pack-objects: break out of want_object loop early
[git] / builtin / pack-objects.c
1 #include "builtin.h"
2 #include "cache.h"
3 #include "attr.h"
4 #include "object.h"
5 #include "blob.h"
6 #include "commit.h"
7 #include "tag.h"
8 #include "tree.h"
9 #include "delta.h"
10 #include "pack.h"
11 #include "pack-revindex.h"
12 #include "csum-file.h"
13 #include "tree-walk.h"
14 #include "diff.h"
15 #include "revision.h"
16 #include "list-objects.h"
17 #include "pack-objects.h"
18 #include "progress.h"
19 #include "refs.h"
20 #include "streaming.h"
21 #include "thread-utils.h"
22 #include "pack-bitmap.h"
23 #include "reachable.h"
24 #include "sha1-array.h"
25 #include "argv-array.h"
26
27 static const char *pack_usage[] = {
28         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
29         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
30         NULL
31 };
32
33 /*
34  * Objects we are going to pack are collected in the `to_pack` structure.
35  * It contains an array (dynamically expanded) of the object data, and a map
36  * that can resolve SHA1s to their position in the array.
37  */
38 static struct packing_data to_pack;
39
40 static struct pack_idx_entry **written_list;
41 static uint32_t nr_result, nr_written;
42
43 static int non_empty;
44 static int reuse_delta = 1, reuse_object = 1;
45 static int keep_unreachable, unpack_unreachable, include_tag;
46 static unsigned long unpack_unreachable_expiration;
47 static int pack_loose_unreachable;
48 static int local;
49 static int incremental;
50 static int ignore_packed_keep;
51 static int allow_ofs_delta;
52 static struct pack_idx_option pack_idx_opts;
53 static const char *base_name;
54 static int progress = 1;
55 static int window = 10;
56 static unsigned long pack_size_limit;
57 static int depth = 50;
58 static int delta_search_threads;
59 static int pack_to_stdout;
60 static int num_preferred_base;
61 static struct progress *progress_state;
62 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
63 static int pack_compression_seen;
64
65 static struct packed_git *reuse_packfile;
66 static uint32_t reuse_packfile_objects;
67 static off_t reuse_packfile_offset;
68
69 static int use_bitmap_index = 1;
70 static int write_bitmap_index;
71 static uint16_t write_bitmap_options;
72
73 static unsigned long delta_cache_size = 0;
74 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
75 static unsigned long cache_max_small_delta_size = 1000;
76
77 static unsigned long window_memory_limit = 0;
78
79 /*
80  * stats
81  */
82 static uint32_t written, written_delta;
83 static uint32_t reused, reused_delta;
84
85 /*
86  * Indexed commits
87  */
88 static struct commit **indexed_commits;
89 static unsigned int indexed_commits_nr;
90 static unsigned int indexed_commits_alloc;
91
92 static void index_commit_for_bitmap(struct commit *commit)
93 {
94         if (indexed_commits_nr >= indexed_commits_alloc) {
95                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
96                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
97         }
98
99         indexed_commits[indexed_commits_nr++] = commit;
100 }
101
102 static void *get_delta(struct object_entry *entry)
103 {
104         unsigned long size, base_size, delta_size;
105         void *buf, *base_buf, *delta_buf;
106         enum object_type type;
107
108         buf = read_sha1_file(entry->idx.sha1, &type, &size);
109         if (!buf)
110                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
111         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
112         if (!base_buf)
113                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
114         delta_buf = diff_delta(base_buf, base_size,
115                                buf, size, &delta_size, 0);
116         if (!delta_buf || delta_size != entry->delta_size)
117                 die("delta size changed");
118         free(buf);
119         free(base_buf);
120         return delta_buf;
121 }
122
123 static unsigned long do_compress(void **pptr, unsigned long size)
124 {
125         git_zstream stream;
126         void *in, *out;
127         unsigned long maxsize;
128
129         git_deflate_init(&stream, pack_compression_level);
130         maxsize = git_deflate_bound(&stream, size);
131
132         in = *pptr;
133         out = xmalloc(maxsize);
134         *pptr = out;
135
136         stream.next_in = in;
137         stream.avail_in = size;
138         stream.next_out = out;
139         stream.avail_out = maxsize;
140         while (git_deflate(&stream, Z_FINISH) == Z_OK)
141                 ; /* nothing */
142         git_deflate_end(&stream);
143
144         free(in);
145         return stream.total_out;
146 }
147
148 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
149                                            const unsigned char *sha1)
150 {
151         git_zstream stream;
152         unsigned char ibuf[1024 * 16];
153         unsigned char obuf[1024 * 16];
154         unsigned long olen = 0;
155
156         git_deflate_init(&stream, pack_compression_level);
157
158         for (;;) {
159                 ssize_t readlen;
160                 int zret = Z_OK;
