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