remote-curl: accept all encodings supported by curl
[git] / packfile.c
1 #include "cache.h"
2 #include "list.h"
3 #include "pack.h"
4 #include "repository.h"
5 #include "dir.h"
6 #include "mergesort.h"
7 #include "packfile.h"
8 #include "delta.h"
9 #include "list.h"
10 #include "streaming.h"
11 #include "sha1-lookup.h"
12 #include "commit.h"
13 #include "object.h"
14 #include "tag.h"
15 #include "tree-walk.h"
16 #include "tree.h"
17 #include "object-store.h"
18
19 char *odb_pack_name(struct strbuf *buf,
20                     const unsigned char *sha1,
21                     const char *ext)
22 {
23         strbuf_reset(buf);
24         strbuf_addf(buf, "%s/pack/pack-%s.%s", get_object_directory(),
25                     sha1_to_hex(sha1), ext);
26         return buf->buf;
27 }
28
29 char *sha1_pack_name(const unsigned char *sha1)
30 {
31         static struct strbuf buf = STRBUF_INIT;
32         return odb_pack_name(&buf, sha1, "pack");
33 }
34
35 char *sha1_pack_index_name(const unsigned char *sha1)
36 {
37         static struct strbuf buf = STRBUF_INIT;
38         return odb_pack_name(&buf, sha1, "idx");
39 }
40
41 static unsigned int pack_used_ctr;
42 static unsigned int pack_mmap_calls;
43 static unsigned int peak_pack_open_windows;
44 static unsigned int pack_open_windows;
45 static unsigned int pack_open_fds;
46 static unsigned int pack_max_fds;
47 static size_t peak_pack_mapped;
48 static size_t pack_mapped;
49
50 #define SZ_FMT PRIuMAX
51 static inline uintmax_t sz_fmt(size_t s) { return s; }
52
53 void pack_report(void)
54 {
55         fprintf(stderr,
56                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
57                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
58                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
59                 sz_fmt(getpagesize()),
60                 sz_fmt(packed_git_window_size),
61                 sz_fmt(packed_git_limit));
62         fprintf(stderr,
63                 "pack_report: pack_used_ctr            = %10u\n"
64                 "pack_report: pack_mmap_calls          = %10u\n"
65                 "pack_report: pack_open_windows        = %10u / %10u\n"
66                 "pack_report: pack_mapped              = "
67                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
68                 pack_used_ctr,
69                 pack_mmap_calls,
70                 pack_open_windows, peak_pack_open_windows,
71                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
72 }
73
74 /*
75  * Open and mmap the index file at path, perform a couple of
76  * consistency checks, then record its information to p.  Return 0 on
77  * success.
78  */
79 static int check_packed_git_idx(const char *path, struct packed_git *p)
80 {
81         void *idx_map;
82         struct pack_idx_header *hdr;
83         size_t idx_size;
84         uint32_t version, nr, i, *index;
85         int fd = git_open(path);
86         struct stat st;
87
88         if (fd < 0)
89                 return -1;
90         if (fstat(fd, &st)) {
91                 close(fd);
92                 return -1;
93         }
94         idx_size = xsize_t(st.st_size);
95         if (idx_size < 4 * 256 + 20 + 20) {
96                 close(fd);
97                 return error("index file %s is too small", path);
98         }
99         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
100         close(fd);
101
102         hdr = idx_map;
103         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
104                 version = ntohl(hdr->idx_version);
105                 if (version < 2 || version > 2) {
106                         munmap(idx_map, idx_size);
107                         return error("index file %s is version %"PRIu32
108                                      " and is not supported by this binary"
109                                      " (try upgrading GIT to a newer version)",
110                                      path, version);
111                 }
112         } else
113                 version = 1;
114
115         nr = 0;
116         index = idx_map;
117         if (version > 1)
118                 index += 2;  /* skip index header */
119         for (i = 0; i < 256; i++) {
120                 uint32_t n = ntohl(index[i]);
121                 if (n < nr) {
122                         munmap(idx_map, idx_size);
123                         return error("non-monotonic index %s", path);
124                 }
125                 nr = n;
126         }
127
128         if (version == 1) {
129                 /*
130                  * Total size:
131                  *  - 256 index entries 4 bytes each
132                  *  - 24-byte entries * nr (20-byte sha1 + 4-byte offset)
133                  *  - 20-byte SHA1 of the packfile
134                  *  - 20-byte SHA1 file checksum
135                  */
136                 if (idx_size != 4*256 + nr * 24 + 20 + 20) {
137                         munmap(idx_map, idx_size);
138                         return error("wrong index v1 file size in %s", path);
139                 }
140         } else if (version == 2) {
141                 /*
142                  * Minimum size:
143                  *  - 8 bytes of header
144                  *  - 256 index entries 4 bytes each
145                  *  - 20-byte sha1 entry * nr
146                  *  - 4-byte crc entry * nr
147                  *  - 4-byte offset entry * nr
148                  *  - 20-byte SHA1 of the packfile
149                  *  - 20-byte SHA1 file checksum
150                  * And after the 4-byte offset table might be a
151                  * variable sized table containing 8-byte entries
152                  * for offsets larger than 2^31.
153                  */
154                 unsigned long min_size = 8 + 4*256 + nr*(20 + 4 + 4) + 20 + 20;
155                 unsigned long max_size = min_size;
156                 if (nr)
157                         max_size += (nr - 1)*8;
158                 if (idx_size < min_size || idx_size > max_size) {
159                         munmap(idx_map, idx_size);
160                         return error("wrong index v2 file size in %s", path);
161                 }
162                 if (idx_size != min_size &&
163                     /*
164                      * make sure we can deal with large pack offsets.
165                      * 31-bit signed offset won't be enough, neither
166                      * 32-bit unsigned one will be.