161                 readlen = read_istream(st, ibuf, sizeof(ibuf));
162                 if (readlen == -1)
163                         die(_("unable to read %s"), sha1_to_hex(sha1));
164
165                 stream.next_in = ibuf;
166                 stream.avail_in = readlen;
167                 while ((stream.avail_in || readlen == 0) &&
168                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
169                         stream.next_out = obuf;
170                         stream.avail_out = sizeof(obuf);
171                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
172                         sha1write(f, obuf, stream.next_out - obuf);
173                         olen += stream.next_out - obuf;
174                 }
175                 if (stream.avail_in)
176                         die(_("deflate error (%d)"), zret);
177                 if (readlen == 0) {
178                         if (zret != Z_STREAM_END)
179                                 die(_("deflate error (%d)"), zret);
180                         break;
181                 }
182         }
183         git_deflate_end(&stream);
184         return olen;
185 }
186
187 /*
188  * we are going to reuse the existing object data as is.  make
189  * sure it is not corrupt.
190  */
191 static int check_pack_inflate(struct packed_git *p,
192                 struct pack_window **w_curs,
193                 off_t offset,
194                 off_t len,
195                 unsigned long expect)
196 {
197         git_zstream stream;
198         unsigned char fakebuf[4096], *in;
199         int st;
200
201         memset(&stream, 0, sizeof(stream));
202         git_inflate_init(&stream);
203         do {
204                 in = use_pack(p, w_curs, offset, &stream.avail_in);
205                 stream.next_in = in;
206                 stream.next_out = fakebuf;
207                 stream.avail_out = sizeof(fakebuf);
208                 st = git_inflate(&stream, Z_FINISH);
209                 offset += stream.next_in - in;
210         } while (st == Z_OK || st == Z_BUF_ERROR);
211         git_inflate_end(&stream);
212         return (st == Z_STREAM_END &&
213                 stream.total_out == expect &&
214                 stream.total_in == len) ? 0 : -1;
215 }
216
217 static void copy_pack_data(struct sha1file *f,
218                 struct packed_git *p,
219                 struct pack_window **w_curs,
220                 off_t offset,
221                 off_t len)
222 {
223         unsigned char *in;
224         unsigned long avail;
225
226         while (len) {
227                 in = use_pack(p, w_curs, offset, &avail);
228                 if (avail > len)
229                         avail = (unsigned long)len;
230                 sha1write(f, in, avail);
231                 offset += avail;
232                 len -= avail;
233         }
234 }
235
236 /* Return 0 if we will bust the pack-size limit */
237 static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
238                                            unsigned long limit, int usable_delta)
239 {
240         unsigned long size, datalen;
241         unsigned char header[10], dheader[10];
242         unsigned hdrlen;
243         enum object_type type;
244         void *buf;
245         struct git_istream *st = NULL;
246
247         if (!usable_delta) {
248                 if (entry->type == OBJ_BLOB &&
249                     entry->size > big_file_threshold &&
250                     (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
251                         buf = NULL;
252                 else {
253                         buf = read_sha1_file(entry->idx.sha1, &type, &size);
254                         if (!buf)
255                                 die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
256                 }
257                 /*
258                  * make sure no cached delta data remains from a
259                  * previous attempt before a pack split occurred.
260                  */
261                 free(entry->delta_data);
262                 entry->delta_data = NULL;
263                 entry->z_delta_size = 0;
264         } else if (entry->delta_data) {
265                 size = entry->delta_size;
266                 buf = entry->delta_data;
267                 entry->delta_data = NULL;
268                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
269                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
270         } else {
271                 buf = get_delta(entry);
272                 size = entry->delta_size;
273                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
274                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
275         }
276
277         if (st) /* large blob case, just assume we don't compress well */
278                 datalen = size;
279         else if (entry->z_delta_size)
280                 datalen = entry->z_delta_size;
281         else
282                 datalen = do_compress(&buf, size);
283
284         /*
285          * The object header is a byte of 'type' followed by zero or
286          * more bytes of length.
