packfile: replace lseek+read with pread
[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 #include "commit-graph.h"
20
21 char *odb_pack_name(struct strbuf *buf,
22                     const unsigned char *sha1,
23                     const char *ext)
24 {
25         strbuf_reset(buf);
26         strbuf_addf(buf, "%s/pack/pack-%s.%s", get_object_directory(),
27                     sha1_to_hex(sha1), ext);
28         return buf->buf;
29 }
30
31 char *sha1_pack_name(const unsigned char *sha1)
32 {
33         static struct strbuf buf = STRBUF_INIT;
34         return odb_pack_name(&buf, sha1, "pack");
35 }
36
37 char *sha1_pack_index_name(const unsigned char *sha1)
38 {
39         static struct strbuf buf = STRBUF_INIT;
40         return odb_pack_name(&buf, sha1, "idx");
41 }
42
43 static unsigned int pack_used_ctr;
44 static unsigned int pack_mmap_calls;
45 static unsigned int peak_pack_open_windows;
46 static unsigned int pack_open_windows;
47 static unsigned int pack_open_fds;
48 static unsigned int pack_max_fds;
49 static size_t peak_pack_mapped;
50 static size_t pack_mapped;
51
52 #define SZ_FMT PRIuMAX
53 static inline uintmax_t sz_fmt(size_t s) { return s; }
54
55 void pack_report(void)
56 {
57         fprintf(stderr,
58                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
59                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
60                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
61                 sz_fmt(getpagesize()),
62                 sz_fmt(packed_git_window_size),
63                 sz_fmt(packed_git_limit));
64         fprintf(stderr,
65                 "pack_report: pack_used_ctr            = %10u\n"
66                 "pack_report: pack_mmap_calls          = %10u\n"
67                 "pack_report: pack_open_windows        = %10u / %10u\n"
68                 "pack_report: pack_mapped              = "
69                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
70                 pack_used_ctr,
71                 pack_mmap_calls,
72                 pack_open_windows, peak_pack_open_windows,
73                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
74 }
75
76 /*
77  * Open and mmap the index file at path, perform a couple of
78  * consistency checks, then record its information to p.  Return 0 on
79  * success.
80  */
81 static int check_packed_git_idx(const char *path, struct packed_git *p)
82 {
83         void *idx_map;
84         size_t idx_size;
85         int fd = git_open(path), ret;
86         struct stat st;
87         const unsigned int hashsz = the_hash_algo->rawsz;
88
89         if (fd < 0)
90                 return -1;
91         if (fstat(fd, &st)) {
92                 close(fd);
93                 return -1;
94         }
95         idx_size = xsize_t(st.st_size);
96         if (idx_size < 4 * 256 + hashsz + hashsz) {
97                 close(fd);
98                 return error("index file %s is too small", path);
99         }
100         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
101         close(fd);
102
103         ret = load_idx(path, hashsz, idx_map, idx_size, p);
104
105         if (ret)
106                 munmap(idx_map, idx_size);
107
108         return ret;
109 }
110
111 int load_idx(const char *path, const unsigned int hashsz, void *idx_map,
112              size_t idx_size, struct packed_git *p)
113 {
114         struct pack_idx_header *hdr = idx_map;
115         uint32_t version, nr, i, *index;
116
117         if (idx_size < 4 * 256 + hashsz + hashsz)
118                 return error("index file %s is too small", path);
119         if (idx_map == NULL)
120                 return error("empty data");
121
122         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
123                 version = ntohl(hdr->idx_version);
124                 if (version < 2 || version > 2)
125                         return error("index file %s is version %"PRIu32
126                                      " and is not supported by this binary"
127                                      " (try upgrading GIT to a newer version)",
128                                      path, version);
129         } else
130                 version = 1;
131
132         nr = 0;
133         index = idx_map;
134         if (version > 1)
135                 index += 2;  /* skip index header */
136         for (i = 0; i < 256; i++) {
137                 uint32_t n = ntohl(index[i]);
138                 if (n < nr)
139                         return error("non-monotonic index %s", path);
140                 nr = n;
141         }
142
143         if (version == 1) {
144                 /*
145                  * Total size:
146                  *  - 256 index entries 4 bytes each
147                  *  - 24-byte entries * nr (object ID + 4-byte offset)
148                  *  - hash of the packfile
149                  *  - file checksum
150                  */
151                 if (idx_size != 4 * 256 + nr * (hashsz + 4) + hashsz + hashsz)
152                         return error("wrong index v1 file size in %s", path);
153         } else if (version == 2) {
154                 /*
155                  * Minimum size:
156                  *  - 8 bytes of header
157                  *  - 256 index entries 4 bytes each
158                  *  - object ID entry * nr
159                  *  - 4-byte crc entry * nr
160                  *  - 4-byte offset entry * nr
161                  *  - hash of the packfile
162                  *  - file checksum
163                  * And after the 4-byte offset table might be a
164                  * variable sized table containing 8-byte entries
165                  * for offsets larger than 2^31.
166                  */
167                 unsigned long min_size = 8 + 4*256 + nr*(hashsz + 4 + 4) + hashsz + hashsz;
168                 unsigned long max_size = min_size;
169                 if (nr)
170                         max_size += (nr - 1)*8;
171                 if (idx_size < min_size || idx_size > max_size)
172                         return error("wrong index v2 file size in %s", path);
173                 if (idx_size != min_size &&
174                     /*
175                      * make sure we can deal with large pack offsets.
176                      * 31-bit signed offset won't be enough, neither
177                      * 32-bit unsigned one will be.
178                      */
179                     (sizeof(off_t) <= 4))
180                         return error("pack too large for current definition of off_t in %s", path);
181         }
182
183         p->index_version = version;
184         p->index_data = idx_map;
185         p->index_size = idx_size;
186         p->num_objects = nr;
187         return 0;
188 }
189
190 int open_pack_index(struct packed_git *p)
191 {
192         char *idx_name;
193         size_t len;
194         int ret;
195
196         if (p->index_data)
197                 return 0;
198
199         if (!strip_suffix(p->pack_name, ".pack", &len))
200                 BUG("pack_name does not end in .pack");
201         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
202         ret = check_packed_git_idx(idx_name, p);
203         free(idx_name);
204         return ret;
205 }
206
207 uint32_t get_pack_fanout(struct packed_git *p, uint32_t value)
208 {
209         const uint32_t *level1_ofs = p->index_data;
210
211         if (!level1_ofs) {
212                 if (open_pack_index(p))
213                         return 0;
214                 level1_ofs = p->index_data;
215         }
216
217         if (p->index_version > 1) {
218                 level1_ofs += 2;
219         }
220
221         return ntohl(level1_ofs[value]);
222 }
223
224 static struct packed_git *alloc_packed_git(int extra)
225 {
226         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
227         memset(p, 0, sizeof(*p));
228         p->pack_fd = -1;
229         return p;
230 }
231
232 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
233 {
234         const char *path = sha1_pack_name(sha1);
235         size_t alloc = st_add(strlen(path), 1);
236         struct packed_git *p = alloc_packed_git(alloc);
237
238         memcpy(p->pack_name, path, alloc); /* includes NUL */
239         hashcpy(p->hash, sha1);
240         if (check_packed_git_idx(idx_path, p)) {
241                 free(p);
242                 return NULL;
243         }
244
245         return p;
246 }
247
248 static void scan_windows(struct packed_git *p,
249         struct packed_git **lru_p,
250         struct pack_window **lru_w,
251         struct pack_window **lru_l)
252 {
253         struct pack_window *w, *w_l;
254
255         for (w_l = NULL, w = p->windows; w; w = w->next) {
256                 if (!w->inuse_cnt) {
257                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
258                                 *lru_p = p;
259                                 *lru_w = w;
260                                 *lru_l = w_l;
261                         }
262                 }
263                 w_l = w;
264         }
265 }
266
267 static int unuse_one_window(struct packed_git *current)
268 {
269         struct packed_git *p, *lru_p = NULL;
270         struct pack_window *lru_w = NULL, *lru_l = NULL;
271
272         if (current)
273                 scan_windows(current, &lru_p, &lru_w, &lru_l);
274         for (p = the_repository->objects->packed_git; p; p = p->next)
275                 scan_windows(p, &lru_p, &lru_w, &lru_l);
276         if (lru_p) {
277                 munmap(lru_w->base, lru_w->len);
278                 pack_mapped -= lru_w->len;
279                 if (lru_l)
280                         lru_l->next = lru_w->next;
281                 else
282                         lru_p->windows = lru_w->next;
283                 free(lru_w);
284                 pack_open_windows--;
285                 return 1;
286         }
287         return 0;
288 }
289
290 void release_pack_memory(size_t need)
291 {
292         size_t cur = pack_mapped;
293         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
294                 ; /* nothing */
295 }
296
297 void close_pack_windows(struct packed_git *p)
298 {
299         while (p->windows) {
300                 struct pack_window *w = p->windows;
301
302                 if (w->inuse_cnt)
303                         die("pack '%s' still has open windows to it",
304                             p->pack_name);
305                 munmap(w->base, w->len);
306                 pack_mapped -= w->len;
307                 pack_open_windows--;
308                 p->windows = w->next;
309                 free(w);
310         }
311 }
312
313 int close_pack_fd(struct packed_git *p)
314 {
315         if (p->pack_fd < 0)
316                 return 0;
317
318         close(p->pack_fd);
319         pack_open_fds--;
320         p->pack_fd = -1;
321
322         return 1;
323 }
324
325 void close_pack_index(struct packed_git *p)
326 {
327         if (p->index_data) {
328                 munmap((void *)p->index_data, p->index_size);
329                 p->index_data = NULL;
330         }
331 }
332
333 void close_pack(struct packed_git *p)
334 {
335         close_pack_windows(p);
336         close_pack_fd(p);
337         close_pack_index(p);
338 }
339
340 void close_object_store(struct raw_object_store *o)
341 {
342         struct packed_git *p;
343
344         for (p = o->packed_git; p; p = p->next)
345                 if (p->do_not_close)
346                         BUG("want to close pack marked 'do-not-close'");
347                 else
348                         close_pack(p);
349
350         if (o->multi_pack_index) {
351                 close_midx(o->multi_pack_index);
352                 o->multi_pack_index = NULL;
353         }
354
355         close_commit_graph(o);
356 }
357
358 /*
359  * The LRU pack is the one with the oldest MRU window, preferring packs
360  * with no used windows, or the oldest mtime if it has no windows allocated.