167                      */
168                     (sizeof(off_t) <= 4)) {
169                         munmap(idx_map, idx_size);
170                         return error("pack too large for current definition of off_t in %s", path);
171                 }
172         }
173
174         p->index_version = version;
175         p->index_data = idx_map;
176         p->index_size = idx_size;
177         p->num_objects = nr;
178         return 0;
179 }
180
181 int open_pack_index(struct packed_git *p)
182 {
183         char *idx_name;
184         size_t len;
185         int ret;
186
187         if (p->index_data)
188                 return 0;
189
190         if (!strip_suffix(p->pack_name, ".pack", &len))
191                 die("BUG: pack_name does not end in .pack");
192         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
193         ret = check_packed_git_idx(idx_name, p);
194         free(idx_name);
195         return ret;
196 }
197
198 static struct packed_git *alloc_packed_git(int extra)
199 {
200         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
201         memset(p, 0, sizeof(*p));
202         p->pack_fd = -1;
203         return p;
204 }
205
206 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
207 {
208         const char *path = sha1_pack_name(sha1);
209         size_t alloc = st_add(strlen(path), 1);
210         struct packed_git *p = alloc_packed_git(alloc);
211
212         memcpy(p->pack_name, path, alloc); /* includes NUL */
213         hashcpy(p->sha1, sha1);
214         if (check_packed_git_idx(idx_path, p)) {
215                 free(p);
216                 return NULL;
217         }
218
219         return p;
220 }
221
222 static void scan_windows(struct packed_git *p,
223         struct packed_git **lru_p,
224         struct pack_window **lru_w,
225         struct pack_window **lru_l)
226 {
227         struct pack_window *w, *w_l;
228
229         for (w_l = NULL, w = p->windows; w; w = w->next) {
230                 if (!w->inuse_cnt) {
231                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
232                                 *lru_p = p;
233                                 *lru_w = w;
234                                 *lru_l = w_l;
235                         }
236                 }
237                 w_l = w;
238         }
239 }
240
241 static int unuse_one_window(struct packed_git *current)
242 {
243         struct packed_git *p, *lru_p = NULL;
244         struct pack_window *lru_w = NULL, *lru_l = NULL;
245
246         if (current)
247                 scan_windows(current, &lru_p, &lru_w, &lru_l);
248         for (p = the_repository->objects->packed_git; p; p = p->next)
249                 scan_windows(p, &lru_p, &lru_w, &lru_l);
250         if (lru_p) {
251                 munmap(lru_w->base, lru_w->len);
252                 pack_mapped -= lru_w->len;
253                 if (lru_l)
254                         lru_l->next = lru_w->next;
255                 else
256                         lru_p->windows = lru_w->next;
257                 free(lru_w);
258                 pack_open_windows--;
259                 return 1;
260         }
261         return 0;
262 }
263
264 void release_pack_memory(size_t need)
265 {
266         size_t cur = pack_mapped;
267         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
268                 ; /* nothing */
269 }
270
271 void close_pack_windows(struct packed_git *p)
272 {
273         while (p->windows) {
274                 struct pack_window *w = p->windows;
275
276                 if (w->inuse_cnt)
277                         die("pack '%s' still has open windows to it",
278                             p->pack_name);
279                 munmap(w->base, w->len);
280                 pack_mapped -= w->len;
281                 pack_open_windows--;
282                 p->windows = w->next;
283                 free(w);
284         }
285 }
286
287 static int close_pack_fd(struct packed_git *p)
288 {
289         if (p->pack_fd < 0)
290                 return 0;
291
292         close(p->pack_fd);
293         pack_open_fds--;
294         p->pack_fd = -1;
295
296         return 1;
297 }
298
299 void close_pack_index(struct packed_git *p)
300 {
301         if (p->index_data) {
302                 munmap((void *)p->index_data, p->index_size);
303                 p->index_data = NULL;
304         }
305 }
306
307 void close_pack(struct packed_git *p)
308 {
309         close_pack_windows(p);
310         close_pack_fd(p);
311         close_pack_index(p);
312 }
313
314 void close_all_packs(struct raw_object_store *o)
315 {
316         struct packed_git *p;
317
318         for (p = o->packed_git; p; p = p->next)
319                 if (p->do_not_close)
320                         die("BUG: want to close pack marked 'do-not-close'");
321                 else
322                         close_pack(p);
323 }
324
325 /*
326  * The LRU pack is the one with the oldest MRU window, preferring packs
327  * with no used windows, or the oldest mtime if it has no windows allocated.
328  */
329 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
330 {
331         struct pack_window *w, *this_mru_w;
332         int has_windows_inuse = 0;
333
334         /*
335          * Reject this pack if it has windows and the previously selected
336          * one does not.  If this pack does not have windows, reject
337          * it if the pack file is newer than the previously selected one.
338          */
339         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
340                 return;
341
342         for (w = this_mru_w = p->windows; w; w = w->next) {
343                 /*
344                  * Reject this pack if any of its windows are in use,
345                  * but the previously selected pack did not have any
346                  * inuse windows.  Otherwise, record that this pack
347                  * has windows in use.
348                  */
349                 if (w->inuse_cnt) {
350                         if (*accept_windows_inuse)
351                                 has_windows_inuse = 1;
352                         else
353                                 return;
354                 }
355
356                 if (w->last_used > this_mru_w->last_used)
357                         this_mru_w = w;
358
359                 /*
360                  * Reject this pack if it has windows that have been
361                  * used more recently than the previously selected pack.
362                  * If the previously selected pack had windows inuse and
363                  * we have not encountered a window in this pack that is
364                  * inuse, skip this check since we prefer a pack with no
365                  * inuse windows to one that has inuse windows.
366                  */
367                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
368                     this_mru_w->last_used > (*mru_w)->last_used)
369                         return;
370         }
371
372         /*
373          * Select this pack.
374          */
375         *mru_w = this_mru_w;
376         *lru_p = p;
377         *accept_windows_inuse = has_windows_inuse;
378 }
379
380 static int close_one_pack(void)
381 {
382         struct packed_git *p, *lru_p = NULL;
383         struct pack_window *mru_w = NULL;
384         int accept_windows_inuse = 1;
385
386         for (p = the_repository->objects->packed_git; p; p = p->next) {
387                 if (p->pack_fd == -1)
388                         continue;
389                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
390         }
391
392         if (lru_p)
393                 return close_pack_fd(lru_p);
394
395         return 0;
396 }
397
398 static unsigned int get_max_fd_limit(void)
399 {
400 #ifdef RLIMIT_NOFILE
401         {
402                 struct rlimit lim;
403
404                 if (!getrlimit(RLIMIT_NOFILE, &lim))
405                         return lim.rlim_cur;
406         }
407 #endif
408
409 #ifdef _SC_OPEN_MAX
410         {
411                 long open_max = sysconf(_SC_OPEN_MAX);
412                 if (0 < open_max)
413                         return open_max;
414                 /*
415                  * Otherwise, we got -1 for one of the two
416                  * reasons:
417                  *
418                  * (1) sysconf() did not understand _SC_OPEN_MAX
419                  *     and signaled an error with -1; or
420                  * (2) sysconf() said there is no limit.
421                  *
422                  * We _could_ clear errno before calling sysconf() to
423                  * tell these two cases apart and return a huge number
424                  * in the latter case to let the caller cap it to a
425                  * value that is not so selfish, but letting the
426                  * fallback OPEN_MAX codepath take care of these cases
427                  * is a lot simpler.
428                  */
429         }
430 #endif
431
432 #ifdef OPEN_MAX
433         return OPEN_MAX;
434 #else
435         return 1; /* see the caller ;-) */
436 #endif
437 }
438
439 /*
440  * Do not call this directly as this leaks p->pack_fd on error return;
441  * call open_packed_git() instead.
442  */
443 static int open_packed_git_1(struct packed_git *p)
444 {
445         struct stat st;
446         struct pack_header hdr;
447         unsigned char sha1[20];
448         unsigned char *idx_sha1;
449         long fd_flag;
450         ssize_t read_result;
451
452         if (!p->index_data && open_pack_index(p))
453                 return error("packfile %s index unavailable", p->pack_name);
454
455         if (!pack_max_fds) {
456                 unsigned int max_fds = get_max_fd_limit();
457
458                 /* Save 3 for stdin/stdout/stderr, 22 for work */
459                 if (25 < max_fds)
460                         pack_max_fds = max_fds - 25;
461                 else
462                         pack_max_fds = 1;
463         }
464
465         while (pack_max_fds <= pack_open_fds && close_one_pack())
466                 ; /* nothing */
467
468         p->pack_fd = git_open(p->pack_name);
469         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
470                 return -1;
471         pack_open_fds++;
472
473         /* If we created the struct before we had the pack we lack size. */
474         if (!p->pack_size) {
475                 if (!S_ISREG(st.st_mode))
476                         return error("packfile %s not a regular file", p->pack_name);
477                 p->pack_size = st.st_size;
478         } else if (p->pack_size != st.st_size)
479                 return error("packfile %s size changed", p->pack_name);
480
481         /* We leave these file descriptors open with sliding mmap;
482          * there is no point keeping them open across exec(), though.