287          */
288         hdrlen = encode_in_pack_object_header(type, size, header);
289
290         if (type == OBJ_OFS_DELTA) {
291                 /*
292                  * Deltas with relative base contain an additional
293                  * encoding of the relative offset for the delta
294                  * base from this object's position in the pack.
295                  */
296                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
297                 unsigned pos = sizeof(dheader) - 1;
298                 dheader[pos] = ofs & 127;
299                 while (ofs >>= 7)
300                         dheader[--pos] = 128 | (--ofs & 127);
301                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
302                         if (st)
303                                 close_istream(st);
304                         free(buf);
305                         return 0;
306                 }
307                 sha1write(f, header, hdrlen);
308                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
309                 hdrlen += sizeof(dheader) - pos;
310         } else if (type == OBJ_REF_DELTA) {
311                 /*
312                  * Deltas with a base reference contain
313                  * an additional 20 bytes for the base sha1.
314                  */
315                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
316                         if (st)
317                                 close_istream(st);
318                         free(buf);
319                         return 0;
320                 }
321                 sha1write(f, header, hdrlen);
322                 sha1write(f, entry->delta->idx.sha1, 20);
323                 hdrlen += 20;
324         } else {
325                 if (limit && hdrlen + datalen + 20 >= limit) {
326                         if (st)
327                                 close_istream(st);
328                         free(buf);
329                         return 0;
330                 }
331                 sha1write(f, header, hdrlen);
332         }
333         if (st) {
334                 datalen = write_large_blob_data(st, f, entry->idx.sha1);
335                 close_istream(st);
336         } else {
337                 sha1write(f, buf, datalen);
338                 free(buf);
339         }
340
341         return hdrlen + datalen;
342 }
343
344 /* Return 0 if we will bust the pack-size limit */
345 static unsigned long write_reuse_object(struct sha1file *f, struct object_entry *entry,
346                                         unsigned long limit, int usable_delta)
347 {
348         struct packed_git *p = entry->in_pack;
349         struct pack_window *w_curs = NULL;
350         struct revindex_entry *revidx;
351         off_t offset;
352         enum object_type type = entry->type;
353         unsigned long datalen;
354         unsigned char header[10], dheader[10];
355         unsigned hdrlen;
356
357         if (entry->delta)
358                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
359                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
360         hdrlen = encode_in_pack_object_header(type, entry->size, header);
361
362         offset = entry->in_pack_offset;
363         revidx = find_pack_revindex(p, offset);
364         datalen = revidx[1].offset - offset;
365         if (!pack_to_stdout && p->index_version > 1 &&
366             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
367                 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
368                 unuse_pack(&w_curs);
369                 return write_no_reuse_object(f, entry, limit, usable_delta);
370         }
371
372         offset += entry->in_pack_header_size;
373         datalen -= entry->in_pack_header_size;
374
375         if (!pack_to_stdout && p->index_version == 1 &&
376             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
377                 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
378                 unuse_pack(&w_curs);
379                 return write_no_reuse_object(f, entry, limit, usable_delta);
380         }
381
382         if (type == OBJ_OFS_DELTA) {
383                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
384                 unsigned pos = sizeof(dheader) - 1;
385                 dheader[pos] = ofs & 127;
386                 while (ofs >>= 7)
387                         dheader[--pos] = 128 | (--ofs & 127);
388                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
389                         unuse_pack(&w_curs);
390                         return 0;
391                 }
392                 sha1write(f, header, hdrlen);
393                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
394                 hdrlen += sizeof(dheader) - pos;
395                 reused_delta++;
396         } else if (type == OBJ_REF_DELTA) {
397                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
398                         unuse_pack(&w_curs);
399                         return 0;
400                 }
401                 sha1write(f, header, hdrlen);
402                 sha1write(f, entry->delta->idx.sha1, 20);
403                 hdrlen += 20;
404                 reused_delta++;
405         } else {
406                 if (limit && hdrlen + datalen + 20 >= limit) {
407                         unuse_pack(&w_curs);
408                         return 0;
409                 }
410                 sha1write(f, header, hdrlen);
411         }
412         copy_pack_data(f, p, &w_curs, offset, datalen);
413         unuse_pack(&w_curs);
414         reused++;
415         return hdrlen + datalen;
416 }
417
418 /* Return 0 if we will bust the pack-size limit */
419 static unsigned long write_object(struct sha1file *f,
420                                   struct object_entry *entry,
421                                   off_t write_offset)
422 {
423         unsigned long limit, len;
424         int usable_delta, to_reuse;
425
426         if (!pack_to_stdout)
427                 crc32_begin(f);
428
429         /* apply size limit if limited packsize and not first object */
430         if (!pack_size_limit || !nr_written)
431                 limit = 0;
432         else if (pack_size_limit <= write_offset)
433                 /*
434                  * the earlier object did not fit the limit; avoid
435                  * mistaking this with unlimited (i.e. limit = 0).
