Merge branch 'bc/tree-walk-oid'
[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 #include "midx.h"
19
20 char *odb_pack_name(struct strbuf *buf,
21                     const unsigned char *sha1,
22                     const char *ext)
23 {
24         strbuf_reset(buf);
25         strbuf_addf(buf, "%s/pack/pack-%s.%s", get_object_directory(),
26                     sha1_to_hex(sha1), ext);
27         return buf->buf;
28 }
29
30 char *sha1_pack_name(const unsigned char *sha1)
31 {
32         static struct strbuf buf = STRBUF_INIT;
33         return odb_pack_name(&buf, sha1, "pack");
34 }
35
36 char *sha1_pack_index_name(const unsigned char *sha1)
37 {
38         static struct strbuf buf = STRBUF_INIT;
39         return odb_pack_name(&buf, sha1, "idx");
40 }
41
42 static unsigned int pack_used_ctr;
43 static unsigned int pack_mmap_calls;
44 static unsigned int peak_pack_open_windows;
45 static unsigned int pack_open_windows;
46 static unsigned int pack_open_fds;
47 static unsigned int pack_max_fds;
48 static size_t peak_pack_mapped;
49 static size_t pack_mapped;
50
51 #define SZ_FMT PRIuMAX
52 static inline uintmax_t sz_fmt(size_t s) { return s; }
53
54 void pack_report(void)
55 {
56         fprintf(stderr,
57                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
58                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
59                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
60                 sz_fmt(getpagesize()),
61                 sz_fmt(packed_git_window_size),
62                 sz_fmt(packed_git_limit));
63         fprintf(stderr,
64                 "pack_report: pack_used_ctr            = %10u\n"
65                 "pack_report: pack_mmap_calls          = %10u\n"
66                 "pack_report: pack_open_windows        = %10u / %10u\n"
67                 "pack_report: pack_mapped              = "
68                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
69                 pack_used_ctr,
70                 pack_mmap_calls,
71                 pack_open_windows, peak_pack_open_windows,
72                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
73 }
74
75 /*
76  * Open and mmap the index file at path, perform a couple of
77  * consistency checks, then record its information to p.  Return 0 on
78  * success.
79  */
80 static int check_packed_git_idx(const char *path, struct packed_git *p)
81 {
82         void *idx_map;
83         size_t idx_size;
84         int fd = git_open(path), ret;
85         struct stat st;
86         const unsigned int hashsz = the_hash_algo->rawsz;
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 + hashsz + hashsz) {
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         ret = load_idx(path, hashsz, idx_map, idx_size, p);
103
104         if (ret)
105                 munmap(idx_map, idx_size);
106
107         return ret;
108 }
109
110 int load_idx(const char *path, const unsigned int hashsz, void *idx_map,
111              size_t idx_size, struct packed_git *p)
112 {
113         struct pack_idx_header *hdr = idx_map;
114         uint32_t version, nr, i, *index;
115
116         if (idx_size < 4 * 256 + hashsz + hashsz)
117                 return error("index file %s is too small", path);
118         if (idx_map == NULL)
119                 return error("empty data");
120
121         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
122                 version = ntohl(hdr->idx_version);
123                 if (version < 2 || version > 2)
124                         return error("index file %s is version %"PRIu32
125                                      " and is not supported by this binary"
126                                      " (try upgrading GIT to a newer version)",
127                                      path, version);
128         } else
129                 version = 1;
130
131         nr = 0;
132         index = idx_map;
133         if (version > 1)
134                 index += 2;  /* skip index header */
135         for (i = 0; i < 256; i++) {
136                 uint32_t n = ntohl(index[i]);
137                 if (n < nr)
138                         return error("non-monotonic index %s", path);
139                 nr = n;
140         }
141
142         if (version == 1) {
143                 /*
144                  * Total size:
145                  *  - 256 index entries 4 bytes each
146                  *  - 24-byte entries * nr (object ID + 4-byte offset)
147                  *  - hash of the packfile
148                  *  - file checksum
149                  */
150                 if (idx_size != 4 * 256 + nr * (hashsz + 4) + hashsz + hashsz)
151                         return error("wrong index v1 file size in %s", path);
152         } else if (version == 2) {
153                 /*
154                  * Minimum size:
155                  *  - 8 bytes of header
156                  *  - 256 index entries 4 bytes each
157                  *  - object ID entry * nr
158                  *  - 4-byte crc entry * nr
159                  *  - 4-byte offset entry * nr
160                  *  - hash of the packfile
161                  *  - file checksum
162                  * And after the 4-byte offset table might be a
163                  * variable sized table containing 8-byte entries
164                  * for offsets larger than 2^31.
165                  */
166                 unsigned long min_size = 8 + 4*256 + nr*(hashsz + 4 + 4) + hashsz + hashsz;
167                 unsigned long max_size = min_size;
168                 if (nr)
169                         max_size += (nr - 1)*8;
170                 if (idx_size < min_size || idx_size > max_size)
171                         return error("wrong index v2 file size in %s", path);
172                 if (idx_size != min_size &&
173                     /*
174                      * make sure we can deal with large pack offsets.
175                      * 31-bit signed offset won't be enough, neither
176                      * 32-bit unsigned one will be.
177                      */
178                     (sizeof(off_t) <= 4))
179                         return error("pack too large for current definition of off_t in %s", path);
180         }
181
182         p->index_version = version;
183         p->index_data = idx_map;
184         p->index_size = idx_size;
185         p->num_objects = nr;
186         return 0;
187 }
188
189 int open_pack_index(struct packed_git *p)
190 {
191         char *idx_name;
192         size_t len;
193         int ret;
194
195         if (p->index_data)
196                 return 0;
197
198         if (!strip_suffix(p->pack_name, ".pack", &len))
199                 BUG("pack_name does not end in .pack");
200         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
201         ret = check_packed_git_idx(idx_name, p);
202         free(idx_name);
203         return ret;
204 }
205
206 uint32_t get_pack_fanout(struct packed_git *p, uint32_t value)
207 {
208         const uint32_t *level1_ofs = p->index_data;
209
210         if (!level1_ofs) {
211                 if (open_pack_index(p))
212                         return 0;
213                 level1_ofs = p->index_data;
214         }
215
216         if (p->index_version > 1) {
217                 level1_ofs += 2;
218         }
219
220         return ntohl(level1_ofs[value]);
221 }
222
223 static struct packed_git *alloc_packed_git(int extra)
224 {
225         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
226         memset(p, 0, sizeof(*p));
227         p->pack_fd = -1;
228         return p;
229 }
230
231 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
232 {
233         const char *path = sha1_pack_name(sha1);
234         size_t alloc = st_add(strlen(path), 1);
235         struct packed_git *p = alloc_packed_git(alloc);
236
237         memcpy(p->pack_name, path, alloc); /* includes NUL */
238         hashcpy(p->sha1, sha1);
239         if (check_packed_git_idx(idx_path, p)) {
240                 free(p);
241                 return NULL;
242         }
243
244         return p;
245 }
246
247 static void scan_windows(struct packed_git *p,
248         struct packed_git **lru_p,
249         struct pack_window **lru_w,
250         struct pack_window **lru_l)
251 {
252         struct pack_window *w, *w_l;
253
254         for (w_l = NULL, w = p->windows; w; w = w->next) {
255                 if (!w->inuse_cnt) {
256                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
257                                 *lru_p = p;
258                                 *lru_w = w;
259                                 *lru_l = w_l;
260                         }
261                 }
262                 w_l = w;
263         }
264 }
265
266 static int unuse_one_window(struct packed_git *current)
267 {
268         struct packed_git *p, *lru_p = NULL;
269         struct pack_window *lru_w = NULL, *lru_l = NULL;
270
271         if (current)
272                 scan_windows(current, &lru_p, &lru_w, &lru_l);
273         for (p = the_repository->objects->packed_git; p; p = p->next)
274                 scan_windows(p, &lru_p, &lru_w, &lru_l);
275         if (lru_p) {
276                 munmap(lru_w->base, lru_w->len);
277                 pack_mapped -= lru_w->len;
278                 if (lru_l)
279                         lru_l->next = lru_w->next;
280                 else
281                         lru_p->windows = lru_w->next;
282                 free(lru_w);
283                 pack_open_windows--;
284                 return 1;
285         }
286         return 0;
287 }
288
289 void release_pack_memory(size_t need)
290 {
291         size_t cur = pack_mapped;
292         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
293                 ; /* nothing */
294 }
295
296 void close_pack_windows(struct packed_git *p)
297 {
298         while (p->windows) {
299                 struct pack_window *w = p->windows;
300
301                 if (w->inuse_cnt)
302                         die("pack '%s' still has open windows to it",
303                             p->pack_name);
304                 munmap(w->base, w->len);
305                 pack_mapped -= w->len;
306                 pack_open_windows--;
307                 p->windows = w->next;
308                 free(w);
309         }
310 }
311
312 static int close_pack_fd(struct packed_git *p)
313 {
314         if (p->pack_fd < 0)
315                 return 0;
316
317         close(p->pack_fd);
318         pack_open_fds--;
319         p->pack_fd = -1;
320
321         return 1;
322 }
323
324 void close_pack_index(struct packed_git *p)
325 {
326         if (p->index_data) {
327                 munmap((void *)p->index_data, p->index_size);
328                 p->index_data = NULL;
329         }
330 }
331
332 void close_pack(struct packed_git *p)
333 {
334         close_pack_windows(p);
335         close_pack_fd(p);
336         close_pack_index(p);
337 }
338
339 void close_all_packs(struct raw_object_store *o)
340 {
341         struct packed_git *p;
342
343         for (p = o->packed_git; p; p = p->next)
344                 if (p->do_not_close)
345                         BUG("want to close pack marked 'do-not-close'");
346                 else
347                         close_pack(p);
348
349         if (o->multi_pack_index) {
350                 close_midx(o->multi_pack_index);
351                 o->multi_pack_index = NULL;
352         }
353 }
354
355 /*
356  * The LRU pack is the one with the oldest MRU window, preferring packs
357  * with no used windows, or the oldest mtime if it has no windows allocated.