483          */
484         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
485         if (fd_flag < 0)
486                 return error("cannot determine file descriptor flags");
487         fd_flag |= FD_CLOEXEC;
488         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
489                 return error("cannot set FD_CLOEXEC");
490
491         /* Verify we recognize this pack file format. */
492         read_result = read_in_full(p->pack_fd, &hdr, sizeof(hdr));
493         if (read_result < 0)
494                 return error_errno("error reading from %s", p->pack_name);
495         if (read_result != sizeof(hdr))
496                 return error("file %s is far too short to be a packfile", p->pack_name);
497         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
498                 return error("file %s is not a GIT packfile", p->pack_name);
499         if (!pack_version_ok(hdr.hdr_version))
500                 return error("packfile %s is version %"PRIu32" and not"
501                         " supported (try upgrading GIT to a newer version)",
502                         p->pack_name, ntohl(hdr.hdr_version));
503
504         /* Verify the pack matches its index. */
505         if (p->num_objects != ntohl(hdr.hdr_entries))
506                 return error("packfile %s claims to have %"PRIu32" objects"
507                              " while index indicates %"PRIu32" objects",
508                              p->pack_name, ntohl(hdr.hdr_entries),
509                              p->num_objects);
510         if (lseek(p->pack_fd, p->pack_size - sizeof(sha1), SEEK_SET) == -1)
511                 return error("end of packfile %s is unavailable", p->pack_name);
512         read_result = read_in_full(p->pack_fd, sha1, sizeof(sha1));
513         if (read_result < 0)
514                 return error_errno("error reading from %s", p->pack_name);
515         if (read_result != sizeof(sha1))
516                 return error("packfile %s signature is unavailable", p->pack_name);
517         idx_sha1 = ((unsigned char *)p->index_data) + p->index_size - 40;
518         if (hashcmp(sha1, idx_sha1))
519                 return error("packfile %s does not match index", p->pack_name);
520         return 0;
521 }
522
523 static int open_packed_git(struct packed_git *p)
524 {
525         if (!open_packed_git_1(p))
526                 return 0;
527         close_pack_fd(p);
528         return -1;
529 }
530
531 static int in_window(struct pack_window *win, off_t offset)
532 {
533         /* We must promise at least 20 bytes (one hash) after the
534          * offset is available from this window, otherwise the offset
535          * is not actually in this window and a different window (which
536          * has that one hash excess) must be used.  This is to support
537          * the object header and delta base parsing routines below.
538          */
539         off_t win_off = win->offset;
540         return win_off <= offset
541                 && (offset + 20) <= (win_off + win->len);
542 }
543
544 unsigned char *use_pack(struct packed_git *p,
545                 struct pack_window **w_cursor,
546                 off_t offset,
547                 unsigned long *left)
548 {
549         struct pack_window *win = *w_cursor;
550
551         /* Since packfiles end in a hash of their content and it's
552          * pointless to ask for an offset into the middle of that
553          * hash, and the in_window function above wouldn't match
554          * don't allow an offset too close to the end of the file.
555          */
556         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
557                 die("packfile %s cannot be accessed", p->pack_name);
558         if (offset > (p->pack_size - 20))
559                 die("offset beyond end of packfile (truncated pack?)");
560         if (offset < 0)
561                 die(_("offset before end of packfile (broken .idx?)"));
562
563         if (!win || !in_window(win, offset)) {
564                 if (win)
565                         win->inuse_cnt--;
566                 for (win = p->windows; win; win = win->next) {
567                         if (in_window(win, offset))
568                                 break;
569                 }
570                 if (!win) {
571                         size_t window_align = packed_git_window_size / 2;
572                         off_t len;
573
574                         if (p->pack_fd == -1 && open_packed_git(p))
575                                 die("packfile %s cannot be accessed", p->pack_name);
576
577                         win = xcalloc(1, sizeof(*win));
578                         win->offset = (offset / window_align) * window_align;
579                         len = p->pack_size - win->offset;
580                         if (len > packed_git_window_size)
581                                 len = packed_git_window_size;
582                         win->len = (size_t)len;
583                         pack_mapped += win->len;
584                         while (packed_git_limit < pack_mapped
585                                 && unuse_one_window(p))
586                                 ; /* nothing */
587                         win->base = xmmap(NULL, win->len,
588                                 PROT_READ, MAP_PRIVATE,
589                                 p->pack_fd, win->offset);
590                         if (win->base == MAP_FAILED)
591                                 die_errno("packfile %s cannot be mapped",
592                                           p->pack_name);
593                         if (!win->offset && win->len == p->pack_size
594                                 && !p->do_not_close)
595                                 close_pack_fd(p);
596                         pack_mmap_calls++;
597                         pack_open_windows++;
598                         if (pack_mapped > peak_pack_mapped)
599                                 peak_pack_mapped = pack_mapped;
600                         if (pack_open_windows > peak_pack_open_windows)
601                                 peak_pack_open_windows = pack_open_windows;
602                         win->next = p->windows;
603                         p->windows = win;
604                 }
605         }
606         if (win != *w_cursor) {
607                 win->last_used = pack_used_ctr++;
608                 win->inuse_cnt++;
609                 *w_cursor = win;
610         }
611         offset -= win->offset;
612         if (left)
613                 *left = win->len - xsize_t(offset);
614         return win->base + offset;
615 }
616
617 void unuse_pack(struct pack_window **w_cursor)
618 {
619         struct pack_window *w = *w_cursor;
620         if (w) {
621                 w->inuse_cnt--;
622                 *w_cursor = NULL;
623         }
624 }
625
626 static void try_to_free_pack_memory(size_t size)
627 {
628         release_pack_memory(size);
629 }
630
631 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
632 {
633         static int have_set_try_to_free_routine;
634         struct stat st;
635         size_t alloc;
636         struct packed_git *p;
637
638         if (!have_set_try_to_free_routine) {
639                 have_set_try_to_free_routine = 1;
640                 set_try_to_free_routine(try_to_free_pack_memory);
641         }
642
643         /*
644          * Make sure a corresponding .pack file exists and that
645          * the index looks sane.
646          */
647         if (!strip_suffix_mem(path, &path_len, ".idx"))
648                 return NULL;
649
650         /*
651          * ".promisor" is long enough to hold any suffix we're adding (and
652          * the use xsnprintf double-checks that)
653          */
654         alloc = st_add3(path_len, strlen(".promisor"), 1);
655         p = alloc_packed_git(alloc);
656         memcpy(p->pack_name, path, path_len);
657
658         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
659         if (!access(p->pack_name, F_OK))
660                 p->pack_keep = 1;
661
662         xsnprintf(p->pack_name + path_len, alloc - path_len, ".promisor");
663         if (!access(p->pack_name, F_OK))
664                 p->pack_promisor = 1;
665
666         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
667         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
668                 free(p);
669                 return NULL;
670         }
671
672         /* ok, it looks sane as far as we can check without
673          * actually mapping the pack file.
674          */
675         p->pack_size = st.st_size;
676         p->pack_local = local;
677         p->mtime = st.st_mtime;
678         if (path_len < 40 || get_sha1_hex(path + path_len - 40, p->sha1))
679                 hashclr(p->sha1);
680         return p;
681 }
682
683 void install_packed_git(struct repository *r, struct packed_git *pack)
684 {
685         if (pack->pack_fd != -1)
686                 pack_open_fds++;
687
688         pack->next = r->objects->packed_git;
689         r->objects->packed_git = pack;
690 }
691
692 void (*report_garbage)(unsigned seen_bits, const char *path);
693
694 static void report_helper(const struct string_list *list,
695                           int seen_bits, int first, int last)
696 {
697         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
698                 return;
699
700         for (; first < last; first++)
701                 report_garbage(seen_bits, list->items[first].string);
702 }
703
704 static void report_pack_garbage(struct string_list *list)
705 {
706         int i, baselen = -1, first = 0, seen_bits = 0;
707
708         if (!report_garbage)
709                 return;
710
711         string_list_sort(list);
712
713         for (i = 0; i < list->nr; i++) {
714                 const char *path = list->items[i].string;
715                 if (baselen != -1 &&
716                     strncmp(path, list->items[first].string, baselen)) {
717                         report_helper(list, seen_bits, first, i);
718                         baselen = -1;
719                         seen_bits = 0;
720                 }
721                 if (baselen == -1) {
722                         const char *dot = strrchr(path, '.');
723                         if (!dot) {
724                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
725                                 continue;
726                         }
727                         baselen = dot - path + 1;
728                         first = i;
729                 }
730                 if (!strcmp(path + baselen, "pack"))
731                         seen_bits |= 1;
732                 else if (!strcmp(path + baselen, "idx"))
733                         seen_bits |= 2;
734         }
735         report_helper(list, seen_bits, first, list->nr);
736 }
737
738 static void prepare_packed_git_one(struct repository *r, char *objdir, int local)
739 {
740         struct strbuf path = STRBUF_INIT;
741         size_t dirnamelen;
742         DIR *dir;
743         struct dirent *de;
744         struct string_list garbage = STRING_LIST_INIT_DUP;
745
746         strbuf_addstr(&path, objdir);
747         strbuf_addstr(&path, "/pack");
748         dir = opendir(path.buf);
749         if (!dir) {
750                 if (errno != ENOENT)
751                         error_errno("unable to open object pack directory: %s",
752                                     path.buf);
753                 strbuf_release(&path);
754                 return;
755         }
756         strbuf_addch(&path, '/');
757         dirnamelen = path.len;
758         while ((de = readdir(dir)) != NULL) {
759                 struct packed_git *p;
760                 size_t base_len;
761
762                 if (is_dot_or_dotdot(de->d_name))
763                         continue;
764
765                 strbuf_setlen(&path, dirnamelen);
766                 strbuf_addstr(&path, de->d_name);
767
768                 base_len = path.len;
769                 if (strip_suffix_mem(path.buf, &base_len, ".idx")) {
770                         /* Don't reopen a pack we already have. */
771                         for (p = r->objects->packed_git; p;
772                              p = p->next) {
773                                 size_t len;
774                                 if (strip_suffix(p->pack_name, ".pack", &len) &&
775                                     len == base_len &&
776                                     !memcmp(p->pack_name, path.buf, len))
777                                         break;
778                         }
779                         if (p == NULL &&
780                             /*
781                              * See if it really is a valid .idx file with
782                              * corresponding .pack file that we can map.