436                  */
437                 limit = 1;
438         else
439                 limit = pack_size_limit - write_offset;
440
441         if (!entry->delta)
442                 usable_delta = 0;       /* no delta */
443         else if (!pack_size_limit)
444                usable_delta = 1;        /* unlimited packfile */
445         else if (entry->delta->idx.offset == (off_t)-1)
446                 usable_delta = 0;       /* base was written to another pack */
447         else if (entry->delta->idx.offset)
448                 usable_delta = 1;       /* base already exists in this pack */
449         else
450                 usable_delta = 0;       /* base could end up in another pack */
451
452         if (!reuse_object)
453                 to_reuse = 0;   /* explicit */
454         else if (!entry->in_pack)
455                 to_reuse = 0;   /* can't reuse what we don't have */
456         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
457                                 /* check_object() decided it for us ... */
458                 to_reuse = usable_delta;
459                                 /* ... but pack split may override that */
460         else if (entry->type != entry->in_pack_type)
461                 to_reuse = 0;   /* pack has delta which is unusable */
462         else if (entry->delta)
463                 to_reuse = 0;   /* we want to pack afresh */
464         else
465                 to_reuse = 1;   /* we have it in-pack undeltified,
466                                  * and we do not need to deltify it.
467                                  */
468
469         if (!to_reuse)
470                 len = write_no_reuse_object(f, entry, limit, usable_delta);
471         else
472                 len = write_reuse_object(f, entry, limit, usable_delta);
473         if (!len)
474                 return 0;
475
476         if (usable_delta)
477                 written_delta++;
478         written++;
479         if (!pack_to_stdout)
480                 entry->idx.crc32 = crc32_end(f);
481         return len;
482 }
483
484 enum write_one_status {
485         WRITE_ONE_SKIP = -1, /* already written */
486         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
487         WRITE_ONE_WRITTEN = 1, /* normal */
488         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
489 };
490
491 static enum write_one_status write_one(struct sha1file *f,
492                                        struct object_entry *e,
493                                        off_t *offset)
494 {
495         unsigned long size;
496         int recursing;
497
498         /*
499          * we set offset to 1 (which is an impossible value) to mark
500          * the fact that this object is involved in "write its base
501          * first before writing a deltified object" recursion.
502          */
503         recursing = (e->idx.offset == 1);
504         if (recursing) {
505                 warning("recursive delta detected for object %s",
506                         sha1_to_hex(e->idx.sha1));
507                 return WRITE_ONE_RECURSIVE;
508         } else if (e->idx.offset || e->preferred_base) {
509                 /* offset is non zero if object is written already. */
510                 return WRITE_ONE_SKIP;
511         }
512
513         /* if we are deltified, write out base object first. */
514         if (e->delta) {
515                 e->idx.offset = 1; /* now recurse */
516                 switch (write_one(f, e->delta, offset)) {
517                 case WRITE_ONE_RECURSIVE:
518                         /* we cannot depend on this one */
519                         e->delta = NULL;
520                         break;
521                 default:
522                         break;
523                 case WRITE_ONE_BREAK:
524                         e->idx.offset = recursing;
525                         return WRITE_ONE_BREAK;
526                 }
527         }
528
529         e->idx.offset = *offset;
530         size = write_object(f, e, *offset);
531         if (!size) {
532                 e->idx.offset = recursing;
533                 return WRITE_ONE_BREAK;
534         }
535         written_list[nr_written++] = &e->idx;
536
537         /* make sure off_t is sufficiently large not to wrap */
538         if (signed_add_overflows(*offset, size))
539                 die("pack too large for current definition of off_t");
540         *offset += size;
541         return WRITE_ONE_WRITTEN;
542 }
543
544 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
545                        void *cb_data)
546 {
547         unsigned char peeled[20];
548         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
549
550         if (entry)
551                 entry->tagged = 1;
552         if (!peel_ref(path, peeled)) {
553                 entry = packlist_find(&to_pack, peeled, NULL);
554                 if (entry)
555                         entry->tagged = 1;
556         }
557         return 0;
558 }
559
560 static inline void add_to_write_order(struct object_entry **wo,
