368         char path[PATH_MAX + 1 + sizeof alt_file_name];
369         int fd;
370
371         sprintf(path, "%s/%s", relative_base, alt_file_name);
372         fd = git_open_noatime(path);
373         if (fd < 0)
374                 return;
375         if (fstat(fd, &st) || (st.st_size == 0)) {
376                 close(fd);
377                 return;
378         }
379         mapsz = xsize_t(st.st_size);
380         map = xmmap(NULL, mapsz, PROT_READ, MAP_PRIVATE, fd, 0);
381         close(fd);
382
383         link_alt_odb_entries(map, mapsz, '\n', relative_base, depth);
384
385         munmap(map, mapsz);
386 }
387
388 void add_to_alternates_file(const char *reference)
389 {
390         struct lock_file *lock = xcalloc(1, sizeof(struct lock_file));
391         int fd = hold_lock_file_for_append(lock, git_path("objects/info/alternates"), LOCK_DIE_ON_ERROR);
392         char *alt = mkpath("%s\n", reference);
393         write_or_die(fd, alt, strlen(alt));
394         if (commit_lock_file(lock))
395                 die("could not close alternates file");
396         if (alt_odb_tail)
397                 link_alt_odb_entries(alt, strlen(alt), '\n', NULL, 0);
398 }
399
400 void foreach_alt_odb(alt_odb_fn fn, void *cb)
401 {
402         struct alternate_object_database *ent;
403
404         prepare_alt_odb();
405         for (ent = alt_odb_list; ent; ent = ent->next)
406                 if (fn(ent, cb))
407                         return;
408 }
409
410 void prepare_alt_odb(void)
411 {
412         const char *alt;
413
414         if (alt_odb_tail)
415                 return;
416
417         alt = getenv(ALTERNATE_DB_ENVIRONMENT);
418         if (!alt) alt = "";
419
420         alt_odb_tail = &alt_odb_list;
421         link_alt_odb_entries(alt, strlen(alt), PATH_SEP, NULL, 0);
422
423         read_info_alternates(get_object_directory(), 0);
424 }
425
426 static int has_loose_object_local(const unsigned char *sha1)
427 {
428         char *name = sha1_file_name(sha1);
429         return !access(name, F_OK);
430 }
431
432 int has_loose_object_nonlocal(const unsigned char *sha1)
433 {
434         struct alternate_object_database *alt;
435         prepare_alt_odb();
436         for (alt = alt_odb_list; alt; alt = alt->next) {
437                 fill_sha1_path(alt->name, sha1);
438                 if (!access(alt->base, F_OK))
439                         return 1;
440         }
441         return 0;
442 }
443
444 static int has_loose_object(const unsigned char *sha1)
445 {
446         return has_loose_object_local(sha1) ||
447                has_loose_object_nonlocal(sha1);
448 }
449
450 static unsigned int pack_used_ctr;
451 static unsigned int pack_mmap_calls;
452 static unsigned int peak_pack_open_windows;
453 static unsigned int pack_open_windows;
454 static unsigned int pack_open_fds;
455 static unsigned int pack_max_fds;
456 static size_t peak_pack_mapped;
457 static size_t pack_mapped;
458 struct packed_git *packed_git;
459
460 void pack_report(void)
461 {
462         fprintf(stderr,
463                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
464                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
465                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
466                 sz_fmt(getpagesize()),
467                 sz_fmt(packed_git_window_size),
468                 sz_fmt(packed_git_limit));
469         fprintf(stderr,
470                 "pack_report: pack_used_ctr            = %10u\n"
471                 "pack_report: pack_mmap_calls          = %10u\n"
472                 "pack_report: pack_open_windows        = %10u / %10u\n"
473                 "pack_report: pack_mapped              = "
474                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
475                 pack_used_ctr,
476                 pack_mmap_calls,
477                 pack_open_windows, peak_pack_open_windows,
478                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
479 }
480
481 static int check_packed_git_idx(const char *path,  struct packed_git *p)
482 {
483         void *idx_map;
484         struct pack_idx_header *hdr;
485         size_t idx_size;
486         uint32_t version, nr, i, *index;
487         int fd = git_open_noatime(path);
488         struct stat st;
489
490         if (fd < 0)
491                 return -1;
492         if (fstat(fd, &st)) {
493                 close(fd);
494                 return -1;
495         }
496         idx_size = xsize_t(st.st_size);
497         if (idx_size < 4 * 256 + 20 + 20) {
498                 close(fd);
499                 return error("index file %s is too small", path);
500         }
501         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
502         close(fd);
503
504         hdr = idx_map;
505         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
506                 version = ntohl(hdr->idx_version);
507                 if (version < 2 || version > 2) {
508                         munmap(idx_map, idx_size);
509                         return error("index file %s is version %"PRIu32
510                                      " and is not supported by this binary"
511                                      " (try upgrading GIT to a newer version)",
512                                      path, version);
513                 }
514         } else
515                 version = 1;
516
517         nr = 0;
518         index = idx_map;
519         if (version > 1)
520                 index += 2;  /* skip index header */
521         for (i = 0; i < 256; i++) {
522                 uint32_t n = ntohl(index[i]);
523                 if (n < nr) {
524                         munmap(idx_map, idx_size);
525                         return error("non-monotonic index %s", path);
526                 }
527                 nr = n;
528         }
529
530         if (version == 1) {
531                 /*
532                  * Total size:
533                  *  - 256 index entries 4 bytes each
534                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
535                  *  - 20-byte SHA1 of the packfile
536                  *  - 20-byte SHA1 file checksum
537                  */
538                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
539                         munmap(idx_map, idx_size);
540                         return error("wrong index v1 file size in %s", path);
541                 }
542         } else if (version == 2) {
543                 /*
544                  * Minimum size:
545                  *  - 8 bytes of header
546                  *  - 256 index entries 4 bytes each
547                  *  - 20-byte sha1 entry * nr
548                  *  - 4-byte crc entry * nr
549                  *  - 4-byte offset entry * nr
550                  *  - 20-byte SHA1 of the packfile
551                  *  - 20-byte SHA1 file checksum
552                  * And after the 4-byte offset table might be a
553                  * variable sized table containing 8-byte entries
554                  * for offsets larger than 2^31.
555                  */
556                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
557                 unsigned long max_size = min_size;
558                 if (nr)
559                         max_size += (nr - 1)*8;
560                 if (idx_size < min_size || idx_size > max_size) {
561                         munmap(idx_map, idx_size);
562                         return error("wrong index v2 file size in %s", path);
563                 }
564                 if (idx_size != min_size &&
565                     /*
566                      * make sure we can deal with large pack offsets.
567                      * 31-bit signed offset won't be enough, neither
568                      * 32-bit unsigned one will be.
569                      */
570                     (sizeof(off_t) <= 4)) {
571                         munmap(idx_map, idx_size);
572                         return error("pack too large for current definition of off_t in %s", path);
573                 }
574         }
575
576         p->index_version = version;
577         p->index_data = idx_map;
578         p->index_size = idx_size;
579         p->num_objects = nr;
580         return 0;
581 }
582
583 int open_pack_index(struct packed_git *p)
584 {
585         char *idx_name;
586         int ret;
587
588         if (p->index_data)
589                 return 0;
590
591         idx_name = xstrdup(p->pack_name);
592         strcpy(idx_name + strlen(idx_name) - strlen(".pack"), ".idx");
593         ret = check_packed_git_idx(idx_name, p);
594         free(idx_name);
595         return ret;
596 }
597
598 static void scan_windows(struct packed_git *p,
599         struct packed_git **lru_p,
600         struct pack_window **lru_w,
601         struct pack_window **lru_l)
602 {
603         struct pack_window *w, *w_l;
604
605         for (w_l = NULL, w = p->windows; w; w = w->next) {
606                 if (!w->inuse_cnt) {
607                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
608                                 *lru_p = p;
609                                 *lru_w = w;
610                                 *lru_l = w_l;
611                         }
612                 }
613                 w_l = w;
614         }
615 }
616
617 static int unuse_one_window(struct packed_git *current)
618 {
619         struct packed_git *p, *lru_p = NULL;
620         struct pack_window *lru_w = NULL, *lru_l = NULL;
621
622         if (current)
623                 scan_windows(current, &lru_p, &lru_w, &lru_l);
624         for (p = packed_git; p; p = p->next)
625                 scan_windows(p, &lru_p, &lru_w, &lru_l);
626         if (lru_p) {
627                 munmap(lru_w->base, lru_w->len);
628                 pack_mapped -= lru_w->len;
629                 if (lru_l)
630                         lru_l->next = lru_w->next;
631                 else
632                         lru_p->windows = lru_w->next;
633                 free(lru_w);
634                 pack_open_windows--;
635                 return 1;
636         }
637         return 0;
638 }
639
640 void release_pack_memory(size_t need)
641 {
642         size_t cur = pack_mapped;
643         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
644                 ; /* nothing */
645 }
646
647 void *xmmap(void *start, size_t length,
648         int prot, int flags, int fd, off_t offset)
649 {
650         void *ret = mmap(start, length, prot, flags, fd, offset);
651         if (ret == MAP_FAILED) {
652                 if (!length)
653                         return NULL;
654                 release_pack_memory(length);
655                 ret = mmap(start, length, prot, flags, fd, offset);
656                 if (ret == MAP_FAILED)
657                         die_errno("Out of memory? mmap failed");
658         }
659         return ret;
660 }
661
662 void close_pack_windows(struct packed_git *p)
663 {
664         while (p->windows) {
665                 struct pack_window *w = p->windows;
666
667                 if (w->inuse_cnt)
668                         die("pack '%s' still has open windows to it",
669                             p->pack_name);
670                 munmap(w->base, w->len);
671                 pack_mapped -= w->len;
672                 pack_open_windows--;
673                 p->windows = w->next;
674                 free(w);
675         }
676 }
677
678 /*
679  * The LRU pack is the one with the oldest MRU window, preferring packs
680  * with no used windows, or the oldest mtime if it has no windows allocated.
681  */
682 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
683 {
684         struct pack_window *w, *this_mru_w;
685         int has_windows_inuse = 0;
686
687         /*
688          * Reject this pack if it has windows and the previously selected
689          * one does not.  If this pack does not have windows, reject
690          * it if the pack file is newer than the previously selected one.
691          */
692         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
693                 return;
694
695         for (w = this_mru_w = p->windows; w; w = w->next) {
696                 /*
697                  * Reject this pack if any of its windows are in use,
698                  * but the previously selected pack did not have any
699                  * inuse windows.  Otherwise, record that this pack
700                  * has windows in use.
701                  */
702                 if (w->inuse_cnt) {
703                         if (*accept_windows_inuse)
704                                 has_windows_inuse = 1;
705                         else
706                                 return;
707                 }
708
709                 if (w->last_used > this_mru_w->last_used)
710                         this_mru_w = w;
711
712                 /*
713                  * Reject this pack if it has windows that have been
714                  * used more recently than the previously selected pack.
715                  * If the previously selected pack had windows inuse and
716                  * we have not encountered a window in this pack that is
717                  * inuse, skip this check since we prefer a pack with no
718                  * inuse windows to one that has inuse windows.
719                  */
720                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
721                     this_mru_w->last_used > (*mru_w)->last_used)
722                         return;
723         }
724
725         /*
726          * Select this pack.
727          */
728         *mru_w = this_mru_w;
729         *lru_p = p;
730         *accept_windows_inuse = has_windows_inuse;
731 }
732
733 static int close_one_pack(void)
734 {
735         struct packed_git *p, *lru_p = NULL;
736         struct pack_window *mru_w = NULL;
737         int accept_windows_inuse = 1;
738
739         for (p = packed_git; p; p = p->next) {
740                 if (p->pack_fd == -1)
741                         continue;
742                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
743         }
744
745         if (lru_p) {
746                 close(lru_p->pack_fd);
747                 pack_open_fds--;
748                 lru_p->pack_fd = -1;
749                 return 1;
750         }
751
752         return 0;
753 }
754
755 void unuse_pack(struct pack_window **w_cursor)
756 {
757         struct pack_window *w = *w_cursor;
758         if (w) {
759                 w->inuse_cnt--;
760                 *w_cursor = NULL;
761         }
762 }
763
764 void close_pack_index(struct packed_git *p)
765 {
766         if (p->index_data) {
767                 munmap((void *)p->index_data, p->index_size);
768                 p->index_data = NULL;
769         }
770 }
771
772 /*
773  * This is used by git-repack in case a newly created pack happens to
774  * contain the same set of objects as an existing one.  In that case
775  * the resulting file might be different even if its name would be the
776  * same.  It is best to close any reference to the old pack before it is
777  * replaced on disk.  Of course no index pointers nor windows for given pack
778  * must subsist at this point.  If ever objects from this pack are requested
779  * again, the new version of the pack will be reinitialized through
780  * reprepare_packed_git().