361  */
362 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
363 {
364         struct pack_window *w, *this_mru_w;
365         int has_windows_inuse = 0;
366
367         /*
368          * Reject this pack if it has windows and the previously selected
369          * one does not.  If this pack does not have windows, reject
370          * it if the pack file is newer than the previously selected one.
371          */
372         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
373                 return;
374
375         for (w = this_mru_w = p->windows; w; w = w->next) {
376                 /*
377                  * Reject this pack if any of its windows are in use,
378                  * but the previously selected pack did not have any
379                  * inuse windows.  Otherwise, record that this pack
380                  * has windows in use.
381                  */
382                 if (w->inuse_cnt) {
383                         if (*accept_windows_inuse)
384                                 has_windows_inuse = 1;
385                         else
386                                 return;
387                 }
388
389                 if (w->last_used > this_mru_w->last_used)
390                         this_mru_w = w;
391
392                 /*
393                  * Reject this pack if it has windows that have been
394                  * used more recently than the previously selected pack.
395                  * If the previously selected pack had windows inuse and
396                  * we have not encountered a window in this pack that is
397                  * inuse, skip this check since we prefer a pack with no
398                  * inuse windows to one that has inuse windows.
399                  */
400                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
401                     this_mru_w->last_used > (*mru_w)->last_used)
402                         return;
403         }
404
405         /*
406          * Select this pack.
407          */
408         *mru_w = this_mru_w;
409         *lru_p = p;
410         *accept_windows_inuse = has_windows_inuse;
411 }
412
413 static int close_one_pack(void)
414 {
415         struct packed_git *p, *lru_p = NULL;
416         struct pack_window *mru_w = NULL;
417         int accept_windows_inuse = 1;
418
419         for (p = the_repository->objects->packed_git; p; p = p->next) {
420                 if (p->pack_fd == -1)
421                         continue;
422                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
423         }
424
425         if (lru_p)
426                 return close_pack_fd(lru_p);
427
428         return 0;
429 }
430
431 static unsigned int get_max_fd_limit(void)
432 {
433 #ifdef RLIMIT_NOFILE
434         {
435                 struct rlimit lim;
436
437                 if (!getrlimit(RLIMIT_NOFILE, &lim))
438                         return lim.rlim_cur;
439         }
440 #endif
441
442 #ifdef _SC_OPEN_MAX
443         {
444                 long open_max = sysconf(_SC_OPEN_MAX);
445                 if (0 < open_max)
446                         return open_max;
447                 /*
448                  * Otherwise, we got -1 for one of the two
449                  * reasons:
450                  *
451                  * (1) sysconf() did not understand _SC_OPEN_MAX
452                  *     and signaled an error with -1; or
453                  * (2) sysconf() said there is no limit.
454                  *
455                  * We _could_ clear errno before calling sysconf() to
456                  * tell these two cases apart and return a huge number
457                  * in the latter case to let the caller cap it to a
458                  * value that is not so selfish, but letting the
459                  * fallback OPEN_MAX codepath take care of these cases
460                  * is a lot simpler.
461                  */
462         }
463 #endif
464
465 #ifdef OPEN_MAX
466         return OPEN_MAX;
467 #else
468         return 1; /* see the caller ;-) */
469 #endif
470 }
471
472 const char *pack_basename(struct packed_git *p)
473 {
474         const char *ret = strrchr(p->pack_name, '/');
475         if (ret)
476                 ret = ret + 1; /* skip past slash */
477         else
478                 ret = p->pack_name; /* we only have a base */
479         return ret;
480 }
481
482 /*
483  * Do not call this directly as this leaks p->pack_fd on error return;
484  * call open_packed_git() instead.
485  */
486 static int open_packed_git_1(struct packed_git *p)
487 {
488         struct stat st;
489         struct pack_header hdr;
490         unsigned char hash[GIT_MAX_RAWSZ];
491         unsigned char *idx_hash;
492         ssize_t read_result;
493         const unsigned hashsz = the_hash_algo->rawsz;
494
495         if (!p->index_data) {
496                 struct multi_pack_index *m;
497                 const char *pack_name = pack_basename(p);
498
499                 for (m = the_repository->objects->multi_pack_index;
500                      m; m = m->next) {
501                         if (midx_contains_pack(m, pack_name))
502                                 break;
503                 }
504
505                 if (!m && open_pack_index(p))
506                         return error("packfile %s index unavailable", p->pack_name);
507         }
508
509         if (!pack_max_fds) {
510                 unsigned int max_fds = get_max_fd_limit();
511
512                 /* Save 3 for stdin/stdout/stderr, 22 for work */
513                 if (25 < max_fds)
514                         pack_max_fds = max_fds - 25;
515                 else
516                         pack_max_fds = 1;
517         }
518
519         while (pack_max_fds <= pack_open_fds && close_one_pack())
520                 ; /* nothing */
521
522         p->pack_fd = git_open(p->pack_name);
523         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
524                 return -1;
525         pack_open_fds++;
526
527         /* If we created the struct before we had the pack we lack size. */
528         if (!p->pack_size) {
529                 if (!S_ISREG(st.st_mode))
530                         return error("packfile %s not a regular file", p->pack_name);
531                 p->pack_size = st.st_size;
532         } else if (p->pack_size != st.st_size)
533                 return error("packfile %s size changed", p->pack_name);
534
535         /* Verify we recognize this pack file format. */
536         read_result = read_in_full(p->pack_fd, &hdr, sizeof(hdr));
537         if (read_result < 0)
538                 return error_errno("error reading from %s", p->pack_name);
539         if (read_result != sizeof(hdr))
540                 return error("file %s is far too short to be a packfile", p->pack_name);
541         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
542                 return error("file %s is not a GIT packfile", p->pack_name);
543         if (!pack_version_ok(hdr.hdr_version))
544                 return error("packfile %s is version %"PRIu32" and not"
545                         " supported (try upgrading GIT to a newer version)",
546                         p->pack_name, ntohl(hdr.hdr_version));
547
548         /* Skip index checking if in multi-pack-index */
549         if (!p->index_data)
550                 return 0;
551
552         /* Verify the pack matches its index. */
553         if (p->num_objects != ntohl(hdr.hdr_entries))
554                 return error("packfile %s claims to have %"PRIu32" objects"
555                              " while index indicates %"PRIu32" objects",
556                              p->pack_name, ntohl(hdr.hdr_entries),
557                              p->num_objects);
558         read_result = pread_in_full(p->pack_fd, hash, hashsz,
559                                         p->pack_size - hashsz);
560         if (read_result < 0)
561                 return error_errno("error reading from %s", p->pack_name);
562         if (read_result != hashsz)
563                 return error("packfile %s signature is unavailable", p->pack_name);
564         idx_hash = ((unsigned char *)p->index_data) + p->index_size - hashsz * 2;
565         if (!hasheq(hash, idx_hash))
566                 return error("packfile %s does not match index", p->pack_name);
567         return 0;
568 }
569
570 static int open_packed_git(struct packed_git *p)
571 {
572         if (!open_packed_git_1(p))
573                 return 0;
574         close_pack_fd(p);
575         return -1;
576 }
577
578 static int in_window(struct pack_window *win, off_t offset)
579 {
580         /* We must promise at least one full hash after the
581          * offset is available from this window, otherwise the offset
582          * is not actually in this window and a different window (which
583          * has that one hash excess) must be used.  This is to support
584          * the object header and delta base parsing routines below.