561                                unsigned int *endp,
562                                struct object_entry *e)
563 {
564         if (e->filled)
565                 return;
566         wo[(*endp)++] = e;
567         e->filled = 1;
568 }
569
570 static void add_descendants_to_write_order(struct object_entry **wo,
571                                            unsigned int *endp,
572                                            struct object_entry *e)
573 {
574         int add_to_order = 1;
575         while (e) {
576                 if (add_to_order) {
577                         struct object_entry *s;
578                         /* add this node... */
579                         add_to_write_order(wo, endp, e);
580                         /* all its siblings... */
581                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
582                                 add_to_write_order(wo, endp, s);
583                         }
584                 }
585                 /* drop down a level to add left subtree nodes if possible */
586                 if (e->delta_child) {
587                         add_to_order = 1;
588                         e = e->delta_child;
589                 } else {
590                         add_to_order = 0;
591                         /* our sibling might have some children, it is next */
592                         if (e->delta_sibling) {
593                                 e = e->delta_sibling;
594                                 continue;
595                         }
596                         /* go back to our parent node */
597                         e = e->delta;
598                         while (e && !e->delta_sibling) {
599                                 /* we're on the right side of a subtree, keep
600                                  * going up until we can go right again */
601                                 e = e->delta;
602                         }
603                         if (!e) {
604                                 /* done- we hit our original root node */
605                                 return;
606                         }
607                         /* pass it off to sibling at this level */
608                         e = e->delta_sibling;
609                 }
610         };
611 }
612
613 static void add_family_to_write_order(struct object_entry **wo,
614                                       unsigned int *endp,
615                                       struct object_entry *e)
616 {
617         struct object_entry *root;
618
619         for (root = e; root->delta; root = root->delta)
620                 ; /* nothing */
621         add_descendants_to_write_order(wo, endp, root);
622 }
623
624 static struct object_entry **compute_write_order(void)
625 {
626         unsigned int i, wo_end, last_untagged;
627
628         struct object_entry **wo;
629         struct object_entry *objects = to_pack.objects;
630
631         for (i = 0; i < to_pack.nr_objects; i++) {
632                 objects[i].tagged = 0;
633                 objects[i].filled = 0;
634                 objects[i].delta_child = NULL;
635                 objects[i].delta_sibling = NULL;
636         }
637
638         /*
639          * Fully connect delta_child/delta_sibling network.
640          * Make sure delta_sibling is sorted in the original
641          * recency order.
642          */
643         for (i = to_pack.nr_objects; i > 0;) {
644                 struct object_entry *e = &objects[--i];
645                 if (!e->delta)
646                         continue;
647                 /* Mark me as the first child */
648                 e->delta_sibling = e->delta->delta_child;
649                 e->delta->delta_child = e;
650         }
651
652         /*
653          * Mark objects that are at the tip of tags.
654          */
655         for_each_tag_ref(mark_tagged, NULL);
656
657         /*
658          * Give the objects in the original recency order until
659          * we see a tagged tip.
660          */
661         ALLOC_ARRAY(wo, to_pack.nr_objects);
662         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
663                 if (objects[i].tagged)
664                         break;
665                 add_to_write_order(wo, &wo_end, &objects[i]);
666         }
667         last_untagged = i;
668
669         /*
670          * Then fill all the tagged tips.
671          */
672         for (; i < to_pack.nr_objects; i++) {
673                 if (objects[i].tagged)
674                         add_to_write_order(wo, &wo_end, &objects[i]);
675         }
676
677         /*
678          * And then all remaining commits and tags.
679          */
680         for (i = last_untagged; i < to_pack.nr_objects; i++) {
681                 if (objects[i].type != OBJ_COMMIT &&
682                     objects[i].type != OBJ_TAG)
683                         continue;
684                 add_to_write_order(wo, &wo_end, &objects[i]);
685         }
686
687         /*
688          * And then all the trees.