358  */
359 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
360 {
361         struct pack_window *w, *this_mru_w;
362         int has_windows_inuse = 0;
363
364         /*
365          * Reject this pack if it has windows and the previously selected
366          * one does not.  If this pack does not have windows, reject
367          * it if the pack file is newer than the previously selected one.
368          */
369         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
370                 return;
371
372         for (w = this_mru_w = p->windows; w; w = w->next) {
373                 /*
374                  * Reject this pack if any of its windows are in use,
375                  * but the previously selected pack did not have any
376                  * inuse windows.  Otherwise, record that this pack
377                  * has windows in use.
378                  */
379                 if (w->inuse_cnt) {
380                         if (*accept_windows_inuse)
381                                 has_windows_inuse = 1;
382                         else
383                                 return;
384                 }
385
386                 if (w->last_used > this_mru_w->last_used)
387                         this_mru_w = w;
388
389                 /*
390                  * Reject this pack if it has windows that have been
391                  * used more recently than the previously selected pack.
392                  * If the previously selected pack had windows inuse and
393                  * we have not encountered a window in this pack that is
394                  * inuse, skip this check since we prefer a pack with no
395                  * inuse windows to one that has inuse windows.
396                  */
397                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
398                     this_mru_w->last_used > (*mru_w)->last_used)
399                         return;
400         }
401
402         /*
403          * Select this pack.
404          */
405         *mru_w = this_mru_w;
406         *lru_p = p;
407         *accept_windows_inuse = has_windows_inuse;
408 }
409
410 static int close_one_pack(void)
411 {
412         struct packed_git *p, *lru_p = NULL;
413         struct pack_window *mru_w = NULL;
414         int accept_windows_inuse = 1;
415
416         for (p = the_repository->objects->packed_git; p; p = p->next) {
417                 if (p->pack_fd == -1)
418                         continue;
419                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
420         }
421
422         if (lru_p)
423                 return close_pack_fd(lru_p);
424
425         return 0;
426 }
427
428 static unsigned int get_max_fd_limit(void)
429 {
430 #ifdef RLIMIT_NOFILE
431         {
432                 struct rlimit lim;
433
434                 if (!getrlimit(RLIMIT_NOFILE, &lim))
435                         return lim.rlim_cur;
436         }
437 #endif
438
439 #ifdef _SC_OPEN_MAX
440         {
441                 long open_max = sysconf(_SC_OPEN_MAX);
442                 if (0 < open_max)
443                         return open_max;
444                 /*
445                  * Otherwise, we got -1 for one of the two
446                  * reasons:
447                  *
448                  * (1) sysconf() did not understand _SC_OPEN_MAX
449                  *     and signaled an error with -1; or
450                  * (2) sysconf() said there is no limit.
451                  *
452                  * We _could_ clear errno before calling sysconf() to
453                  * tell these two cases apart and return a huge number
454                  * in the latter case to let the caller cap it to a
455                  * value that is not so selfish, but letting the
456                  * fallback OPEN_MAX codepath take care of these cases
457                  * is a lot simpler.
458                  */
459         }
460 #endif
461
462 #ifdef OPEN_MAX
463         return OPEN_MAX;
464 #else
465         return 1; /* see the caller ;-) */
466 #endif
467 }
468
469 /*
470  * Do not call this directly as this leaks p->pack_fd on error return;
471  * call open_packed_git() instead.
472  */
473 static int open_packed_git_1(struct packed_git *p)
474 {
475         struct stat st;
476         struct pack_header hdr;
477         unsigned char hash[GIT_MAX_RAWSZ];
478         unsigned char *idx_hash;
479         long fd_flag;
480         ssize_t read_result;
481         const unsigned hashsz = the_hash_algo->rawsz;
482
483         if (!p->index_data) {
484                 struct multi_pack_index *m;
485                 const char *pack_name = strrchr(p->pack_name, '/');
486
487                 for (m = the_repository->objects->multi_pack_index;
488                      m; m = m->next) {
489                         if (midx_contains_pack(m, pack_name))
490                                 break;
491                 }
492
493                 if (!m && open_pack_index(p))
494                         return error("packfile %s index unavailable", p->pack_name);
495         }
496
497         if (!pack_max_fds) {
498                 unsigned int max_fds = get_max_fd_limit();
499
500                 /* Save 3 for stdin/stdout/stderr, 22 for work */
501                 if (25 < max_fds)
502                         pack_max_fds = max_fds - 25;
503                 else
504                         pack_max_fds = 1;
505         }
506
507         while (pack_max_fds <= pack_open_fds && close_one_pack())
508                 ; /* nothing */
509
510         p->pack_fd = git_open(p->pack_name);
511         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
512                 return -1;
513         pack_open_fds++;
514
515         /* If we created the struct before we had the pack we lack size. */
516         if (!p->pack_size) {
517                 if (!S_ISREG(st.st_mode))
518                         return error("packfile %s not a regular file", p->pack_name);
519                 p->pack_size = st.st_size;
520         } else if (p->pack_size != st.st_size)
521                 return error("packfile %s size changed", p->pack_name);
522
523         /* We leave these file descriptors open with sliding mmap;
524          * there is no point keeping them open across exec(), though.
525          */
526         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
527         if (fd_flag < 0)
528                 return error("cannot determine file descriptor flags");
529         fd_flag |= FD_CLOEXEC;
530         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
531                 return error("cannot set FD_CLOEXEC");
532
533         /* Verify we recognize this pack file format. */
534         read_result = read_in_full(p->pack_fd, &hdr, sizeof(hdr));
535         if (read_result < 0)
536                 return error_errno("error reading from %s", p->pack_name);
537         if (read_result != sizeof(hdr))
538                 return error("file %s is far too short to be a packfile", p->pack_name);
539         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
540                 return error("file %s is not a GIT packfile", p->pack_name);
541         if (!pack_version_ok(hdr.hdr_version))
542                 return error("packfile %s is version %"PRIu32" and not"
543                         " supported (try upgrading GIT to a newer version)",
544                         p->pack_name, ntohl(hdr.hdr_version));
545
546         /* Skip index checking if in multi-pack-index */
547         if (!p->index_data)
548                 return 0;
549
550         /* Verify the pack matches its index. */
551         if (p->num_objects != ntohl(hdr.hdr_entries))
552                 return error("packfile %s claims to have %"PRIu32" objects"
553                              " while index indicates %"PRIu32" objects",
554                              p->pack_name, ntohl(hdr.hdr_entries),
555                              p->num_objects);
556         if (lseek(p->pack_fd, p->pack_size - hashsz, SEEK_SET) == -1)
557                 return error("end of packfile %s is unavailable", p->pack_name);
558         read_result = read_in_full(p->pack_fd, hash, hashsz);
559         if (read_result < 0)
560                 return error_errno("error reading from %s", p->pack_name);
561         if (read_result != hashsz)
562                 return error("packfile %s signature is unavailable", p->pack_name);
563         idx_hash = ((unsigned char *)p->index_data) + p->index_size - hashsz * 2;
564         if (!hasheq(hash, idx_hash))
565                 return error("packfile %s does not match index", p->pack_name);
566         return 0;
567 }
568
569 static int open_packed_git(struct packed_git *p)
570 {
571         if (!open_packed_git_1(p))
572                 return 0;
573         close_pack_fd(p);
574         return -1;
575 }
576
577 static int in_window(struct pack_window *win, off_t offset)
578 {
579         /* We must promise at least one full hash after the
580          * offset is available from this window, otherwise the offset
581          * is not actually in this window and a different window (which
582          * has that one hash excess) must be used.  This is to support
583          * the object header and delta base parsing routines below.