585          */
586         off_t win_off = win->offset;
587         return win_off <= offset
588                 && (offset + the_hash_algo->rawsz) <= (win_off + win->len);
589 }
590
591 unsigned char *use_pack(struct packed_git *p,
592                 struct pack_window **w_cursor,
593                 off_t offset,
594                 unsigned long *left)
595 {
596         struct pack_window *win = *w_cursor;
597
598         /* Since packfiles end in a hash of their content and it's
599          * pointless to ask for an offset into the middle of that
600          * hash, and the in_window function above wouldn't match
601          * don't allow an offset too close to the end of the file.
602          */
603         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
604                 die("packfile %s cannot be accessed", p->pack_name);
605         if (offset > (p->pack_size - the_hash_algo->rawsz))
606                 die("offset beyond end of packfile (truncated pack?)");
607         if (offset < 0)
608                 die(_("offset before end of packfile (broken .idx?)"));
609
610         if (!win || !in_window(win, offset)) {
611                 if (win)
612                         win->inuse_cnt--;
613                 for (win = p->windows; win; win = win->next) {
614                         if (in_window(win, offset))
615                                 break;
616                 }
617                 if (!win) {
618                         size_t window_align = packed_git_window_size / 2;
619                         off_t len;
620
621                         if (p->pack_fd == -1 && open_packed_git(p))
622                                 die("packfile %s cannot be accessed", p->pack_name);
623
624                         win = xcalloc(1, sizeof(*win));
625                         win->offset = (offset / window_align) * window_align;
626                         len = p->pack_size - win->offset;
627                         if (len > packed_git_window_size)
628                                 len = packed_git_window_size;
629                         win->len = (size_t)len;
630                         pack_mapped += win->len;
631                         while (packed_git_limit < pack_mapped
632                                 && unuse_one_window(p))
633                                 ; /* nothing */
634                         win->base = xmmap_gently(NULL, win->len,
635                                 PROT_READ, MAP_PRIVATE,
636                                 p->pack_fd, win->offset);
637                         if (win->base == MAP_FAILED)
638                                 die_errno("packfile %s cannot be mapped",
639                                           p->pack_name);
640                         if (!win->offset && win->len == p->pack_size
641                                 && !p->do_not_close)
642                                 close_pack_fd(p);
643                         pack_mmap_calls++;
644                         pack_open_windows++;
645                         if (pack_mapped > peak_pack_mapped)
646                                 peak_pack_mapped = pack_mapped;
647                         if (pack_open_windows > peak_pack_open_windows)
648                                 peak_pack_open_windows = pack_open_windows;
649                         win->next = p->windows;
650                         p->windows = win;
651                 }
652         }
653         if (win != *w_cursor) {
654                 win->last_used = pack_used_ctr++;
655                 win->inuse_cnt++;
656                 *w_cursor = win;
657         }
658         offset -= win->offset;
659         if (left)
660                 *left = win->len - xsize_t(offset);
661         return win->base + offset;
662 }
663
664 void unuse_pack(struct pack_window **w_cursor)
665 {
666         struct pack_window *w = *w_cursor;
667         if (w) {
668                 w->inuse_cnt--;
669                 *w_cursor = NULL;
670         }
671 }
672
673 static void try_to_free_pack_memory(size_t size)
674 {
675         release_pack_memory(size);
676 }
677
678 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
679 {
680         static int have_set_try_to_free_routine;
681         struct stat st;
682         size_t alloc;
683         struct packed_git *p;
684
685         if (!have_set_try_to_free_routine) {
686                 have_set_try_to_free_routine = 1;
687                 set_try_to_free_routine(try_to_free_pack_memory);
688         }
689
690         /*
691          * Make sure a corresponding .pack file exists and that
692          * the index looks sane.
693          */
694         if (!strip_suffix_mem(path, &path_len, ".idx"))
695                 return NULL;
696
697         /*
698          * ".promisor" is long enough to hold any suffix we're adding (and
699          * the use xsnprintf double-checks that)
700          */
701         alloc = st_add3(path_len, strlen(".promisor"), 1);
702         p = alloc_packed_git(alloc);
703         memcpy(p->pack_name, path, path_len);
704
705         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
706         if (!access(p->pack_name, F_OK))
707                 p->pack_keep = 1;
708
709         xsnprintf(p->pack_name + path_len, alloc - path_len, ".promisor");
710         if (!access(p->pack_name, F_OK))
711                 p->pack_promisor = 1;
712
713         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
714         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
715                 free(p);
716                 return NULL;
717         }
718
719         /* ok, it looks sane as far as we can check without
720          * actually mapping the pack file.
721          */
722         p->pack_size = st.st_size;
723         p->pack_local = local;
724         p->mtime = st.st_mtime;
725         if (path_len < the_hash_algo->hexsz ||
726             get_sha1_hex(path + path_len - the_hash_algo->hexsz, p->hash))
727                 hashclr(p->hash);
728         return p;
729 }
730
731 void install_packed_git(struct repository *r, struct packed_git *pack)
732 {
733         if (pack->pack_fd != -1)
734                 pack_open_fds++;
735
736         pack->next = r->objects->packed_git;
737         r->objects->packed_git = pack;
738 }
739
740 void (*report_garbage)(unsigned seen_bits, const char *path);
741
742 static void report_helper(const struct string_list *list,
743                           int seen_bits, int first, int last)
744 {
745         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
746                 return;
747
748         for (; first < last; first++)
749                 report_garbage(seen_bits, list->items[first].string);
750 }
751
752 static void report_pack_garbage(struct string_list *list)
753 {
754         int i, baselen = -1, first = 0, seen_bits = 0;
755
756         if (!report_garbage)
757                 return;
758
759         string_list_sort(list);
760
761         for (i = 0; i < list->nr; i++) {
762                 const char *path = list->items[i].string;
763                 if (baselen != -1 &&
764                     strncmp(path, list->items[first].string, baselen)) {
765                         report_helper(list, seen_bits, first, i);
766                         baselen = -1;
767                         seen_bits = 0;
768                 }
769                 if (baselen == -1) {
770                         const char *dot = strrchr(path, '.');
771                         if (!dot) {
772                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
773                                 continue;
774                         }
775                         baselen = dot - path + 1;
776                         first = i;
777                 }
778                 if (!strcmp(path + baselen, "pack"))
779                         seen_bits |= 1;
780                 else if (!strcmp(path + baselen, "idx"))
781                         seen_bits |= 2;
782         }
783         report_helper(list, seen_bits, first, list->nr);
784 }
785
786 void for_each_file_in_pack_dir(const char *objdir,
787                                each_file_in_pack_dir_fn fn,
788                                void *data)
789 {
790         struct strbuf path = STRBUF_INIT;
791         size_t dirnamelen;
792         DIR *dir;
793         struct dirent *de;
794
795         strbuf_addstr(&path, objdir);
796         strbuf_addstr(&path, "/pack");
797         dir = opendir(path.buf);
798         if (!dir) {
799                 if (errno != ENOENT)
800                         error_errno("unable to open object pack directory: %s",
801                                     path.buf);
802                 strbuf_release(&path);
803                 return;
804         }
805         strbuf_addch(&path, '/');
806         dirnamelen = path.len;
807         while ((de = readdir(dir)) != NULL) {
808                 if (is_dot_or_dotdot(de->d_name))
809                         continue;
810
811                 strbuf_setlen(&path, dirnamelen);
812                 strbuf_addstr(&path, de->d_name);
813
814                 fn(path.buf, path.len, de->d_name, data);
815         }
816
817         closedir(dir);
818         strbuf_release(&path);
819 }
820
821 struct prepare_pack_data {
822         struct repository *r;
823         struct string_list *garbage;
824         int local;
825         struct multi_pack_index *m;
826 };
827
