packfile: rename close_all_packs to close_object_store
[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->sha1, 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 static 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 /*
473  * Do not call this directly as this leaks p->pack_fd on error return;
474  * call open_packed_git() instead.
475  */
476 static int open_packed_git_1(struct packed_git *p)
477 {
478         struct stat st;
479         struct pack_header hdr;
480         unsigned char hash[GIT_MAX_RAWSZ];
481         unsigned char *idx_hash;
482         long fd_flag;
483         ssize_t read_result;
484         const unsigned hashsz = the_hash_algo->rawsz;
485
486         if (!p->index_data) {
487                 struct multi_pack_index *m;
488                 const char *pack_name = strrchr(p->pack_name, '/');
489
490                 for (m = the_repository->objects->multi_pack_index;
491                      m; m = m->next) {
492                         if (midx_contains_pack(m, pack_name))
493                                 break;
494                 }
495
496                 if (!m && open_pack_index(p))
497                         return error("packfile %s index unavailable", p->pack_name);
498         }
499
500         if (!pack_max_fds) {
501                 unsigned int max_fds = get_max_fd_limit();
502
503                 /* Save 3 for stdin/stdout/stderr, 22 for work */
504                 if (25 < max_fds)
505                         pack_max_fds = max_fds - 25;
506                 else
507                         pack_max_fds = 1;
508         }
509
510         while (pack_max_fds <= pack_open_fds && close_one_pack())
511                 ; /* nothing */
512
513         p->pack_fd = git_open(p->pack_name);
514         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
515                 return -1;
516         pack_open_fds++;
517
518         /* If we created the struct before we had the pack we lack size. */
519         if (!p->pack_size) {
520                 if (!S_ISREG(st.st_mode))
521                         return error("packfile %s not a regular file", p->pack_name);
522                 p->pack_size = st.st_size;
523         } else if (p->pack_size != st.st_size)
524                 return error("packfile %s size changed", p->pack_name);
525
526         /* We leave these file descriptors open with sliding mmap;
527          * there is no point keeping them open across exec(), though.
528          */
529         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
530         if (fd_flag < 0)
531                 return error("cannot determine file descriptor flags");
532         fd_flag |= FD_CLOEXEC;
533         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
534                 return error("cannot set FD_CLOEXEC");
535
536         /* Verify we recognize this pack file format. */
537         read_result = read_in_full(p->pack_fd, &hdr, sizeof(hdr));
538         if (read_result < 0)
539                 return error_errno("error reading from %s", p->pack_name);
540         if (read_result != sizeof(hdr))
541                 return error("file %s is far too short to be a packfile", p->pack_name);
542         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
543                 return error("file %s is not a GIT packfile", p->pack_name);
544         if (!pack_version_ok(hdr.hdr_version))
545                 return error("packfile %s is version %"PRIu32" and not"
546                         " supported (try upgrading GIT to a newer version)",
547                         p->pack_name, ntohl(hdr.hdr_version));
548
549         /* Skip index checking if in multi-pack-index */
550         if (!p->index_data)
551                 return 0;
552
553         /* Verify the pack matches its index. */
554         if (p->num_objects != ntohl(hdr.hdr_entries))
555                 return error("packfile %s claims to have %"PRIu32" objects"
556                              " while index indicates %"PRIu32" objects",
557                              p->pack_name, ntohl(hdr.hdr_entries),
558                              p->num_objects);
559         if (lseek(p->pack_fd, p->pack_size - hashsz, SEEK_SET) == -1)
560                 return error("end of packfile %s is unavailable", p->pack_name);
561         read_result = read_in_full(p->pack_fd, hash, hashsz);
562         if (read_result < 0)
563                 return error_errno("error reading from %s", p->pack_name);
564         if (read_result != hashsz)
565                 return error("packfile %s signature is unavailable", p->pack_name);
566         idx_hash = ((unsigned char *)p->index_data) + p->index_size - hashsz * 2;
567         if (!hasheq(hash, idx_hash))
568                 return error("packfile %s does not match index", p->pack_name);
569         return 0;
570 }
571
572 static int open_packed_git(struct packed_git *p)
573 {
574         if (!open_packed_git_1(p))
575                 return 0;
576         close_pack_fd(p);
577         return -1;
578 }
579
580 static int in_window(struct pack_window *win, off_t offset)
581 {
582         /* We must promise at least one full hash after the
583          * offset is available from this window, otherwise the offset
584          * is not actually in this window and a different window (which
585          * has that one hash excess) must be used.  This is to support
586          * the object header and delta base parsing routines below.
587          */
588         off_t win_off = win->offset;
589         return win_off <= offset
590                 && (offset + the_hash_algo->rawsz) <= (win_off + win->len);
591 }
592
593 unsigned char *use_pack(struct packed_git *p,
594                 struct pack_window **w_cursor,
595                 off_t offset,
596                 unsigned long *left)
597 {
598         struct pack_window *win = *w_cursor;
599
600         /* Since packfiles end in a hash of their content and it's
601          * pointless to ask for an offset into the middle of that
602          * hash, and the in_window function above wouldn't match
603          * don't allow an offset too close to the end of the file.
604          */
605         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
606                 die("packfile %s cannot be accessed", p->pack_name);
607         if (offset > (p->pack_size - the_hash_algo->rawsz))
608                 die("offset beyond end of packfile (truncated pack?)");
609         if (offset < 0)
610                 die(_("offset before end of packfile (broken .idx?)"));
611
612         if (!win || !in_window(win, offset)) {
613                 if (win)
614                         win->inuse_cnt--;
615                 for (win = p->windows; win; win = win->next) {
616                         if (in_window(win, offset))
617                                 break;
618                 }
619                 if (!win) {
620                         size_t window_align = packed_git_window_size / 2;
621                         off_t len;
622
623                         if (p->pack_fd == -1 && open_packed_git(p))
624                                 die("packfile %s cannot be accessed", p->pack_name);
625
626                         win = xcalloc(1, sizeof(*win));
627                         win->offset = (offset / window_align) * window_align;
628                         len = p->pack_size - win->offset;
629                         if (len > packed_git_window_size)
630                                 len = packed_git_window_size;
631                         win->len = (size_t)len;
632                         pack_mapped += win->len;
633                         while (packed_git_limit < pack_mapped
634                                 && unuse_one_window(p))
635                                 ; /* nothing */
636                         win->base = xmmap(NULL, win->len,
637                                 PROT_READ, MAP_PRIVATE,
638                                 p->pack_fd, win->offset);
639                         if (win->base == MAP_FAILED)
640                                 die_errno("packfile %s cannot be mapped",
641                                           p->pack_name);
642                         if (!win->offset && win->len == p->pack_size
643                                 && !p->do_not_close)
644                                 close_pack_fd(p);
645                         pack_mmap_calls++;
646                         pack_open_windows++;
647                         if (pack_mapped > peak_pack_mapped)
648                                 peak_pack_mapped = pack_mapped;
649                         if (pack_open_windows > peak_pack_open_windows)
650                                 peak_pack_open_windows = pack_open_windows;
651                         win->next = p->windows;
652                         p->windows = win;
653                 }
654         }
655         if (win != *w_cursor) {
656                 win->last_used = pack_used_ctr++;
657                 win->inuse_cnt++;
658                 *w_cursor = win;
659         }
660         offset -= win->offset;
661         if (left)
662                 *left = win->len - xsize_t(offset);
663         return win->base + offset;
664 }
665
666 void unuse_pack(struct pack_window **w_cursor)
667 {
668         struct pack_window *w = *w_cursor;
669         if (w) {
670                 w->inuse_cnt--;
671                 *w_cursor = NULL;
672         }
673 }
674
675 static void try_to_free_pack_memory(size_t size)
676 {
677         release_pack_memory(size);
678 }
679
680 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
681 {
682         static int have_set_try_to_free_routine;
683         struct stat st;
684         size_t alloc;
685         struct packed_git *p;
686
687         if (!have_set_try_to_free_routine) {
688                 have_set_try_to_free_routine = 1;
689                 set_try_to_free_routine(try_to_free_pack_memory);
690         }
691
692         /*
693          * Make sure a corresponding .pack file exists and that
694          * the index looks sane.