783                              */
784                             (p = add_packed_git(path.buf, path.len, local)) != NULL)
785                                 install_packed_git(r, p);
786                 }
787
788                 if (!report_garbage)
789                         continue;
790
791                 if (ends_with(de->d_name, ".idx") ||
792                     ends_with(de->d_name, ".pack") ||
793                     ends_with(de->d_name, ".bitmap") ||
794                     ends_with(de->d_name, ".keep") ||
795                     ends_with(de->d_name, ".promisor"))
796                         string_list_append(&garbage, path.buf);
797                 else
798                         report_garbage(PACKDIR_FILE_GARBAGE, path.buf);
799         }
800         closedir(dir);
801         report_pack_garbage(&garbage);
802         string_list_clear(&garbage, 0);
803         strbuf_release(&path);
804 }
805
806 static void prepare_packed_git(struct repository *r);
807 /*
808  * Give a fast, rough count of the number of objects in the repository. This
809  * ignores loose objects completely. If you have a lot of them, then either
810  * you should repack because your performance will be awful, or they are
811  * all unreachable objects about to be pruned, in which case they're not really
812  * interesting as a measure of repo size in the first place.
813  */
814 unsigned long approximate_object_count(void)
815 {
816         if (!the_repository->objects->approximate_object_count_valid) {
817                 unsigned long count;
818                 struct packed_git *p;
819
820                 prepare_packed_git(the_repository);
821                 count = 0;
822                 for (p = the_repository->objects->packed_git; p; p = p->next) {
823                         if (open_pack_index(p))
824                                 continue;
825                         count += p->num_objects;
826                 }
827                 the_repository->objects->approximate_object_count = count;
828         }
829         return the_repository->objects->approximate_object_count;
830 }
831
832 static void *get_next_packed_git(const void *p)
833 {
834         return ((const struct packed_git *)p)->next;
835 }
836
837 static void set_next_packed_git(void *p, void *next)
838 {
839         ((struct packed_git *)p)->next = next;
840 }
841
842 static int sort_pack(const void *a_, const void *b_)
843 {
844         const struct packed_git *a = a_;
845         const struct packed_git *b = b_;
846         int st;
847
848         /*
849          * Local packs tend to contain objects specific to our
850          * variant of the project than remote ones.  In addition,
851          * remote ones could be on a network mounted filesystem.
852          * Favor local ones for these reasons.
853          */
854         st = a->pack_local - b->pack_local;
855         if (st)
856                 return -st;
857
858         /*
859          * Younger packs tend to contain more recent objects,
860          * and more recent objects tend to get accessed more
861          * often.
862          */
863         if (a->mtime < b->mtime)
864                 return 1;
865         else if (a->mtime == b->mtime)
866                 return 0;
867         return -1;
868 }
869
870 static void rearrange_packed_git(struct repository *r)
871 {
872         r->objects->packed_git = llist_mergesort(
873                 r->objects->packed_git, get_next_packed_git,
874                 set_next_packed_git, sort_pack);
875 }
876
877 static void prepare_packed_git_mru(struct repository *r)
878 {
879         struct packed_git *p;
880
881         INIT_LIST_HEAD(&r->objects->packed_git_mru);
882
883         for (p = r->objects->packed_git; p; p = p->next)
884                 list_add_tail(&p->mru, &r->objects->packed_git_mru);
885 }
886
887 static void prepare_packed_git(struct repository *r)
888 {
889         struct alternate_object_database *alt;
890
891         if (r->objects->packed_git_initialized)
892                 return;
893         prepare_packed_git_one(r, r->objects->objectdir, 1);
894         prepare_alt_odb(r);
895         for (alt = r->objects->alt_odb_list; alt; alt = alt->next)
896                 prepare_packed_git_one(r, alt->path, 0);
897         rearrange_packed_git(r);
898         prepare_packed_git_mru(r);
899         r->objects->packed_git_initialized = 1;
900 }
901
902 void reprepare_packed_git(struct repository *r)
903 {
904         r->objects->approximate_object_count_valid = 0;
905         r->objects->packed_git_initialized = 0;
906         prepare_packed_git(r);
907 }
908
909 struct packed_git *get_packed_git(struct repository *r)
910 {
911         prepare_packed_git(r);
912         return r->objects->packed_git;
913 }
914
915 struct list_head *get_packed_git_mru(struct repository *r)
916 {
917         prepare_packed_git(r);
918         return &r->objects->packed_git_mru;
919 }
920
921 unsigned long unpack_object_header_buffer(const unsigned char *buf,
922                 unsigned long len, enum object_type *type, unsigned long *sizep)
923 {
924         unsigned shift;
925         unsigned long size, c;
926         unsigned long used = 0;
927
928         c = buf[used++];
929         *type = (c >> 4) & 7;
930         size = c & 15;
931         shift = 4;
932         while (c & 0x80) {
933                 if (len <= used || bitsizeof(long) <= shift) {
934                         error("bad object header");
935                         size = used = 0;
936                         break;
937                 }
938                 c = buf[used++];
939                 size += (c & 0x7f) << shift;
940                 shift += 7;
941         }
942         *sizep = size;
943         return used;
944 }
945
946 unsigned long get_size_from_delta(struct packed_git *p,
947                                   struct pack_window **w_curs,
948                                   off_t curpos)
949 {
950         const unsigned char *data;
951         unsigned char delta_head[20], *in;
952         git_zstream stream;
953         int st;
954
955         memset(&stream, 0, sizeof(stream));
956         stream.next_out = delta_head;
957         stream.avail_out = sizeof(delta_head);
958
959         git_inflate_init(&stream);
960         do {
961                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
962                 stream.next_in = in;
963                 st = git_inflate(&stream, Z_FINISH);
964                 curpos += stream.next_in - in;
965         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
966                  stream.total_out < sizeof(delta_head));
967         git_inflate_end(&stream);
968         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
969                 error("delta data unpack-initial failed");
970                 return 0;
971         }
972
973         /* Examine the initial part of the delta to figure out
974          * the result size.