828 static void prepare_pack(const char *full_name, size_t full_name_len,
829                          const char *file_name, void *_data)
830 {
831         struct prepare_pack_data *data = (struct prepare_pack_data *)_data;
832         struct packed_git *p;
833         size_t base_len = full_name_len;
834
835         if (strip_suffix_mem(full_name, &base_len, ".idx") &&
836             !(data->m && midx_contains_pack(data->m, file_name))) {
837                 /* Don't reopen a pack we already have. */
838                 for (p = data->r->objects->packed_git; p; p = p->next) {
839                         size_t len;
840                         if (strip_suffix(p->pack_name, ".pack", &len) &&
841                             len == base_len &&
842                             !memcmp(p->pack_name, full_name, len))
843                                 break;
844                 }
845
846                 if (!p) {
847                         p = add_packed_git(full_name, full_name_len, data->local);
848                         if (p)
849                                 install_packed_git(data->r, p);
850                 }
851         }
852
853         if (!report_garbage)
854                 return;
855
856         if (!strcmp(file_name, "multi-pack-index"))
857                 return;
858         if (ends_with(file_name, ".idx") ||
859             ends_with(file_name, ".pack") ||
860             ends_with(file_name, ".bitmap") ||
861             ends_with(file_name, ".keep") ||
862             ends_with(file_name, ".promisor"))
863                 string_list_append(data->garbage, full_name);
864         else
865                 report_garbage(PACKDIR_FILE_GARBAGE, full_name);
866 }
867
868 static void prepare_packed_git_one(struct repository *r, char *objdir, int local)
869 {
870         struct prepare_pack_data data;
871         struct string_list garbage = STRING_LIST_INIT_DUP;
872
873         data.m = r->objects->multi_pack_index;
874
875         /* look for the multi-pack-index for this object directory */
876         while (data.m && strcmp(data.m->object_dir, objdir))
877                 data.m = data.m->next;
878
879         data.r = r;
880         data.garbage = &garbage;
881         data.local = local;
882
883         for_each_file_in_pack_dir(objdir, prepare_pack, &data);
884
885         report_pack_garbage(data.garbage);
886         string_list_clear(data.garbage, 0);
887 }
888
889 static void prepare_packed_git(struct repository *r);
890 /*
891  * Give a fast, rough count of the number of objects in the repository. This
892  * ignores loose objects completely. If you have a lot of them, then either
893  * you should repack because your performance will be awful, or they are
894  * all unreachable objects about to be pruned, in which case they're not really
895  * interesting as a measure of repo size in the first place.
896  */
897 unsigned long repo_approximate_object_count(struct repository *r)
898 {
899         if (!r->objects->approximate_object_count_valid) {
900                 unsigned long count;
901                 struct multi_pack_index *m;
902                 struct packed_git *p;
903
904                 prepare_packed_git(r);
905                 count = 0;
906                 for (m = get_multi_pack_index(r); m; m = m->next)
907                         count += m->num_objects;
908                 for (p = r->objects->packed_git; p; p = p->next) {
909                         if (open_pack_index(p))
910                                 continue;
911                         count += p->num_objects;
912                 }
913                 r->objects->approximate_object_count = count;
914         }
915         return r->objects->approximate_object_count;
916 }
917
918 static void *get_next_packed_git(const void *p)
919 {
920         return ((const struct packed_git *)p)->next;
921 }
922
923 static void set_next_packed_git(void *p, void *next)
924 {
925         ((struct packed_git *)p)->next = next;
926 }
927
928 static int sort_pack(const void *a_, const void *b_)
929 {
930         const struct packed_git *a = a_;
931         const struct packed_git *b = b_;
932         int st;
933
934         /*
935          * Local packs tend to contain objects specific to our
936          * variant of the project than remote ones.  In addition,
937          * remote ones could be on a network mounted filesystem.
938          * Favor local ones for these reasons.
939          */
940         st = a->pack_local - b->pack_local;
941         if (st)
942                 return -st;
943
944         /*
945          * Younger packs tend to contain more recent objects,
946          * and more recent objects tend to get accessed more
947          * often.
948          */
949         if (a->mtime < b->mtime)
950                 return 1;
951         else if (a->mtime == b->mtime)
952                 return 0;
953         return -1;
954 }
955
956 static void rearrange_packed_git(struct repository *r)
957 {
958         r->objects->packed_git = llist_mergesort(
959                 r->objects->packed_git, get_next_packed_git,
960                 set_next_packed_git, sort_pack);
961 }
962
963 static void prepare_packed_git_mru(struct repository *r)
964 {
965         struct packed_git *p;
966
967         INIT_LIST_HEAD(&r->objects->packed_git_mru);
968
969         for (p = r->objects->packed_git; p; p = p->next)
970                 list_add_tail(&p->mru, &r->objects->packed_git_mru);
971 }
972
973 static void prepare_packed_git(struct repository *r)
974 {
975         struct object_directory *odb;
976
977         if (r->objects->packed_git_initialized)
978                 return;
979
980         prepare_alt_odb(r);
981         for (odb = r->objects->odb; odb; odb = odb->next) {
982                 int local = (odb == r->objects->odb);
983                 prepare_multi_pack_index_one(r, odb->path, local);
984                 prepare_packed_git_one(r, odb->path, local);
985         }
986         rearrange_packed_git(r);
987
988         prepare_packed_git_mru(r);
989         r->objects->packed_git_initialized = 1;
990 }
991
992 void reprepare_packed_git(struct repository *r)
993 {
994         struct object_directory *odb;
995
996         for (odb = r->objects->odb; odb; odb = odb->next)
997                 odb_clear_loose_cache(odb);
998
999         r->objects->approximate_object_count_valid = 0;
1000         r->objects->packed_git_initialized = 0;
1001         prepare_packed_git(r);
1002 }
1003
1004 struct packed_git *get_packed_git(struct repository *r)
1005 {
1006         prepare_packed_git(r);
1007         return r->objects->packed_git;
1008 }
1009
1010 struct multi_pack_index *get_multi_pack_index(struct repository *r)
1011 {
1012         prepare_packed_git(r);
1013         return r->objects->multi_pack_index;
1014 }
1015
1016 struct packed_git *get_all_packs(struct repository *r)
1017 {
1018         struct multi_pack_index *m;
1019
1020         prepare_packed_git(r);
1021         for (m = r->objects->multi_pack_index; m; m = m->next) {
1022                 uint32_t i;
1023                 for (i = 0; i < m->num_packs; i++)
1024                         prepare_midx_pack(r, m, i);
1025         }
1026
1027         return r->objects->packed_git;
1028 }
1029
1030 struct list_head *get_packed_git_mru(struct repository *r)
1031 {
1032         prepare_packed_git(r);
1033         return &r->objects->packed_git_mru;
1034 }
1035
1036 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1037                 unsigned long len, enum object_type *type, unsigned long *sizep)
1038 {
1039         unsigned shift;
1040         unsigned long size, c;
1041         unsigned long used = 0;
1042
1043         c = buf[used++];
1044         *type = (c >> 4) & 7;
1045         size = c & 15;
1046         shift = 4;
1047         while (c & 0x80) {
1048                 if (len <= used || bitsizeof(long) <= shift) {
1049                         error("bad object header");
1050                         size = used = 0;
1051                         break;
1052                 }
1053                 c = buf[used++];
1054                 size += (c & 0x7f) << shift;
1055                 shift += 7;
1056         }
1057         *sizep = size;
1058         return used;
1059 }
1060
1061 unsigned long get_size_from_delta(struct packed_git *p,
1062                                   struct pack_window **w_curs,
1063                                   off_t curpos)
1064 {
1065         const unsigned char *data;
1066         unsigned char delta_head[20], *in;
1067         git_zstream stream;
1068         int st;
1069
1070         memset(&stream, 0, sizeof(stream));
1071         stream.next_out = delta_head;
1072         stream.avail_out = sizeof(delta_head);
1073
1074         git_inflate_init(&stream);
1075         do {
1076                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1077                 stream.next_in = in;
1078                 st = git_inflate(&stream, Z_FINISH);
1079                 curpos += stream.next_in - in;
1080         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1081                  stream.total_out < sizeof(delta_head));
1082         git_inflate_end(&stream);
1083         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1084                 error("delta data unpack-initial failed");
1085                 return 0;
1086         }
1087
1088         /* Examine the initial part of the delta to figure out
1089          * the result size.