781  */
782 void free_pack_by_name(const char *pack_name)
783 {
784         struct packed_git *p, **pp = &packed_git;
785
786         while (*pp) {
787                 p = *pp;
788                 if (strcmp(pack_name, p->pack_name) == 0) {
789                         clear_delta_base_cache();
790                         close_pack_windows(p);
791                         if (p->pack_fd != -1) {
792                                 close(p->pack_fd);
793                                 pack_open_fds--;
794                         }
795                         close_pack_index(p);
796                         free(p->bad_object_sha1);
797                         *pp = p->next;
798                         if (last_found_pack == p)
799                                 last_found_pack = NULL;
800                         free(p);
801                         return;
802                 }
803                 pp = &p->next;
804         }
805 }
806
807 static unsigned int get_max_fd_limit(void)
808 {
809 #ifdef RLIMIT_NOFILE
810         struct rlimit lim;
811
812         if (getrlimit(RLIMIT_NOFILE, &lim))
813                 die_errno("cannot get RLIMIT_NOFILE");
814
815         return lim.rlim_cur;
816 #elif defined(_SC_OPEN_MAX)
817         return sysconf(_SC_OPEN_MAX);
818 #elif defined(OPEN_MAX)
819         return OPEN_MAX;
820 #else
821         return 1; /* see the caller ;-) */
822 #endif
823 }
824
825 /*
826  * Do not call this directly as this leaks p->pack_fd on error return;
827  * call open_packed_git() instead.
828  */
829 static int open_packed_git_1(struct packed_git *p)
830 {
831         struct stat st;
832         struct pack_header hdr;
833         unsigned char sha1[20];
834         unsigned char *idx_sha1;
835         long fd_flag;
836
837         if (!p->index_data && open_pack_index(p))
838                 return error("packfile %s index unavailable", p->pack_name);
839
840         if (!pack_max_fds) {
841                 unsigned int max_fds = get_max_fd_limit();
842
843                 /* Save 3 for stdin/stdout/stderr, 22 for work */
844                 if (25 < max_fds)
845                         pack_max_fds = max_fds - 25;
846                 else
847                         pack_max_fds = 1;
848         }
849
850         while (pack_max_fds <= pack_open_fds && close_one_pack())
851                 ; /* nothing */
852
853         p->pack_fd = git_open_noatime(p->pack_name);
854         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
855                 return -1;
856         pack_open_fds++;
857
858         /* If we created the struct before we had the pack we lack size. */
859         if (!p->pack_size) {
860                 if (!S_ISREG(st.st_mode))
861                         return error("packfile %s not a regular file", p->pack_name);
862                 p->pack_size = st.st_size;
863         } else if (p->pack_size != st.st_size)
864                 return error("packfile %s size changed", p->pack_name);
865
866         /* We leave these file descriptors open with sliding mmap;
867          * there is no point keeping them open across exec(), though.
868          */
869         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
870         if (fd_flag < 0)
871                 return error("cannot determine file descriptor flags");
872         fd_flag |= FD_CLOEXEC;
873         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
874                 return error("cannot set FD_CLOEXEC");
875
876         /* Verify we recognize this pack file format. */
877         if (read_in_full(p->pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
878                 return error("file %s is far too short to be a packfile", p->pack_name);
879         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
880                 return error("file %s is not a GIT packfile", p->pack_name);
881         if (!pack_version_ok(hdr.hdr_version))
882                 return error("packfile %s is version %"PRIu32" and not"
883                         " supported (try upgrading GIT to a newer version)",
884                         p->pack_name, ntohl(hdr.hdr_version));
885
886         /* Verify the pack matches its index. */
887         if (p->num_objects != ntohl(hdr.hdr_entries))
888                 return error("packfile %s claims to have %"PRIu32" objects"
889                              " while index indicates %"PRIu32" objects",
890                              p->pack_name, ntohl(hdr.hdr_entries),
891                              p->num_objects);
892         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
893                 return error("end of packfile %s is unavailable", p->pack_name);
894         if (read_in_full(p->pack_fd, sha1, sizeof(sha1)) != sizeof(sha1))
895                 return error("packfile %s signature is unavailable", p->pack_name);
896         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
897         if (hashcmp(sha1, idx_sha1))
898                 return error("packfile %s does not match index", p->pack_name);
899         return 0;
900 }
901
902 static int open_packed_git(struct packed_git *p)
903 {
904         if (!open_packed_git_1(p))
905                 return 0;
906         if (p->pack_fd != -1) {
907                 close(p->pack_fd);
908                 pack_open_fds--;
909                 p->pack_fd = -1;
910         }
911         return -1;
912 }
913
914 static int in_window(struct pack_window *win, off_t offset)
915 {
916         /* We must promise at least 20 bytes (one hash) after the
917          * offset is available from this window, otherwise the offset
918          * is not actually in this window and a different window (which
919          * has that one hash excess) must be used.  This is to support
920          * the object header and delta base parsing routines below.
921          */
922         off_t win_off = win->offset;
923         return win_off <= offset
924                 && (offset + 20) <= (win_off + win->len);
925 }
926
927 unsigned char *use_pack(struct packed_git *p,
928                 struct pack_window **w_cursor,
929                 off_t offset,
930                 unsigned long *left)
931 {
932         struct pack_window *win = *w_cursor;
933
934         /* Since packfiles end in a hash of their content and it's
935          * pointless to ask for an offset into the middle of that
936          * hash, and the in_window function above wouldn't match
937          * don't allow an offset too close to the end of the file.
938          */
939         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
940                 die("packfile %s cannot be accessed", p->pack_name);
941         if (offset > (p->pack_size - 20))
942                 die("offset beyond end of packfile (truncated pack?)");
943
944         if (!win || !in_window(win, offset)) {
945                 if (win)
946                         win->inuse_cnt--;
947                 for (win = p->windows; win; win = win->next) {
948                         if (in_window(win, offset))
949                                 break;
950                 }
951                 if (!win) {
952                         size_t window_align = packed_git_window_size / 2;
953                         off_t len;
954
955                         if (p->pack_fd == -1 && open_packed_git(p))
956                                 die("packfile %s cannot be accessed", p->pack_name);
957
958                         win = xcalloc(1, sizeof(*win));
959                         win->offset = (offset / window_align) * window_align;
960                         len = p->pack_size - win->offset;
961                         if (len > packed_git_window_size)
962                                 len = packed_git_window_size;
963                         win->len = (size_t)len;
964                         pack_mapped += win->len;
965                         while (packed_git_limit < pack_mapped
966                                 && unuse_one_window(p))
967                                 ; /* nothing */
968                         win->base = xmmap(NULL, win->len,
969                                 PROT_READ, MAP_PRIVATE,
970                                 p->pack_fd, win->offset);
971                         if (win->base == MAP_FAILED)
972                                 die("packfile %s cannot be mapped: %s",
973                                         p->pack_name,
974                                         strerror(errno));
975                         if (!win->offset && win->len == p->pack_size
976                                 && !p->do_not_close) {
977                                 close(p->pack_fd);
978                                 pack_open_fds--;
979                                 p->pack_fd = -1;
980                         }
981                         pack_mmap_calls++;
982                         pack_open_windows++;
983                         if (pack_mapped > peak_pack_mapped)
984                                 peak_pack_mapped = pack_mapped;
985                         if (pack_open_windows > peak_pack_open_windows)
986                                 peak_pack_open_windows = pack_open_windows;
987                         win->next = p->windows;
988                         p->windows = win;
989                 }
990         }
991         if (win != *w_cursor) {
992                 win->last_used = pack_used_ctr++;
993                 win->inuse_cnt++;
994                 *w_cursor = win;
995         }
996         offset -= win->offset;
997         if (left)
998                 *left = win->len - xsize_t(offset);
999         return win->base + offset;
1000 }
1001
1002 static struct packed_git *alloc_packed_git(int extra)
1003 {
1004         struct packed_git *p = xmalloc(sizeof(*p) + extra);
1005         memset(p, 0, sizeof(*p));
1006         p->pack_fd = -1;
1007         return p;
1008 }
1009
1010 static void try_to_free_pack_memory(size_t size)
1011 {
1012         release_pack_memory(size);
1013 }
1014
1015 struct packed_git *add_packed_git(const char *path, int path_len, int local)
1016 {
1017         static int have_set_try_to_free_routine;
1018         struct stat st;
1019         struct packed_git *p = alloc_packed_git(path_len + 2);
1020
1021         if (!have_set_try_to_free_routine) {
1022                 have_set_try_to_free_routine = 1;
1023                 set_try_to_free_routine(try_to_free_pack_memory);
1024         }
1025
1026         /*
1027          * Make sure a corresponding .pack file exists and that
1028          * the index looks sane.
1029          */
1030         path_len -= strlen(".idx");
1031         if (path_len < 1) {
1032                 free(p);
1033                 return NULL;
1034         }
1035         memcpy(p->pack_name, path, path_len);
1036
1037         strcpy(p->pack_name + path_len, ".keep");
1038         if (!access(p->pack_name, F_OK))
1039                 p->pack_keep = 1;
1040
1041         strcpy(p->pack_name + path_len, ".pack");
1042         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
1043                 free(p);
1044                 return NULL;
1045         }
1046
1047         /* ok, it looks sane as far as we can check without
1048          * actually mapping the pack file.