584          */
585         off_t win_off = win->offset;
586         return win_off <= offset
587                 && (offset + the_hash_algo->rawsz) <= (win_off + win->len);
588 }
589
590 unsigned char *use_pack(struct packed_git *p,
591                 struct pack_window **w_cursor,
592                 off_t offset,
593                 unsigned long *left)
594 {
595         struct pack_window *win = *w_cursor;
596
597         /* Since packfiles end in a hash of their content and it's
598          * pointless to ask for an offset into the middle of that
599          * hash, and the in_window function above wouldn't match
600          * don't allow an offset too close to the end of the file.
601          */
602         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
603                 die("packfile %s cannot be accessed", p->pack_name);
604         if (offset > (p->pack_size - the_hash_algo->rawsz))
605                 die("offset beyond end of packfile (truncated pack?)");
606         if (offset < 0)
607                 die(_("offset before end of packfile (broken .idx?)"));
608
609         if (!win || !in_window(win, offset)) {
610                 if (win)
611                         win->inuse_cnt--;
612                 for (win = p->windows; win; win = win->next) {
613                         if (in_window(win, offset))
614                                 break;
615                 }
616                 if (!win) {
617                         size_t window_align = packed_git_window_size / 2;
618                         off_t len;
619
620                         if (p->pack_fd == -1 && open_packed_git(p))
621                                 die("packfile %s cannot be accessed", p->pack_name);
622
623                         win = xcalloc(1, sizeof(*win));
624                         win->offset = (offset / window_align) * window_align;
625                         len = p->pack_size - win->offset;
626                         if (len > packed_git_window_size)
627                                 len = packed_git_window_size;
628                         win->len = (size_t)len;
629                         pack_mapped += win->len;
630                         while (packed_git_limit < pack_mapped
631                                 && unuse_one_window(p))
632                                 ; /* nothing */
633                         win->base = xmmap(NULL, win->len,
634                                 PROT_READ, MAP_PRIVATE,
635                                 p->pack_fd, win->offset);
636                         if (win->base == MAP_FAILED)
637                                 die_errno("packfile %s cannot be mapped",
638                                           p->pack_name);
639                         if (!win->offset && win->len == p->pack_size
640                                 && !p->do_not_close)
641                                 close_pack_fd(p);
642                         pack_mmap_calls++;
643                         pack_open_windows++;
644                         if (pack_mapped > peak_pack_mapped)
645                                 peak_pack_mapped = pack_mapped;
646                         if (pack_open_windows > peak_pack_open_windows)
647                                 peak_pack_open_windows = pack_open_windows;
648                         win->next = p->windows;
649                         p->windows = win;
650                 }
651         }
652         if (win != *w_cursor) {
653                 win->last_used = pack_used_ctr++;
654                 win->inuse_cnt++;
655                 *w_cursor = win;
656         }
657         offset -= win->offset;
658         if (left)
659                 *left = win->len - xsize_t(offset);
660         return win->base + offset;
661 }
662
663 void unuse_pack(struct pack_window **w_cursor)
664 {
665         struct pack_window *w = *w_cursor;
666         if (w) {
667                 w->inuse_cnt--;
668                 *w_cursor = NULL;
669         }
670 }
671
672 static void try_to_free_pack_memory(size_t size)
673 {
674         release_pack_memory(size);
675 }
676
677 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
678 {
679         static int have_set_try_to_free_routine;
680         struct stat st;
681         size_t alloc;
682         struct packed_git *p;
683
684         if (!have_set_try_to_free_routine) {
685                 have_set_try_to_free_routine = 1;
686                 set_try_to_free_routine(try_to_free_pack_memory);
687         }
688
689         /*
690          * Make sure a corresponding .pack file exists and that
691          * the index looks sane.
692          */
693         if (!strip_suffix_mem(path, &path_len, ".idx"))
694                 return NULL;
695
696         /*
697          * ".promisor" is long enough to hold any suffix we're adding (and
698          * the use xsnprintf double-checks that)
699          */
700         alloc = st_add3(path_len, strlen(".promisor"), 1);
701         p = alloc_packed_git(alloc);
702         memcpy(p->pack_name, path, path_len);
703
704         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
705         if (!access(p->pack_name, F_OK))
706                 p->pack_keep = 1;
707
708         xsnprintf(p->pack_name + path_len, alloc - path_len, ".promisor");
709         if (!access(p->pack_name, F_OK))
710                 p->pack_promisor = 1;
711
712         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
713         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
714                 free(p);
715                 return NULL;
716         }
717
718         /* ok, it looks sane as far as we can check without
719          * actually mapping the pack file.
720          */
721         p->pack_size = st.st_size;
722         p->pack_local = local;
723         p->mtime = st.st_mtime;
724         if (path_len < the_hash_algo->hexsz ||
725             get_sha1_hex(path + path_len - the_hash_algo->hexsz, p->sha1))
726                 hashclr(p->sha1);
727         return p;
728 }
729
730 void install_packed_git(struct repository *r, struct packed_git *pack)
731 {
732         if (pack->pack_fd != -1)
733                 pack_open_fds++;
734
735         pack->next = r->objects->packed_git;
736         r->objects->packed_git = pack;
737 }
738
739 void (*report_garbage)(unsigned seen_bits, const char *path);
740
741 static void report_helper(const struct string_list *list,
742                           int seen_bits, int first, int last)
743 {
744         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
745                 return;
746
747         for (; first < last; first++)
748                 report_garbage(seen_bits, list->items[first].string);
749 }
750
751 static void report_pack_garbage(struct string_list *list)
752 {
753         int i, baselen = -1, first = 0, seen_bits = 0;
754
755         if (!report_garbage)
756                 return;
757
758         string_list_sort(list);
759
760         for (i = 0; i < list->nr; i++) {
761                 const char *path = list->items[i].string;
762                 if (baselen != -1 &&
763                     strncmp(path, list->items[first].string, baselen)) {
764                         report_helper(list, seen_bits, first, i);
765                         baselen = -1;
766                         seen_bits = 0;
767                 }
768                 if (baselen == -1) {
769                         const char *dot = strrchr(path, '.');
770                         if (!dot) {
771                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
772                                 continue;
773                         }
774                         baselen = dot - path + 1;
775                         first = i;
776                 }
777                 if (!strcmp(path + baselen, "pack"))
778                         seen_bits |= 1;
779                 else if (!strcmp(path + baselen, "idx"))
780                         seen_bits |= 2;
781         }
782         report_helper(list, seen_bits, first, list->nr);
783 }
784
785 void for_each_file_in_pack_dir(const char *objdir,
786                                each_file_in_pack_dir_fn fn,
787                                void *data)
788 {
789         struct strbuf path = STRBUF_INIT;
790         size_t dirnamelen;
791         DIR *dir;
792         struct dirent *de;
793
794         strbuf_addstr(&path, objdir);
795         strbuf_addstr(&path, "/pack");
796         dir = opendir(path.buf);
797         if (!dir) {
798                 if (errno != ENOENT)
799                         error_errno("unable to open object pack directory: %s",
800                                     path.buf);
801                 strbuf_release(&path);
802                 return;
803         }
804         strbuf_addch(&path, '/');
805         dirnamelen = path.len;
806         while ((de = readdir(dir)) != NULL) {
807                 if (is_dot_or_dotdot(de->d_name))
808                         continue;
809
810                 strbuf_setlen(&path, dirnamelen);
811                 strbuf_addstr(&path, de->d_name);
812
813                 fn(path.buf, path.len, de->d_name, data);
814         }
815
816         closedir(dir);
817         strbuf_release(&path);
818 }
819
820 struct prepare_pack_data {
821         struct repository *r;
822         struct string_list *garbage;
823         int local;
824         struct multi_pack_index *m;
825 };
826
827 static void prepare_pack(const char *full_name, size_t full_name_len,
828                          const char *file_name, void *_data)
829 {
830         struct prepare_pack_data *data = (struct prepare_pack_data *)_data;
831         struct packed_git *p;
832         size_t base_len = full_name_len;