695          */
696         if (!strip_suffix_mem(path, &path_len, ".idx"))
697                 return NULL;
698
699         /*
700          * ".promisor" is long enough to hold any suffix we're adding (and
701          * the use xsnprintf double-checks that)
702          */
703         alloc = st_add3(path_len, strlen(".promisor"), 1);
704         p = alloc_packed_git(alloc);
705         memcpy(p->pack_name, path, path_len);
706
707         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
708         if (!access(p->pack_name, F_OK))
709                 p->pack_keep = 1;
710
711         xsnprintf(p->pack_name + path_len, alloc - path_len, ".promisor");
712         if (!access(p->pack_name, F_OK))
713                 p->pack_promisor = 1;
714
715         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
716         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
717                 free(p);
718                 return NULL;
719         }
720
721         /* ok, it looks sane as far as we can check without
722          * actually mapping the pack file.
723          */
724         p->pack_size = st.st_size;
725         p->pack_local = local;
726         p->mtime = st.st_mtime;
727         if (path_len < the_hash_algo->hexsz ||
728             get_sha1_hex(path + path_len - the_hash_algo->hexsz, p->sha1))
729                 hashclr(p->sha1);
730         return p;
731 }
732
733 void install_packed_git(struct repository *r, struct packed_git *pack)
734 {
735         if (pack->pack_fd != -1)
736                 pack_open_fds++;
737
738         pack->next = r->objects->packed_git;
739         r->objects->packed_git = pack;
740 }
741
742 void (*report_garbage)(unsigned seen_bits, const char *path);
743
744 static void report_helper(const struct string_list *list,
745                           int seen_bits, int first, int last)
746 {
747         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
748                 return;
749
750         for (; first < last; first++)
751                 report_garbage(seen_bits, list->items[first].string);
752 }
753
754 static void report_pack_garbage(struct string_list *list)
755 {
756         int i, baselen = -1, first = 0, seen_bits = 0;
757
758         if (!report_garbage)
759                 return;
760
761         string_list_sort(list);
762
763         for (i = 0; i < list->nr; i++) {
764                 const char *path = list->items[i].string;
765                 if (baselen != -1 &&
766                     strncmp(path, list->items[first].string, baselen)) {
767                         report_helper(list, seen_bits, first, i);
768                         baselen = -1;
769                         seen_bits = 0;
770                 }
771                 if (baselen == -1) {
772                         const char *dot = strrchr(path, '.');
773                         if (!dot) {
774                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
775                                 continue;
776                         }
777                         baselen = dot - path + 1;
778                         first = i;
779                 }
780                 if (!strcmp(path + baselen, "pack"))
781                         seen_bits |= 1;
782                 else if (!strcmp(path + baselen, "idx"))
783                         seen_bits |= 2;
784         }
785         report_helper(list, seen_bits, first, list->nr);
786 }
787
788 void for_each_file_in_pack_dir(const char *objdir,
789                                each_file_in_pack_dir_fn fn,
790                                void *data)
791 {
792         struct strbuf path = STRBUF_INIT;
793         size_t dirnamelen;
794         DIR *dir;
795         struct dirent *de;
796
797         strbuf_addstr(&path, objdir);
798         strbuf_addstr(&path, "/pack");
799         dir = opendir(path.buf);
800         if (!dir) {
801                 if (errno != ENOENT)
802                         error_errno("unable to open object pack directory: %s",
803                                     path.buf);
804                 strbuf_release(&path);
805                 return;
806         }
807         strbuf_addch(&path, '/');
808         dirnamelen = path.len;
809         while ((de = readdir(dir)) != NULL) {
810                 if (is_dot_or_dotdot(de->d_name))
811                         continue;
812
813                 strbuf_setlen(&path, dirnamelen);
814                 strbuf_addstr(&path, de->d_name);
815
816                 fn(path.buf, path.len, de->d_name, data);
817         }
818
819         closedir(dir);
820         strbuf_release(&path);
821 }
822
823 struct prepare_pack_data {
824         struct repository *r;
825         struct string_list *garbage;
826         int local;
827         struct multi_pack_index *m;
828 };
829
830 static void prepare_pack(const char *full_name, size_t full_name_len,
831                          const char *file_name, void *_data)
832 {
833         struct prepare_pack_data *data = (struct prepare_pack_data *)_data;
834         struct packed_git *p;
835         size_t base_len = full_name_len;
836
837         if (strip_suffix_mem(full_name, &base_len, ".idx") &&
838             !(data->m && midx_contains_pack(data->m, file_name))) {
839                 /* Don't reopen a pack we already have. */
840                 for (p = data->r->objects->packed_git; p; p = p->next) {
841                         size_t len;
842                         if (strip_suffix(p->pack_name, ".pack", &len) &&
843                             len == base_len &&
844                             !memcmp(p->pack_name, full_name, len))
845                                 break;
846                 }
847
848                 if (!p) {
849                         p = add_packed_git(full_name, full_name_len, data->local);
850                         if (p)
851                                 install_packed_git(data->r, p);
852                 }
853         }
854
855         if (!report_garbage)
856                 return;
857
858         if (!strcmp(file_name, "multi-pack-index"))
859                 return;
860         if (ends_with(file_name, ".idx") ||
861             ends_with(file_name, ".pack") ||
862             ends_with(file_name, ".bitmap") ||
863             ends_with(file_name, ".keep") ||
864             ends_with(file_name, ".promisor"))
865                 string_list_append(data->garbage, full_name);
866         else
867                 report_garbage(PACKDIR_FILE_GARBAGE, full_name);
868 }
869
870 static void prepare_packed_git_one(struct repository *r, char *objdir, int local)
871 {
872         struct prepare_pack_data data;
873         struct string_list garbage = STRING_LIST_INIT_DUP;
874
875         data.m = r->objects->multi_pack_index;
876
877         /* look for the multi-pack-index for this object directory */
878         while (data.m && strcmp(data.m->object_dir, objdir))
879                 data.m = data.m->next;
880
881         data.r = r;
882         data.garbage = &garbage;
883         data.local = local;
884
885         for_each_file_in_pack_dir(objdir, prepare_pack, &data);
886
887         report_pack_garbage(data.garbage);
888         string_list_clear(data.garbage, 0);
889 }
890
891 static void prepare_packed_git(struct repository *r);
892 /*
893  * Give a fast, rough count of the number of objects in the repository. This
894  * ignores loose objects completely. If you have a lot of them, then either
895  * you should repack because your performance will be awful, or they are
896  * all unreachable objects about to be pruned, in which case they're not really
897  * interesting as a measure of repo size in the first place.