1049          */
1050         p->pack_size = st.st_size;
1051         p->pack_local = local;
1052         p->mtime = st.st_mtime;
1053         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
1054                 hashclr(p->sha1);
1055         return p;
1056 }
1057
1058 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
1059 {
1060         const char *path = sha1_pack_name(sha1);
1061         struct packed_git *p = alloc_packed_git(strlen(path) + 1);
1062
1063         strcpy(p->pack_name, path);
1064         hashcpy(p->sha1, sha1);
1065         if (check_packed_git_idx(idx_path, p)) {
1066                 free(p);
1067                 return NULL;
1068         }
1069
1070         return p;
1071 }
1072
1073 void install_packed_git(struct packed_git *pack)
1074 {
1075         if (pack->pack_fd != -1)
1076                 pack_open_fds++;
1077
1078         pack->next = packed_git;
1079         packed_git = pack;
1080 }
1081
1082 void (*report_garbage)(const char *desc, const char *path);
1083
1084 static void report_helper(const struct string_list *list,
1085                           int seen_bits, int first, int last)
1086 {
1087         const char *msg;
1088         switch (seen_bits) {
1089         case 0:
1090                 msg = "no corresponding .idx nor .pack";
1091                 break;
1092         case 1:
1093                 msg = "no corresponding .idx";
1094                 break;
1095         case 2:
1096                 msg = "no corresponding .pack";
1097                 break;
1098         default:
1099                 return;
1100         }
1101         for (; first < last; first++)
1102                 report_garbage(msg, list->items[first].string);
1103 }
1104
1105 static void report_pack_garbage(struct string_list *list)
1106 {
1107         int i, baselen = -1, first = 0, seen_bits = 0;
1108
1109         if (!report_garbage)
1110                 return;
1111
1112         sort_string_list(list);
1113
1114         for (i = 0; i < list->nr; i++) {
1115                 const char *path = list->items[i].string;
1116                 if (baselen != -1 &&
1117                     strncmp(path, list->items[first].string, baselen)) {
1118                         report_helper(list, seen_bits, first, i);
1119                         baselen = -1;
1120                         seen_bits = 0;
1121                 }
1122                 if (baselen == -1) {
1123                         const char *dot = strrchr(path, '.');
1124                         if (!dot) {
1125                                 report_garbage("garbage found", path);
1126                                 continue;
1127                         }
1128                         baselen = dot - path + 1;
1129                         first = i;
1130                 }
1131                 if (!strcmp(path + baselen, "pack"))
1132                         seen_bits |= 1;
1133                 else if (!strcmp(path + baselen, "idx"))
1134                         seen_bits |= 2;
1135         }
1136         report_helper(list, seen_bits, first, list->nr);
1137 }
1138
1139 static void prepare_packed_git_one(char *objdir, int local)
1140 {
1141         /* Ensure that this buffer is large enough so that we can
1142            append "/pack/" without clobbering the stack even if
1143            strlen(objdir) were PATH_MAX.  */
1144         char path[PATH_MAX + 1 + 4 + 1 + 1];
1145         int len;
1146         DIR *dir;
1147         struct dirent *de;
1148         struct string_list garbage = STRING_LIST_INIT_DUP;
1149
1150         sprintf(path, "%s/pack", objdir);
1151         len = strlen(path);
1152         dir = opendir(path);
1153         if (!dir) {
1154                 if (errno != ENOENT)
1155                         error("unable to open object pack directory: %s: %s",
1156                               path, strerror(errno));
1157                 return;
1158         }
1159         path[len++] = '/';
1160         while ((de = readdir(dir)) != NULL) {
1161                 int namelen = strlen(de->d_name);
1162                 struct packed_git *p;
1163
1164                 if (len + namelen + 1 > sizeof(path)) {
1165                         if (report_garbage) {
1166                                 struct strbuf sb = STRBUF_INIT;
1167                                 strbuf_addf(&sb, "%.*s/%s", len - 1, path, de->d_name);
1168                                 report_garbage("path too long", sb.buf);
1169                                 strbuf_release(&sb);
1170                         }
1171                         continue;
1172                 }
1173
1174                 if (is_dot_or_dotdot(de->d_name))
1175                         continue;
1176
1177                 strcpy(path + len, de->d_name);
1178
1179                 if (has_extension(de->d_name, ".idx")) {
1180                         /* Don't reopen a pack we already have. */
1181                         for (p = packed_git; p; p = p->next) {
1182                                 if (!memcmp(path, p->pack_name, len + namelen - 4))
1183                                         break;
1184                         }
1185                         if (p == NULL &&
1186                             /*
1187                              * See if it really is a valid .idx file with
1188                              * corresponding .pack file that we can map.
1189                              */
1190                             (p = add_packed_git(path, len + namelen, local)) != NULL)
1191                                 install_packed_git(p);
1192                 }
1193
1194                 if (!report_garbage)
1195                         continue;
1196
1197                 if (has_extension(de->d_name, ".idx") ||
1198                     has_extension(de->d_name, ".pack") ||
1199                     has_extension(de->d_name, ".keep"))
1200                         string_list_append(&garbage, path);
1201                 else
1202                         report_garbage("garbage found", path);
1203         }
1204         closedir(dir);
1205         report_pack_garbage(&garbage);
1206         string_list_clear(&garbage, 0);
1207 }
1208
1209 static int sort_pack(const void *a_, const void *b_)
1210 {
1211         struct packed_git *a = *((struct packed_git **)a_);
1212         struct packed_git *b = *((struct packed_git **)b_);
1213         int st;
1214
1215         /*
1216          * Local packs tend to contain objects specific to our
1217          * variant of the project than remote ones.  In addition,
1218          * remote ones could be on a network mounted filesystem.
1219          * Favor local ones for these reasons.
1220          */
1221         st = a->pack_local - b->pack_local;
1222         if (st)
1223                 return -st;
1224
1225         /*
1226          * Younger packs tend to contain more recent objects,
1227          * and more recent objects tend to get accessed more
1228          * often.
1229          */
1230         if (a->mtime < b->mtime)
1231                 return 1;
1232         else if (a->mtime == b->mtime)
1233                 return 0;
1234         return -1;
1235 }
1236
1237 static void rearrange_packed_git(void)
1238 {
1239         struct packed_git **ary, *p;
1240         int i, n;
1241
1242         for (n = 0, p = packed_git; p; p = p->next)
1243                 n++;
1244         if (n < 2)
1245                 return;
1246
1247         /* prepare an array of packed_git for easier sorting */
1248         ary = xcalloc(n, sizeof(struct packed_git *));
1249         for (n = 0, p = packed_git; p; p = p->next)
1250                 ary[n++] = p;
1251
1252         qsort(ary, n, sizeof(struct packed_git *), sort_pack);
1253
1254         /* link them back again */
1255         for (i = 0; i < n - 1; i++)
1256                 ary[i]->next = ary[i + 1];
1257         ary[n - 1]->next = NULL;
1258         packed_git = ary[0];
1259
1260         free(ary);
1261 }
1262
1263 static int prepare_packed_git_run_once = 0;
1264 void prepare_packed_git(void)
1265 {
1266         struct alternate_object_database *alt;
1267
1268         if (prepare_packed_git_run_once)
1269                 return;
1270         prepare_packed_git_one(get_object_directory(), 1);
1271         prepare_alt_odb();
1272         for (alt = alt_odb_list; alt; alt = alt->next) {
1273                 alt->name[-1] = 0;
1274                 prepare_packed_git_one(alt->base, 0);
1275                 alt->name[-1] = '/';
1276         }
1277         rearrange_packed_git();
1278         prepare_packed_git_run_once = 1;
1279 }
1280
1281 void reprepare_packed_git(void)
1282 {
1283         discard_revindex();
1284         prepare_packed_git_run_once = 0;
1285         prepare_packed_git();
1286 }
1287
1288 static void mark_bad_packed_object(struct packed_git *p,
1289                                    const unsigned char *sha1)
1290 {
1291         unsigned i;
1292         for (i = 0; i < p->num_bad_objects; i++)
1293                 if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1294                         return;
1295         p->bad_object_sha1 = xrealloc(p->bad_object_sha1, 20 * (p->num_bad_objects + 1));
1296         hashcpy(p->bad_object_sha1 + 20 * p->num_bad_objects, sha1);
1297         p->num_bad_objects++;
1298 }
1299
1300 static const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1301 {
1302         struct packed_git *p;
1303         unsigned i;
1304
1305         for (p = packed_git; p; p = p->next)
1306                 for (i = 0; i < p->num_bad_objects; i++)
1307                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1308                                 return p;
1309         return NULL;
1310 }
1311
1312 /*
1313  * With an in-core object data in "map", rehash it to make sure the
1314  * object name actually matches "sha1" to detect object corruption.
1315  * With "map" == NULL, try reading the object named with "sha1" using
1316  * the streaming interface and rehash it to do the same.
1317  */
1318 int check_sha1_signature(const unsigned char *sha1, void *map,
1319                          unsigned long size, const char *type)
1320 {
1321         unsigned char real_sha1[20];
1322         enum object_type obj_type;
1323         struct git_istream *st;
1324         git_SHA_CTX c;
1325         char hdr[32];
1326         int hdrlen;
1327
1328         if (map) {
1329                 hash_sha1_file(map, size, type, real_sha1);
1330                 return hashcmp(sha1, real_sha1) ? -1 : 0;
1331         }
1332
1333         st = open_istream(sha1, &obj_type, &size, NULL);
1334         if (!st)
1335                 return -1;
1336
1337         /* Generate the header */
1338         hdrlen = sprintf(hdr, "%s %lu", typename(obj_type), size) + 1;
1339
1340         /* Sha1.. */
1341         git_SHA1_Init(&c);
1342         git_SHA1_Update(&c, hdr, hdrlen);
1343         for (;;) {
1344                 char buf[1024 * 16];
1345                 ssize_t readlen = read_istream(st, buf, sizeof(buf));
1346
1347                 if (readlen < 0) {
1348                         close_istream(st);
1349                         return -1;
1350                 }
1351                 if (!readlen)
1352                         break;
1353                 git_SHA1_Update(&c, buf, readlen);
1354         }
1355         git_SHA1_Final(real_sha1, &c);
1356         close_istream(st);
1357         return hashcmp(sha1, real_sha1) ? -1 : 0;
1358 }
1359
1360 static int git_open_noatime(const char *name)
1361 {
1362         static int sha1_file_open_flag = O_NOATIME;
1363
1364         for (;;) {
1365                 int fd = open(name, O_RDONLY | sha1_file_open_flag);
1366                 if (fd >= 0)
1367                         return fd;
1368
1369                 /* Might the failure be due to O_NOATIME? */
1370                 if (errno != ENOENT && sha1_file_open_flag) {
1371                         sha1_file_open_flag = 0;
1372                         continue;
1373                 }
1374
1375                 return -1;
1376         }
1377 }
1378
1379 static int stat_sha1_file(const unsigned char *sha1, struct stat *st)
1380 {
1381         char *name = sha1_file_name(sha1);
1382         struct alternate_object_database *alt;
1383
1384         if (!lstat(name, st))
1385                 return 0;
1386
1387         prepare_alt_odb();
1388         errno = ENOENT;
1389         for (alt = alt_odb_list; alt; alt = alt->next) {
1390                 name = alt->name;
1391                 fill_sha1_path(name, sha1);
1392                 if (!lstat(alt->base, st))
1393                         return 0;
1394         }
1395
1396         return -1;
1397 }
1398
1399 static int open_sha1_file(const unsigned char *sha1)
1400 {
1401         int fd;
1402         char *name = sha1_file_name(sha1);
1403         struct alternate_object_database *alt;
1404
1405         fd = git_open_noatime(name);
1406         if (fd >= 0)
1407                 return fd;
1408
1409         prepare_alt_odb();
1410         errno = ENOENT;
1411         for (alt = alt_odb_list; alt; alt = alt->next) {
1412                 name = alt->name;
1413                 fill_sha1_path(name, sha1);
1414                 fd = git_open_noatime(alt->base);
1415                 if (fd >= 0)
1416                         return fd;
1417         }
1418         return -1;
1419 }
1420
1421 void *map_sha1_file(const unsigned char *sha1, unsigned long *size)
1422 {
1423         void *map;
1424         int fd;
1425
1426         fd = open_sha1_file(sha1);
1427         map = NULL;
1428         if (fd >= 0) {
1429                 struct stat st;
1430
1431                 if (!fstat(fd, &st)) {
1432                         *size = xsize_t(st.st_size);
1433                         if (!*size) {
1434                                 /* mmap() is forbidden on empty files */
1435                                 error("object file %s is empty", sha1_file_name(sha1));
1436                                 return NULL;
1437                         }
1438                         map = xmmap(NULL, *size, PROT_READ, MAP_PRIVATE, fd, 0);
1439                 }
1440                 close(fd);
1441         }
1442         return map;
1443 }
1444
1445 /*
1446  * There used to be a second loose object header format which
1447  * was meant to mimic the in-pack format, allowing for direct
1448  * copy of the object data.  This format turned up not to be
1449  * really worth it and we no longer write loose objects in that
1450  * format.
1451  */
1452 static int experimental_loose_object(unsigned char *map)
1453 {
1454         unsigned int word;
1455
1456         /*
1457          * We must determine if the buffer contains the standard
1458          * zlib-deflated stream or the experimental format based
1459          * on the in-pack object format. Compare the header byte
1460          * for each format:
1461          *
1462          * RFC1950 zlib w/ deflate : 0www1000 : 0 <= www <= 7
1463          * Experimental pack-based : Stttssss : ttt = 1,2,3,4
1464          *
1465          * If bit 7 is clear and bits 0-3 equal 8, the buffer MUST be
1466          * in standard loose-object format, UNLESS it is a Git-pack
1467          * format object *exactly* 8 bytes in size when inflated.
1468          *
1469          * However, RFC1950 also specifies that the 1st 16-bit word
1470          * must be divisible by 31 - this checksum tells us our buffer
1471          * is in the standard format, giving a false positive only if
1472          * the 1st word of the Git-pack format object happens to be
1473          * divisible by 31, ie:
1474          *      ((byte0 * 256) + byte1) % 31 = 0
1475          *   =>        0ttt10000www1000 % 31 = 0
1476          *
1477          * As it happens, this case can only arise for www=3 & ttt=1
1478          * - ie, a Commit object, which would have to be 8 bytes in
1479          * size. As no Commit can be that small, we find that the
1480          * combination of these two criteria (bitmask & checksum)
1481          * can always correctly determine the buffer format.