1090          */
1091         data = delta_head;
1092
1093         /* ignore base size */
1094         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1095
1096         /* Read the result size */
1097         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1098 }
1099
1100 int unpack_object_header(struct packed_git *p,
1101                          struct pack_window **w_curs,
1102                          off_t *curpos,
1103                          unsigned long *sizep)
1104 {
1105         unsigned char *base;
1106         unsigned long left;
1107         unsigned long used;
1108         enum object_type type;
1109
1110         /* use_pack() assures us we have [base, base + 20) available
1111          * as a range that we can look at.  (Its actually the hash
1112          * size that is assured.)  With our object header encoding
1113          * the maximum deflated object size is 2^137, which is just
1114          * insane, so we know won't exceed what we have been given.
1115          */
1116         base = use_pack(p, w_curs, *curpos, &left);
1117         used = unpack_object_header_buffer(base, left, &type, sizep);
1118         if (!used) {
1119                 type = OBJ_BAD;
1120         } else
1121                 *curpos += used;
1122
1123         return type;
1124 }
1125
1126 void mark_bad_packed_object(struct packed_git *p, const unsigned char *sha1)
1127 {
1128         unsigned i;
1129         const unsigned hashsz = the_hash_algo->rawsz;
1130         for (i = 0; i < p->num_bad_objects; i++)
1131                 if (hasheq(sha1, p->bad_object_sha1 + hashsz * i))
1132                         return;
1133         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1134                                       st_mult(GIT_MAX_RAWSZ,
1135                                               st_add(p->num_bad_objects, 1)));
1136         hashcpy(p->bad_object_sha1 + hashsz * p->num_bad_objects, sha1);
1137         p->num_bad_objects++;
1138 }
1139
1140 const struct packed_git *has_packed_and_bad(struct repository *r,
1141                                             const unsigned char *sha1)
1142 {
1143         struct packed_git *p;
1144         unsigned i;
1145
1146         for (p = r->objects->packed_git; p; p = p->next)
1147                 for (i = 0; i < p->num_bad_objects; i++)
1148                         if (hasheq(sha1,
1149                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
1150                                 return p;
1151         return NULL;
1152 }
1153
1154 static off_t get_delta_base(struct packed_git *p,
1155                                     struct pack_window **w_curs,
1156                                     off_t *curpos,
1157                                     enum object_type type,
1158                                     off_t delta_obj_offset)
1159 {
1160         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1161         off_t base_offset;
1162
1163         /* use_pack() assured us we have [base_info, base_info + 20)
1164          * as a range that we can look at without walking off the
1165          * end of the mapped window.  Its actually the hash size
1166          * that is assured.  An OFS_DELTA longer than the hash size
1167          * is stupid, as then a REF_DELTA would be smaller to store.
1168          */
1169         if (type == OBJ_OFS_DELTA) {
1170                 unsigned used = 0;
1171                 unsigned char c = base_info[used++];
1172                 base_offset = c & 127;
1173                 while (c & 128) {
1174                         base_offset += 1;
1175                         if (!base_offset || MSB(base_offset, 7))
1176                                 return 0;  /* overflow */
1177                         c = base_info[used++];
1178                         base_offset = (base_offset << 7) + (c & 127);
1179                 }
1180                 base_offset = delta_obj_offset - base_offset;
1181                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1182                         return 0;  /* out of bound */
1183                 *curpos += used;
1184         } else if (type == OBJ_REF_DELTA) {
1185                 /* The base entry _must_ be in the same pack */
1186                 base_offset = find_pack_entry_one(base_info, p);
1187                 *curpos += the_hash_algo->rawsz;
1188         } else
1189                 die("I am totally screwed");
1190         return base_offset;
1191 }
1192
1193 /*
1194  * Like get_delta_base above, but we return the sha1 instead of the pack
1195  * offset. This means it is cheaper for REF deltas (we do not have to do
1196  * the final object lookup), but more expensive for OFS deltas (we
1197  * have to load the revidx to convert the offset back into a sha1).
1198  */
1199 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1200                                                 struct pack_window **w_curs,
1201                                                 off_t curpos,
1202                                                 enum object_type type,
1203                                                 off_t delta_obj_offset)
1204 {
1205         if (type == OBJ_REF_DELTA) {
1206                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1207                 return base;
1208         } else if (type == OBJ_OFS_DELTA) {
1209                 struct revindex_entry *revidx;
1210                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1211                                                    type, delta_obj_offset);
1212
1213                 if (!base_offset)
1214                         return NULL;
1215
1216                 revidx = find_pack_revindex(p, base_offset);
1217                 if (!revidx)
1218                         return NULL;
1219
1220                 return nth_packed_object_sha1(p, revidx->nr);
1221         } else
1222                 return NULL;
1223 }
1224
1225 static int retry_bad_packed_offset(struct repository *r,
1226                                    struct packed_git *p,
1227                                    off_t obj_offset)
1228 {
1229         int type;
1230         struct revindex_entry *revidx;
1231         struct object_id oid;
1232         revidx = find_pack_revindex(p, obj_offset);
1233         if (!revidx)
1234                 return OBJ_BAD;
1235         nth_packed_object_oid(&oid, p, revidx->nr);
1236         mark_bad_packed_object(p, oid.hash);
1237         type = oid_object_info(r, &oid, NULL);
1238         if (type <= OBJ_NONE)
1239                 return OBJ_BAD;
1240         return type;
1241 }
1242
1243 #define POI_STACK_PREALLOC 64
1244
1245 static enum object_type packed_to_object_type(struct repository *r,
1246                                               struct packed_git *p,
1247                                               off_t obj_offset,
1248                                               enum object_type type,
1249                                               struct pack_window **w_curs,
1250                                               off_t curpos)
1251 {
1252         off_t small_poi_stack[POI_STACK_PREALLOC];
1253         off_t *poi_stack = small_poi_stack;
1254         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1255
1256         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1257                 off_t base_offset;
1258                 unsigned long size;
1259                 /* Push the object we're going to leave behind */
1260                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1261                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1262                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
1263                         COPY_ARRAY(poi_stack, small_poi_stack, poi_stack_nr);
1264                 } else {
1265                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1266                 }
1267                 poi_stack[poi_stack_nr++] = obj_offset;
1268                 /* If parsing the base offset fails, just unwind */
1269                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1270                 if (!base_offset)
1271                         goto unwind;
1272                 curpos = obj_offset = base_offset;
1273                 type = unpack_object_header(p, w_curs, &curpos, &size);
1274                 if (type <= OBJ_NONE) {
1275                         /* If getting the base itself fails, we first
1276                          * retry the base, otherwise unwind */
1277                         type = retry_bad_packed_offset(r, p, base_offset);
1278                         if (type > OBJ_NONE)
1279                                 goto out;
1280                         goto unwind;
1281                 }
1282         }
1283
1284         switch (type) {
1285         case OBJ_BAD:
1286         case OBJ_COMMIT:
1287         case OBJ_TREE:
1288         case OBJ_BLOB:
1289         case OBJ_TAG:
1290                 break;
1291         default:
1292                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1293                       type, (uintmax_t)obj_offset, p->pack_name);
1294                 type = OBJ_BAD;
1295         }
1296
1297 out:
1298         if (poi_stack != small_poi_stack)
1299                 free(poi_stack);
1300         return type;
1301
1302 unwind:
1303         while (poi_stack_nr) {
1304                 obj_offset = poi_stack[--poi_stack_nr];
1305                 type = retry_bad_packed_offset(r, p, obj_offset);
1306                 if (type > OBJ_NONE)
1307                         goto out;
1308         }
1309         type = OBJ_BAD;
1310         goto out;
1311 }
1312
1313 static struct hashmap delta_base_cache;
1314 static size_t delta_base_cached;
1315
1316 static LIST_HEAD(delta_base_cache_lru);
1317
1318 struct delta_base_cache_key {
1319         struct packed_git *p;
1320         off_t base_offset;
1321 };
1322
1323 struct delta_base_cache_entry {
1324         struct hashmap hash;
1325         struct delta_base_cache_key key;
1326         struct list_head lru;
1327         void *data;
1328         unsigned long size;
1329         enum object_type type;
1330 };
1331
1332 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
1333 {
1334         unsigned int hash;
1335
1336         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
1337         hash += (hash >> 8) + (hash >> 16);
1338         return hash;
1339 }
1340
1341 static struct delta_base_cache_entry *
1342 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
1343 {
1344         struct hashmap_entry entry;
1345         struct delta_base_cache_key key;
1346
1347         if (!delta_base_cache.cmpfn)
1348                 return NULL;
1349
1350         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
1351         key.p = p;
1352         key.base_offset = base_offset;
1353         return hashmap_get(&delta_base_cache, &entry, &key);
1354 }
1355
1356 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
1357                                    const struct delta_base_cache_key *b)
1358 {
1359         return a->p == b->p && a->base_offset == b->base_offset;
1360 }
1361
1362 static int delta_base_cache_hash_cmp(const void *unused_cmp_data,
1363                                      const void *va, const void *vb,
1364                                      const void *vkey)
1365 {
1366         const struct delta_base_cache_entry *a = va, *b = vb;
1367         const struct delta_base_cache_key *key = vkey;
1368         if (key)
1369                 return !delta_base_cache_key_eq(&a->key, key);
1370         else
1371                 return !delta_base_cache_key_eq(&a->key, &b->key);
1372 }
1373
1374 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
1375 {
1376         return !!get_delta_base_cache_entry(p, base_offset);
1377 }
1378
1379 /*
1380  * Remove the entry from the cache, but do _not_ free the associated
1381  * entry data. The caller takes ownership of the "data" buffer, and
1382  * should copy out any fields it wants before detaching.
