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