1482          */
1483         word = (map[0] << 8) + map[1];
1484         if ((map[0] & 0x8F) == 0x08 && !(word % 31))
1485                 return 0;
1486         else
1487                 return 1;
1488 }
1489
1490 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1491                 unsigned long len, enum object_type *type, unsigned long *sizep)
1492 {
1493         unsigned shift;
1494         unsigned long size, c;
1495         unsigned long used = 0;
1496
1497         c = buf[used++];
1498         *type = (c >> 4) & 7;
1499         size = c & 15;
1500         shift = 4;
1501         while (c & 0x80) {
1502                 if (len <= used || bitsizeof(long) <= shift) {
1503                         error("bad object header");
1504                         size = used = 0;
1505                         break;
1506                 }
1507                 c = buf[used++];
1508                 size += (c & 0x7f) << shift;
1509                 shift += 7;
1510         }
1511         *sizep = size;
1512         return used;
1513 }
1514
1515 int unpack_sha1_header(git_zstream *stream, unsigned char *map, unsigned long mapsize, void *buffer, unsigned long bufsiz)
1516 {
1517         unsigned long size, used;
1518         static const char valid_loose_object_type[8] = {
1519                 0, /* OBJ_EXT */
1520                 1, 1, 1, 1, /* "commit", "tree", "blob", "tag" */
1521                 0, /* "delta" and others are invalid in a loose object */
1522         };
1523         enum object_type type;
1524
1525         /* Get the data stream */
1526         memset(stream, 0, sizeof(*stream));
1527         stream->next_in = map;
1528         stream->avail_in = mapsize;
1529         stream->next_out = buffer;
1530         stream->avail_out = bufsiz;
1531
1532         if (experimental_loose_object(map)) {
1533                 /*
1534                  * The old experimental format we no longer produce;
1535                  * we can still read it.
1536                  */
1537                 used = unpack_object_header_buffer(map, mapsize, &type, &size);
1538                 if (!used || !valid_loose_object_type[type])
1539                         return -1;
1540                 map += used;
1541                 mapsize -= used;
1542
1543                 /* Set up the stream for the rest.. */
1544                 stream->next_in = map;
1545                 stream->avail_in = mapsize;
1546                 git_inflate_init(stream);
1547
1548                 /* And generate the fake traditional header */
1549                 stream->total_out = 1 + snprintf(buffer, bufsiz, "%s %lu",
1550                                                  typename(type), size);
1551                 return 0;
1552         }
1553         git_inflate_init(stream);
1554         return git_inflate(stream, 0);
1555 }
1556
1557 static void *unpack_sha1_rest(git_zstream *stream, void *buffer, unsigned long size, const unsigned char *sha1)
1558 {
1559         int bytes = strlen(buffer) + 1;
1560         unsigned char *buf = xmallocz(size);
1561         unsigned long n;
1562         int status = Z_OK;
1563
1564         n = stream->total_out - bytes;
1565         if (n > size)
1566                 n = size;
1567         memcpy(buf, (char *) buffer + bytes, n);
1568         bytes = n;
1569         if (bytes <= size) {
1570                 /*
1571                  * The above condition must be (bytes <= size), not
1572                  * (bytes < size).  In other words, even though we
1573                  * expect no more output and set avail_out to zero,
1574                  * the input zlib stream may have bytes that express
1575                  * "this concludes the stream", and we *do* want to
1576                  * eat that input.
1577                  *
1578                  * Otherwise we would not be able to test that we
1579                  * consumed all the input to reach the expected size;
1580                  * we also want to check that zlib tells us that all
1581                  * went well with status == Z_STREAM_END at the end.
1582                  */
1583                 stream->next_out = buf + bytes;
1584                 stream->avail_out = size - bytes;
1585                 while (status == Z_OK)
1586                         status = git_inflate(stream, Z_FINISH);
1587         }
1588         if (status == Z_STREAM_END && !stream->avail_in) {
1589                 git_inflate_end(stream);
1590                 return buf;
1591         }
1592
1593         if (status < 0)
1594                 error("corrupt loose object '%s'", sha1_to_hex(sha1));
1595         else if (stream->avail_in)
1596                 error("garbage at end of loose object '%s'",
1597                       sha1_to_hex(sha1));
1598         free(buf);
1599         return NULL;
1600 }
1601
1602 /*
1603  * We used to just use "sscanf()", but that's actually way
1604  * too permissive for what we want to check. So do an anal
1605  * object header parse by hand.
1606  */
1607 int parse_sha1_header(const char *hdr, unsigned long *sizep)
1608 {
1609         char type[10];
1610         int i;
1611         unsigned long size;
1612
1613         /*
1614          * The type can be at most ten bytes (including the
1615          * terminating '\0' that we add), and is followed by
1616          * a space.
1617          */
1618         i = 0;
1619         for (;;) {
1620                 char c = *hdr++;
1621                 if (c == ' ')
1622                         break;
1623                 type[i++] = c;
1624                 if (i >= sizeof(type))
1625                         return -1;
1626         }
1627         type[i] = 0;
1628
1629         /*
1630          * The length must follow immediately, and be in canonical
1631          * decimal format (ie "010" is not valid).
1632          */
1633         size = *hdr++ - '0';
1634         if (size > 9)
1635                 return -1;
1636         if (size) {
1637                 for (;;) {
1638                         unsigned long c = *hdr - '0';
1639                         if (c > 9)
1640                                 break;
1641                         hdr++;
1642                         size = size * 10 + c;
1643                 }
1644         }
1645         *sizep = size;
1646
1647         /*
1648          * The length must be followed by a zero byte
1649          */
1650         return *hdr ? -1 : type_from_string(type);
1651 }
1652
1653 static void *unpack_sha1_file(void *map, unsigned long mapsize, enum object_type *type, unsigned long *size, const unsigned char *sha1)
1654 {
1655         int ret;
1656         git_zstream stream;
1657         char hdr[8192];
1658
1659         ret = unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr));
1660         if (ret < Z_OK || (*type = parse_sha1_header(hdr, size)) < 0)
1661                 return NULL;
1662
1663         return unpack_sha1_rest(&stream, hdr, *size, sha1);
1664 }
1665
1666 unsigned long get_size_from_delta(struct packed_git *p,
1667                                   struct pack_window **w_curs,
1668                                   off_t curpos)
1669 {
1670         const unsigned char *data;
1671         unsigned char delta_head[20], *in;
1672         git_zstream stream;
1673         int st;
1674
1675         memset(&stream, 0, sizeof(stream));
1676         stream.next_out = delta_head;
1677         stream.avail_out = sizeof(delta_head);
1678
1679         git_inflate_init(&stream);
1680         do {
1681                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1682                 stream.next_in = in;
1683                 st = git_inflate(&stream, Z_FINISH);
1684                 curpos += stream.next_in - in;
1685         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1686                  stream.total_out < sizeof(delta_head));
1687         git_inflate_end(&stream);
1688         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1689                 error("delta data unpack-initial failed");
1690                 return 0;
1691         }
1692
1693         /* Examine the initial part of the delta to figure out
1694          * the result size.
1695          */
1696         data = delta_head;
1697
1698         /* ignore base size */
1699         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1700
1701         /* Read the result size */
1702         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1703 }
1704
1705 static off_t get_delta_base(struct packed_git *p,
1706                                     struct pack_window **w_curs,
1707                                     off_t *curpos,
1708                                     enum object_type type,
1709                                     off_t delta_obj_offset)
1710 {
1711         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1712         off_t base_offset;
1713
1714         /* use_pack() assured us we have [base_info, base_info + 20)
1715          * as a range that we can look at without walking off the
1716          * end of the mapped window.  Its actually the hash size
1717          * that is assured.  An OFS_DELTA longer than the hash size
1718          * is stupid, as then a REF_DELTA would be smaller to store.
1719          */
1720         if (type == OBJ_OFS_DELTA) {
1721                 unsigned used = 0;
1722                 unsigned char c = base_info[used++];
1723                 base_offset = c & 127;
1724                 while (c & 128) {
1725                         base_offset += 1;
1726                         if (!base_offset || MSB(base_offset, 7))
1727                                 return 0;  /* overflow */
1728                         c = base_info[used++];
1729                         base_offset = (base_offset << 7) + (c & 127);
1730                 }
1731                 base_offset = delta_obj_offset - base_offset;
1732                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1733                         return 0;  /* out of bound */
1734                 *curpos += used;
1735         } else if (type == OBJ_REF_DELTA) {
1736                 /* The base entry _must_ be in the same pack */
1737                 base_offset = find_pack_entry_one(base_info, p);
1738                 *curpos += 20;
1739         } else
1740                 die("I am totally screwed");
1741         return base_offset;
1742 }
1743
1744 int unpack_object_header(struct packed_git *p,
1745                          struct pack_window **w_curs,
1746                          off_t *curpos,
1747                          unsigned long *sizep)
1748 {
1749         unsigned char *base;
1750         unsigned long left;
1751         unsigned long used;
1752         enum object_type type;
1753
1754         /* use_pack() assures us we have [base, base + 20) available
1755          * as a range that we can look at.  (Its actually the hash
1756          * size that is assured.)  With our object header encoding
1757          * the maximum deflated object size is 2^137, which is just
1758          * insane, so we know won't exceed what we have been given.
1759          */
1760         base = use_pack(p, w_curs, *curpos, &left);
1761         used = unpack_object_header_buffer(base, left, &type, sizep);
1762         if (!used) {
1763                 type = OBJ_BAD;
1764         } else
1765                 *curpos += used;
1766
1767         return type;
1768 }
1769
1770 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
1771 {
1772         int type;
1773         struct revindex_entry *revidx;
1774         const unsigned char *sha1;
1775         revidx = find_pack_revindex(p, obj_offset);
1776         if (!revidx)
1777                 return OBJ_BAD;
1778         sha1 = nth_packed_object_sha1(p, revidx->nr);
1779         mark_bad_packed_object(p, sha1);
1780         type = sha1_object_info(sha1, NULL);
1781         if (type <= OBJ_NONE)
1782                 return OBJ_BAD;
1783         return type;
1784 }
1785
1786 #define POI_STACK_PREALLOC 64
1787
1788 static enum object_type packed_to_object_type(struct packed_git *p,
1789                                               off_t obj_offset,
1790                                               enum object_type type,
1791                                               struct pack_window **w_curs,
1792                                               off_t curpos)
1793 {
1794         off_t small_poi_stack[POI_STACK_PREALLOC];
1795         off_t *poi_stack = small_poi_stack;
1796         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1797
1798         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1799                 off_t base_offset;
1800                 unsigned long size;
1801                 /* Push the object we're going to leave behind */
1802                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1803                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1804                         poi_stack = xmalloc(sizeof(off_t)*poi_stack_alloc);
1805                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1806                 } else {
1807                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1808                 }
1809                 poi_stack[poi_stack_nr++] = obj_offset;
1810                 /* If parsing the base offset fails, just unwind */
1811                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1812                 if (!base_offset)
1813                         goto unwind;
1814                 curpos = obj_offset = base_offset;
1815                 type = unpack_object_header(p, w_curs, &curpos, &size);
1816                 if (type <= OBJ_NONE) {
1817                         /* If getting the base itself fails, we first
1818                          * retry the base, otherwise unwind */
1819                         type = retry_bad_packed_offset(p, base_offset);
1820                         if (type > OBJ_NONE)
1821                                 goto out;
1822                         goto unwind;
1823                 }
1824         }
1825
1826         switch (type) {
1827         case OBJ_BAD:
1828         case OBJ_COMMIT:
1829         case OBJ_TREE:
1830         case OBJ_BLOB:
1831         case OBJ_TAG:
1832                 break;
1833         default:
1834                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1835                       type, (uintmax_t)obj_offset, p->pack_name);
1836                 type = OBJ_BAD;
1837         }
1838
1839 out:
1840         if (poi_stack != small_poi_stack)
1841                 free(poi_stack);
1842         return type;
1843
1844 unwind:
1845         while (poi_stack_nr) {
1846                 obj_offset = poi_stack[--poi_stack_nr];
1847                 type = retry_bad_packed_offset(p, obj_offset);
1848                 if (type > OBJ_NONE)
1849                         goto out;
1850         }
1851         type = OBJ_BAD;
1852         goto out;
1853 }
1854
1855 static int packed_object_info(struct packed_git *p, off_t obj_offset,
1856                               struct object_info *oi)
1857 {
1858         struct pack_window *w_curs = NULL;
1859         unsigned long size;
1860         off_t curpos = obj_offset;
1861         enum object_type type;
1862
1863         /*
1864          * We always get the representation type, but only convert it to
1865          * a "real" type later if the caller is interested.