898  */
899 unsigned long approximate_object_count(void)
900 {
901         if (!the_repository->objects->approximate_object_count_valid) {
902                 unsigned long count;
903                 struct multi_pack_index *m;
904                 struct packed_git *p;
905
906                 prepare_packed_git(the_repository);
907                 count = 0;
908                 for (m = get_multi_pack_index(the_repository); m; m = m->next)
909                         count += m->num_objects;
910                 for (p = the_repository->objects->packed_git; p; p = p->next) {
911                         if (open_pack_index(p))
912                                 continue;
913                         count += p->num_objects;
914                 }
915                 the_repository->objects->approximate_object_count = count;
916         }
917         return the_repository->objects->approximate_object_count;
918 }
919
920 static void *get_next_packed_git(const void *p)
921 {
922         return ((const struct packed_git *)p)->next;
923 }
924
925 static void set_next_packed_git(void *p, void *next)
926 {
927         ((struct packed_git *)p)->next = next;
928 }
929
930 static int sort_pack(const void *a_, const void *b_)
931 {
932         const struct packed_git *a = a_;
933         const struct packed_git *b = b_;
934         int st;
935
936         /*
937          * Local packs tend to contain objects specific to our
938          * variant of the project than remote ones.  In addition,
939          * remote ones could be on a network mounted filesystem.
940          * Favor local ones for these reasons.
941          */
942         st = a->pack_local - b->pack_local;
943         if (st)
944                 return -st;
945
946         /*
947          * Younger packs tend to contain more recent objects,
948          * and more recent objects tend to get accessed more
949          * often.
950          */
951         if (a->mtime < b->mtime)
952                 return 1;
953         else if (a->mtime == b->mtime)
954                 return 0;
955         return -1;
956 }
957
958 static void rearrange_packed_git(struct repository *r)
959 {
960         r->objects->packed_git = llist_mergesort(
961                 r->objects->packed_git, get_next_packed_git,
962                 set_next_packed_git, sort_pack);
963 }
964
965 static void prepare_packed_git_mru(struct repository *r)
966 {
967         struct packed_git *p;
968
969         INIT_LIST_HEAD(&r->objects->packed_git_mru);
970
971         for (p = r->objects->packed_git; p; p = p->next)
972                 list_add_tail(&p->mru, &r->objects->packed_git_mru);
973 }
974
975 static void prepare_packed_git(struct repository *r)
976 {
977         struct object_directory *odb;
978
979         if (r->objects->packed_git_initialized)
980                 return;
981
982         prepare_alt_odb(r);
983         for (odb = r->objects->odb; odb; odb = odb->next) {
984                 int local = (odb == r->objects->odb);
985                 prepare_multi_pack_index_one(r, odb->path, local);
986                 prepare_packed_git_one(r, odb->path, local);
987         }
988         rearrange_packed_git(r);
989
990         r->objects->all_packs = NULL;
991
992         prepare_packed_git_mru(r);
993         r->objects->packed_git_initialized = 1;
994 }
995
996 void reprepare_packed_git(struct repository *r)
997 {
998         struct object_directory *odb;
999
1000         for (odb = r->objects->odb; odb; odb = odb->next)
1001                 odb_clear_loose_cache(odb);
1002
1003         r->objects->approximate_object_count_valid = 0;
1004         r->objects->packed_git_initialized = 0;
1005         prepare_packed_git(r);
1006 }
1007
1008 struct packed_git *get_packed_git(struct repository *r)
1009 {
1010         prepare_packed_git(r);
1011         return r->objects->packed_git;
1012 }
1013
1014 struct multi_pack_index *get_multi_pack_index(struct repository *r)
1015 {
1016         prepare_packed_git(r);
1017         return r->objects->multi_pack_index;
1018 }
1019
1020 struct packed_git *get_all_packs(struct repository *r)
1021 {
1022         prepare_packed_git(r);
1023
1024         if (!r->objects->all_packs) {
1025                 struct packed_git *p = r->objects->packed_git;
1026                 struct multi_pack_index *m;
1027
1028                 for (m = r->objects->multi_pack_index; m; m = m->next) {
1029                         uint32_t i;
1030                         for (i = 0; i < m->num_packs; i++) {
1031                                 if (!prepare_midx_pack(m, i)) {
1032                                         m->packs[i]->next = p;
1033                                         p = m->packs[i];
1034                                 }
1035                         }
1036                 }
1037
1038                 r->objects->all_packs = p;
1039         }
1040
1041         return r->objects->all_packs;
1042 }
1043
1044 struct list_head *get_packed_git_mru(struct repository *r)
1045 {
1046         prepare_packed_git(r);
1047         return &r->objects->packed_git_mru;
1048 }
1049
1050 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1051                 unsigned long len, enum object_type *type, unsigned long *sizep)
1052 {
1053         unsigned shift;
1054         unsigned long size, c;
1055         unsigned long used = 0;
1056
1057         c = buf[used++];
1058         *type = (c >> 4) & 7;
1059         size = c & 15;
1060         shift = 4;
1061         while (c & 0x80) {
1062                 if (len <= used || bitsizeof(long) <= shift) {
1063                         error("bad object header");
1064                         size = used = 0;
1065                         break;
1066                 }
1067                 c = buf[used++];
1068                 size += (c & 0x7f) << shift;
1069                 shift += 7;
1070         }
1071         *sizep = size;
1072         return used;
1073 }
1074
1075 unsigned long get_size_from_delta(struct packed_git *p,
1076                                   struct pack_window **w_curs,
1077                                   off_t curpos)
1078 {
1079         const unsigned char *data;
1080         unsigned char delta_head[20], *in;
1081         git_zstream stream;
1082         int st;
1083
1084         memset(&stream, 0, sizeof(stream));
1085         stream.next_out = delta_head;
1086         stream.avail_out = sizeof(delta_head);
1087
1088         git_inflate_init(&stream);
1089         do {
1090                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1091                 stream.next_in = in;
1092                 st = git_inflate(&stream, Z_FINISH);
1093                 curpos += stream.next_in - in;
1094         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1095                  stream.total_out < sizeof(delta_head));
1096         git_inflate_end(&stream);
1097         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1098                 error("delta data unpack-initial failed");
1099                 return 0;
1100         }
1101
1102         /* Examine the initial part of the delta to figure out
1103          * the result size.