833
834         if (strip_suffix_mem(full_name, &base_len, ".idx") &&
835             !(data->m && midx_contains_pack(data->m, file_name))) {
836                 /* Don't reopen a pack we already have. */
837                 for (p = data->r->objects->packed_git; p; p = p->next) {
838                         size_t len;
839                         if (strip_suffix(p->pack_name, ".pack", &len) &&
840                             len == base_len &&
841                             !memcmp(p->pack_name, full_name, len))
842                                 break;
843                 }
844
845                 if (!p) {
846                         p = add_packed_git(full_name, full_name_len, data->local);
847                         if (p)
848                                 install_packed_git(data->r, p);
849                 }
850         }
851
852         if (!report_garbage)
853                 return;
854
855         if (!strcmp(file_name, "multi-pack-index"))
856                 return;
857         if (ends_with(file_name, ".idx") ||
858             ends_with(file_name, ".pack") ||
859             ends_with(file_name, ".bitmap") ||
860             ends_with(file_name, ".keep") ||
861             ends_with(file_name, ".promisor"))
862                 string_list_append(data->garbage, full_name);
863         else
864                 report_garbage(PACKDIR_FILE_GARBAGE, full_name);
865 }
866
867 static void prepare_packed_git_one(struct repository *r, char *objdir, int local)
868 {
869         struct prepare_pack_data data;
870         struct string_list garbage = STRING_LIST_INIT_DUP;
871
872         data.m = r->objects->multi_pack_index;
873
874         /* look for the multi-pack-index for this object directory */
875         while (data.m && strcmp(data.m->object_dir, objdir))
876                 data.m = data.m->next;
877
878         data.r = r;
879         data.garbage = &garbage;
880         data.local = local;
881
882         for_each_file_in_pack_dir(objdir, prepare_pack, &data);
883
884         report_pack_garbage(data.garbage);
885         string_list_clear(data.garbage, 0);
886 }
887
888 static void prepare_packed_git(struct repository *r);
889 /*
890  * Give a fast, rough count of the number of objects in the repository. This
891  * ignores loose objects completely. If you have a lot of them, then either
892  * you should repack because your performance will be awful, or they are
893  * all unreachable objects about to be pruned, in which case they're not really
894  * interesting as a measure of repo size in the first place.
895  */
896 unsigned long approximate_object_count(void)
897 {
898         if (!the_repository->objects->approximate_object_count_valid) {
899                 unsigned long count;
900                 struct multi_pack_index *m;
901                 struct packed_git *p;
902
903                 prepare_packed_git(the_repository);
904                 count = 0;
905                 for (m = get_multi_pack_index(the_repository); m; m = m->next)
906                         count += m->num_objects;
907                 for (p = the_repository->objects->packed_git; p; p = p->next) {
908                         if (open_pack_index(p))
909                                 continue;
910                         count += p->num_objects;
911                 }
912                 the_repository->objects->approximate_object_count = count;
913         }
914         return the_repository->objects->approximate_object_count;
915 }
916
917 static void *get_next_packed_git(const void *p)
918 {
919         return ((const struct packed_git *)p)->next;
920 }
921
922 static void set_next_packed_git(void *p, void *next)
923 {
924         ((struct packed_git *)p)->next = next;
925 }
926
927 static int sort_pack(const void *a_, const void *b_)
928 {
929         const struct packed_git *a = a_;
930         const struct packed_git *b = b_;
931         int st;
932
933         /*
934          * Local packs tend to contain objects specific to our
935          * variant of the project than remote ones.  In addition,
936          * remote ones could be on a network mounted filesystem.
937          * Favor local ones for these reasons.
938          */
939         st = a->pack_local - b->pack_local;
940         if (st)
941                 return -st;
942
943         /*
944          * Younger packs tend to contain more recent objects,
945          * and more recent objects tend to get accessed more
946          * often.
947          */
948         if (a->mtime < b->mtime)
949                 return 1;
950         else if (a->mtime == b->mtime)
951                 return 0;
952         return -1;
953 }
954
955 static void rearrange_packed_git(struct repository *r)
956 {
957         r->objects->packed_git = llist_mergesort(
958                 r->objects->packed_git, get_next_packed_git,
959                 set_next_packed_git, sort_pack);
960 }
961
962 static void prepare_packed_git_mru(struct repository *r)
963 {
964         struct packed_git *p;
965
966         INIT_LIST_HEAD(&r->objects->packed_git_mru);
967
968         for (p = r->objects->packed_git; p; p = p->next)
969                 list_add_tail(&p->mru, &r->objects->packed_git_mru);
970 }
971
972 static void prepare_packed_git(struct repository *r)
973 {
974         struct object_directory *odb;
975
976         if (r->objects->packed_git_initialized)
977                 return;
978
979         prepare_alt_odb(r);
980         for (odb = r->objects->odb; odb; odb = odb->next) {
981                 int local = (odb == r->objects->odb);
982                 prepare_multi_pack_index_one(r, odb->path, local);
983                 prepare_packed_git_one(r, odb->path, local);
984         }
985         rearrange_packed_git(r);
986
987         r->objects->all_packs = NULL;
988
989         prepare_packed_git_mru(r);
990         r->objects->packed_git_initialized = 1;
991 }
992
993 void reprepare_packed_git(struct repository *r)
994 {
995         struct object_directory *odb;
996
997         for (odb = r->objects->odb; odb; odb = odb->next)
998                 odb_clear_loose_cache(odb);
999
1000         r->objects->approximate_object_count_valid = 0;
1001         r->objects->packed_git_initialized = 0;
1002         prepare_packed_git(r);
1003 }
1004
1005 struct packed_git *get_packed_git(struct repository *r)
1006 {
1007         prepare_packed_git(r);
1008         return r->objects->packed_git;
1009 }
1010
1011 struct multi_pack_index *get_multi_pack_index(struct repository *r)
1012 {
1013         prepare_packed_git(r);
1014         return r->objects->multi_pack_index;
1015 }
1016
1017 struct packed_git *get_all_packs(struct repository *r)
1018 {
1019         prepare_packed_git(r);
1020
1021         if (!r->objects->all_packs) {
1022                 struct packed_git *p = r->objects->packed_git;
1023                 struct multi_pack_index *m;
1024
1025                 for (m = r->objects->multi_pack_index; m; m = m->next) {
1026                         uint32_t i;
1027                         for (i = 0; i < m->num_packs; i++) {
1028                                 if (!prepare_midx_pack(m, i)) {
1029                                         m->packs[i]->next = p;
1030                                         p = m->packs[i];
1031                                 }
1032                         }
1033                 }
1034
1035                 r->objects->all_packs = p;
1036         }
1037
1038         return r->objects->all_packs;
1039 }
1040
1041 struct list_head *get_packed_git_mru(struct repository *r)
1042 {
1043         prepare_packed_git(r);
1044         return &r->objects->packed_git_mru;
1045 }
1046
1047 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1048                 unsigned long len, enum object_type *type, unsigned long *sizep)
1049 {
1050         unsigned shift;
1051         unsigned long size, c;
1052         unsigned long used = 0;
1053
1054         c = buf[used++];
1055         *type = (c >> 4) & 7;
1056         size = c & 15;
1057         shift = 4;
1058         while (c & 0x80) {
1059                 if (len <= used || bitsizeof(long) <= shift) {
1060                         error("bad object header");
1061                         size = used = 0;
1062                         break;
1063                 }
1064                 c = buf[used++];
1065                 size += (c & 0x7f) << shift;
1066                 shift += 7;
1067         }
1068         *sizep = size;
1069         return used;
1070 }
1071
1072 unsigned long get_size_from_delta(struct packed_git *p,
1073                                   struct pack_window **w_curs,
1074                                   off_t curpos)
1075 {
1076         const unsigned char *data;
1077         unsigned char delta_head[20], *in;
1078         git_zstream stream;
1079         int st;
1080
1081         memset(&stream, 0, sizeof(stream));
1082         stream.next_out = delta_head;
1083         stream.avail_out = sizeof(delta_head);
1084
1085         git_inflate_init(&stream);
1086         do {
1087                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1088                 stream.next_in = in;
1089                 st = git_inflate(&stream, Z_FINISH);
1090                 curpos += stream.next_in - in;
1091         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1092                  stream.total_out < sizeof(delta_head));
1093         git_inflate_end(&stream);
1094         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1095                 error("delta data unpack-initial failed");
1096                 return 0;
1097         }
1098
1099         /* Examine the initial part of the delta to figure out
1100          * the result size.