689          */
690         for (i = last_untagged; i < to_pack.nr_objects; i++) {
691                 if (objects[i].type != OBJ_TREE)
692                         continue;
693                 add_to_write_order(wo, &wo_end, &objects[i]);
694         }
695
696         /*
697          * Finally all the rest in really tight order
698          */
699         for (i = last_untagged; i < to_pack.nr_objects; i++) {
700                 if (!objects[i].filled)
701                         add_family_to_write_order(wo, &wo_end, &objects[i]);
702         }
703
704         if (wo_end != to_pack.nr_objects)
705                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
706
707         return wo;
708 }
709
710 static off_t write_reused_pack(struct sha1file *f)
711 {
712         unsigned char buffer[8192];
713         off_t to_write, total;
714         int fd;
715
716         if (!is_pack_valid(reuse_packfile))
717                 die("packfile is invalid: %s", reuse_packfile->pack_name);
718
719         fd = git_open_noatime(reuse_packfile->pack_name);
720         if (fd < 0)
721                 die_errno("unable to open packfile for reuse: %s",
722                           reuse_packfile->pack_name);
723
724         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
725                 die_errno("unable to seek in reused packfile");
726
727         if (reuse_packfile_offset < 0)
728                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
729
730         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
731
732         while (to_write) {
733                 int read_pack = xread(fd, buffer, sizeof(buffer));
734
735                 if (read_pack <= 0)
736                         die_errno("unable to read from reused packfile");
737
738                 if (read_pack > to_write)
739                         read_pack = to_write;
740
741                 sha1write(f, buffer, read_pack);
742                 to_write -= read_pack;
743
744                 /*
745                  * We don't know the actual number of objects written,
746                  * only how many bytes written, how many bytes total, and
747                  * how many objects total. So we can fake it by pretending all
748                  * objects we are writing are the same size. This gives us a
749                  * smooth progress meter, and at the end it matches the true
750                  * answer.
751                  */
752                 written = reuse_packfile_objects *
753                                 (((double)(total - to_write)) / total);
754                 display_progress(progress_state, written);
755         }
756
757         close(fd);
758         written = reuse_packfile_objects;
759         display_progress(progress_state, written);
760         return reuse_packfile_offset - sizeof(struct pack_header);
761 }
762
763 static const char no_split_warning[] = N_(
764 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
765 );
766
767 static void write_pack_file(void)
768 {
769         uint32_t i = 0, j;
770         struct sha1file *f;
771         off_t offset;
772         uint32_t nr_remaining = nr_result;
773         time_t last_mtime = 0;
774         struct object_entry **write_order;
775
776         if (progress > pack_to_stdout)
777                 progress_state = start_progress(_("Writing objects"), nr_result);
778         ALLOC_ARRAY(written_list, to_pack.nr_objects);
779         write_order = compute_write_order();
780
781         do {
782                 unsigned char sha1[20];
783                 char *pack_tmp_name = NULL;
784
785                 if (pack_to_stdout)
786                         f = sha1fd_throughput(1, "<stdout>", progress_state);
787                 else
788                         f = create_tmp_packfile(&pack_tmp_name);
789
790                 offset = write_pack_header(f, nr_remaining);
791
792                 if (reuse_packfile) {
793                         off_t packfile_size;
794                         assert(pack_to_stdout);
795
796                         packfile_size = write_reused_pack(f);
797                         offset += packfile_size;
798                 }
799
800                 nr_written = 0;
801                 for (; i < to_pack.nr_objects; i++) {
802                         struct object_entry *e = write_order[i];
803                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
804                                 break;
805                         display_progress(progress_state, written);
806                 }
807
808                 /*
809                  * Did we write the wrong # entries in the header?
810                  * If so, rewrite it like in fast-import
811                  */
812                 if (pack_to_stdout) {
813                         sha1close(f, sha1, CSUM_CLOSE);
814                 } else if (nr_written == nr_remaining) {
815                         sha1close(f, sha1, CSUM_FSYNC);
816                 } else {
817                         int fd = sha1close(f, sha1, 0);
818                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
819                                                  nr_written, sha1, offset);
820                         close(fd);
821                         if (write_bitmap_index) {
822                                 warning(_(no_split_warning));
823                                 write_bitmap_index = 0;
824                         }
825                 }
826
827                 if (!pack_to_stdout) {
828                         struct stat st;
829                         struct strbuf tmpname = STRBUF_INIT;
830
831                         /*
832                          * Packs are runtime accessed in their mtime
833                          * order since newer packs are more likely to contain
834                          * younger objects.  So if we are creating multiple
835                          * packs then we should modify the mtime of later ones
836                          * to preserve this property.