1104          */
1105         data = delta_head;
1106
1107         /* ignore base size */
1108         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1109
1110         /* Read the result size */
1111         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1112 }
1113
1114 int unpack_object_header(struct packed_git *p,
1115                          struct pack_window **w_curs,
1116                          off_t *curpos,
1117                          unsigned long *sizep)
1118 {
1119         unsigned char *base;
1120         unsigned long left;
1121         unsigned long used;
1122         enum object_type type;
1123
1124         /* use_pack() assures us we have [base, base + 20) available
1125          * as a range that we can look at.  (Its actually the hash
1126          * size that is assured.)  With our object header encoding
1127          * the maximum deflated object size is 2^137, which is just
1128          * insane, so we know won't exceed what we have been given.
1129          */
1130         base = use_pack(p, w_curs, *curpos, &left);
1131         used = unpack_object_header_buffer(base, left, &type, sizep);
1132         if (!used) {
1133                 type = OBJ_BAD;
1134         } else
1135                 *curpos += used;
1136
1137         return type;
1138 }
1139
1140 void mark_bad_packed_object(struct packed_git *p, const unsigned char *sha1)
1141 {
1142         unsigned i;
1143         const unsigned hashsz = the_hash_algo->rawsz;
1144         for (i = 0; i < p->num_bad_objects; i++)
1145                 if (hasheq(sha1, p->bad_object_sha1 + hashsz * i))
1146                         return;
1147         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1148                                       st_mult(GIT_MAX_RAWSZ,
1149                                               st_add(p->num_bad_objects, 1)));
1150         hashcpy(p->bad_object_sha1 + hashsz * p->num_bad_objects, sha1);
1151         p->num_bad_objects++;
1152 }
1153
1154 const struct packed_git *has_packed_and_bad(struct repository *r,
1155                                             const unsigned char *sha1)
1156 {
1157         struct packed_git *p;
1158         unsigned i;
1159
1160         for (p = r->objects->packed_git; p; p = p->next)
1161                 for (i = 0; i < p->num_bad_objects; i++)
1162                         if (hasheq(sha1,
1163                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
1164                                 return p;
1165         return NULL;
1166 }
1167
1168 static off_t get_delta_base(struct packed_git *p,
1169                                     struct pack_window **w_curs,
1170                                     off_t *curpos,
1171                                     enum object_type type,
1172                                     off_t delta_obj_offset)
1173 {
1174         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1175         off_t base_offset;
1176
1177         /* use_pack() assured us we have [base_info, base_info + 20)
1178          * as a range that we can look at without walking off the
1179          * end of the mapped window.  Its actually the hash size
1180          * that is assured.  An OFS_DELTA longer than the hash size
1181          * is stupid, as then a REF_DELTA would be smaller to store.
1182          */
1183         if (type == OBJ_OFS_DELTA) {
1184                 unsigned used = 0;
1185                 unsigned char c = base_info[used++];
1186                 base_offset = c & 127;
1187                 while (c & 128) {
1188                         base_offset += 1;
1189                         if (!base_offset || MSB(base_offset, 7))
1190                                 return 0;  /* overflow */
1191                         c = base_info[used++];
1192                         base_offset = (base_offset << 7) + (c & 127);
1193                 }
1194                 base_offset = delta_obj_offset - base_offset;
1195                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1196                         return 0;  /* out of bound */
1197                 *curpos += used;
1198         } else if (type == OBJ_REF_DELTA) {
1199                 /* The base entry _must_ be in the same pack */
1200                 base_offset = find_pack_entry_one(base_info, p);
1201                 *curpos += the_hash_algo->rawsz;
1202         } else
1203                 die("I am totally screwed");
1204         return base_offset;
1205 }
1206
1207 /*
1208  * Like get_delta_base above, but we return the sha1 instead of the pack
1209  * offset. This means it is cheaper for REF deltas (we do not have to do
1210  * the final object lookup), but more expensive for OFS deltas (we
1211  * have to load the revidx to convert the offset back into a sha1).
1212  */
1213 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1214                                                 struct pack_window **w_curs,
1215                                                 off_t curpos,
1216                                                 enum object_type type,
1217                                                 off_t delta_obj_offset)
1218 {
1219         if (type == OBJ_REF_DELTA) {
1220                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1221                 return base;
1222         } else if (type == OBJ_OFS_DELTA) {
1223                 struct revindex_entry *revidx;
1224                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1225                                                    type, delta_obj_offset);
1226
1227                 if (!base_offset)
1228                         return NULL;
1229
1230                 revidx = find_pack_revindex(p, base_offset);
1231                 if (!revidx)
1232                         return NULL;
1233
1234                 return nth_packed_object_sha1(p, revidx->nr);
1235         } else
1236                 return NULL;
1237 }
1238
1239 static int retry_bad_packed_offset(struct repository *r,
1240                                    struct packed_git *p,
1241                                    off_t obj_offset)
1242 {
1243         int type;
1244         struct revindex_entry *revidx;
1245         struct object_id oid;
1246         revidx = find_pack_revindex(p, obj_offset);
1247         if (!revidx)
1248                 return OBJ_BAD;
1249         nth_packed_object_oid(&oid, p, revidx->nr);
1250         mark_bad_packed_object(p, oid.hash);
1251         type = oid_object_info(r, &oid, NULL);
1252         if (type <= OBJ_NONE)
1253                 return OBJ_BAD;
1254         return type;
1255 }
1256
1257 #define POI_STACK_PREALLOC 64
1258
1259 static enum object_type packed_to_object_type(struct repository *r,
1260                                               struct packed_git *p,
1261                                               off_t obj_offset,
1262                                               enum object_type type,
1263                                               struct pack_window **w_curs,
1264                                               off_t curpos)
1265 {
1266         off_t small_poi_stack[POI_STACK_PREALLOC];
1267         off_t *poi_stack = small_poi_stack;
1268         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1269
1270         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1271                 off_t base_offset;
1272                 unsigned long size;
1273                 /* Push the object we're going to leave behind */