1866          */
1867         type = unpack_object_header(p, &w_curs, &curpos, &size);
1868
1869         if (oi->sizep) {
1870                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1871                         off_t tmp_pos = curpos;
1872                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1873                                                            type, obj_offset);
1874                         if (!base_offset) {
1875                                 type = OBJ_BAD;
1876                                 goto out;
1877                         }
1878                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1879                         if (*oi->sizep == 0) {
1880                                 type = OBJ_BAD;
1881                                 goto out;
1882                         }
1883                 } else {
1884                         *oi->sizep = size;
1885                 }
1886         }
1887
1888         if (oi->disk_sizep) {
1889                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1890                 *oi->disk_sizep = revidx[1].offset - obj_offset;
1891         }
1892
1893         if (oi->typep) {
1894                 *oi->typep = packed_to_object_type(p, obj_offset, type, &w_curs, curpos);
1895                 if (*oi->typep < 0) {
1896                         type = OBJ_BAD;
1897                         goto out;
1898                 }
1899         }
1900
1901 out:
1902         unuse_pack(&w_curs);
1903         return type;
1904 }
1905
1906 static void *unpack_compressed_entry(struct packed_git *p,
1907                                     struct pack_window **w_curs,
1908                                     off_t curpos,
1909                                     unsigned long size)
1910 {
1911         int st;
1912         git_zstream stream;
1913         unsigned char *buffer, *in;
1914
1915         buffer = xmallocz(size);
1916         memset(&stream, 0, sizeof(stream));
1917         stream.next_out = buffer;
1918         stream.avail_out = size + 1;
1919
1920         git_inflate_init(&stream);
1921         do {
1922                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1923                 stream.next_in = in;
1924                 st = git_inflate(&stream, Z_FINISH);
1925                 if (!stream.avail_out)
1926                         break; /* the payload is larger than it should be */
1927                 curpos += stream.next_in - in;
1928         } while (st == Z_OK || st == Z_BUF_ERROR);
1929         git_inflate_end(&stream);
1930         if ((st != Z_STREAM_END) || stream.total_out != size) {
1931                 free(buffer);
1932                 return NULL;
1933         }
1934
1935         return buffer;
1936 }
1937
1938 #define MAX_DELTA_CACHE (256)
1939
1940 static size_t delta_base_cached;
1941
1942 static struct delta_base_cache_lru_list {
1943         struct delta_base_cache_lru_list *prev;
1944         struct delta_base_cache_lru_list *next;
1945 } delta_base_cache_lru = { &delta_base_cache_lru, &delta_base_cache_lru };
1946
1947 static struct delta_base_cache_entry {
1948         struct delta_base_cache_lru_list lru;
1949         void *data;
1950         struct packed_git *p;
1951         off_t base_offset;
1952         unsigned long size;
1953         enum object_type type;
1954 } delta_base_cache[MAX_DELTA_CACHE];
1955
1956 static unsigned long pack_entry_hash(struct packed_git *p, off_t base_offset)
1957 {
1958         unsigned long hash;
1959
1960         hash = (unsigned long)p + (unsigned long)base_offset;
1961         hash += (hash >> 8) + (hash >> 16);
1962         return hash % MAX_DELTA_CACHE;
1963 }
1964
1965 static struct delta_base_cache_entry *
1966 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
1967 {
1968         unsigned long hash = pack_entry_hash(p, base_offset);
1969         return delta_base_cache + hash;
1970 }
1971
1972 static int eq_delta_base_cache_entry(struct delta_base_cache_entry *ent,
1973                                      struct packed_git *p, off_t base_offset)
1974 {
1975         return (ent->data && ent->p == p && ent->base_offset == base_offset);
1976 }
1977
1978 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
1979 {
1980         struct delta_base_cache_entry *ent;
1981         ent = get_delta_base_cache_entry(p, base_offset);
1982         return eq_delta_base_cache_entry(ent, p, base_offset);
1983 }
1984
1985 static void clear_delta_base_cache_entry(struct delta_base_cache_entry *ent)
1986 {
1987         ent->data = NULL;
1988         ent->lru.next->prev = ent->lru.prev;
1989         ent->lru.prev->next = ent->lru.next;
1990         delta_base_cached -= ent->size;
1991 }
1992
1993 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
1994         unsigned long *base_size, enum object_type *type, int keep_cache)
1995 {
1996         struct delta_base_cache_entry *ent;
1997         void *ret;
1998
1999         ent = get_delta_base_cache_entry(p, base_offset);
2000
2001         if (!eq_delta_base_cache_entry(ent, p, base_offset))
2002                 return unpack_entry(p, base_offset, type, base_size);
2003
2004         ret = ent->data;
2005
2006         if (!keep_cache)
2007                 clear_delta_base_cache_entry(ent);
2008         else
2009                 ret = xmemdupz(ent->data, ent->size);
2010         *type = ent->type;
2011         *base_size = ent->size;
2012         return ret;
2013 }
2014
2015 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
2016 {
2017         if (ent->data) {
2018                 free(ent->data);
2019                 ent->data = NULL;
2020                 ent->lru.next->prev = ent->lru.prev;
2021                 ent->lru.prev->next = ent->lru.next;
2022                 delta_base_cached -= ent->size;
2023         }
2024 }
2025
2026 void clear_delta_base_cache(void)
2027 {
2028         unsigned long p;
2029         for (p = 0; p < MAX_DELTA_CACHE; p++)
2030                 release_delta_base_cache(&delta_base_cache[p]);
2031 }
2032
2033 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
2034         void *base, unsigned long base_size, enum object_type type)
2035 {
2036         unsigned long hash = pack_entry_hash(p, base_offset);
2037         struct delta_base_cache_entry *ent = delta_base_cache + hash;
2038         struct delta_base_cache_lru_list *lru;
2039
2040         release_delta_base_cache(ent);
2041         delta_base_cached += base_size;
2042
2043         for (lru = delta_base_cache_lru.next;
2044              delta_base_cached > delta_base_cache_limit
2045              && lru != &delta_base_cache_lru;
2046              lru = lru->next) {
2047                 struct delta_base_cache_entry *f = (void *)lru;
2048                 if (f->type == OBJ_BLOB)
2049                         release_delta_base_cache(f);
2050         }
2051         for (lru = delta_base_cache_lru.next;
2052              delta_base_cached > delta_base_cache_limit
2053              && lru != &delta_base_cache_lru;
2054              lru = lru->next) {
2055                 struct delta_base_cache_entry *f = (void *)lru;
2056                 release_delta_base_cache(f);
2057         }
2058
2059         ent->p = p;
2060         ent->base_offset = base_offset;
2061         ent->type = type;
2062         ent->data = base;
2063         ent->size = base_size;
2064         ent->lru.next = &delta_base_cache_lru;
2065         ent->lru.prev = delta_base_cache_lru.prev;
2066         delta_base_cache_lru.prev->next = &ent->lru;
2067         delta_base_cache_lru.prev = &ent->lru;
2068 }
2069
2070 static void *read_object(const unsigned char *sha1, enum object_type *type,
2071                          unsigned long *size);
2072
2073 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
2074 {
2075         static FILE *log_file;
2076
2077         if (!log_pack_access)
2078                 log_pack_access = getenv("GIT_TRACE_PACK_ACCESS");
2079         if (!log_pack_access)
2080                 log_pack_access = no_log_pack_access;
2081         if (log_pack_access == no_log_pack_access)
2082                 return;
2083
2084         if (!log_file) {
2085                 log_file = fopen(log_pack_access, "w");
2086                 if (!log_file) {
2087                         error("cannot open pack access log '%s' for writing: %s",
2088                               log_pack_access, strerror(errno));
2089                         log_pack_access = no_log_pack_access;
2090                         return;
2091                 }
2092         }
2093         fprintf(log_file, "%s %"PRIuMAX"\n",
2094                 p->pack_name, (uintmax_t)obj_offset);
2095         fflush(log_file);
2096 }
2097
2098 int do_check_packed_object_crc;
2099
2100 #define UNPACK_ENTRY_STACK_PREALLOC 64
2101 struct unpack_entry_stack_ent {
2102         off_t obj_offset;
2103         off_t curpos;
2104         unsigned long size;
2105 };
2106
2107 void *unpack_entry(struct packed_git *p, off_t obj_offset,
2108                    enum object_type *final_type, unsigned long *final_size)
2109 {
2110         struct pack_window *w_curs = NULL;
2111         off_t curpos = obj_offset;
2112         void *data = NULL;
2113         unsigned long size;
2114         enum object_type type;
2115         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
2116         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
2117         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
2118         int base_from_cache = 0;
2119
2120         if (log_pack_access != no_log_pack_access)
2121                 write_pack_access_log(p, obj_offset);
2122
2123         /* PHASE 1: drill down to the innermost base object */
2124         for (;;) {
2125                 off_t base_offset;
2126                 int i;
2127                 struct delta_base_cache_entry *ent;
2128
2129                 ent = get_delta_base_cache_entry(p, curpos);
2130                 if (eq_delta_base_cache_entry(ent, p, curpos)) {
2131                         type = ent->type;
2132                         data = ent->data;
2133                         size = ent->size;
2134                         clear_delta_base_cache_entry(ent);
2135                         base_from_cache = 1;
2136                         break;
2137                 }
2138
2139                 if (do_check_packed_object_crc && p->index_version > 1) {
2140                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
2141                         unsigned long len = revidx[1].offset - obj_offset;
2142                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
2143                                 const unsigned char *sha1 =
2144                                         nth_packed_object_sha1(p, revidx->nr);
2145                                 error("bad packed object CRC for %s",
2146                                       sha1_to_hex(sha1));
2147                                 mark_bad_packed_object(p, sha1);
2148                                 unuse_pack(&w_curs);
2149                                 return NULL;
2150                         }
2151                 }
2152
2153                 type = unpack_object_header(p, &w_curs, &curpos, &size);
2154                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
2155                         break;
2156
2157                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
2158                 if (!base_offset) {
2159                         error("failed to validate delta base reference "
2160                               "at offset %"PRIuMAX" from %s",
2161                               (uintmax_t)curpos, p->pack_name);
2162                         /* bail to phase 2, in hopes of recovery */
2163                         data = NULL;
2164                         break;
2165                 }
2166
2167                 /* push object, proceed to base */
2168                 if (delta_stack_nr >= delta_stack_alloc
2169                     && delta_stack == small_delta_stack) {
2170                         delta_stack_alloc = alloc_nr(delta_stack_nr);
2171                         delta_stack = xmalloc(sizeof(*delta_stack)*delta_stack_alloc);
2172                         memcpy(delta_stack, small_delta_stack,
2173                                sizeof(*delta_stack)*delta_stack_nr);
2174                 } else {
2175                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
2176                 }
2177                 i = delta_stack_nr++;
2178                 delta_stack[i].obj_offset = obj_offset;
2179                 delta_stack[i].curpos = curpos;
2180                 delta_stack[i].size = size;
2181
2182                 curpos = obj_offset = base_offset;
2183         }
2184
2185         /* PHASE 2: handle the base */
2186         switch (type) {
2187         case OBJ_OFS_DELTA:
2188         case OBJ_REF_DELTA:
2189                 if (data)
2190                         die("BUG in unpack_entry: left loop at a valid delta");
2191                 break;
2192         case OBJ_COMMIT:
2193         case OBJ_TREE:
2194         case OBJ_BLOB:
2195         case OBJ_TAG:
2196                 if (!base_from_cache)
2197                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
2198                 break;
2199         default:
2200                 data = NULL;
2201                 error("unknown object type %i at offset %"PRIuMAX" in %s",
2202                       type, (uintmax_t)obj_offset, p->pack_name);
2203         }
2204
2205         /* PHASE 3: apply deltas in order */
2206
2207         /* invariants:
2208          *   'data' holds the base data, or NULL if there was corruption
2209          */
2210         while (delta_stack_nr) {
2211                 void *delta_data;
2212                 void *base = data;
2213                 unsigned long delta_size, base_size = size;
2214                 int i;
2215
2216                 data = NULL;
2217
2218                 if (base)
2219                         add_delta_base_cache(p, obj_offset, base, base_size, type);
2220
2221                 if (!base) {
2222                         /*
2223                          * We're probably in deep shit, but let's try to fetch
2224                          * the required base anyway from another pack or loose.