1383  */
1384 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
1385 {
1386         hashmap_remove(&delta_base_cache, ent, &ent->key);
1387         list_del(&ent->lru);
1388         delta_base_cached -= ent->size;
1389         free(ent);
1390 }
1391
1392 static void *cache_or_unpack_entry(struct repository *r, struct packed_git *p,
1393                                    off_t base_offset, unsigned long *base_size,
1394                                    enum object_type *type)
1395 {
1396         struct delta_base_cache_entry *ent;
1397
1398         ent = get_delta_base_cache_entry(p, base_offset);
1399         if (!ent)
1400                 return unpack_entry(r, p, base_offset, type, base_size);
1401
1402         if (type)
1403                 *type = ent->type;
1404         if (base_size)
1405                 *base_size = ent->size;
1406         return xmemdupz(ent->data, ent->size);
1407 }
1408
1409 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1410 {
1411         free(ent->data);
1412         detach_delta_base_cache_entry(ent);
1413 }
1414
1415 void clear_delta_base_cache(void)
1416 {
1417         struct list_head *lru, *tmp;
1418         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1419                 struct delta_base_cache_entry *entry =
1420                         list_entry(lru, struct delta_base_cache_entry, lru);
1421                 release_delta_base_cache(entry);
1422         }
1423 }
1424
1425 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1426         void *base, unsigned long base_size, enum object_type type)
1427 {
1428         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
1429         struct list_head *lru, *tmp;
1430
1431         delta_base_cached += base_size;
1432
1433         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1434                 struct delta_base_cache_entry *f =
1435                         list_entry(lru, struct delta_base_cache_entry, lru);
1436                 if (delta_base_cached <= delta_base_cache_limit)
1437                         break;
1438                 release_delta_base_cache(f);
1439         }
1440
1441         ent->key.p = p;
1442         ent->key.base_offset = base_offset;
1443         ent->type = type;
1444         ent->data = base;
1445         ent->size = base_size;
1446         list_add_tail(&ent->lru, &delta_base_cache_lru);
1447
1448         if (!delta_base_cache.cmpfn)
1449                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, NULL, 0);
1450         hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
1451         hashmap_add(&delta_base_cache, ent);
1452 }
1453
1454 int packed_object_info(struct repository *r, struct packed_git *p,
1455                        off_t obj_offset, struct object_info *oi)
1456 {
1457         struct pack_window *w_curs = NULL;
1458         unsigned long size;
1459         off_t curpos = obj_offset;
1460         enum object_type type;
1461
1462         /*
1463          * We always get the representation type, but only convert it to
1464          * a "real" type later if the caller is interested.
1465          */
1466         if (oi->contentp) {
1467                 *oi->contentp = cache_or_unpack_entry(r, p, obj_offset, oi->sizep,
1468                                                       &type);
1469                 if (!*oi->contentp)
1470                         type = OBJ_BAD;
1471         } else {
1472                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1473         }
1474
1475         if (!oi->contentp && oi->sizep) {
1476                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1477                         off_t tmp_pos = curpos;
1478                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1479                                                            type, obj_offset);
1480                         if (!base_offset) {
1481                                 type = OBJ_BAD;
1482                                 goto out;
1483                         }
1484                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1485                         if (*oi->sizep == 0) {
1486                                 type = OBJ_BAD;
1487                                 goto out;
1488                         }
1489                 } else {
1490                         *oi->sizep = size;
1491                 }
1492         }
1493
1494         if (oi->disk_sizep) {
1495                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1496                 *oi->disk_sizep = revidx[1].offset - obj_offset;
1497         }
1498
1499         if (oi->typep || oi->type_name) {
1500                 enum object_type ptot;
1501                 ptot = packed_to_object_type(r, p, obj_offset,
1502                                              type, &w_curs, curpos);
1503                 if (oi->typep)
1504                         *oi->typep = ptot;
1505                 if (oi->type_name) {
1506                         const char *tn = type_name(ptot);
1507                         if (tn)
1508                                 strbuf_addstr(oi->type_name, tn);
1509                 }
1510                 if (ptot < 0) {
1511                         type = OBJ_BAD;
1512                         goto out;
1513                 }
1514         }
1515
1516         if (oi->delta_base_sha1) {
1517                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1518                         const unsigned char *base;
1519
1520                         base = get_delta_base_sha1(p, &w_curs, curpos,
1521                                                    type, obj_offset);
1522                         if (!base) {
1523                                 type = OBJ_BAD;
1524                                 goto out;
1525                         }
1526
1527                         hashcpy(oi->delta_base_sha1, base);
1528                 } else
1529                         hashclr(oi->delta_base_sha1);
1530         }
1531
1532         oi->whence = in_delta_base_cache(p, obj_offset) ? OI_DBCACHED :
1533                                                           OI_PACKED;
1534
1535 out:
1536         unuse_pack(&w_curs);
1537         return type;
1538 }
1539
1540 static void *unpack_compressed_entry(struct packed_git *p,
1541                                     struct pack_window **w_curs,
1542                                     off_t curpos,
1543                                     unsigned long size)
1544 {
1545         int st;
1546         git_zstream stream;
1547         unsigned char *buffer, *in;
1548
1549         buffer = xmallocz_gently(size);
1550         if (!buffer)
1551                 return NULL;
1552         memset(&stream, 0, sizeof(stream));
1553         stream.next_out = buffer;
1554         stream.avail_out = size + 1;
1555
1556         git_inflate_init(&stream);
1557         do {
1558                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1559                 stream.next_in = in;
1560                 st = git_inflate(&stream, Z_FINISH);
1561                 if (!stream.avail_out)
1562                         break; /* the payload is larger than it should be */
1563                 curpos += stream.next_in - in;
1564         } while (st == Z_OK || st == Z_BUF_ERROR);
1565         git_inflate_end(&stream);
1566         if ((st != Z_STREAM_END) || stream.total_out != size) {
1567                 free(buffer);
1568                 return NULL;
1569         }
1570
1571         /* versions of zlib can clobber unconsumed portion of outbuf */
1572         buffer[size] = '\0';
1573
1574         return buffer;
1575 }
1576
1577 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
1578 {
1579         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
1580         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
1581                          p->pack_name, (uintmax_t)obj_offset);
1582 }
1583
1584 int do_check_packed_object_crc;
1585
1586 #define UNPACK_ENTRY_STACK_PREALLOC 64
1587 struct unpack_entry_stack_ent {
1588         off_t obj_offset;
1589         off_t curpos;
1590         unsigned long size;
1591 };
1592
1593 static void *read_object(struct repository *r,
1594                          const struct object_id *oid,
1595                          enum object_type *type,
1596                          unsigned long *size)
1597 {
1598         struct object_info oi = OBJECT_INFO_INIT;
1599         void *content;
1600         oi.typep = type;
1601         oi.sizep = size;
1602         oi.contentp = &content;
1603
1604         if (oid_object_info_extended(r, oid, &oi, 0) < 0)
1605                 return NULL;
1606         return content;
1607 }
1608
1609 void *unpack_entry(struct repository *r, struct packed_git *p, off_t obj_offset,
1610                    enum object_type *final_type, unsigned long *final_size)
1611 {
1612         struct pack_window *w_curs = NULL;
1613         off_t curpos = obj_offset;
1614         void *data = NULL;
1615         unsigned long size;
1616         enum object_type type;
1617         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
1618         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
1619         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
1620         int base_from_cache = 0;
1621
1622         write_pack_access_log(p, obj_offset);
1623
1624         /* PHASE 1: drill down to the innermost base object */
1625         for (;;) {
1626                 off_t base_offset;
1627                 int i;
1628                 struct delta_base_cache_entry *ent;
1629
1630                 ent = get_delta_base_cache_entry(p, curpos);
1631                 if (ent) {
1632                         type = ent->type;
1633                         data = ent->data;
1634                         size = ent->size;
1635                         detach_delta_base_cache_entry(ent);
1636                         base_from_cache = 1;
1637                         break;
1638                 }
1639
1640                 if (do_check_packed_object_crc && p->index_version > 1) {
1641                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1642                         off_t len = revidx[1].offset - obj_offset;
1643                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1644                                 struct object_id oid;
1645                                 nth_packed_object_oid(&oid, p, revidx->nr);
1646                                 error("bad packed object CRC for %s",
1647                                       oid_to_hex(&oid));
1648                                 mark_bad_packed_object(p, oid.hash);
1649                                 data = NULL;
1650                                 goto out;
1651                         }
1652                 }
1653