1274                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1275                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1276                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
1277                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1278                 } else {
1279                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1280                 }
1281                 poi_stack[poi_stack_nr++] = obj_offset;
1282                 /* If parsing the base offset fails, just unwind */
1283                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1284                 if (!base_offset)
1285                         goto unwind;
1286                 curpos = obj_offset = base_offset;
1287                 type = unpack_object_header(p, w_curs, &curpos, &size);
1288                 if (type <= OBJ_NONE) {
1289                         /* If getting the base itself fails, we first
1290                          * retry the base, otherwise unwind */
1291                         type = retry_bad_packed_offset(r, p, base_offset);
1292                         if (type > OBJ_NONE)
1293                                 goto out;
1294                         goto unwind;
1295                 }
1296         }
1297
1298         switch (type) {
1299         case OBJ_BAD:
1300         case OBJ_COMMIT:
1301         case OBJ_TREE:
1302         case OBJ_BLOB:
1303         case OBJ_TAG:
1304                 break;
1305         default:
1306                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1307                       type, (uintmax_t)obj_offset, p->pack_name);
1308                 type = OBJ_BAD;
1309         }
1310
1311 out:
1312         if (poi_stack != small_poi_stack)
1313                 free(poi_stack);
1314         return type;
1315
1316 unwind:
1317         while (poi_stack_nr) {
1318                 obj_offset = poi_stack[--poi_stack_nr];
1319                 type = retry_bad_packed_offset(r, p, obj_offset);
1320                 if (type > OBJ_NONE)
1321                         goto out;
1322         }
1323         type = OBJ_BAD;
1324         goto out;
1325 }
1326
1327 static struct hashmap delta_base_cache;
1328 static size_t delta_base_cached;
1329
1330 static LIST_HEAD(delta_base_cache_lru);
1331
1332 struct delta_base_cache_key {
1333         struct packed_git *p;
1334         off_t base_offset;
1335 };
1336
1337 struct delta_base_cache_entry {
1338         struct hashmap hash;
1339         struct delta_base_cache_key key;
1340         struct list_head lru;
1341         void *data;
1342         unsigned long size;
1343         enum object_type type;
1344 };
1345
1346 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
1347 {
1348         unsigned int hash;
1349
1350         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
1351         hash += (hash >> 8) + (hash >> 16);
1352         return hash;
1353 }
1354
1355 static struct delta_base_cache_entry *
1356 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
1357 {
1358         struct hashmap_entry entry;
1359         struct delta_base_cache_key key;
1360
1361         if (!delta_base_cache.cmpfn)
1362                 return NULL;
1363
1364         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
1365         key.p = p;
1366         key.base_offset = base_offset;
1367         return hashmap_get(&delta_base_cache, &entry, &key);
1368 }
1369
1370 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
1371                                    const struct delta_base_cache_key *b)
1372 {
1373         return a->p == b->p && a->base_offset == b->base_offset;
1374 }
1375
1376 static int delta_base_cache_hash_cmp(const void *unused_cmp_data,
1377                                      const void *va, const void *vb,
1378                                      const void *vkey)
1379 {
1380         const struct delta_base_cache_entry *a = va, *b = vb;
1381         const struct delta_base_cache_key *key = vkey;
1382         if (key)
1383                 return !delta_base_cache_key_eq(&a->key, key);
1384         else
1385                 return !delta_base_cache_key_eq(&a->key, &b->key);
1386 }
1387
1388 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
1389 {
1390         return !!get_delta_base_cache_entry(p, base_offset);
1391 }
1392
1393 /*
1394  * Remove the entry from the cache, but do _not_ free the associated
1395  * entry data. The caller takes ownership of the "data" buffer, and
1396  * should copy out any fields it wants before detaching.
1397  */
1398 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
1399 {
1400         hashmap_remove(&delta_base_cache, ent, &ent->key);
1401         list_del(&ent->lru);
1402         delta_base_cached -= ent->size;
1403         free(ent);
1404 }
1405
1406 static void *cache_or_unpack_entry(struct repository *r, struct packed_git *p,
1407                                    off_t base_offset, unsigned long *base_size,
1408                                    enum object_type *type)
1409 {
1410         struct delta_base_cache_entry *ent;
1411
1412         ent = get_delta_base_cache_entry(p, base_offset);
1413         if (!ent)
1414                 return unpack_entry(r, p, base_offset, type, base_size);
1415
1416         if (type)
1417                 *type = ent->type;
1418         if (base_size)
1419                 *base_size = ent->size;
1420         return xmemdupz(ent->data, ent->size);
1421 }
1422
1423 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1424 {
1425         free(ent->data);
1426         detach_delta_base_cache_entry(ent);
1427 }
1428
1429 void clear_delta_base_cache(void)
1430 {
1431         struct list_head *lru, *tmp;
1432         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1433                 struct delta_base_cache_entry *entry =
1434                         list_entry(lru, struct delta_base_cache_entry, lru);
1435                 release_delta_base_cache(entry);
1436         }
1437 }
1438
1439 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1440         void *base, unsigned long base_size, enum object_type type)
1441 {
1442         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
1443         struct list_head *lru, *tmp;
1444
1445         delta_base_cached += base_size;
1446
1447         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1448                 struct delta_base_cache_entry *f =
1449                         list_entry(lru, struct delta_base_cache_entry, lru);
1450                 if (delta_base_cached <= delta_base_cache_limit)
1451                         break;
1452                 release_delta_base_cache(f);
1453         }
1454
1455         ent->key.p = p;
1456         ent->key.base_offset = base_offset;
1457         ent->type = type;
1458         ent->data = base;
1459         ent->size = base_size;
1460         list_add_tail(&ent->lru, &delta_base_cache_lru);
1461
1462         if (!delta_base_cache.cmpfn)
1463                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, NULL, 0);
1464         hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
1465         hashmap_add(&delta_base_cache, ent);
1466 }
1467
1468 int packed_object_info(struct repository *r, struct packed_git *p,
1469                        off_t obj_offset, struct object_info *oi)
1470 {
1471         struct pack_window *w_curs = NULL;
1472         unsigned long size;
1473         off_t curpos = obj_offset;
1474         enum object_type type;
1475
1476         /*
1477          * We always get the representation type, but only convert it to
1478          * a "real" type later if the caller is interested.