975          */
976         data = delta_head;
977
978         /* ignore base size */
979         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
980
981         /* Read the result size */
982         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
983 }
984
985 int unpack_object_header(struct packed_git *p,
986                          struct pack_window **w_curs,
987                          off_t *curpos,
988                          unsigned long *sizep)
989 {
990         unsigned char *base;
991         unsigned long left;
992         unsigned long used;
993         enum object_type type;
994
995         /* use_pack() assures us we have [base, base + 20) available
996          * as a range that we can look at.  (Its actually the hash
997          * size that is assured.)  With our object header encoding
998          * the maximum deflated object size is 2^137, which is just
999          * insane, so we know won't exceed what we have been given.
1000          */
1001         base = use_pack(p, w_curs, *curpos, &left);
1002         used = unpack_object_header_buffer(base, left, &type, sizep);
1003         if (!used) {
1004                 type = OBJ_BAD;
1005         } else
1006                 *curpos += used;
1007
1008         return type;
1009 }
1010
1011 void mark_bad_packed_object(struct packed_git *p, const unsigned char *sha1)
1012 {
1013         unsigned i;
1014         for (i = 0; i < p->num_bad_objects; i++)
1015                 if (!hashcmp(sha1, p->bad_object_sha1 + GIT_SHA1_RAWSZ * i))
1016                         return;
1017         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1018                                       st_mult(GIT_MAX_RAWSZ,
1019                                               st_add(p->num_bad_objects, 1)));
1020         hashcpy(p->bad_object_sha1 + GIT_SHA1_RAWSZ * p->num_bad_objects, sha1);
1021         p->num_bad_objects++;
1022 }
1023
1024 const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1025 {
1026         struct packed_git *p;
1027         unsigned i;
1028
1029         for (p = the_repository->objects->packed_git; p; p = p->next)
1030                 for (i = 0; i < p->num_bad_objects; i++)
1031                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1032                                 return p;
1033         return NULL;
1034 }
1035
1036 static off_t get_delta_base(struct packed_git *p,
1037                                     struct pack_window **w_curs,
1038                                     off_t *curpos,
1039                                     enum object_type type,
1040                                     off_t delta_obj_offset)
1041 {
1042         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1043         off_t base_offset;
1044
1045         /* use_pack() assured us we have [base_info, base_info + 20)
1046          * as a range that we can look at without walking off the
1047          * end of the mapped window.  Its actually the hash size
1048          * that is assured.  An OFS_DELTA longer than the hash size
1049          * is stupid, as then a REF_DELTA would be smaller to store.
1050          */
1051         if (type == OBJ_OFS_DELTA) {
1052                 unsigned used = 0;
1053                 unsigned char c = base_info[used++];
1054                 base_offset = c & 127;
1055                 while (c & 128) {
1056                         base_offset += 1;
1057                         if (!base_offset || MSB(base_offset, 7))
1058                                 return 0;  /* overflow */
1059                         c = base_info[used++];
1060                         base_offset = (base_offset << 7) + (c & 127);
1061                 }
1062                 base_offset = delta_obj_offset - base_offset;
1063                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1064                         return 0;  /* out of bound */
1065                 *curpos += used;
1066         } else if (type == OBJ_REF_DELTA) {
1067                 /* The base entry _must_ be in the same pack */
1068                 base_offset = find_pack_entry_one(base_info, p);
1069                 *curpos += 20;
1070         } else
1071                 die("I am totally screwed");
1072         return base_offset;
1073 }
1074
1075 /*
1076  * Like get_delta_base above, but we return the sha1 instead of the pack
1077  * offset. This means it is cheaper for REF deltas (we do not have to do
1078  * the final object lookup), but more expensive for OFS deltas (we
1079  * have to load the revidx to convert the offset back into a sha1).
1080  */
1081 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1082                                                 struct pack_window **w_curs,
1083                                                 off_t curpos,
1084                                                 enum object_type type,
1085                                                 off_t delta_obj_offset)
1086 {
1087         if (type == OBJ_REF_DELTA) {
1088                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1089                 return base;
1090         } else if (type == OBJ_OFS_DELTA) {
1091                 struct revindex_entry *revidx;
1092                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1093                                                    type, delta_obj_offset);
1094
1095                 if (!base_offset)
1096                         return NULL;
1097
1098                 revidx = find_pack_revindex(p, base_offset);
1099                 if (!revidx)
1100                         return NULL;
1101
1102                 return nth_packed_object_sha1(p, revidx->nr);
1103         } else
1104                 return NULL;
1105 }
1106
1107 static int retry_bad_packed_offset(struct packed_git *p, off_t obj_offset)
1108 {
1109         int type;
1110         struct revindex_entry *revidx;
1111         struct object_id oid;
1112         revidx = find_pack_revindex(p, obj_offset);
1113         if (!revidx)
1114                 return OBJ_BAD;
1115         nth_packed_object_oid(&oid, p, revidx->nr);
1116         mark_bad_packed_object(p, oid.hash);
1117         type = oid_object_info(&oid, NULL);
1118         if (type <= OBJ_NONE)
1119                 return OBJ_BAD;
1120         return type;
1121 }
1122
1123 #define POI_STACK_PREALLOC 64
1124
1125 static enum object_type packed_to_object_type(struct packed_git *p,
1126                                               off_t obj_offset,
1127                                               enum object_type type,
1128                                               struct pack_window **w_curs,
1129                                               off_t curpos)
1130 {
1131         off_t small_poi_stack[POI_STACK_PREALLOC];
1132         off_t *poi_stack = small_poi_stack;
1133         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1134
1135         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1136                 off_t base_offset;
1137                 unsigned long size;
1138                 /* Push the object we're going to leave behind */
1139                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1140                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1141                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
1142                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1143                 } else {
1144                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1145                 }
1146                 poi_stack[poi_stack_nr++] = obj_offset;
1147                 /* If parsing the base offset fails, just unwind */
1148                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1149                 if (!base_offset)
1150                         goto unwind;
1151                 curpos = obj_offset = base_offset;
1152                 type = unpack_object_header(p, w_curs, &curpos, &size);
1153                 if (type <= OBJ_NONE) {
1154                         /* If getting the base itself fails, we first
1155                          * retry the base, otherwise unwind */
1156                         type = retry_bad_packed_offset(p, base_offset);
1157                         if (type > OBJ_NONE)
1158                                 goto out;
1159                         goto unwind;
1160                 }
1161         }
1162
1163         switch (type) {
1164         case OBJ_BAD:
1165         case OBJ_COMMIT:
1166         case OBJ_TREE:
1167         case OBJ_BLOB:
1168         case OBJ_TAG:
1169                 break;
1170         default:
1171                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1172                       type, (uintmax_t)obj_offset, p->pack_name);
1173                 type = OBJ_BAD;
1174         }
1175
1176 out:
1177         if (poi_stack != small_poi_stack)
1178                 free(poi_stack);
1179         return type;
1180
1181 unwind:
1182         while (poi_stack_nr) {
1183                 obj_offset = poi_stack[--poi_stack_nr];
1184                 type = retry_bad_packed_offset(p, obj_offset);
1185                 if (type > OBJ_NONE)
1186                         goto out;
1187         }
1188         type = OBJ_BAD;
1189         goto out;
1190 }
1191
1192 static struct hashmap delta_base_cache;
1193 static size_t delta_base_cached;
1194
1195 static LIST_HEAD(delta_base_cache_lru);
1196
1197 struct delta_base_cache_key {
1198         struct packed_git *p;
1199         off_t base_offset;
1200 };
1201
1202 struct delta_base_cache_entry {
1203         struct hashmap hash;
1204         struct delta_base_cache_key key;
1205         struct list_head lru;
1206         void *data;
1207         unsigned long size;
1208         enum object_type type;
1209 };
1210
1211 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
1212 {
1213         unsigned int hash;
1214
1215         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
1216         hash += (hash >> 8) + (hash >> 16);
1217         return hash;
1218 }
1219
1220 static struct delta_base_cache_entry *
1221 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
1222 {
1223         struct hashmap_entry entry;
1224         struct delta_base_cache_key key;
1225
1226         if (!delta_base_cache.cmpfn)
1227                 return NULL;
1228
1229         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
1230         key.p = p;
1231         key.base_offset = base_offset;
1232         return hashmap_get(&delta_base_cache, &entry, &key);
1233 }
1234
1235 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
1236                                    const struct delta_base_cache_key *b)
1237 {
1238         return a->p == b->p && a->base_offset == b->base_offset;
1239 }
1240
1241 static int delta_base_cache_hash_cmp(const void *unused_cmp_data,
1242                                      const void *va, const void *vb,
1243                                      const void *vkey)
1244 {
1245         const struct delta_base_cache_entry *a = va, *b = vb;
1246         const struct delta_base_cache_key *key = vkey;
1247         if (key)
1248                 return !delta_base_cache_key_eq(&a->key, key);
1249         else
1250                 return !delta_base_cache_key_eq(&a->key, &b->key);
1251 }
1252
1253 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
1254 {
1255         return !!get_delta_base_cache_entry(p, base_offset);
1256 }
1257
1258 /*
1259  * Remove the entry from the cache, but do _not_ free the associated
1260  * entry data. The caller takes ownership of the "data" buffer, and
1261  * should copy out any fields it wants before detaching.