1654                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1655                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
1656                         break;
1657
1658                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1659                 if (!base_offset) {
1660                         error("failed to validate delta base reference "
1661                               "at offset %"PRIuMAX" from %s",
1662                               (uintmax_t)curpos, p->pack_name);
1663                         /* bail to phase 2, in hopes of recovery */
1664                         data = NULL;
1665                         break;
1666                 }
1667
1668                 /* push object, proceed to base */
1669                 if (delta_stack_nr >= delta_stack_alloc
1670                     && delta_stack == small_delta_stack) {
1671                         delta_stack_alloc = alloc_nr(delta_stack_nr);
1672                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
1673                         COPY_ARRAY(delta_stack, small_delta_stack,
1674                                    delta_stack_nr);
1675                 } else {
1676                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
1677                 }
1678                 i = delta_stack_nr++;
1679                 delta_stack[i].obj_offset = obj_offset;
1680                 delta_stack[i].curpos = curpos;
1681                 delta_stack[i].size = size;
1682
1683                 curpos = obj_offset = base_offset;
1684         }
1685
1686         /* PHASE 2: handle the base */
1687         switch (type) {
1688         case OBJ_OFS_DELTA:
1689         case OBJ_REF_DELTA:
1690                 if (data)
1691                         BUG("unpack_entry: left loop at a valid delta");
1692                 break;
1693         case OBJ_COMMIT:
1694         case OBJ_TREE:
1695         case OBJ_BLOB:
1696         case OBJ_TAG:
1697                 if (!base_from_cache)
1698                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
1699                 break;
1700         default:
1701                 data = NULL;
1702                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1703                       type, (uintmax_t)obj_offset, p->pack_name);
1704         }
1705
1706         /* PHASE 3: apply deltas in order */
1707
1708         /* invariants:
1709          *   'data' holds the base data, or NULL if there was corruption
1710          */
1711         while (delta_stack_nr) {
1712                 void *delta_data;
1713                 void *base = data;
1714                 void *external_base = NULL;
1715                 unsigned long delta_size, base_size = size;
1716                 int i;
1717
1718                 data = NULL;
1719
1720                 if (base)
1721                         add_delta_base_cache(p, obj_offset, base, base_size, type);
1722
1723                 if (!base) {
1724                         /*
1725                          * We're probably in deep shit, but let's try to fetch
1726                          * the required base anyway from another pack or loose.
1727                          * This is costly but should happen only in the presence
1728                          * of a corrupted pack, and is better than failing outright.
1729                          */
1730                         struct revindex_entry *revidx;
1731                         struct object_id base_oid;
1732                         revidx = find_pack_revindex(p, obj_offset);
1733                         if (revidx) {
1734                                 nth_packed_object_oid(&base_oid, p, revidx->nr);
1735                                 error("failed to read delta base object %s"
1736                                       " at offset %"PRIuMAX" from %s",
1737                                       oid_to_hex(&base_oid), (uintmax_t)obj_offset,
1738                                       p->pack_name);
1739                                 mark_bad_packed_object(p, base_oid.hash);
1740                                 base = read_object(r, &base_oid, &type, &base_size);
1741                                 external_base = base;
1742                         }
1743                 }
1744
1745                 i = --delta_stack_nr;
1746                 obj_offset = delta_stack[i].obj_offset;
1747                 curpos = delta_stack[i].curpos;
1748                 delta_size = delta_stack[i].size;
1749
1750                 if (!base)
1751                         continue;
1752
1753                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
1754
1755                 if (!delta_data) {
1756                         error("failed to unpack compressed delta "
1757                               "at offset %"PRIuMAX" from %s",
1758                               (uintmax_t)curpos, p->pack_name);
1759                         data = NULL;
1760                         free(external_base);
1761                         continue;
1762                 }
1763
1764                 data = patch_delta(base, base_size,
1765                                    delta_data, delta_size,
1766                                    &size);
1767
1768                 /*
1769                  * We could not apply the delta; warn the user, but keep going.
1770                  * Our failure will be noticed either in the next iteration of
1771                  * the loop, or if this is the final delta, in the caller when
1772                  * we return NULL. Those code paths will take care of making
1773                  * a more explicit warning and retrying with another copy of
1774                  * the object.
1775                  */
1776                 if (!data)
1777                         error("failed to apply delta");
1778
1779                 free(delta_data);
1780                 free(external_base);
1781         }
1782
1783         if (final_type)
1784                 *final_type = type;
1785         if (final_size)
1786                 *final_size = size;
1787
1788 out:
1789         unuse_pack(&w_curs);
1790
1791         if (delta_stack != small_delta_stack)
1792                 free(delta_stack);
1793
1794         return data;
1795 }
1796
1797 int bsearch_pack(const struct object_id *oid, const struct packed_git *p, uint32_t *result)
1798 {
1799         const unsigned char *index_fanout = p->index_data;
1800         const unsigned char *index_lookup;
1801         const unsigned int hashsz = the_hash_algo->rawsz;
1802         int index_lookup_width;
1803
1804         if (!index_fanout)
1805                 BUG("bsearch_pack called without a valid pack-index");
1806
1807         index_lookup = index_fanout + 4 * 256;
1808         if (p->index_version == 1) {
1809                 index_lookup_width = hashsz + 4;
1810                 index_lookup += 4;
1811         } else {
1812                 index_lookup_width = hashsz;
1813                 index_fanout += 8;
1814                 index_lookup += 8;
1815         }
1816
1817         return bsearch_hash(oid->hash, (const uint32_t*)index_fanout,
1818                             index_lookup, index_lookup_width, result);
1819 }
1820
1821 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1822                                             uint32_t n)
1823 {
1824         const unsigned char *index = p->index_data;
1825         const unsigned int hashsz = the_hash_algo->rawsz;
1826         if (!index) {
1827                 if (open_pack_index(p))
1828                         return NULL;
1829                 index = p->index_data;
1830         }
1831         if (n >= p->num_objects)
1832                 return NULL;
1833         index += 4 * 256;
1834         if (p->index_version == 1) {
1835                 return index + (hashsz + 4) * n + 4;
1836         } else {
1837                 index += 8;
1838                 return index + hashsz * n;
1839         }
1840 }
1841
1842 const struct object_id *nth_packed_object_oid(struct object_id *oid,
1843                                               struct packed_git *p,
1844                                               uint32_t n)
1845 {
1846         const unsigned char *hash = nth_packed_object_sha1(p, n);
1847         if (!hash)
1848                 return NULL;
1849         hashcpy(oid->hash, hash);
1850         return oid;
1851 }
1852
1853 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
1854 {
1855         const unsigned char *ptr = vptr;
1856         const unsigned char *start = p->index_data;
1857         const unsigned char *end = start + p->index_size;
1858         if (ptr < start)
1859                 die(_("offset before start of pack index for %s (corrupt index?)"),
1860                     p->pack_name);
1861         /* No need to check for underflow; .idx files must be at least 8 bytes */
1862         if (ptr >= end - 8)
1863                 die(_("offset beyond end of pack index for %s (truncated index?)"),
1864                     p->pack_name);
1865 }
1866
1867 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1868 {
1869         const unsigned char *index = p->index_data;
1870         const unsigned int hashsz = the_hash_algo->rawsz;
1871         index += 4 * 256;
1872         if (p->index_version == 1) {
1873                 return ntohl(*((uint32_t *)(index + (hashsz + 4) * n)));
1874         } else {
1875                 uint32_t off;
1876                 index += 8 + p->num_objects * (hashsz + 4);
1877                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1878                 if (!(off & 0x80000000))
1879                         return off;
1880                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1881                 check_pack_index_ptr(p, index);
1882                 return get_be64(index);
1883         }
1884 }
1885
1886 off_t find_pack_entry_one(const unsigned char *sha1,
1887                                   struct packed_git *p)
1888 {
1889         const unsigned char *index = p->index_data;
1890         struct object_id oid;
1891         uint32_t result;
1892
1893         if (!index) {
1894                 if (open_pack_index(p))
1895                         return 0;
1896         }
1897
1898         hashcpy(oid.hash, sha1);
1899         if (bsearch_pack(&oid, p, &result))
1900                 return nth_packed_object_offset(p, result);
1901         return 0;
1902 }
1903
1904 int is_pack_valid(struct packed_git *p)
1905 {
1906         /* An already open pack is known to be valid. */
1907         if (p->pack_fd != -1)
1908                 return 1;
1909
1910         /* If the pack has one window completely covering the
1911          * file size, the pack is known to be valid even if
1912          * the descriptor is not currently open.