1101          */
1102         data = delta_head;
1103
1104         /* ignore base size */
1105         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1106
1107         /* Read the result size */
1108         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1109 }
1110
1111 int unpack_object_header(struct packed_git *p,
1112                          struct pack_window **w_curs,
1113                          off_t *curpos,
1114                          unsigned long *sizep)
1115 {
1116         unsigned char *base;
1117         unsigned long left;
1118         unsigned long used;
1119         enum object_type type;
1120
1121         /* use_pack() assures us we have [base, base + 20) available
1122          * as a range that we can look at.  (Its actually the hash
1123          * size that is assured.)  With our object header encoding
1124          * the maximum deflated object size is 2^137, which is just
1125          * insane, so we know won't exceed what we have been given.
1126          */
1127         base = use_pack(p, w_curs, *curpos, &left);
1128         used = unpack_object_header_buffer(base, left, &type, sizep);
1129         if (!used) {
1130                 type = OBJ_BAD;
1131         } else
1132                 *curpos += used;
1133
1134         return type;
1135 }
1136
1137 void mark_bad_packed_object(struct packed_git *p, const unsigned char *sha1)
1138 {
1139         unsigned i;
1140         const unsigned hashsz = the_hash_algo->rawsz;
1141         for (i = 0; i < p->num_bad_objects; i++)
1142                 if (hasheq(sha1, p->bad_object_sha1 + hashsz * i))
1143                         return;
1144         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1145                                       st_mult(GIT_MAX_RAWSZ,
1146                                               st_add(p->num_bad_objects, 1)));
1147         hashcpy(p->bad_object_sha1 + hashsz * p->num_bad_objects, sha1);
1148         p->num_bad_objects++;
1149 }
1150
1151 const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1152 {
1153         struct packed_git *p;
1154         unsigned i;
1155
1156         for (p = the_repository->objects->packed_git; p; p = p->next)
1157                 for (i = 0; i < p->num_bad_objects; i++)
1158                         if (hasheq(sha1,
1159                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
1160                                 return p;
1161         return NULL;
1162 }
1163
1164 static off_t get_delta_base(struct packed_git *p,
1165                                     struct pack_window **w_curs,
1166                                     off_t *curpos,
1167                                     enum object_type type,
1168                                     off_t delta_obj_offset)
1169 {
1170         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1171         off_t base_offset;
1172
1173         /* use_pack() assured us we have [base_info, base_info + 20)
1174          * as a range that we can look at without walking off the
1175          * end of the mapped window.  Its actually the hash size
1176          * that is assured.  An OFS_DELTA longer than the hash size
1177          * is stupid, as then a REF_DELTA would be smaller to store.
1178          */
1179         if (type == OBJ_OFS_DELTA) {
1180                 unsigned used = 0;
1181                 unsigned char c = base_info[used++];
1182                 base_offset = c & 127;
1183                 while (c & 128) {
1184                         base_offset += 1;
1185                         if (!base_offset || MSB(base_offset, 7))
1186                                 return 0;  /* overflow */
1187                         c = base_info[used++];
1188                         base_offset = (base_offset << 7) + (c & 127);
1189                 }
1190                 base_offset = delta_obj_offset - base_offset;
1191                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1192                         return 0;  /* out of bound */
1193                 *curpos += used;
1194         } else if (type == OBJ_REF_DELTA) {
1195                 /* The base entry _must_ be in the same pack */
1196                 base_offset = find_pack_entry_one(base_info, p);
1197                 *curpos += the_hash_algo->rawsz;
1198         } else
1199                 die("I am totally screwed");
1200         return base_offset;
1201 }
1202
1203 /*
1204  * Like get_delta_base above, but we return the sha1 instead of the pack
1205  * offset. This means it is cheaper for REF deltas (we do not have to do
1206  * the final object lookup), but more expensive for OFS deltas (we
1207  * have to load the revidx to convert the offset back into a sha1).
1208  */
1209 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1210                                                 struct pack_window **w_curs,
1211                                                 off_t curpos,
1212                                                 enum object_type type,
1213                                                 off_t delta_obj_offset)
1214 {
1215         if (type == OBJ_REF_DELTA) {
1216                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1217                 return base;
1218         } else if (type == OBJ_OFS_DELTA) {
1219                 struct revindex_entry *revidx;
1220                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1221                                                    type, delta_obj_offset);
1222
1223                 if (!base_offset)
1224                         return NULL;
1225
1226                 revidx = find_pack_revindex(p, base_offset);
1227                 if (!revidx)
1228                         return NULL;
1229
1230                 return nth_packed_object_sha1(p, revidx->nr);
1231         } else
1232                 return NULL;
1233 }
1234
1235 static int retry_bad_packed_offset(struct repository *r,
1236                                    struct packed_git *p,
1237                                    off_t obj_offset)
1238 {
1239         int type;
1240         struct revindex_entry *revidx;
1241         struct object_id oid;
1242         revidx = find_pack_revindex(p, obj_offset);
1243         if (!revidx)
1244                 return OBJ_BAD;
1245         nth_packed_object_oid(&oid, p, revidx->nr);
1246         mark_bad_packed_object(p, oid.hash);
1247         type = oid_object_info(r, &oid, NULL);
1248         if (type <= OBJ_NONE)
1249                 return OBJ_BAD;
1250         return type;
1251 }
1252
1253 #define POI_STACK_PREALLOC 64
1254
1255 static enum object_type packed_to_object_type(struct repository *r,
1256                                               struct packed_git *p,
1257                                               off_t obj_offset,
1258                                               enum object_type type,
1259                                               struct pack_window **w_curs,
1260                                               off_t curpos)
1261 {
1262         off_t small_poi_stack[POI_STACK_PREALLOC];
1263         off_t *poi_stack = small_poi_stack;
1264         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1265
1266         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1267                 off_t base_offset;
1268                 unsigned long size;
1269                 /* Push the object we're going to leave behind */
1270                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1271                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1272                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
1273                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1274                 } else {
1275                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1276                 }
1277                 poi_stack[poi_stack_nr++] = obj_offset;
1278                 /* If parsing the base offset fails, just unwind */
1279                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1280                 if (!base_offset)
1281                         goto unwind;
1282                 curpos = obj_offset = base_offset;
1283                 type = unpack_object_header(p, w_curs, &curpos, &size);
1284                 if (type <= OBJ_NONE) {
1285                         /* If getting the base itself fails, we first
1286                          * retry the base, otherwise unwind */
1287                         type = retry_bad_packed_offset(r, p, base_offset);
1288                         if (type > OBJ_NONE)
1289                                 goto out;
1290                         goto unwind;
1291                 }
1292         }
1293
1294         switch (type) {
1295         case OBJ_BAD:
1296         case OBJ_COMMIT:
1297         case OBJ_TREE:
1298         case OBJ_BLOB:
1299         case OBJ_TAG:
1300                 break;
1301         default:
1302                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1303                       type, (uintmax_t)obj_offset, p->pack_name);
1304                 type = OBJ_BAD;
1305         }
1306
1307 out:
1308         if (poi_stack != small_poi_stack)
1309                 free(poi_stack);
1310         return type;
1311
1312 unwind:
1313         while (poi_stack_nr) {
1314                 obj_offset = poi_stack[--poi_stack_nr];
1315                 type = retry_bad_packed_offset(r, p, obj_offset);
1316                 if (type > OBJ_NONE)
1317                         goto out;
1318         }
1319         type = OBJ_BAD;
1320         goto out;
1321 }
1322
1323 static struct hashmap delta_base_cache;
1324 static size_t delta_base_cached;
1325
1326 static LIST_HEAD(delta_base_cache_lru);
1327
1328 struct delta_base_cache_key {
1329         struct packed_git *p;
1330         off_t base_offset;
1331 };
1332
1333 struct delta_base_cache_entry {
1334         struct hashmap hash;
1335         struct delta_base_cache_key key;
1336         struct list_head lru;
1337         void *data;
1338         unsigned long size;
1339         enum object_type type;
1340 };
1341
1342 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
1343 {
1344         unsigned int hash;
1345
1346         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
1347         hash += (hash >> 8) + (hash >> 16);
1348         return hash;
1349 }
1350
1351 static struct delta_base_cache_entry *
1352 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
1353 {
1354         struct hashmap_entry entry;
1355         struct delta_base_cache_key key;
1356
1357         if (!delta_base_cache.cmpfn)
1358                 return NULL;
1359
1360         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
1361         key.p = p;
1362         key.base_offset = base_offset;
1363         return hashmap_get(&delta_base_cache, &entry, &key);
1364 }
1365
1366 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
1367                                    const struct delta_base_cache_key *b)
1368 {
1369         return a->p == b->p && a->base_offset == b->base_offset;
1370 }
1371
1372 static int delta_base_cache_hash_cmp(const void *unused_cmp_data,
1373                                      const void *va, const void *vb,
1374                                      const void *vkey)
1375 {
1376         const struct delta_base_cache_entry *a = va, *b = vb;
1377         const struct delta_base_cache_key *key = vkey;
1378         if (key)
1379                 return !delta_base_cache_key_eq(&a->key, key);
1380         else
1381                 return !delta_base_cache_key_eq(&a->key, &b->key);
1382 }
1383
1384 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
1385 {
1386         return !!get_delta_base_cache_entry(p, base_offset);
1387 }
1388
1389 /*
1390  * Remove the entry from the cache, but do _not_ free the associated
1391  * entry data. The caller takes ownership of the "data" buffer, and
1392  * should copy out any fields it wants before detaching.