2225                          * This is costly but should happen only in the presence
2226                          * of a corrupted pack, and is better than failing outright.
2227                          */
2228                         struct revindex_entry *revidx;
2229                         const unsigned char *base_sha1;
2230                         revidx = find_pack_revindex(p, obj_offset);
2231                         if (revidx) {
2232                                 base_sha1 = nth_packed_object_sha1(p, revidx->nr);
2233                                 error("failed to read delta base object %s"
2234                                       " at offset %"PRIuMAX" from %s",
2235                                       sha1_to_hex(base_sha1), (uintmax_t)obj_offset,
2236                                       p->pack_name);
2237                                 mark_bad_packed_object(p, base_sha1);
2238                                 base = read_object(base_sha1, &type, &base_size);
2239                         }
2240                 }
2241
2242                 i = --delta_stack_nr;
2243                 obj_offset = delta_stack[i].obj_offset;
2244                 curpos = delta_stack[i].curpos;
2245                 delta_size = delta_stack[i].size;
2246
2247                 if (!base)
2248                         continue;
2249
2250                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
2251
2252                 if (!delta_data) {
2253                         error("failed to unpack compressed delta "
2254                               "at offset %"PRIuMAX" from %s",
2255                               (uintmax_t)curpos, p->pack_name);
2256                         data = NULL;
2257                         continue;
2258                 }
2259
2260                 data = patch_delta(base, base_size,
2261                                    delta_data, delta_size,
2262                                    &size);
2263
2264                 /*
2265                  * We could not apply the delta; warn the user, but keep going.
2266                  * Our failure will be noticed either in the next iteration of
2267                  * the loop, or if this is the final delta, in the caller when
2268                  * we return NULL. Those code paths will take care of making
2269                  * a more explicit warning and retrying with another copy of
2270                  * the object.
2271                  */
2272                 if (!data)
2273                         error("failed to apply delta");
2274
2275                 free(delta_data);
2276         }
2277
2278         *final_type = type;
2279         *final_size = size;
2280
2281         unuse_pack(&w_curs);
2282         return data;
2283 }
2284
2285 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
2286                                             uint32_t n)
2287 {
2288         const unsigned char *index = p->index_data;
2289         if (!index) {
2290                 if (open_pack_index(p))
2291                         return NULL;
2292                 index = p->index_data;
2293         }
2294         if (n >= p->num_objects)
2295                 return NULL;
2296         index += 4 * 256;
2297         if (p->index_version == 1) {
2298                 return index + 24 * n + 4;
2299         } else {
2300                 index += 8;
2301                 return index + 20 * n;
2302         }
2303 }
2304
2305 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
2306 {
2307         const unsigned char *index = p->index_data;
2308         index += 4 * 256;
2309         if (p->index_version == 1) {
2310                 return ntohl(*((uint32_t *)(index + 24 * n)));
2311         } else {
2312                 uint32_t off;
2313                 index += 8 + p->num_objects * (20 + 4);
2314                 off = ntohl(*((uint32_t *)(index + 4 * n)));
2315                 if (!(off & 0x80000000))
2316                         return off;
2317                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
2318                 return (((uint64_t)ntohl(*((uint32_t *)(index + 0)))) << 32) |
2319                                    ntohl(*((uint32_t *)(index + 4)));
2320         }
2321 }
2322
2323 off_t find_pack_entry_one(const unsigned char *sha1,
2324                                   struct packed_git *p)
2325 {
2326         const uint32_t *level1_ofs = p->index_data;
2327         const unsigned char *index = p->index_data;
2328         unsigned hi, lo, stride;
2329         static int use_lookup = -1;
2330         static int debug_lookup = -1;
2331
2332         if (debug_lookup < 0)
2333                 debug_lookup = !!getenv("GIT_DEBUG_LOOKUP");
2334
2335         if (!index) {
2336                 if (open_pack_index(p))
2337                         return 0;
2338                 level1_ofs = p->index_data;
2339                 index = p->index_data;
2340         }
2341         if (p->index_version > 1) {
2342                 level1_ofs += 2;
2343                 index += 8;
2344         }
2345         index += 4 * 256;
2346         hi = ntohl(level1_ofs[*sha1]);
2347         lo = ((*sha1 == 0x0) ? 0 : ntohl(level1_ofs[*sha1 - 1]));
2348         if (p->index_version > 1) {
2349                 stride = 20;
2350         } else {
2351                 stride = 24;
2352                 index += 4;
2353         }
2354
2355         if (debug_lookup)
2356                 printf("%02x%02x%02x... lo %u hi %u nr %"PRIu32"\n",
2357                        sha1[0], sha1[1], sha1[2], lo, hi, p->num_objects);
2358
2359         if (use_lookup < 0)
2360                 use_lookup = !!getenv("GIT_USE_LOOKUP");
2361         if (use_lookup) {
2362                 int pos = sha1_entry_pos(index, stride, 0,
2363                                          lo, hi, p->num_objects, sha1);
2364                 if (pos < 0)
2365                         return 0;
2366                 return nth_packed_object_offset(p, pos);
2367         }
2368
2369         do {
2370                 unsigned mi = (lo + hi) / 2;
2371                 int cmp = hashcmp(index + mi * stride, sha1);
2372
2373                 if (debug_lookup)
2374                         printf("lo %u hi %u rg %u mi %u\n",
2375                                lo, hi, hi - lo, mi);
2376                 if (!cmp)
2377                         return nth_packed_object_offset(p, mi);
2378                 if (cmp > 0)
2379                         hi = mi;
2380                 else
2381                         lo = mi+1;
2382         } while (lo < hi);
2383         return 0;
2384 }
2385
2386 int is_pack_valid(struct packed_git *p)
2387 {
2388         /* An already open pack is known to be valid. */
2389         if (p->pack_fd != -1)
2390                 return 1;
2391
2392         /* If the pack has one window completely covering the
2393          * file size, the pack is known to be valid even if
2394          * the descriptor is not currently open.
2395          */
2396         if (p->windows) {
2397                 struct pack_window *w = p->windows;
2398
2399                 if (!w->offset && w->len == p->pack_size)
2400                         return 1;
2401         }
2402
2403         /* Force the pack to open to prove its valid. */
2404         return !open_packed_git(p);
2405 }
2406
2407 static int fill_pack_entry(const unsigned char *sha1,
2408                            struct pack_entry *e,
2409                            struct packed_git *p)
2410 {
2411         off_t offset;
2412
2413         if (p->num_bad_objects) {
2414                 unsigned i;
2415                 for (i = 0; i < p->num_bad_objects; i++)
2416                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
2417                                 return 0;
2418         }
2419
2420         offset = find_pack_entry_one(sha1, p);
2421         if (!offset)
2422                 return 0;
2423
2424         /*
2425          * We are about to tell the caller where they can locate the
2426          * requested object.  We better make sure the packfile is
2427          * still here and can be accessed before supplying that
2428          * answer, as it may have been deleted since the index was
2429          * loaded!
2430          */
2431         if (!is_pack_valid(p)) {
2432                 warning("packfile %s cannot be accessed", p->pack_name);
2433                 return 0;
2434         }
2435         e->offset = offset;
2436         e->p = p;
2437         hashcpy(e->sha1, sha1);
2438         return 1;
2439 }
2440
2441 static int find_pack_entry(const unsigned char *sha1, struct pack_entry *e)
2442 {
2443         struct packed_git *p;
2444
2445         prepare_packed_git();
2446         if (!packed_git)
2447                 return 0;
2448
2449         if (last_found_pack && fill_pack_entry(sha1, e, last_found_pack))
2450                 return 1;
2451
2452         for (p = packed_git; p; p = p->next) {
2453                 if (p == last_found_pack || !fill_pack_entry(sha1, e, p))
2454                         continue;
2455
2456                 last_found_pack = p;
2457                 return 1;
2458         }
2459         return 0;
2460 }
2461
2462 struct packed_git *find_sha1_pack(const unsigned char *sha1,
2463                                   struct packed_git *packs)
2464 {
2465         struct packed_git *p;
2466
2467         for (p = packs; p; p = p->next) {
2468                 if (find_pack_entry_one(sha1, p))
2469                         return p;
2470         }
2471         return NULL;
2472
2473 }
2474
2475 static int sha1_loose_object_info(const unsigned char *sha1,
2476                                   struct object_info *oi)
2477 {
2478         int status;
2479         unsigned long mapsize, size;
2480         void *map;
2481         git_zstream stream;
2482         char hdr[32];
2483
2484         /*
2485          * If we don't care about type or size, then we don't
2486          * need to look inside the object at all.
2487          */
2488         if (!oi->typep && !oi->sizep) {
2489                 if (oi->disk_sizep) {
2490                         struct stat st;
2491                         if (stat_sha1_file(sha1, &st) < 0)
2492                                 return -1;
2493                         *oi->disk_sizep = st.st_size;
2494                 }
2495                 return 0;
2496         }
2497
2498         map = map_sha1_file(sha1, &mapsize);
2499         if (!map)
2500                 return -1;
2501         if (oi->disk_sizep)
2502                 *oi->disk_sizep = mapsize;
2503         if (unpack_sha1_header(&stream, map, mapsize, hdr, sizeof(hdr)) < 0)
2504                 status = error("unable to unpack %s header",
2505                                sha1_to_hex(sha1));
2506         else if ((status = parse_sha1_header(hdr, &size)) < 0)
2507                 status = error("unable to parse %s header", sha1_to_hex(sha1));
2508         else if (oi->sizep)
2509                 *oi->sizep = size;
2510         git_inflate_end(&stream);
2511         munmap(map, mapsize);
2512         if (oi->typep)
2513                 *oi->typep = status;
2514         return 0;
2515 }
2516
2517 /* returns enum object_type or negative */
2518 int sha1_object_info_extended(const unsigned char *sha1, struct object_info *oi)
2519 {
2520         struct cached_object *co;
2521         struct pack_entry e;
2522         int rtype;
2523
2524         co = find_cached_object(sha1);
2525         if (co) {
2526                 if (oi->typep)
2527                         *(oi->typep) = co->type;
2528                 if (oi->sizep)
2529                         *(oi->sizep) = co->size;
2530                 if (oi->disk_sizep)
2531                         *(oi->disk_sizep) = 0;
2532                 oi->whence = OI_CACHED;
2533                 return 0;
2534         }
2535
2536         if (!find_pack_entry(sha1, &e)) {
2537                 /* Most likely it's a loose object. */
2538                 if (!sha1_loose_object_info(sha1, oi)) {
2539                         oi->whence = OI_LOOSE;
2540                         return 0;
2541                 }
2542
2543                 /* Not a loose object; someone else may have just packed it. */
2544                 reprepare_packed_git();
2545                 if (!find_pack_entry(sha1, &e))
2546                         return -1;
2547         }
2548
2549         rtype = packed_object_info(e.p, e.offset, oi);
2550         if (rtype < 0) {
2551                 mark_bad_packed_object(e.p, sha1);
2552                 return sha1_object_info_extended(sha1, oi);
2553         } else if (in_delta_base_cache(e.p, e.offset)) {
2554                 oi->whence = OI_DBCACHED;
2555         } else {
2556                 oi->whence = OI_PACKED;
2557                 oi->u.packed.offset = e.offset;
2558                 oi->u.packed.pack = e.p;
2559                 oi->u.packed.is_delta = (rtype == OBJ_REF_DELTA ||
2560                                          rtype == OBJ_OFS_DELTA);
2561         }
2562
2563         return 0;
2564 }
2565
2566 int sha1_object_info(const unsigned char *sha1, unsigned long *sizep)
2567 {
2568         enum object_type type;
2569         struct object_info oi = {NULL};
2570
2571         oi.typep = &type;
2572         oi.sizep = sizep;
2573         if (sha1_object_info_extended(sha1, &oi) < 0)
2574                 return -1;
2575         return type;
2576 }
2577
2578 static void *read_packed_sha1(const unsigned char *sha1,
2579                               enum object_type *type, unsigned long *size)
2580 {
2581         struct pack_entry e;
2582         void *data;
2583
2584         if (!find_pack_entry(sha1, &e))
2585                 return NULL;
2586         data = cache_or_unpack_entry(e.p, e.offset, size, type, 1);
2587         if (!data) {
2588                 /*
2589                  * We're probably in deep shit, but let's try to fetch
2590                  * the required object anyway from another pack or loose.