1479          */
1480         if (oi->contentp) {
1481                 *oi->contentp = cache_or_unpack_entry(r, p, obj_offset, oi->sizep,
1482                                                       &type);
1483                 if (!*oi->contentp)
1484                         type = OBJ_BAD;
1485         } else {
1486                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1487         }
1488
1489         if (!oi->contentp && oi->sizep) {
1490                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1491                         off_t tmp_pos = curpos;
1492                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1493                                                            type, obj_offset);
1494                         if (!base_offset) {
1495                                 type = OBJ_BAD;
1496                                 goto out;
1497                         }
1498                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1499                         if (*oi->sizep == 0) {
1500                                 type = OBJ_BAD;
1501                                 goto out;
1502                         }
1503                 } else {
1504                         *oi->sizep = size;
1505                 }
1506         }
1507
1508         if (oi->disk_sizep) {
1509                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1510                 *oi->disk_sizep = revidx[1].offset - obj_offset;
1511         }
1512
1513         if (oi->typep || oi->type_name) {
1514                 enum object_type ptot;
1515                 ptot = packed_to_object_type(r, p, obj_offset,
1516                                              type, &w_curs, curpos);
1517                 if (oi->typep)
1518                         *oi->typep = ptot;
1519                 if (oi->type_name) {
1520                         const char *tn = type_name(ptot);
1521                         if (tn)
1522                                 strbuf_addstr(oi->type_name, tn);
1523                 }
1524                 if (ptot < 0) {
1525                         type = OBJ_BAD;
1526                         goto out;
1527                 }
1528         }
1529
1530         if (oi->delta_base_sha1) {
1531                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1532                         const unsigned char *base;
1533
1534                         base = get_delta_base_sha1(p, &w_curs, curpos,
1535                                                    type, obj_offset);
1536                         if (!base) {
1537                                 type = OBJ_BAD;
1538                                 goto out;
1539                         }
1540
1541                         hashcpy(oi->delta_base_sha1, base);
1542                 } else
1543                         hashclr(oi->delta_base_sha1);
1544         }
1545
1546         oi->whence = in_delta_base_cache(p, obj_offset) ? OI_DBCACHED :
1547                                                           OI_PACKED;
1548
1549 out:
1550         unuse_pack(&w_curs);
1551         return type;
1552 }
1553
1554 static void *unpack_compressed_entry(struct packed_git *p,
1555                                     struct pack_window **w_curs,
1556                                     off_t curpos,
1557                                     unsigned long size)
1558 {
1559         int st;
1560         git_zstream stream;
1561         unsigned char *buffer, *in;
1562
1563         buffer = xmallocz_gently(size);
1564         if (!buffer)
1565                 return NULL;
1566         memset(&stream, 0, sizeof(stream));
1567         stream.next_out = buffer;
1568         stream.avail_out = size + 1;
1569
1570         git_inflate_init(&stream);
1571         do {
1572                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1573                 stream.next_in = in;
1574                 st = git_inflate(&stream, Z_FINISH);
1575                 if (!stream.avail_out)
1576                         break; /* the payload is larger than it should be */
1577                 curpos += stream.next_in - in;
1578         } while (st == Z_OK || st == Z_BUF_ERROR);
1579         git_inflate_end(&stream);
1580         if ((st != Z_STREAM_END) || stream.total_out != size) {
1581                 free(buffer);
1582                 return NULL;
1583         }
1584
1585         /* versions of zlib can clobber unconsumed portion of outbuf */
1586         buffer[size] = '\0';
1587
1588         return buffer;
1589 }
1590
1591 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
1592 {
1593         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
1594         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
1595                          p->pack_name, (uintmax_t)obj_offset);
1596 }
1597
1598 int do_check_packed_object_crc;
1599
1600 #define UNPACK_ENTRY_STACK_PREALLOC 64
1601 struct unpack_entry_stack_ent {
1602         off_t obj_offset;
1603         off_t curpos;
1604         unsigned long size;
1605 };
1606
1607 static void *read_object(struct repository *r,
1608                          const struct object_id *oid,
1609                          enum object_type *type,
1610                          unsigned long *size)
1611 {
1612         struct object_info oi = OBJECT_INFO_INIT;
1613         void *content;
1614         oi.typep = type;
1615         oi.sizep = size;
1616         oi.contentp = &content;
1617
1618         if (oid_object_info_extended(r, oid, &oi, 0) < 0)
1619                 return NULL;
1620         return content;
1621 }
1622
1623 void *unpack_entry(struct repository *r, struct packed_git *p, off_t obj_offset,
1624                    enum object_type *final_type, unsigned long *final_size)
1625 {
1626         struct pack_window *w_curs = NULL;
1627         off_t curpos = obj_offset;
1628         void *data = NULL;
1629         unsigned long size;
1630         enum object_type type;
1631         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
1632         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
1633         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
1634         int base_from_cache = 0;
1635
1636         write_pack_access_log(p, obj_offset);
1637
1638         /* PHASE 1: drill down to the innermost base object */
1639         for (;;) {
1640                 off_t base_offset;
1641                 int i;
1642                 struct delta_base_cache_entry *ent;
1643
1644                 ent = get_delta_base_cache_entry(p, curpos);
1645                 if (ent) {
1646                         type = ent->type;
1647                         data = ent->data;
1648                         size = ent->size;
1649                         detach_delta_base_cache_entry(ent);
1650                         base_from_cache = 1;
1651                         break;
1652                 }
1653
1654                 if (do_check_packed_object_crc && p->index_version > 1) {
1655                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1656                         off_t len = revidx[1].offset - obj_offset;
1657                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1658                                 struct object_id oid;
1659                                 nth_packed_object_oid(&oid, p, revidx->nr);
1660                                 error("bad packed object CRC for %s",
1661                                       oid_to_hex(&oid));
1662                                 mark_bad_packed_object(p, oid.hash);
1663                                 data = NULL;
1664                                 goto out;
1665                         }
1666                 }
1667
1668                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1669                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
1670                         break;
1671
1672                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1673                 if (!base_offset) {
1674                         error("failed to validate delta base reference "
1675                               "at offset %"PRIuMAX" from %s",
1676                               (uintmax_t)curpos, p->pack_name);
1677                         /* bail to phase 2, in hopes of recovery */
1678                         data = NULL;
1679                         break;
1680                 }
1681
1682                 /* push object, proceed to base */
1683                 if (delta_stack_nr >= delta_stack_alloc
1684                     && delta_stack == small_delta_stack) {
1685                         delta_stack_alloc = alloc_nr(delta_stack_nr);
1686                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
1687                         memcpy(delta_stack, small_delta_stack,
1688                                sizeof(*delta_stack)*delta_stack_nr);
1689                 } else {
1690                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
1691                 }
1692                 i = delta_stack_nr++;
1693                 delta_stack[i].obj_offset = obj_offset;
1694                 delta_stack[i].curpos = curpos;
1695                 delta_stack[i].size = size;
1696
1697                 curpos = obj_offset = base_offset;
1698         }
1699
1700         /* PHASE 2: handle the base */
1701         switch (type) {
1702         case OBJ_OFS_DELTA:
1703         case OBJ_REF_DELTA:
1704                 if (data)
1705                         BUG("unpack_entry: left loop at a valid delta");
1706                 break;
1707         case OBJ_COMMIT:
1708         case OBJ_TREE:
1709         case OBJ_BLOB:
1710         case OBJ_TAG:
1711                 if (!base_from_cache)
1712                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
1713                 break;
1714         default:
1715                 data = NULL;
1716                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1717                       type, (uintmax_t)obj_offset, p->pack_name);
1718         }
1719
1720         /* PHASE 3: apply deltas in order */
1721
1722         /* invariants:
1723          *   'data' holds the base data, or NULL if there was corruption
1724          */
1725         while (delta_stack_nr) {
1726                 void *delta_data;
1727                 void *base = data;
1728                 void *external_base = NULL;
1729                 unsigned long delta_size, base_size = size;
1730                 int i;
1731
1732                 data = NULL;
1733
1734                 if (base)
1735                         add_delta_base_cache(p, obj_offset, base, base_size, type);
1736
1737                 if (!base) {
1738                         /*
1739                          * We're probably in deep shit, but let's try to fetch
1740                          * the required base anyway from another pack or loose.