1393  */
1394 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
1395 {
1396         hashmap_remove(&delta_base_cache, ent, &ent->key);
1397         list_del(&ent->lru);
1398         delta_base_cached -= ent->size;
1399         free(ent);
1400 }
1401
1402 static void *cache_or_unpack_entry(struct repository *r, struct packed_git *p,
1403                                    off_t base_offset, unsigned long *base_size,
1404                                    enum object_type *type)
1405 {
1406         struct delta_base_cache_entry *ent;
1407
1408         ent = get_delta_base_cache_entry(p, base_offset);
1409         if (!ent)
1410                 return unpack_entry(r, p, base_offset, type, base_size);
1411
1412         if (type)
1413                 *type = ent->type;
1414         if (base_size)
1415                 *base_size = ent->size;
1416         return xmemdupz(ent->data, ent->size);
1417 }
1418
1419 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1420 {
1421         free(ent->data);
1422         detach_delta_base_cache_entry(ent);
1423 }
1424
1425 void clear_delta_base_cache(void)
1426 {
1427         struct list_head *lru, *tmp;
1428         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1429                 struct delta_base_cache_entry *entry =
1430                         list_entry(lru, struct delta_base_cache_entry, lru);
1431                 release_delta_base_cache(entry);
1432         }
1433 }
1434
1435 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1436         void *base, unsigned long base_size, enum object_type type)
1437 {
1438         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
1439         struct list_head *lru, *tmp;
1440
1441         delta_base_cached += base_size;
1442
1443         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1444                 struct delta_base_cache_entry *f =
1445                         list_entry(lru, struct delta_base_cache_entry, lru);
1446                 if (delta_base_cached <= delta_base_cache_limit)
1447                         break;
1448                 release_delta_base_cache(f);
1449         }
1450
1451         ent->key.p = p;
1452         ent->key.base_offset = base_offset;
1453         ent->type = type;
1454         ent->data = base;
1455         ent->size = base_size;
1456         list_add_tail(&ent->lru, &delta_base_cache_lru);
1457
1458         if (!delta_base_cache.cmpfn)
1459                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, NULL, 0);
1460         hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
1461         hashmap_add(&delta_base_cache, ent);
1462 }
1463
1464 int packed_object_info(struct repository *r, struct packed_git *p,
1465                        off_t obj_offset, struct object_info *oi)
1466 {
1467         struct pack_window *w_curs = NULL;
1468         unsigned long size;
1469         off_t curpos = obj_offset;
1470         enum object_type type;
1471
1472         /*
1473          * We always get the representation type, but only convert it to
1474          * a "real" type later if the caller is interested.
1475          */
1476         if (oi->contentp) {
1477                 *oi->contentp = cache_or_unpack_entry(r, p, obj_offset, oi->sizep,
1478                                                       &type);
1479                 if (!*oi->contentp)
1480                         type = OBJ_BAD;
1481         } else {
1482                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1483         }
1484
1485         if (!oi->contentp && oi->sizep) {
1486                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1487                         off_t tmp_pos = curpos;
1488                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1489                                                            type, obj_offset);
1490                         if (!base_offset) {
1491                                 type = OBJ_BAD;
1492                                 goto out;
1493                         }
1494                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1495                         if (*oi->sizep == 0) {
1496                                 type = OBJ_BAD;
1497                                 goto out;
1498                         }
1499                 } else {
1500                         *oi->sizep = size;
1501                 }
1502         }
1503
1504         if (oi->disk_sizep) {
1505                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1506                 *oi->disk_sizep = revidx[1].offset - obj_offset;
1507         }
1508
1509         if (oi->typep || oi->type_name) {
1510                 enum object_type ptot;
1511                 ptot = packed_to_object_type(r, p, obj_offset,
1512                                              type, &w_curs, curpos);
1513                 if (oi->typep)
1514                         *oi->typep = ptot;
1515                 if (oi->type_name) {
1516                         const char *tn = type_name(ptot);
1517                         if (tn)
1518                                 strbuf_addstr(oi->type_name, tn);
1519                 }
1520                 if (ptot < 0) {
1521                         type = OBJ_BAD;
1522                         goto out;
1523                 }
1524         }
1525
1526         if (oi->delta_base_sha1) {
1527                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1528                         const unsigned char *base;
1529
1530                         base = get_delta_base_sha1(p, &w_curs, curpos,
1531                                                    type, obj_offset);
1532                         if (!base) {
1533                                 type = OBJ_BAD;
1534                                 goto out;
1535                         }
1536
1537                         hashcpy(oi->delta_base_sha1, base);
1538                 } else
1539                         hashclr(oi->delta_base_sha1);
1540         }
1541
1542         oi->whence = in_delta_base_cache(p, obj_offset) ? OI_DBCACHED :
1543                                                           OI_PACKED;
1544
1545 out:
1546         unuse_pack(&w_curs);
1547         return type;
1548 }
1549
1550 static void *unpack_compressed_entry(struct packed_git *p,
1551                                     struct pack_window **w_curs,
1552                                     off_t curpos,
1553                                     unsigned long size)
1554 {
1555         int st;
1556         git_zstream stream;
1557         unsigned char *buffer, *in;
1558
1559         buffer = xmallocz_gently(size);
1560         if (!buffer)
1561                 return NULL;
1562         memset(&stream, 0, sizeof(stream));
1563         stream.next_out = buffer;
1564         stream.avail_out = size + 1;
1565
1566         git_inflate_init(&stream);
1567         do {
1568                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1569                 stream.next_in = in;
1570                 st = git_inflate(&stream, Z_FINISH);
1571                 if (!stream.avail_out)
1572                         break; /* the payload is larger than it should be */
1573                 curpos += stream.next_in - in;
1574         } while (st == Z_OK || st == Z_BUF_ERROR);
1575         git_inflate_end(&stream);
1576         if ((st != Z_STREAM_END) || stream.total_out != size) {
1577                 free(buffer);
1578                 return NULL;
1579         }
1580
1581         /* versions of zlib can clobber unconsumed portion of outbuf */
1582         buffer[size] = '\0';
1583
1584         return buffer;
1585 }
1586
1587 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
1588 {
1589         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
1590         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
1591                          p->pack_name, (uintmax_t)obj_offset);
1592 }
1593
1594 int do_check_packed_object_crc;
1595
1596 #define UNPACK_ENTRY_STACK_PREALLOC 64
1597 struct unpack_entry_stack_ent {
1598         off_t obj_offset;
1599         off_t curpos;
1600         unsigned long size;
1601 };
1602
1603 static void *read_object(struct repository *r,
1604                          const struct object_id *oid,
1605                          enum object_type *type,
1606                          unsigned long *size)
1607 {
1608         struct object_info oi = OBJECT_INFO_INIT;
1609         void *content;
1610         oi.typep = type;
1611         oi.sizep = size;
1612         oi.contentp = &content;
1613
1614         if (oid_object_info_extended(r, oid, &oi, 0) < 0)
1615                 return NULL;
1616         return content;
1617 }
1618
1619 void *unpack_entry(struct repository *r, struct packed_git *p, off_t obj_offset,
1620                    enum object_type *final_type, unsigned long *final_size)
1621 {
1622         struct pack_window *w_curs = NULL;
1623         off_t curpos = obj_offset;
1624         void *data = NULL;
1625         unsigned long size;
1626         enum object_type type;
1627         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
1628         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
1629         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
1630         int base_from_cache = 0;
1631
1632         write_pack_access_log(p, obj_offset);
1633
1634         /* PHASE 1: drill down to the innermost base object */
1635         for (;;) {
1636                 off_t base_offset;
1637                 int i;
1638                 struct delta_base_cache_entry *ent;
1639
1640                 ent = get_delta_base_cache_entry(p, curpos);
1641                 if (ent) {
1642                         type = ent->type;
1643                         data = ent->data;
1644                         size = ent->size;
1645                         detach_delta_base_cache_entry(ent);
1646                         base_from_cache = 1;
1647                         break;
1648                 }
1649
1650                 if (do_check_packed_object_crc && p->index_version > 1) {
1651                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1652                         off_t len = revidx[1].offset - obj_offset;
1653                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1654                                 struct object_id oid;
1655                                 nth_packed_object_oid(&oid, p, revidx->nr);
1656                                 error("bad packed object CRC for %s",
1657                                       oid_to_hex(&oid));
1658                                 mark_bad_packed_object(p, oid.hash);
1659                                 data = NULL;
1660                                 goto out;
1661                         }
1662                 }
1663