2591                  * This should happen only in the presence of a corrupted
2592                  * pack, and is better than failing outright.
2593                  */
2594                 error("failed to read object %s at offset %"PRIuMAX" from %s",
2595                       sha1_to_hex(sha1), (uintmax_t)e.offset, e.p->pack_name);
2596                 mark_bad_packed_object(e.p, sha1);
2597                 data = read_object(sha1, type, size);
2598         }
2599         return data;
2600 }
2601
2602 int pretend_sha1_file(void *buf, unsigned long len, enum object_type type,
2603                       unsigned char *sha1)
2604 {
2605         struct cached_object *co;
2606
2607         hash_sha1_file(buf, len, typename(type), sha1);
2608         if (has_sha1_file(sha1) || find_cached_object(sha1))
2609                 return 0;
2610         if (cached_object_alloc <= cached_object_nr) {
2611                 cached_object_alloc = alloc_nr(cached_object_alloc);
2612                 cached_objects = xrealloc(cached_objects,
2613                                           sizeof(*cached_objects) *
2614                                           cached_object_alloc);
2615         }
2616         co = &cached_objects[cached_object_nr++];
2617         co->size = len;
2618         co->type = type;
2619         co->buf = xmalloc(len);
2620         memcpy(co->buf, buf, len);
2621         hashcpy(co->sha1, sha1);
2622         return 0;
2623 }
2624
2625 static void *read_object(const unsigned char *sha1, enum object_type *type,
2626                          unsigned long *size)
2627 {
2628         unsigned long mapsize;
2629         void *map, *buf;
2630         struct cached_object *co;
2631
2632         co = find_cached_object(sha1);
2633         if (co) {
2634                 *type = co->type;
2635                 *size = co->size;
2636                 return xmemdupz(co->buf, co->size);
2637         }
2638
2639         buf = read_packed_sha1(sha1, type, size);
2640         if (buf)
2641                 return buf;
2642         map = map_sha1_file(sha1, &mapsize);
2643         if (map) {
2644                 buf = unpack_sha1_file(map, mapsize, type, size, sha1);
2645                 munmap(map, mapsize);
2646                 return buf;
2647         }
2648         reprepare_packed_git();
2649         return read_packed_sha1(sha1, type, size);
2650 }
2651
2652 /*
2653  * This function dies on corrupt objects; the callers who want to
2654  * deal with them should arrange to call read_object() and give error
2655  * messages themselves.
2656  */
2657 void *read_sha1_file_extended(const unsigned char *sha1,
2658                               enum object_type *type,
2659                               unsigned long *size,
2660                               unsigned flag)
2661 {
2662         void *data;
2663         char *path;
2664         const struct packed_git *p;
2665         const unsigned char *repl = (flag & READ_SHA1_FILE_REPLACE)
2666                 ? lookup_replace_object(sha1) : sha1;
2667
2668         errno = 0;
2669         data = read_object(repl, type, size);
2670         if (data)
2671                 return data;
2672
2673         if (errno && errno != ENOENT)
2674                 die_errno("failed to read object %s", sha1_to_hex(sha1));
2675
2676         /* die if we replaced an object with one that does not exist */
2677         if (repl != sha1)
2678                 die("replacement %s not found for %s",
2679                     sha1_to_hex(repl), sha1_to_hex(sha1));
2680
2681         if (has_loose_object(repl)) {
2682                 path = sha1_file_name(sha1);
2683                 die("loose object %s (stored in %s) is corrupt",
2684                     sha1_to_hex(repl), path);
2685         }
2686
2687         if ((p = has_packed_and_bad(repl)) != NULL)
2688                 die("packed object %s (stored in %s) is corrupt",
2689                     sha1_to_hex(repl), p->pack_name);
2690
2691         return NULL;
2692 }
2693
2694 void *read_object_with_reference(const unsigned char *sha1,
2695                                  const char *required_type_name,
2696                                  unsigned long *size,
2697                                  unsigned char *actual_sha1_return)
2698 {
2699         enum object_type type, required_type;
2700         void *buffer;
2701         unsigned long isize;
2702         unsigned char actual_sha1[20];
2703
2704         required_type = type_from_string(required_type_name);
2705         hashcpy(actual_sha1, sha1);
2706         while (1) {
2707                 int ref_length = -1;
2708                 const char *ref_type = NULL;
2709
2710                 buffer = read_sha1_file(actual_sha1, &type, &isize);
2711                 if (!buffer)
2712                         return NULL;
2713                 if (type == required_type) {
2714                         *size = isize;
2715                         if (actual_sha1_return)
2716                                 hashcpy(actual_sha1_return, actual_sha1);
2717                         return buffer;
2718                 }
2719                 /* Handle references */
2720                 else if (type == OBJ_COMMIT)
2721                         ref_type = "tree ";
2722                 else if (type == OBJ_TAG)
2723                         ref_type = "object ";
2724                 else {
2725                         free(buffer);
2726                         return NULL;
2727                 }
2728                 ref_length = strlen(ref_type);
2729
2730                 if (ref_length + 40 > isize ||
2731                     memcmp(buffer, ref_type, ref_length) ||
2732                     get_sha1_hex((char *) buffer + ref_length, actual_sha1)) {
2733                         free(buffer);
2734                         return NULL;
2735                 }
2736                 free(buffer);
2737                 /* Now we have the ID of the referred-to object in
2738                  * actual_sha1.  Check again. */
2739         }
2740 }
2741
2742 static void write_sha1_file_prepare(const void *buf, unsigned long len,
2743                                     const char *type, unsigned char *sha1,
2744                                     char *hdr, int *hdrlen)
2745 {
2746         git_SHA_CTX c;
2747
2748         /* Generate the header */
2749         *hdrlen = sprintf(hdr, "%s %lu", type, len)+1;
2750
2751         /* Sha1.. */
2752         git_SHA1_Init(&c);
2753         git_SHA1_Update(&c, hdr, *hdrlen);
2754         git_SHA1_Update(&c, buf, len);
2755         git_SHA1_Final(sha1, &c);
2756 }
2757
2758 /*
2759  * Move the just written object into its final resting place.
2760  * NEEDSWORK: this should be renamed to finalize_temp_file() as
2761  * "moving" is only a part of what it does, when no patch between
2762  * master to pu changes the call sites of this function.
2763  */
2764 int move_temp_to_file(const char *tmpfile, const char *filename)
2765 {
2766         int ret = 0;
2767
2768         if (object_creation_mode == OBJECT_CREATION_USES_RENAMES)
2769                 goto try_rename;
2770         else if (link(tmpfile, filename))
2771                 ret = errno;
2772
2773         /*
2774          * Coda hack - coda doesn't like cross-directory links,
2775          * so we fall back to a rename, which will mean that it
2776          * won't be able to check collisions, but that's not a
2777          * big deal.
2778          *
2779          * The same holds for FAT formatted media.
2780          *
2781          * When this succeeds, we just return.  We have nothing
2782          * left to unlink.
2783          */
2784         if (ret && ret != EEXIST) {
2785         try_rename:
2786                 if (!rename(tmpfile, filename))
2787                         goto out;
2788                 ret = errno;
2789         }
2790         unlink_or_warn(tmpfile);
2791         if (ret) {
2792                 if (ret != EEXIST) {
2793                         return error("unable to write sha1 filename %s: %s", filename, strerror(ret));
2794                 }
2795                 /* FIXME!!! Collision check here ? */
2796         }
2797
2798 out:
2799         if (adjust_shared_perm(filename))
2800                 return error("unable to set permission to '%s'", filename);
2801         return 0;
2802 }
2803
2804 static int write_buffer(int fd, const void *buf, size_t len)
2805 {
2806         if (write_in_full(fd, buf, len) < 0)
2807                 return error("file write error (%s)", strerror(errno));
2808         return 0;
2809 }
2810
2811 int hash_sha1_file(const void *buf, unsigned long len, const char *type,
2812                    unsigned char *sha1)
2813 {
2814         char hdr[32];
2815         int hdrlen;
2816         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2817         return 0;
2818 }
2819
2820 /* Finalize a file on disk, and close it. */
2821 static void close_sha1_file(int fd)
2822 {
2823         if (fsync_object_files)
2824                 fsync_or_die(fd, "sha1 file");
2825         if (close(fd) != 0)
2826                 die_errno("error when closing sha1 file");
2827 }
2828
2829 /* Size of directory component, including the ending '/' */
2830 static inline int directory_size(const char *filename)
2831 {
2832         const char *s = strrchr(filename, '/');
2833         if (!s)
2834                 return 0;
2835         return s - filename + 1;
2836 }
2837
2838 /*
2839  * This creates a temporary file in the same directory as the final
2840  * 'filename'
2841  *
2842  * We want to avoid cross-directory filename renames, because those
2843  * can have problems on various filesystems (FAT, NFS, Coda).