837                          */
838                         if (stat(pack_tmp_name, &st) < 0) {
839                                 warning_errno("failed to stat %s", pack_tmp_name);
840                         } else if (!last_mtime) {
841                                 last_mtime = st.st_mtime;
842                         } else {
843                                 struct utimbuf utb;
844                                 utb.actime = st.st_atime;
845                                 utb.modtime = --last_mtime;
846                                 if (utime(pack_tmp_name, &utb) < 0)
847                                         warning_errno("failed utime() on %s", pack_tmp_name);
848                         }
849
850                         strbuf_addf(&tmpname, "%s-", base_name);
851
852                         if (write_bitmap_index) {
853                                 bitmap_writer_set_checksum(sha1);
854                                 bitmap_writer_build_type_index(written_list, nr_written);
855                         }
856
857                         finish_tmp_packfile(&tmpname, pack_tmp_name,
858                                             written_list, nr_written,
859                                             &pack_idx_opts, sha1);
860
861                         if (write_bitmap_index) {
862                                 strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
863
864                                 stop_progress(&progress_state);
865
866                                 bitmap_writer_show_progress(progress);
867                                 bitmap_writer_reuse_bitmaps(&to_pack);
868                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
869                                 bitmap_writer_build(&to_pack);
870                                 bitmap_writer_finish(written_list, nr_written,
871                                                      tmpname.buf, write_bitmap_options);
872                                 write_bitmap_index = 0;
873                         }
874
875                         strbuf_release(&tmpname);
876                         free(pack_tmp_name);
877                         puts(sha1_to_hex(sha1));
878                 }
879
880                 /* mark written objects as written to previous pack */
881                 for (j = 0; j < nr_written; j++) {
882                         written_list[j]->offset = (off_t)-1;
883                 }
884                 nr_remaining -= nr_written;
885         } while (nr_remaining && i < to_pack.nr_objects);
886
887         free(written_list);
888         free(write_order);
889         stop_progress(&progress_state);
890         if (written != nr_result)
891                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
892                         written, nr_result);
893 }
894
895 static void setup_delta_attr_check(struct git_attr_check *check)
896 {
897         static struct git_attr *attr_delta;
898
899         if (!attr_delta)
900                 attr_delta = git_attr("delta");
901
902         check[0].attr = attr_delta;
903 }
904
905 static int no_try_delta(const char *path)
906 {
907         struct git_attr_check check[1];
908
909         setup_delta_attr_check(check);
910         if (git_check_attr(path, ARRAY_SIZE(check), check))
911                 return 0;
912         if (ATTR_FALSE(check->value))
913                 return 1;
914         return 0;
915 }
916
917 /*
918  * When adding an object, check whether we have already added it
919  * to our packing list. If so, we can skip. However, if we are
920  * being asked to excludei t, but the previous mention was to include
921  * it, make sure to adjust its flags and tweak our numbers accordingly.
922  *
923  * As an optimization, we pass out the index position where we would have
924  * found the item, since that saves us from having to look it up again a
925  * few lines later when we want to add the new entry.
926  */
927 static int have_duplicate_entry(const unsigned char *sha1,
928                                 int exclude,
929                                 uint32_t *index_pos)
930 {
931         struct object_entry *entry;
932
933         entry = packlist_find(&to_pack, sha1, index_pos);
934         if (!entry)
935                 return 0;
936
937         if (exclude) {
938                 if (!entry->preferred_base)
939                         nr_result--;
940                 entry->preferred_base = 1;
941         }
942
943         return 1;
944 }
945
946 /*
947  * Check whether we want the object in the pack (e.g., we do not want
948  * objects found in non-local stores if the "--local" option was used).
949  *
950  * As a side effect of this check, we will find the packed version of this
951  * object, if any. We therefore pass out the pack information to avoid having
952  * to look it up again later.
953  */
954 static int want_object_in_pack(const unsigned char *sha1,
955                                int exclude,
956                                struct packed_git **found_pack,
957                                off_t *found_offset)
958 {
959         struct packed_git *p;
960
961         if (!exclude && local && has_loose_object_nonlocal(sha1))
962                 return 0;
963
964         *found_pack = NULL;
965         *found_offset = 0;
966
967         for (p = packed_git; p; p = p->next) {
968                 off_t offset = find_pack_entry_one(sha1, p);
969                 if (offset) {
970                         if (!*found_pack) {
971                                 if (!is_pack_valid(p))
972                                         continue;
973                                 *found_offset = offset;
974                                 *found_pack = p;
975                         }
976                         if (exclude)
977                                 return 1;
978                         if (incremental)
979                                 return 0;
980
981                         /*
982                          * When asked to do --local (do not include an
983                          * object that appears in a pack we borrow
984                          * from elsewhere) or --honor-pack-keep (do not
985                          * include an object that appears in a pack marked
986                          * with .keep), we need to make sure no copy of this
987                          * object come from in _any_ pack that causes us to
988                          * omit it, and need to complete this loop.  When
989                          * neither option is in effect, we know the object
990                          * we just found is going to be packed, so break
991                          * out of the loop to return 1 now.