1741                          * This is costly but should happen only in the presence
1742                          * of a corrupted pack, and is better than failing outright.
1743                          */
1744                         struct revindex_entry *revidx;
1745                         struct object_id base_oid;
1746                         revidx = find_pack_revindex(p, obj_offset);
1747                         if (revidx) {
1748                                 nth_packed_object_oid(&base_oid, p, revidx->nr);
1749                                 error("failed to read delta base object %s"
1750                                       " at offset %"PRIuMAX" from %s",
1751                                       oid_to_hex(&base_oid), (uintmax_t)obj_offset,
1752                                       p->pack_name);
1753                                 mark_bad_packed_object(p, base_oid.hash);
1754                                 base = read_object(r, &base_oid, &type, &base_size);
1755                                 external_base = base;
1756                         }
1757                 }
1758
1759                 i = --delta_stack_nr;
1760                 obj_offset = delta_stack[i].obj_offset;
1761                 curpos = delta_stack[i].curpos;
1762                 delta_size = delta_stack[i].size;
1763
1764                 if (!base)
1765                         continue;
1766
1767                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
1768
1769                 if (!delta_data) {
1770                         error("failed to unpack compressed delta "
1771                               "at offset %"PRIuMAX" from %s",
1772                               (uintmax_t)curpos, p->pack_name);
1773                         data = NULL;
1774                         free(external_base);
1775                         continue;
1776                 }
1777
1778                 data = patch_delta(base, base_size,
1779                                    delta_data, delta_size,
1780                                    &size);
1781
1782                 /*
1783                  * We could not apply the delta; warn the user, but keep going.
1784                  * Our failure will be noticed either in the next iteration of
1785                  * the loop, or if this is the final delta, in the caller when
1786                  * we return NULL. Those code paths will take care of making
1787                  * a more explicit warning and retrying with another copy of
1788                  * the object.
1789                  */
1790                 if (!data)
1791                         error("failed to apply delta");
1792
1793                 free(delta_data);
1794                 free(external_base);
1795         }
1796
1797         if (final_type)
1798                 *final_type = type;
1799         if (final_size)
1800                 *final_size = size;
1801
1802 out:
1803         unuse_pack(&w_curs);
1804
1805         if (delta_stack != small_delta_stack)
1806                 free(delta_stack);
1807
1808         return data;
1809 }
1810
1811 int bsearch_pack(const struct object_id *oid, const struct packed_git *p, uint32_t *result)
1812 {
1813         const unsigned char *index_fanout = p->index_data;
1814         const unsigned char *index_lookup;
1815         const unsigned int hashsz = the_hash_algo->rawsz;
1816         int index_lookup_width;
1817
1818         if (!index_fanout)
1819                 BUG("bsearch_pack called without a valid pack-index");
1820
1821         index_lookup = index_fanout + 4 * 256;
1822         if (p->index_version == 1) {
1823                 index_lookup_width = hashsz + 4;
1824                 index_lookup += 4;
1825         } else {
1826                 index_lookup_width = hashsz;
1827                 index_fanout += 8;
1828                 index_lookup += 8;
1829         }
1830
1831         return bsearch_hash(oid->hash, (const uint32_t*)index_fanout,
1832                             index_lookup, index_lookup_width, result);
1833 }
1834
1835 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1836                                             uint32_t n)
1837 {
1838         const unsigned char *index = p->index_data;
1839         const unsigned int hashsz = the_hash_algo->rawsz;
1840         if (!index) {
1841                 if (open_pack_index(p))
1842                         return NULL;
1843                 index = p->index_data;
1844         }
1845         if (n >= p->num_objects)
1846                 return NULL;
1847         index += 4 * 256;
1848         if (p->index_version == 1) {
1849                 return index + (hashsz + 4) * n + 4;
1850         } else {
1851                 index += 8;
1852                 return index + hashsz * n;
1853         }
1854 }
1855
1856 const struct object_id *nth_packed_object_oid(struct object_id *oid,
1857                                               struct packed_git *p,
1858                                               uint32_t n)
1859 {
1860         const unsigned char *hash = nth_packed_object_sha1(p, n);
1861         if (!hash)
1862                 return NULL;
1863         hashcpy(oid->hash, hash);
1864         return oid;
1865 }
1866
1867 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
1868 {
1869         const unsigned char *ptr = vptr;
1870         const unsigned char *start = p->index_data;
1871         const unsigned char *end = start + p->index_size;
1872         if (ptr < start)
1873                 die(_("offset before start of pack index for %s (corrupt index?)"),
1874                     p->pack_name);
1875         /* No need to check for underflow; .idx files must be at least 8 bytes */
1876         if (ptr >= end - 8)
1877                 die(_("offset beyond end of pack index for %s (truncated index?)"),
1878                     p->pack_name);
1879 }
1880
1881 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1882 {
1883         const unsigned char *index = p->index_data;
1884         const unsigned int hashsz = the_hash_algo->rawsz;
1885         index += 4 * 256;
1886         if (p->index_version == 1) {
1887                 return ntohl(*((uint32_t *)(index + (hashsz + 4) * n)));
1888         } else {
1889                 uint32_t off;
1890                 index += 8 + p->num_objects * (hashsz + 4);
1891                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1892                 if (!(off & 0x80000000))
1893                         return off;
1894                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1895                 check_pack_index_ptr(p, index);
1896                 return get_be64(index);
1897         }
1898 }
1899
1900 off_t find_pack_entry_one(const unsigned char *sha1,
1901                                   struct packed_git *p)
1902 {
1903         const unsigned char *index = p->index_data;
1904         struct object_id oid;
1905         uint32_t result;
1906
1907         if (!index) {
1908                 if (open_pack_index(p))
1909                         return 0;
1910         }
1911
1912         hashcpy(oid.hash, sha1);
1913         if (bsearch_pack(&oid, p, &result))
1914                 return nth_packed_object_offset(p, result);
1915         return 0;
1916 }
1917
1918 int is_pack_valid(struct packed_git *p)
1919 {
1920         /* An already open pack is known to be valid. */
1921         if (p->pack_fd != -1)
1922                 return 1;
1923
1924         /* If the pack has one window completely covering the
1925          * file size, the pack is known to be valid even if
1926          * the descriptor is not currently open.