2844  */
2845 static int create_tmpfile(char *buffer, size_t bufsiz, const char *filename)
2846 {
2847         int fd, dirlen = directory_size(filename);
2848
2849         if (dirlen + 20 > bufsiz) {
2850                 errno = ENAMETOOLONG;
2851                 return -1;
2852         }
2853         memcpy(buffer, filename, dirlen);
2854         strcpy(buffer + dirlen, "tmp_obj_XXXXXX");
2855         fd = git_mkstemp_mode(buffer, 0444);
2856         if (fd < 0 && dirlen && errno == ENOENT) {
2857                 /* Make sure the directory exists */
2858                 memcpy(buffer, filename, dirlen);
2859                 buffer[dirlen-1] = 0;
2860                 if (mkdir(buffer, 0777) || adjust_shared_perm(buffer))
2861                         return -1;
2862
2863                 /* Try again */
2864                 strcpy(buffer + dirlen - 1, "/tmp_obj_XXXXXX");
2865                 fd = git_mkstemp_mode(buffer, 0444);
2866         }
2867         return fd;
2868 }
2869
2870 static int write_loose_object(const unsigned char *sha1, char *hdr, int hdrlen,
2871                               const void *buf, unsigned long len, time_t mtime)
2872 {
2873         int fd, ret;
2874         unsigned char compressed[4096];
2875         git_zstream stream;
2876         git_SHA_CTX c;
2877         unsigned char parano_sha1[20];
2878         char *filename;
2879         static char tmp_file[PATH_MAX];
2880
2881         filename = sha1_file_name(sha1);
2882         fd = create_tmpfile(tmp_file, sizeof(tmp_file), filename);
2883         if (fd < 0) {
2884                 if (errno == EACCES)
2885                         return error("insufficient permission for adding an object to repository database %s", get_object_directory());
2886                 else
2887                         return error("unable to create temporary file: %s", strerror(errno));
2888         }
2889
2890         /* Set it up */
2891         memset(&stream, 0, sizeof(stream));
2892         git_deflate_init(&stream, zlib_compression_level);
2893         stream.next_out = compressed;
2894         stream.avail_out = sizeof(compressed);
2895         git_SHA1_Init(&c);
2896
2897         /* First header.. */
2898         stream.next_in = (unsigned char *)hdr;
2899         stream.avail_in = hdrlen;
2900         while (git_deflate(&stream, 0) == Z_OK)
2901                 ; /* nothing */
2902         git_SHA1_Update(&c, hdr, hdrlen);
2903
2904         /* Then the data itself.. */
2905         stream.next_in = (void *)buf;
2906         stream.avail_in = len;
2907         do {
2908                 unsigned char *in0 = stream.next_in;
2909                 ret = git_deflate(&stream, Z_FINISH);
2910                 git_SHA1_Update(&c, in0, stream.next_in - in0);
2911                 if (write_buffer(fd, compressed, stream.next_out - compressed) < 0)
2912                         die("unable to write sha1 file");
2913                 stream.next_out = compressed;
2914                 stream.avail_out = sizeof(compressed);
2915         } while (ret == Z_OK);
2916
2917         if (ret != Z_STREAM_END)
2918                 die("unable to deflate new object %s (%d)", sha1_to_hex(sha1), ret);
2919         ret = git_deflate_end_gently(&stream);
2920         if (ret != Z_OK)
2921                 die("deflateEnd on object %s failed (%d)", sha1_to_hex(sha1), ret);
2922         git_SHA1_Final(parano_sha1, &c);
2923         if (hashcmp(sha1, parano_sha1) != 0)
2924                 die("confused by unstable object source data for %s", sha1_to_hex(sha1));
2925
2926         close_sha1_file(fd);
2927
2928         if (mtime) {
2929                 struct utimbuf utb;
2930                 utb.actime = mtime;
2931                 utb.modtime = mtime;
2932                 if (utime(tmp_file, &utb) < 0)
2933                         warning("failed utime() on %s: %s",
2934                                 tmp_file, strerror(errno));
2935         }
2936
2937         return move_temp_to_file(tmp_file, filename);
2938 }
2939
2940 int write_sha1_file(const void *buf, unsigned long len, const char *type, unsigned char *returnsha1)
2941 {
2942         unsigned char sha1[20];
2943         char hdr[32];
2944         int hdrlen;
2945
2946         /* Normally if we have it in the pack then we do not bother writing
2947          * it out into .git/objects/??/?{38} file.
2948          */
2949         write_sha1_file_prepare(buf, len, type, sha1, hdr, &hdrlen);
2950         if (returnsha1)
2951                 hashcpy(returnsha1, sha1);
2952         if (has_sha1_file(sha1))
2953                 return 0;
2954         return write_loose_object(sha1, hdr, hdrlen, buf, len, 0);
2955 }
2956
2957 int force_object_loose(const unsigned char *sha1, time_t mtime)
2958 {
2959         void *buf;
2960         unsigned long len;
2961         enum object_type type;
2962         char hdr[32];
2963         int hdrlen;
2964         int ret;
2965
2966         if (has_loose_object(sha1))
2967                 return 0;
2968         buf = read_packed_sha1(sha1, &type, &len);
2969         if (!buf)
2970                 return error("cannot read sha1_file for %s", sha1_to_hex(sha1));
2971         hdrlen = sprintf(hdr, "%s %lu", typename(type), len) + 1;
2972         ret = write_loose_object(sha1, hdr, hdrlen, buf, len, mtime);
2973         free(buf);
2974
2975         return ret;
2976 }
2977
2978 int has_pack_index(const unsigned char *sha1)
2979 {
2980         struct stat st;
2981         if (stat(sha1_pack_index_name(sha1), &st))
2982                 return 0;
2983         return 1;
2984 }
2985
2986 int has_sha1_pack(const unsigned char *sha1)
2987 {
2988         struct pack_entry e;
2989         return find_pack_entry(sha1, &e);
2990 }
2991
2992 int has_sha1_file(const unsigned char *sha1)
2993 {
2994         struct pack_entry e;
2995
2996         if (find_pack_entry(sha1, &e))
2997                 return 1;
2998         if (has_loose_object(sha1))
2999                 return 1;
3000         reprepare_packed_git();
3001         return find_pack_entry(sha1, &e);
3002 }
3003
3004 static void check_tree(const void *buf, size_t size)
3005 {
3006         struct tree_desc desc;
3007         struct name_entry entry;
3008
3009         init_tree_desc(&desc, buf, size);
3010         while (tree_entry(&desc, &entry))
3011                 /* do nothing
3012                  * tree_entry() will die() on malformed entries */
3013                 ;
3014 }
3015
3016 static void check_commit(const void *buf, size_t size)
3017 {
3018         struct commit c;
3019         memset(&c, 0, sizeof(c));
3020         if (parse_commit_buffer(&c, buf, size))
3021                 die("corrupt commit");
3022 }
3023
3024 static void check_tag(const void *buf, size_t size)
3025 {
3026         struct tag t;
3027         memset(&t, 0, sizeof(t));
3028         if (parse_tag_buffer(&t, buf, size))
3029                 die("corrupt tag");
3030 }
3031
3032 static int index_mem(unsigned char *sha1, void *buf, size_t size,
3033                      enum object_type type,
3034                      const char *path, unsigned flags)
3035 {
3036         int ret, re_allocated = 0;
3037         int write_object = flags & HASH_WRITE_OBJECT;
3038
3039         if (!type)
3040                 type = OBJ_BLOB;
3041
3042         /*
3043          * Convert blobs to git internal format
3044          */
3045         if ((type == OBJ_BLOB) && path) {
3046                 struct strbuf nbuf = STRBUF_INIT;
3047                 if (convert_to_git(path, buf, size, &nbuf,
3048                                    write_object ? safe_crlf : SAFE_CRLF_FALSE)) {
3049                         buf = strbuf_detach(&nbuf, &size);
3050                         re_allocated = 1;
3051                 }
3052         }
3053         if (flags & HASH_FORMAT_CHECK) {
3054                 if (type == OBJ_TREE)
3055                         check_tree(buf, size);
3056                 if (type == OBJ_COMMIT)
3057                         check_commit(buf, size);
3058                 if (type == OBJ_TAG)
3059                         check_tag(buf, size);
3060         }
3061
3062         if (write_object)
3063                 ret = write_sha1_file(buf, size, typename(type), sha1);
3064         else
3065                 ret = hash_sha1_file(buf, size, typename(type), sha1);
3066         if (re_allocated)
3067                 free(buf);
3068         return ret;
3069 }
3070
3071 static int index_pipe(unsigned char *sha1, int fd, enum object_type type,
3072                       const char *path, unsigned flags)
3073 {
3074         struct strbuf sbuf = STRBUF_INIT;
3075         int ret;
3076
3077         if (strbuf_read(&sbuf, fd, 4096) >= 0)
3078                 ret = index_mem(sha1, sbuf.buf, sbuf.len, type, path, flags);
3079         else
3080                 ret = -1;
3081         strbuf_release(&sbuf);
3082         return ret;
3083 }
3084
3085 #define SMALL_FILE_SIZE (32*1024)
3086
3087 static int index_core(unsigned char *sha1, int fd, size_t size,
3088                       enum object_type type, const char *path,
3089                       unsigned flags)
3090 {
3091         int ret;
3092
3093         if (!size) {
3094                 ret = index_mem(sha1, NULL, size, type, path, flags);
3095         } else if (size <= SMALL_FILE_SIZE) {
3096                 char *buf = xmalloc(size);
3097                 if (size == read_in_full(fd, buf, size))
3098                         ret = index_mem(sha1, buf, size, type, path, flags);
3099                 else
3100                         ret = error("short read %s", strerror(errno));
3101                 free(buf);
3102         } else {
3103                 void *buf = xmmap(NULL, size, PROT_READ, MAP_PRIVATE, fd, 0);
3104                 ret = index_mem(sha1, buf, size, type, path, flags);
3105                 munmap(buf, size);
3106         }
3107         return ret;
3108 }
3109
3110 /*
3111  * This creates one packfile per large blob unless bulk-checkin
3112  * machinery is "plugged".
3113  *
3114  * This also bypasses the usual "convert-to-git" dance, and that is on
3115  * purpose. We could write a streaming version of the converting
3116  * functions and insert that before feeding the data to fast-import
3117  * (or equivalent in-core API described above). However, that is
3118  * somewhat complicated, as we do not know the size of the filter
3119  * result, which we need to know beforehand when writing a git object.
3120  * Since the primary motivation for trying to stream from the working
3121  * tree file and to avoid mmaping it in core is to deal with large
3122  * binary blobs, they generally do not want to get any conversion, and
3123  * callers should avoid this code path when filters are requested.
3124  */
3125 static int index_stream(unsigned char *sha1, int fd, size_t size,
3126                         enum object_type type, const char *path,
3127                         unsigned flags)
3128 {
3129         return index_bulk_checkin(sha1, fd, size, type, path, flags);
3130 }
3131
3132 int index_fd(unsigned char *sha1, int fd, struct stat *st,
3133              enum object_type type, const char *path, unsigned flags)
3134 {
3135         int ret;
3136         size_t size = xsize_t(st->st_size);
3137
3138         if (!S_ISREG(st->st_mode))
3139                 ret = index_pipe(sha1, fd, type, path, flags);
3140         else if (size <= big_file_threshold || type != OBJ_BLOB ||
3141                  (path && would_convert_to_git(path, NULL, 0, 0)))
3142                 ret = index_core(sha1, fd, size, type, path, flags);
3143         else
3144                 ret = index_stream(sha1, fd, size, type, path, flags);
3145         close(fd);
3146         return ret;
3147 }
3148
3149 int index_path(unsigned char *sha1, const char *path, struct stat *st, unsigned flags)
3150 {
3151         int fd;
3152         struct strbuf sb = STRBUF_INIT;
3153
3154         switch (st->st_mode & S_IFMT) {
3155         case S_IFREG:
3156                 fd = open(path, O_RDONLY);
3157                 if (fd < 0)
3158                         return error("open(\"%s\"): %s", path,
3159                                      strerror(errno));
3160                 if (index_fd(sha1, fd, st, OBJ_BLOB, path, flags) < 0)
3161                         return error("%s: failed to insert into database",
3162                                      path);
3163                 break;
3164         case S_IFLNK:
3165                 if (strbuf_readlink(&sb, path, st->st_size)) {
3166                         char *errstr = strerror(errno);
3167                         return error("readlink(\"%s\"): %s", path,
3168                                      errstr);
3169                 }
3170                 if (!(flags & HASH_WRITE_OBJECT))
3171                         hash_sha1_file(sb.buf, sb.len, blob_type, sha1);
3172                 else if (write_sha1_file(sb.buf, sb.len, blob_type, sha1))
3173                         return error("%s: failed to insert into database",
3174                                      path);
3175                 strbuf_release(&sb);
3176                 break;
3177         case S_IFDIR:
3178                 return resolve_gitlink_ref(path, "HEAD", sha1);
3179         default:
3180                 return error("%s: unsupported file type", path);
3181         }
3182         return 0;
3183 }
3184
3185 int read_pack_header(int fd, struct pack_header *header)
3186 {
3187         if (read_in_full(fd, header, sizeof(*header)) < sizeof(*header))
3188                 /* "eof before pack header was fully read" */
3189                 return PH_ERROR_EOF;
3190
3191         if (header->hdr_signature != htonl(PACK_SIGNATURE))
3192                 /* "protocol error (pack signature mismatch detected)" */
3193                 return PH_ERROR_PACK_SIGNATURE;
3194         if (!pack_version_ok(header->hdr_version))
3195                 /* "protocol error (pack version unsupported)" */
3196                 return PH_ERROR_PROTOCOL;
3197         return 0;
3198 }
3199
3200 void assert_sha1_type(const unsigned char *sha1, enum object_type expect)
3201 {
3202         enum object_type type = sha1_object_info(sha1, NULL);
3203         if (type < 0)
3204                 die("%s is not a valid object", sha1_to_hex(sha1));
3205         if (type != expect)
3206                 die("%s is not a valid '%s' object", sha1_to_hex(sha1),
3207                     typename(expect));
3208 }