1913          */
1914         if (p->windows) {
1915                 struct pack_window *w = p->windows;
1916
1917                 if (!w->offset && w->len == p->pack_size)
1918                         return 1;
1919         }
1920
1921         /* Force the pack to open to prove its valid. */
1922         return !open_packed_git(p);
1923 }
1924
1925 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1926                                   struct packed_git *packs)
1927 {
1928         struct packed_git *p;
1929
1930         for (p = packs; p; p = p->next) {
1931                 if (find_pack_entry_one(sha1, p))
1932                         return p;
1933         }
1934         return NULL;
1935
1936 }
1937
1938 static int fill_pack_entry(const struct object_id *oid,
1939                            struct pack_entry *e,
1940                            struct packed_git *p)
1941 {
1942         off_t offset;
1943
1944         if (p->num_bad_objects) {
1945                 unsigned i;
1946                 for (i = 0; i < p->num_bad_objects; i++)
1947                         if (hasheq(oid->hash,
1948                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
1949                                 return 0;
1950         }
1951
1952         offset = find_pack_entry_one(oid->hash, p);
1953         if (!offset)
1954                 return 0;
1955
1956         /*
1957          * We are about to tell the caller where they can locate the
1958          * requested object.  We better make sure the packfile is
1959          * still here and can be accessed before supplying that
1960          * answer, as it may have been deleted since the index was
1961          * loaded!
1962          */
1963         if (!is_pack_valid(p))
1964                 return 0;
1965         e->offset = offset;
1966         e->p = p;
1967         return 1;
1968 }
1969
1970 int find_pack_entry(struct repository *r, const struct object_id *oid, struct pack_entry *e)
1971 {
1972         struct list_head *pos;
1973         struct multi_pack_index *m;
1974
1975         prepare_packed_git(r);
1976         if (!r->objects->packed_git && !r->objects->multi_pack_index)
1977                 return 0;
1978
1979         for (m = r->objects->multi_pack_index; m; m = m->next) {
1980                 if (fill_midx_entry(r, oid, e, m))
1981                         return 1;
1982         }
1983
1984         list_for_each(pos, &r->objects->packed_git_mru) {
1985                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
1986                 if (!p->multi_pack_index && fill_pack_entry(oid, e, p)) {
1987                         list_move(&p->mru, &r->objects->packed_git_mru);
1988                         return 1;
1989                 }
1990         }
1991         return 0;
1992 }
1993
1994 int has_object_pack(const struct object_id *oid)
1995 {
1996         struct pack_entry e;
1997         return find_pack_entry(the_repository, oid, &e);
1998 }
1999
2000 int has_pack_index(const unsigned char *sha1)
2001 {
2002         struct stat st;
2003         if (stat(sha1_pack_index_name(sha1), &st))
2004                 return 0;
2005         return 1;
2006 }
2007
2008 int for_each_object_in_pack(struct packed_git *p,
2009                             each_packed_object_fn cb, void *data,
2010                             enum for_each_object_flags flags)
2011 {
2012         uint32_t i;
2013         int r = 0;
2014
2015         if (flags & FOR_EACH_OBJECT_PACK_ORDER) {
2016                 if (load_pack_revindex(p))
2017                         return -1;
2018         }
2019
2020         for (i = 0; i < p->num_objects; i++) {
2021                 uint32_t pos;
2022                 struct object_id oid;
2023
2024                 if (flags & FOR_EACH_OBJECT_PACK_ORDER)
2025                         pos = p->revindex[i].nr;
2026                 else
2027                         pos = i;
2028
2029                 if (!nth_packed_object_oid(&oid, p, pos))
2030                         return error("unable to get sha1 of object %u in %s",
2031                                      pos, p->pack_name);
2032
2033                 r = cb(&oid, p, pos, data);
2034                 if (r)
2035                         break;
2036         }
2037         return r;
2038 }
2039
2040 int for_each_packed_object(each_packed_object_fn cb, void *data,
2041                            enum for_each_object_flags flags)
2042 {
2043         struct packed_git *p;
2044         int r = 0;
2045         int pack_errors = 0;
2046
2047         prepare_packed_git(the_repository);
2048         for (p = get_all_packs(the_repository); p; p = p->next) {
2049                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
2050                         continue;
2051                 if ((flags & FOR_EACH_OBJECT_PROMISOR_ONLY) &&
2052                     !p->pack_promisor)
2053                         continue;
2054                 if (open_pack_index(p)) {
2055                         pack_errors = 1;
2056                         continue;
2057                 }
2058                 r = for_each_object_in_pack(p, cb, data, flags);
2059                 if (r)
2060                         break;
2061         }
2062         return r ? r : pack_errors;
2063 }
2064
2065 static int add_promisor_object(const struct object_id *oid,
2066                                struct packed_git *pack,
2067                                uint32_t pos,
2068                                void *set_)
2069 {
2070         struct oidset *set = set_;
2071         struct object *obj = parse_object(the_repository, oid);
2072         if (!obj)
2073                 return 1;
2074
2075         oidset_insert(set, oid);
2076
2077         /*
2078          * If this is a tree, commit, or tag, the objects it refers
2079          * to are also promisor objects. (Blobs refer to no objects->)
2080          */
2081         if (obj->type == OBJ_TREE) {
2082                 struct tree *tree = (struct tree *)obj;
2083                 struct tree_desc desc;
2084                 struct name_entry entry;
2085                 if (init_tree_desc_gently(&desc, tree->buffer, tree->size))
2086                         /*
2087                          * Error messages are given when packs are
2088                          * verified, so do not print any here.
2089                          */
2090                         return 0;
2091                 while (tree_entry_gently(&desc, &entry))
2092                         oidset_insert(set, &entry.oid);
2093         } else if (obj->type == OBJ_COMMIT) {
2094                 struct commit *commit = (struct commit *) obj;
2095                 struct commit_list *parents = commit->parents;
2096
2097                 oidset_insert(set, get_commit_tree_oid(commit));
2098                 for (; parents; parents = parents->next)
2099                         oidset_insert(set, &parents->item->object.oid);
2100         } else if (obj->type == OBJ_TAG) {
2101                 struct tag *tag = (struct tag *) obj;
2102                 oidset_insert(set, &tag->tagged->oid);
2103         }
2104         return 0;
2105 }
2106
2107 int is_promisor_object(const struct object_id *oid)
2108 {
2109         static struct oidset promisor_objects;
2110         static int promisor_objects_prepared;
2111
2112         if (!promisor_objects_prepared) {
2113                 if (repository_format_partial_clone) {
2114                         for_each_packed_object(add_promisor_object,
2115                                                &promisor_objects,
2116                                                FOR_EACH_OBJECT_PROMISOR_ONLY);
2117                 }
2118                 promisor_objects_prepared = 1;
2119         }
2120         return oidset_contains(&promisor_objects, oid);
2121 }