1262  */
1263 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
1264 {
1265         hashmap_remove(&delta_base_cache, ent, &ent->key);
1266         list_del(&ent->lru);
1267         delta_base_cached -= ent->size;
1268         free(ent);
1269 }
1270
1271 static void *cache_or_unpack_entry(struct packed_git *p, off_t base_offset,
1272         unsigned long *base_size, enum object_type *type)
1273 {
1274         struct delta_base_cache_entry *ent;
1275
1276         ent = get_delta_base_cache_entry(p, base_offset);
1277         if (!ent)
1278                 return unpack_entry(p, base_offset, type, base_size);
1279
1280         if (type)
1281                 *type = ent->type;
1282         if (base_size)
1283                 *base_size = ent->size;
1284         return xmemdupz(ent->data, ent->size);
1285 }
1286
1287 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1288 {
1289         free(ent->data);
1290         detach_delta_base_cache_entry(ent);
1291 }
1292
1293 void clear_delta_base_cache(void)
1294 {
1295         struct list_head *lru, *tmp;
1296         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1297                 struct delta_base_cache_entry *entry =
1298                         list_entry(lru, struct delta_base_cache_entry, lru);
1299                 release_delta_base_cache(entry);
1300         }
1301 }
1302
1303 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1304         void *base, unsigned long base_size, enum object_type type)
1305 {
1306         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
1307         struct list_head *lru, *tmp;
1308
1309         delta_base_cached += base_size;
1310
1311         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1312                 struct delta_base_cache_entry *f =
1313                         list_entry(lru, struct delta_base_cache_entry, lru);
1314                 if (delta_base_cached <= delta_base_cache_limit)
1315                         break;
1316                 release_delta_base_cache(f);
1317         }
1318
1319         ent->key.p = p;
1320         ent->key.base_offset = base_offset;
1321         ent->type = type;
1322         ent->data = base;
1323         ent->size = base_size;
1324         list_add_tail(&ent->lru, &delta_base_cache_lru);
1325
1326         if (!delta_base_cache.cmpfn)
1327                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, NULL, 0);
1328         hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
1329         hashmap_add(&delta_base_cache, ent);
1330 }
1331
1332 int packed_object_info(struct packed_git *p, off_t obj_offset,
1333                        struct object_info *oi)
1334 {
1335         struct pack_window *w_curs = NULL;
1336         unsigned long size;
1337         off_t curpos = obj_offset;
1338         enum object_type type;
1339
1340         /*
1341          * We always get the representation type, but only convert it to
1342          * a "real" type later if the caller is interested.
1343          */
1344         if (oi->contentp) {
1345                 *oi->contentp = cache_or_unpack_entry(p, obj_offset, oi->sizep,
1346                                                       &type);
1347                 if (!*oi->contentp)
1348                         type = OBJ_BAD;
1349         } else {
1350                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1351         }
1352
1353         if (!oi->contentp && oi->sizep) {
1354                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1355                         off_t tmp_pos = curpos;
1356                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1357                                                            type, obj_offset);
1358                         if (!base_offset) {
1359                                 type = OBJ_BAD;
1360                                 goto out;
1361                         }
1362                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1363                         if (*oi->sizep == 0) {
1364                                 type = OBJ_BAD;
1365                                 goto out;
1366                         }
1367                 } else {
1368                         *oi->sizep = size;
1369                 }
1370         }
1371
1372         if (oi->disk_sizep) {
1373                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1374                 *oi->disk_sizep = revidx[1].offset - obj_offset;
1375         }
1376
1377         if (oi->typep || oi->type_name) {
1378                 enum object_type ptot;
1379                 ptot = packed_to_object_type(p, obj_offset, type, &w_curs,
1380                                              curpos);
1381                 if (oi->typep)
1382                         *oi->typep = ptot;
1383                 if (oi->type_name) {
1384                         const char *tn = type_name(ptot);
1385                         if (tn)
1386                                 strbuf_addstr(oi->type_name, tn);
1387                 }
1388                 if (ptot < 0) {
1389                         type = OBJ_BAD;
1390                         goto out;
1391                 }
1392         }
1393
1394         if (oi->delta_base_sha1) {
1395                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1396                         const unsigned char *base;
1397
1398                         base = get_delta_base_sha1(p, &w_curs, curpos,
1399                                                    type, obj_offset);
1400                         if (!base) {
1401                                 type = OBJ_BAD;
1402                                 goto out;
1403                         }
1404
1405                         hashcpy(oi->delta_base_sha1, base);
1406                 } else
1407                         hashclr(oi->delta_base_sha1);
1408         }
1409
1410         oi->whence = in_delta_base_cache(p, obj_offset) ? OI_DBCACHED :
1411                                                           OI_PACKED;
1412
1413 out:
1414         unuse_pack(&w_curs);
1415         return type;
1416 }
1417
1418 static void *unpack_compressed_entry(struct packed_git *p,
1419                                     struct pack_window **w_curs,
1420                                     off_t curpos,
1421                                     unsigned long size)
1422 {
1423         int st;
1424         git_zstream stream;
1425         unsigned char *buffer, *in;
1426
1427         buffer = xmallocz_gently(size);
1428         if (!buffer)
1429                 return NULL;
1430         memset(&stream, 0, sizeof(stream));
1431         stream.next_out = buffer;
1432         stream.avail_out = size + 1;
1433
1434         git_inflate_init(&stream);
1435         do {
1436                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1437                 stream.next_in = in;
1438                 st = git_inflate(&stream, Z_FINISH);
1439                 if (!stream.avail_out)
1440                         break; /* the payload is larger than it should be */
1441                 curpos += stream.next_in - in;
1442         } while (st == Z_OK || st == Z_BUF_ERROR);
1443         git_inflate_end(&stream);
1444         if ((st != Z_STREAM_END) || stream.total_out != size) {
1445                 free(buffer);
1446                 return NULL;
1447         }
1448
1449         return buffer;
1450 }
1451
1452 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
1453 {
1454         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
1455         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
1456                          p->pack_name, (uintmax_t)obj_offset);
1457 }
1458
1459 int do_check_packed_object_crc;
1460
1461 #define UNPACK_ENTRY_STACK_PREALLOC 64
1462 struct unpack_entry_stack_ent {
1463         off_t obj_offset;
1464         off_t curpos;
1465         unsigned long size;
1466 };
1467
1468 static void *read_object(const struct object_id *oid, enum object_type *type,
1469                          unsigned long *size)
1470 {
1471         struct object_info oi = OBJECT_INFO_INIT;
1472         void *content;
1473         oi.typep = type;
1474         oi.sizep = size;
1475         oi.contentp = &content;
1476
1477         if (oid_object_info_extended(oid, &oi, 0) < 0)
1478                 return NULL;
1479         return content;
1480 }
1481
1482 void *unpack_entry(struct packed_git *p, off_t obj_offset,
1483                    enum object_type *final_type, unsigned long *final_size)
1484 {
1485         struct pack_window *w_curs = NULL;
1486         off_t curpos = obj_offset;
1487         void *data = NULL;
1488         unsigned long size;
1489         enum object_type type;
1490         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
1491         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
1492         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
1493         int base_from_cache = 0;
1494
1495         write_pack_access_log(p, obj_offset);
1496
1497         /* PHASE 1: drill down to the innermost base object */
1498         for (;;) {
1499                 off_t base_offset;
1500                 int i;
1501                 struct delta_base_cache_entry *ent;
1502
1503                 ent = get_delta_base_cache_entry(p, curpos);
1504                 if (ent) {
1505                         type = ent->type;
1506                         data = ent->data;
1507                         size = ent->size;
1508                         detach_delta_base_cache_entry(ent);
1509                         base_from_cache = 1;
1510                         break;
1511                 }
1512
1513                 if (do_check_packed_object_crc && p->index_version > 1) {
1514                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1515                         off_t len = revidx[1].offset - obj_offset;
1516                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1517                                 struct object_id oid;
1518                                 nth_packed_object_oid(&oid, p, revidx->nr);
1519                                 error("bad packed object CRC for %s",
1520                                       oid_to_hex(&oid));
1521                                 mark_bad_packed_object(p, oid.hash);
1522                                 data = NULL;
1523                                 goto out;
1524                         }
1525                 }
1526