1927          */
1928         if (p->windows) {
1929                 struct pack_window *w = p->windows;
1930
1931                 if (!w->offset && w->len == p->pack_size)
1932                         return 1;
1933         }
1934
1935         /* Force the pack to open to prove its valid. */
1936         return !open_packed_git(p);
1937 }
1938
1939 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1940                                   struct packed_git *packs)
1941 {
1942         struct packed_git *p;
1943
1944         for (p = packs; p; p = p->next) {
1945                 if (find_pack_entry_one(sha1, p))
1946                         return p;
1947         }
1948         return NULL;
1949
1950 }
1951
1952 static int fill_pack_entry(const struct object_id *oid,
1953                            struct pack_entry *e,
1954                            struct packed_git *p)
1955 {
1956         off_t offset;
1957
1958         if (p->num_bad_objects) {
1959                 unsigned i;
1960                 for (i = 0; i < p->num_bad_objects; i++)
1961                         if (hasheq(oid->hash,
1962                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
1963                                 return 0;
1964         }
1965
1966         offset = find_pack_entry_one(oid->hash, p);
1967         if (!offset)
1968                 return 0;
1969
1970         /*
1971          * We are about to tell the caller where they can locate the
1972          * requested object.  We better make sure the packfile is
1973          * still here and can be accessed before supplying that
1974          * answer, as it may have been deleted since the index was
1975          * loaded!
1976          */
1977         if (!is_pack_valid(p))
1978                 return 0;
1979         e->offset = offset;
1980         e->p = p;
1981         return 1;
1982 }
1983
1984 int find_pack_entry(struct repository *r, const struct object_id *oid, struct pack_entry *e)
1985 {
1986         struct list_head *pos;
1987         struct multi_pack_index *m;
1988
1989         prepare_packed_git(r);
1990         if (!r->objects->packed_git && !r->objects->multi_pack_index)
1991                 return 0;
1992
1993         for (m = r->objects->multi_pack_index; m; m = m->next) {
1994                 if (fill_midx_entry(oid, e, m))
1995                         return 1;
1996         }
1997
1998         list_for_each(pos, &r->objects->packed_git_mru) {
1999                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
2000                 if (fill_pack_entry(oid, e, p)) {
2001                         list_move(&p->mru, &r->objects->packed_git_mru);
2002                         return 1;
2003                 }
2004         }
2005         return 0;
2006 }
2007
2008 int has_object_pack(const struct object_id *oid)
2009 {
2010         struct pack_entry e;
2011         return find_pack_entry(the_repository, oid, &e);
2012 }
2013
2014 int has_pack_index(const unsigned char *sha1)
2015 {
2016         struct stat st;
2017         if (stat(sha1_pack_index_name(sha1), &st))
2018                 return 0;
2019         return 1;
2020 }
2021
2022 int for_each_object_in_pack(struct packed_git *p,
2023                             each_packed_object_fn cb, void *data,
2024                             enum for_each_object_flags flags)
2025 {
2026         uint32_t i;
2027         int r = 0;
2028
2029         if (flags & FOR_EACH_OBJECT_PACK_ORDER)
2030                 load_pack_revindex(p);
2031
2032         for (i = 0; i < p->num_objects; i++) {
2033                 uint32_t pos;
2034                 struct object_id oid;
2035
2036                 if (flags & FOR_EACH_OBJECT_PACK_ORDER)
2037                         pos = p->revindex[i].nr;
2038                 else
2039                         pos = i;
2040
2041                 if (!nth_packed_object_oid(&oid, p, pos))
2042                         return error("unable to get sha1 of object %u in %s",
2043                                      pos, p->pack_name);
2044
2045                 r = cb(&oid, p, pos, data);
2046                 if (r)
2047                         break;
2048         }
2049         return r;
2050 }
2051
2052 int for_each_packed_object(each_packed_object_fn cb, void *data,
2053                            enum for_each_object_flags flags)
2054 {
2055         struct packed_git *p;
2056         int r = 0;
2057         int pack_errors = 0;
2058
2059         prepare_packed_git(the_repository);
2060         for (p = get_all_packs(the_repository); p; p = p->next) {
2061                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
2062                         continue;
2063                 if ((flags & FOR_EACH_OBJECT_PROMISOR_ONLY) &&
2064                     !p->pack_promisor)
2065                         continue;
2066                 if (open_pack_index(p)) {
2067                         pack_errors = 1;
2068                         continue;
2069                 }
2070                 r = for_each_object_in_pack(p, cb, data, flags);
2071                 if (r)
2072                         break;
2073         }
2074         return r ? r : pack_errors;
2075 }
2076
2077 static int add_promisor_object(const struct object_id *oid,
2078                                struct packed_git *pack,
2079                                uint32_t pos,
2080                                void *set_)
2081 {
2082         struct oidset *set = set_;
2083         struct object *obj = parse_object(the_repository, oid);
2084         if (!obj)
2085                 return 1;
2086
2087         oidset_insert(set, oid);
2088
2089         /*
2090          * If this is a tree, commit, or tag, the objects it refers
2091          * to are also promisor objects. (Blobs refer to no objects->)
2092          */
2093         if (obj->type == OBJ_TREE) {
2094                 struct tree *tree = (struct tree *)obj;
2095                 struct tree_desc desc;
2096                 struct name_entry entry;
2097                 if (init_tree_desc_gently(&desc, tree->buffer, tree->size))
2098                         /*
2099                          * Error messages are given when packs are
2100                          * verified, so do not print any here.
2101                          */
2102                         return 0;
2103                 while (tree_entry_gently(&desc, &entry))
2104                         oidset_insert(set, &entry.oid);
2105         } else if (obj->type == OBJ_COMMIT) {
2106                 struct commit *commit = (struct commit *) obj;
2107                 struct commit_list *parents = commit->parents;
2108
2109                 oidset_insert(set, get_commit_tree_oid(commit));
2110                 for (; parents; parents = parents->next)
2111                         oidset_insert(set, &parents->item->object.oid);
2112         } else if (obj->type == OBJ_TAG) {
2113                 struct tag *tag = (struct tag *) obj;
2114                 oidset_insert(set, &tag->tagged->oid);
2115         }
2116         return 0;
2117 }
2118
2119 int is_promisor_object(const struct object_id *oid)
2120 {
2121         static struct oidset promisor_objects;
2122         static int promisor_objects_prepared;
2123
2124         if (!promisor_objects_prepared) {
2125                 if (repository_format_partial_clone) {
2126                         for_each_packed_object(add_promisor_object,
2127                                                &promisor_objects,
2128                                                FOR_EACH_OBJECT_PROMISOR_ONLY);
2129                 }
2130                 promisor_objects_prepared = 1;
2131         }
2132         return oidset_contains(&promisor_objects, oid);
2133 }