1664                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1665                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
1666                         break;
1667
1668                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1669                 if (!base_offset) {
1670                         error("failed to validate delta base reference "
1671                               "at offset %"PRIuMAX" from %s",
1672                               (uintmax_t)curpos, p->pack_name);
1673                         /* bail to phase 2, in hopes of recovery */
1674                         data = NULL;
1675                         break;
1676                 }
1677
1678                 /* push object, proceed to base */
1679                 if (delta_stack_nr >= delta_stack_alloc
1680                     && delta_stack == small_delta_stack) {
1681                         delta_stack_alloc = alloc_nr(delta_stack_nr);
1682                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
1683                         memcpy(delta_stack, small_delta_stack,
1684                                sizeof(*delta_stack)*delta_stack_nr);
1685                 } else {
1686                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
1687                 }
1688                 i = delta_stack_nr++;
1689                 delta_stack[i].obj_offset = obj_offset;
1690                 delta_stack[i].curpos = curpos;
1691                 delta_stack[i].size = size;
1692
1693                 curpos = obj_offset = base_offset;
1694         }
1695
1696         /* PHASE 2: handle the base */
1697         switch (type) {
1698         case OBJ_OFS_DELTA:
1699         case OBJ_REF_DELTA:
1700                 if (data)
1701                         BUG("unpack_entry: left loop at a valid delta");
1702                 break;
1703         case OBJ_COMMIT:
1704         case OBJ_TREE:
1705         case OBJ_BLOB:
1706         case OBJ_TAG:
1707                 if (!base_from_cache)
1708                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
1709                 break;
1710         default:
1711                 data = NULL;
1712                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1713                       type, (uintmax_t)obj_offset, p->pack_name);
1714         }
1715
1716         /* PHASE 3: apply deltas in order */
1717
1718         /* invariants:
1719          *   'data' holds the base data, or NULL if there was corruption
1720          */
1721         while (delta_stack_nr) {
1722                 void *delta_data;
1723                 void *base = data;
1724                 void *external_base = NULL;
1725                 unsigned long delta_size, base_size = size;
1726                 int i;
1727
1728                 data = NULL;
1729
1730                 if (base)
1731                         add_delta_base_cache(p, obj_offset, base, base_size, type);
1732
1733                 if (!base) {
1734                         /*
1735                          * We're probably in deep shit, but let's try to fetch
1736                          * the required base anyway from another pack or loose.
1737                          * This is costly but should happen only in the presence
1738                          * of a corrupted pack, and is better than failing outright.
1739                          */
1740                         struct revindex_entry *revidx;
1741                         struct object_id base_oid;
1742                         revidx = find_pack_revindex(p, obj_offset);
1743                         if (revidx) {
1744                                 nth_packed_object_oid(&base_oid, p, revidx->nr);
1745                                 error("failed to read delta base object %s"
1746                                       " at offset %"PRIuMAX" from %s",
1747                                       oid_to_hex(&base_oid), (uintmax_t)obj_offset,
1748                                       p->pack_name);
1749                                 mark_bad_packed_object(p, base_oid.hash);
1750                                 base = read_object(r, &base_oid, &type, &base_size);
1751                                 external_base = base;
1752                         }
1753                 }
1754
1755                 i = --delta_stack_nr;
1756                 obj_offset = delta_stack[i].obj_offset;
1757                 curpos = delta_stack[i].curpos;
1758                 delta_size = delta_stack[i].size;
1759
1760                 if (!base)
1761                         continue;
1762
1763                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
1764
1765                 if (!delta_data) {
1766                         error("failed to unpack compressed delta "
1767                               "at offset %"PRIuMAX" from %s",
1768                               (uintmax_t)curpos, p->pack_name);
1769                         data = NULL;
1770                         free(external_base);
1771                         continue;
1772                 }
1773
1774                 data = patch_delta(base, base_size,
1775                                    delta_data, delta_size,
1776                                    &size);
1777
1778                 /*
1779                  * We could not apply the delta; warn the user, but keep going.
1780                  * Our failure will be noticed either in the next iteration of
1781                  * the loop, or if this is the final delta, in the caller when
1782                  * we return NULL. Those code paths will take care of making
1783                  * a more explicit warning and retrying with another copy of
1784                  * the object.
1785                  */
1786                 if (!data)
1787                         error("failed to apply delta");
1788
1789                 free(delta_data);
1790                 free(external_base);
1791         }
1792
1793         if (final_type)
1794                 *final_type = type;
1795         if (final_size)
1796                 *final_size = size;
1797
1798 out:
1799         unuse_pack(&w_curs);
1800
1801         if (delta_stack != small_delta_stack)
1802                 free(delta_stack);
1803
1804         return data;
1805 }
1806
1807 int bsearch_pack(const struct object_id *oid, const struct packed_git *p, uint32_t *result)
1808 {
1809         const unsigned char *index_fanout = p->index_data;
1810         const unsigned char *index_lookup;
1811         const unsigned int hashsz = the_hash_algo->rawsz;
1812         int index_lookup_width;
1813
1814         if (!index_fanout)
1815                 BUG("bsearch_pack called without a valid pack-index");
1816
1817         index_lookup = index_fanout + 4 * 256;
1818         if (p->index_version == 1) {
1819                 index_lookup_width = hashsz + 4;
1820                 index_lookup += 4;
1821         } else {
1822                 index_lookup_width = hashsz;
1823                 index_fanout += 8;
1824                 index_lookup += 8;
1825         }
1826
1827         return bsearch_hash(oid->hash, (const uint32_t*)index_fanout,
1828                             index_lookup, index_lookup_width, result);
1829 }
1830
1831 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1832                                             uint32_t n)
1833 {
1834         const unsigned char *index = p->index_data;
1835         const unsigned int hashsz = the_hash_algo->rawsz;
1836         if (!index) {
1837                 if (open_pack_index(p))
1838                         return NULL;
1839                 index = p->index_data;
1840         }
1841         if (n >= p->num_objects)
1842                 return NULL;
1843         index += 4 * 256;
1844         if (p->index_version == 1) {
1845                 return index + (hashsz + 4) * n + 4;
1846         } else {
1847                 index += 8;
1848                 return index + hashsz * n;
1849         }
1850 }
1851
1852 const struct object_id *nth_packed_object_oid(struct object_id *oid,
1853                                               struct packed_git *p,
1854                                               uint32_t n)
1855 {
1856         const unsigned char *hash = nth_packed_object_sha1(p, n);
1857         if (!hash)
1858                 return NULL;
1859         hashcpy(oid->hash, hash);
1860         return oid;
1861 }
1862
1863 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
1864 {
1865         const unsigned char *ptr = vptr;
1866         const unsigned char *start = p->index_data;
1867         const unsigned char *end = start + p->index_size;
1868         if (ptr < start)
1869                 die(_("offset before start of pack index for %s (corrupt index?)"),
1870                     p->pack_name);
1871         /* No need to check for underflow; .idx files must be at least 8 bytes */
1872         if (ptr >= end - 8)
1873                 die(_("offset beyond end of pack index for %s (truncated index?)"),
1874                     p->pack_name);
1875 }
1876
1877 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1878 {
1879         const unsigned char *index = p->index_data;
1880         const unsigned int hashsz = the_hash_algo->rawsz;
1881         index += 4 * 256;
1882         if (p->index_version == 1) {
1883                 return ntohl(*((uint32_t *)(index + (hashsz + 4) * n)));
1884         } else {
1885                 uint32_t off;
1886                 index += 8 + p->num_objects * (hashsz + 4);
1887                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1888                 if (!(off & 0x80000000))
1889                         return off;
1890                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1891                 check_pack_index_ptr(p, index);
1892                 return get_be64(index);
1893         }
1894 }
1895
1896 off_t find_pack_entry_one(const unsigned char *sha1,
1897                                   struct packed_git *p)
1898 {
1899         const unsigned char *index = p->index_data;
1900         struct object_id oid;
1901         uint32_t result;
1902
1903         if (!index) {
1904                 if (open_pack_index(p))
1905                         return 0;
1906         }
1907
1908         hashcpy(oid.hash, sha1);
1909         if (bsearch_pack(&oid, p, &result))
1910                 return nth_packed_object_offset(p, result);
1911         return 0;
1912 }
1913
1914 int is_pack_valid(struct packed_git *p)
1915 {
1916         /* An already open pack is known to be valid. */
1917         if (p->pack_fd != -1)
1918                 return 1;
1919
1920         /* If the pack has one window completely covering the
1921          * file size, the pack is known to be valid even if
1922          * the descriptor is not currently open.
