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