1527                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1528                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
1529                         break;
1530
1531                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1532                 if (!base_offset) {
1533                         error("failed to validate delta base reference "
1534                               "at offset %"PRIuMAX" from %s",
1535                               (uintmax_t)curpos, p->pack_name);
1536                         /* bail to phase 2, in hopes of recovery */
1537                         data = NULL;
1538                         break;
1539                 }
1540
1541                 /* push object, proceed to base */
1542                 if (delta_stack_nr >= delta_stack_alloc
1543                     && delta_stack == small_delta_stack) {
1544                         delta_stack_alloc = alloc_nr(delta_stack_nr);
1545                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
1546                         memcpy(delta_stack, small_delta_stack,
1547                                sizeof(*delta_stack)*delta_stack_nr);
1548                 } else {
1549                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
1550                 }
1551                 i = delta_stack_nr++;
1552                 delta_stack[i].obj_offset = obj_offset;
1553                 delta_stack[i].curpos = curpos;
1554                 delta_stack[i].size = size;
1555
1556                 curpos = obj_offset = base_offset;
1557         }
1558
1559         /* PHASE 2: handle the base */
1560         switch (type) {
1561         case OBJ_OFS_DELTA:
1562         case OBJ_REF_DELTA:
1563                 if (data)
1564                         die("BUG: unpack_entry: left loop at a valid delta");
1565                 break;
1566         case OBJ_COMMIT:
1567         case OBJ_TREE:
1568         case OBJ_BLOB:
1569         case OBJ_TAG:
1570                 if (!base_from_cache)
1571                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
1572                 break;
1573         default:
1574                 data = NULL;
1575                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1576                       type, (uintmax_t)obj_offset, p->pack_name);
1577         }
1578
1579         /* PHASE 3: apply deltas in order */
1580
1581         /* invariants:
1582          *   'data' holds the base data, or NULL if there was corruption
1583          */
1584         while (delta_stack_nr) {
1585                 void *delta_data;
1586                 void *base = data;
1587                 void *external_base = NULL;
1588                 unsigned long delta_size, base_size = size;
1589                 int i;
1590
1591                 data = NULL;
1592
1593                 if (base)
1594                         add_delta_base_cache(p, obj_offset, base, base_size, type);
1595
1596                 if (!base) {
1597                         /*
1598                          * We're probably in deep shit, but let's try to fetch
1599                          * the required base anyway from another pack or loose.
1600                          * This is costly but should happen only in the presence
1601                          * of a corrupted pack, and is better than failing outright.
1602                          */
1603                         struct revindex_entry *revidx;
1604                         struct object_id base_oid;
1605                         revidx = find_pack_revindex(p, obj_offset);
1606                         if (revidx) {
1607                                 nth_packed_object_oid(&base_oid, p, revidx->nr);
1608                                 error("failed to read delta base object %s"
1609                                       " at offset %"PRIuMAX" from %s",
1610                                       oid_to_hex(&base_oid), (uintmax_t)obj_offset,
1611                                       p->pack_name);
1612                                 mark_bad_packed_object(p, base_oid.hash);
1613                                 base = read_object(&base_oid, &type, &base_size);
1614                                 external_base = base;
1615                         }
1616                 }
1617
1618                 i = --delta_stack_nr;
1619                 obj_offset = delta_stack[i].obj_offset;
1620                 curpos = delta_stack[i].curpos;
1621                 delta_size = delta_stack[i].size;
1622
1623                 if (!base)
1624                         continue;
1625
1626                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
1627
1628                 if (!delta_data) {
1629                         error("failed to unpack compressed delta "
1630                               "at offset %"PRIuMAX" from %s",
1631                               (uintmax_t)curpos, p->pack_name);
1632                         data = NULL;
1633                         free(external_base);
1634                         continue;
1635                 }
1636
1637                 data = patch_delta(base, base_size,
1638                                    delta_data, delta_size,
1639                                    &size);
1640
1641                 /*
1642                  * We could not apply the delta; warn the user, but keep going.
1643                  * Our failure will be noticed either in the next iteration of
1644                  * the loop, or if this is the final delta, in the caller when
1645                  * we return NULL. Those code paths will take care of making
1646                  * a more explicit warning and retrying with another copy of
1647                  * the object.
1648                  */
1649                 if (!data)
1650                         error("failed to apply delta");
1651
1652                 free(delta_data);
1653                 free(external_base);
1654         }
1655
1656         if (final_type)
1657                 *final_type = type;
1658         if (final_size)
1659                 *final_size = size;
1660
1661 out:
1662         unuse_pack(&w_curs);
1663
1664         if (delta_stack != small_delta_stack)
1665                 free(delta_stack);
1666
1667         return data;
1668 }
1669
1670 int bsearch_pack(const struct object_id *oid, const struct packed_git *p, uint32_t *result)
1671 {
1672         const unsigned char *index_fanout = p->index_data;
1673         const unsigned char *index_lookup;
1674         int index_lookup_width;
1675
1676         if (!index_fanout)
1677                 BUG("bsearch_pack called without a valid pack-index");
1678
1679         index_lookup = index_fanout + 4 * 256;
1680         if (p->index_version == 1) {
1681                 index_lookup_width = 24;
1682                 index_lookup += 4;
1683         } else {
1684                 index_lookup_width = 20;
1685                 index_fanout += 8;
1686                 index_lookup += 8;
1687         }
1688
1689         return bsearch_hash(oid->hash, (const uint32_t*)index_fanout,
1690                             index_lookup, index_lookup_width, result);
1691 }
1692
1693 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1694                                             uint32_t n)
1695 {
1696         const unsigned char *index = p->index_data;
1697         if (!index) {
1698                 if (open_pack_index(p))
1699                         return NULL;
1700                 index = p->index_data;
1701         }
1702         if (n >= p->num_objects)
1703                 return NULL;
1704         index += 4 * 256;
1705         if (p->index_version == 1) {
1706                 return index + 24 * n + 4;
1707         } else {
1708                 index += 8;
1709                 return index + 20 * n;
1710         }
1711 }
1712
1713 const struct object_id *nth_packed_object_oid(struct object_id *oid,
1714                                               struct packed_git *p,
1715                                               uint32_t n)
1716 {
1717         const unsigned char *hash = nth_packed_object_sha1(p, n);
1718         if (!hash)
1719                 return NULL;
1720         hashcpy(oid->hash, hash);
1721         return oid;
1722 }
1723
1724 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
1725 {
1726         const unsigned char *ptr = vptr;
1727         const unsigned char *start = p->index_data;
1728         const unsigned char *end = start + p->index_size;
1729         if (ptr < start)
1730                 die(_("offset before start of pack index for %s (corrupt index?)"),
1731                     p->pack_name);
1732         /* No need to check for underflow; .idx files must be at least 8 bytes */
1733         if (ptr >= end - 8)
1734                 die(_("offset beyond end of pack index for %s (truncated index?)"),
1735                     p->pack_name);
1736 }
1737
1738 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1739 {
1740         const unsigned char *index = p->index_data;
1741         index += 4 * 256;
1742         if (p->index_version == 1) {
1743                 return ntohl(*((uint32_t *)(index + 24 * n)));
1744         } else {
1745                 uint32_t off;
1746                 index += 8 + p->num_objects * (20 + 4);
1747                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1748                 if (!(off & 0x80000000))
1749                         return off;
1750                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1751                 check_pack_index_ptr(p, index);
1752                 return get_be64(index);
1753         }
1754 }
1755
1756 off_t find_pack_entry_one(const unsigned char *sha1,
1757                                   struct packed_git *p)
1758 {
1759         const unsigned char *index = p->index_data;
1760         struct object_id oid;
1761         uint32_t result;
1762
1763         if (!index) {
1764                 if (open_pack_index(p))
1765                         return 0;
1766         }
1767
1768         hashcpy(oid.hash, sha1);
1769         if (bsearch_pack(&oid, p, &result))
1770                 return nth_packed_object_offset(p, result);
1771         return 0;
1772 }
1773
1774 int is_pack_valid(struct packed_git *p)
1775 {
1776         /* An already open pack is known to be valid. */
1777         if (p->pack_fd != -1)
1778                 return 1;
1779
1780         /* If the pack has one window completely covering the
1781          * file size, the pack is known to be valid even if
1782          * the descriptor is not currently open.