1923          */
1924         if (p->windows) {
1925                 struct pack_window *w = p->windows;
1926
1927                 if (!w->offset && w->len == p->pack_size)
1928                         return 1;
1929         }
1930
1931         /* Force the pack to open to prove its valid. */
1932         return !open_packed_git(p);
1933 }
1934
1935 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1936                                   struct packed_git *packs)
1937 {
1938         struct packed_git *p;
1939
1940         for (p = packs; p; p = p->next) {
1941                 if (find_pack_entry_one(sha1, p))
1942                         return p;
1943         }
1944         return NULL;
1945
1946 }
1947
1948 static int fill_pack_entry(const struct object_id *oid,
1949                            struct pack_entry *e,
1950                            struct packed_git *p)
1951 {
1952         off_t offset;
1953
1954         if (p->num_bad_objects) {
1955                 unsigned i;
1956                 for (i = 0; i < p->num_bad_objects; i++)
1957                         if (hasheq(oid->hash,
1958                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
1959                                 return 0;
1960         }
1961
1962         offset = find_pack_entry_one(oid->hash, p);
1963         if (!offset)
1964                 return 0;
1965
1966         /*
1967          * We are about to tell the caller where they can locate the
1968          * requested object.  We better make sure the packfile is
1969          * still here and can be accessed before supplying that
1970          * answer, as it may have been deleted since the index was
1971          * loaded!
1972          */
1973         if (!is_pack_valid(p))
1974                 return 0;
1975         e->offset = offset;
1976         e->p = p;
1977         return 1;
1978 }
1979
1980 int find_pack_entry(struct repository *r, const struct object_id *oid, struct pack_entry *e)
1981 {
1982         struct list_head *pos;
1983         struct multi_pack_index *m;
1984
1985         prepare_packed_git(r);
1986         if (!r->objects->packed_git && !r->objects->multi_pack_index)
1987                 return 0;
1988
1989         for (m = r->objects->multi_pack_index; m; m = m->next) {
1990                 if (fill_midx_entry(oid, e, m))
1991                         return 1;
1992         }
1993
1994         list_for_each(pos, &r->objects->packed_git_mru) {
1995                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
1996                 if (fill_pack_entry(oid, e, p)) {
1997                         list_move(&p->mru, &r->objects->packed_git_mru);
1998                         return 1;
1999                 }
2000         }
2001         return 0;
2002 }
2003
2004 int has_object_pack(const struct object_id *oid)
2005 {
2006         struct pack_entry e;
2007         return find_pack_entry(the_repository, oid, &e);
2008 }
2009
2010 int has_pack_index(const unsigned char *sha1)
2011 {
2012         struct stat st;
2013         if (stat(sha1_pack_index_name(sha1), &st))
2014                 return 0;
2015         return 1;
2016 }
2017
2018 int for_each_object_in_pack(struct packed_git *p,
2019                             each_packed_object_fn cb, void *data,
2020                             enum for_each_object_flags flags)
2021 {
2022         uint32_t i;
2023         int r = 0;
2024
2025         if (flags & FOR_EACH_OBJECT_PACK_ORDER)
2026                 load_pack_revindex(p);
2027
2028         for (i = 0; i < p->num_objects; i++) {
2029                 uint32_t pos;
2030                 struct object_id oid;
2031
2032                 if (flags & FOR_EACH_OBJECT_PACK_ORDER)
2033                         pos = p->revindex[i].nr;
2034                 else
2035                         pos = i;
2036
2037                 if (!nth_packed_object_oid(&oid, p, pos))
2038                         return error("unable to get sha1 of object %u in %s",
2039                                      pos, p->pack_name);
2040
2041                 r = cb(&oid, p, pos, data);
2042                 if (r)
2043                         break;
2044         }
2045         return r;
2046 }
2047
2048 int for_each_packed_object(each_packed_object_fn cb, void *data,
2049                            enum for_each_object_flags flags)
2050 {
2051         struct packed_git *p;
2052         int r = 0;
2053         int pack_errors = 0;
2054
2055         prepare_packed_git(the_repository);
2056         for (p = get_all_packs(the_repository); p; p = p->next) {
2057                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
2058                         continue;
2059                 if ((flags & FOR_EACH_OBJECT_PROMISOR_ONLY) &&
2060                     !p->pack_promisor)
2061                         continue;
2062                 if (open_pack_index(p)) {
2063                         pack_errors = 1;
2064                         continue;
2065                 }
2066                 r = for_each_object_in_pack(p, cb, data, flags);
2067                 if (r)
2068                         break;
2069         }
2070         return r ? r : pack_errors;
2071 }
2072
2073 static int add_promisor_object(const struct object_id *oid,
2074                                struct packed_git *pack,
2075                                uint32_t pos,
2076                                void *set_)
2077 {
2078         struct oidset *set = set_;
2079         struct object *obj = parse_object(the_repository, oid);
2080         if (!obj)
2081                 return 1;
2082
2083         oidset_insert(set, oid);
2084
2085         /*
2086          * If this is a tree, commit, or tag, the objects it refers
2087          * to are also promisor objects. (Blobs refer to no objects->)
2088          */
2089         if (obj->type == OBJ_TREE) {
2090                 struct tree *tree = (struct tree *)obj;
2091                 struct tree_desc desc;
2092                 struct name_entry entry;
2093                 if (init_tree_desc_gently(&desc, tree->buffer, tree->size))
2094                         /*
2095                          * Error messages are given when packs are
2096                          * verified, so do not print any here.
2097                          */
2098                         return 0;
2099                 while (tree_entry_gently(&desc, &entry))
2100                         oidset_insert(set, &entry.oid);
2101         } else if (obj->type == OBJ_COMMIT) {
2102                 struct commit *commit = (struct commit *) obj;
2103                 struct commit_list *parents = commit->parents;
2104
2105                 oidset_insert(set, get_commit_tree_oid(commit));
2106                 for (; parents; parents = parents->next)
2107                         oidset_insert(set, &parents->item->object.oid);
2108         } else if (obj->type == OBJ_TAG) {
2109                 struct tag *tag = (struct tag *) obj;
2110                 oidset_insert(set, &tag->tagged->oid);
2111         }
2112         return 0;
2113 }
2114
2115 int is_promisor_object(const struct object_id *oid)
2116 {
2117         static struct oidset promisor_objects;
2118         static int promisor_objects_prepared;
2119
2120         if (!promisor_objects_prepared) {
2121                 if (repository_format_partial_clone) {
2122                         for_each_packed_object(add_promisor_object,
2123                                                &promisor_objects,
2124                                                FOR_EACH_OBJECT_PROMISOR_ONLY);
2125                 }
2126                 promisor_objects_prepared = 1;
2127         }
2128         return oidset_contains(&promisor_objects, oid);
2129 }