1783          */
1784         if (p->windows) {
1785                 struct pack_window *w = p->windows;
1786
1787                 if (!w->offset && w->len == p->pack_size)
1788                         return 1;
1789         }
1790
1791         /* Force the pack to open to prove its valid. */
1792         return !open_packed_git(p);
1793 }
1794
1795 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1796                                   struct packed_git *packs)
1797 {
1798         struct packed_git *p;
1799
1800         for (p = packs; p; p = p->next) {
1801                 if (find_pack_entry_one(sha1, p))
1802                         return p;
1803         }
1804         return NULL;
1805
1806 }
1807
1808 static int fill_pack_entry(const unsigned char *sha1,
1809                            struct pack_entry *e,
1810                            struct packed_git *p)
1811 {
1812         off_t offset;
1813
1814         if (p->num_bad_objects) {
1815                 unsigned i;
1816                 for (i = 0; i < p->num_bad_objects; i++)
1817                         if (!hashcmp(sha1, p->bad_object_sha1 + 20 * i))
1818                                 return 0;
1819         }
1820
1821         offset = find_pack_entry_one(sha1, p);
1822         if (!offset)
1823                 return 0;
1824
1825         /*
1826          * We are about to tell the caller where they can locate the
1827          * requested object.  We better make sure the packfile is
1828          * still here and can be accessed before supplying that
1829          * answer, as it may have been deleted since the index was
1830          * loaded!
1831          */
1832         if (!is_pack_valid(p))
1833                 return 0;
1834         e->offset = offset;
1835         e->p = p;
1836         hashcpy(e->sha1, sha1);
1837         return 1;
1838 }
1839
1840 int find_pack_entry(struct repository *r, const unsigned char *sha1, struct pack_entry *e)
1841 {
1842         struct list_head *pos;
1843
1844         prepare_packed_git(r);
1845         if (!r->objects->packed_git)
1846                 return 0;
1847
1848         list_for_each(pos, &r->objects->packed_git_mru) {
1849                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
1850                 if (fill_pack_entry(sha1, e, p)) {
1851                         list_move(&p->mru, &r->objects->packed_git_mru);
1852                         return 1;
1853                 }
1854         }
1855         return 0;
1856 }
1857
1858 int has_sha1_pack(const unsigned char *sha1)
1859 {
1860         struct pack_entry e;
1861         return find_pack_entry(the_repository, sha1, &e);
1862 }
1863
1864 int has_pack_index(const unsigned char *sha1)
1865 {
1866         struct stat st;
1867         if (stat(sha1_pack_index_name(sha1), &st))
1868                 return 0;
1869         return 1;
1870 }
1871
1872 int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
1873 {
1874         uint32_t i;
1875         int r = 0;
1876
1877         for (i = 0; i < p->num_objects; i++) {
1878                 struct object_id oid;
1879
1880                 if (!nth_packed_object_oid(&oid, p, i))
1881                         return error("unable to get sha1 of object %u in %s",
1882                                      i, p->pack_name);
1883
1884                 r = cb(&oid, p, i, data);
1885                 if (r)
1886                         break;
1887         }
1888         return r;
1889 }
1890
1891 int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
1892 {
1893         struct packed_git *p;
1894         int r = 0;
1895         int pack_errors = 0;
1896
1897         prepare_packed_git(the_repository);
1898         for (p = the_repository->objects->packed_git; p; p = p->next) {
1899                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
1900                         continue;
1901                 if ((flags & FOR_EACH_OBJECT_PROMISOR_ONLY) &&
1902                     !p->pack_promisor)
1903                         continue;
1904                 if (open_pack_index(p)) {
1905                         pack_errors = 1;
1906                         continue;
1907                 }
1908                 r = for_each_object_in_pack(p, cb, data);
1909                 if (r)
1910                         break;
1911         }
1912         return r ? r : pack_errors;
1913 }
1914
1915 static int add_promisor_object(const struct object_id *oid,
1916                                struct packed_git *pack,
1917                                uint32_t pos,
1918                                void *set_)
1919 {
1920         struct oidset *set = set_;
1921         struct object *obj = parse_object(oid);
1922         if (!obj)
1923                 return 1;
1924
1925         oidset_insert(set, oid);
1926
1927         /*
1928          * If this is a tree, commit, or tag, the objects it refers
1929          * to are also promisor objects. (Blobs refer to no objects->)
1930          */
1931         if (obj->type == OBJ_TREE) {
1932                 struct tree *tree = (struct tree *)obj;
1933                 struct tree_desc desc;
1934                 struct name_entry entry;
1935                 if (init_tree_desc_gently(&desc, tree->buffer, tree->size))
1936                         /*
1937                          * Error messages are given when packs are
1938                          * verified, so do not print any here.
1939                          */
1940                         return 0;
1941                 while (tree_entry_gently(&desc, &entry))
1942                         oidset_insert(set, entry.oid);
1943         } else if (obj->type == OBJ_COMMIT) {
1944                 struct commit *commit = (struct commit *) obj;
1945                 struct commit_list *parents = commit->parents;
1946
1947                 oidset_insert(set, &commit->tree->object.oid);
1948                 for (; parents; parents = parents->next)
1949                         oidset_insert(set, &parents->item->object.oid);
1950         } else if (obj->type == OBJ_TAG) {
1951                 struct tag *tag = (struct tag *) obj;
1952                 oidset_insert(set, &tag->tagged->oid);
1953         }
1954         return 0;
1955 }
1956
1957 int is_promisor_object(const struct object_id *oid)
1958 {
1959         static struct oidset promisor_objects;
1960         static int promisor_objects_prepared;
1961
1962         if (!promisor_objects_prepared) {
1963                 if (repository_format_partial_clone) {
1964                         for_each_packed_object(add_promisor_object,
1965                                                &promisor_objects,
1966                                                FOR_EACH_OBJECT_PROMISOR_ONLY);
1967                 }
1968                 promisor_objects_prepared = 1;
1969         }
1970         return oidset_contains(&promisor_objects, oid);
1971 }