992                          */
993                         if (!ignore_packed_keep && !local)
994                                 break;
995
996                         if (local && !p->pack_local)
997                                 return 0;
998                         if (ignore_packed_keep && p->pack_local && p->pack_keep)
999                                 return 0;
1000                 }
1001         }
1002
1003         return 1;
1004 }
1005
1006 static void create_object_entry(const unsigned char *sha1,
1007                                 enum object_type type,
1008                                 uint32_t hash,
1009                                 int exclude,
1010                                 int no_try_delta,
1011                                 uint32_t index_pos,
1012                                 struct packed_git *found_pack,
1013                                 off_t found_offset)
1014 {
1015         struct object_entry *entry;
1016
1017         entry = packlist_alloc(&to_pack, sha1, index_pos);
1018         entry->hash = hash;
1019         if (type)
1020                 entry->type = type;
1021         if (exclude)
1022                 entry->preferred_base = 1;
1023         else
1024                 nr_result++;
1025         if (found_pack) {
1026                 entry->in_pack = found_pack;
1027                 entry->in_pack_offset = found_offset;
1028         }
1029
1030         entry->no_try_delta = no_try_delta;
1031 }
1032
1033 static const char no_closure_warning[] = N_(
1034 "disabling bitmap writing, as some objects are not being packed"
1035 );
1036
1037 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1038                             const char *name, int exclude)
1039 {
1040         struct packed_git *found_pack;
1041         off_t found_offset;
1042         uint32_t index_pos;
1043
1044         if (have_duplicate_entry(sha1, exclude, &index_pos))
1045                 return 0;
1046
1047         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1048                 /* The pack is missing an object, so it will not have closure */
1049                 if (write_bitmap_index) {
1050                         warning(_(no_closure_warning));
1051                         write_bitmap_index = 0;
1052                 }
1053                 return 0;
1054         }
1055
1056         create_object_entry(sha1, type, pack_name_hash(name),
1057                             exclude, name && no_try_delta(name),
1058                             index_pos, found_pack, found_offset);
1059
1060         display_progress(progress_state, nr_result);
1061         return 1;
1062 }
1063
1064 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1065                                         enum object_type type,
1066                                         int flags, uint32_t name_hash,
1067                                         struct packed_git *pack, off_t offset)
1068 {
1069         uint32_t index_pos;
1070
1071         if (have_duplicate_entry(sha1, 0, &index_pos))
1072                 return 0;
1073
1074         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1075
1076         display_progress(progress_state, nr_result);
1077         return 1;
1078 }
1079
1080 struct pbase_tree_cache {
1081         unsigned char sha1[20];
1082         int ref;
1083         int temporary;
1084         void *tree_data;
1085         unsigned long tree_size;
1086 };
1087
1088 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1089 static int pbase_tree_cache_ix(const unsigned char *sha1)
1090 {
1091         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1092 }
1093 static int pbase_tree_cache_ix_incr(int ix)
1094 {
1095         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1096 }
1097
1098 static struct pbase_tree {
1099         struct pbase_tree *next;
1100         /* This is a phony "cache" entry; we are not
1101          * going to evict it or find it through _get()
1102          * mechanism -- this is for the toplevel node that
1103          * would almost always change with any commit.
1104          */
1105         struct pbase_tree_cache pcache;
1106 } *pbase_tree;
1107
1108 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1109 {
1110         struct pbase_tree_cache *ent, *nent;
1111         void *data;
1112         unsigned long size;
1113         enum object_type type;
1114         int neigh;
1115         int my_ix = pbase_tree_cache_ix(sha1);
1116         int available_ix = -1;
1117
1118         /* pbase-tree-cache acts as a limited hashtable.
1119          * your object will be found at your index or within a few
1120          * slots after that slot if it is cached.
1121          */
1122         for (neigh = 0; neigh < 8; neigh++) {
1123                 ent = pbase_tree_cache[my_ix];
1124                 if (ent && !hashcmp(ent->sha1, sha1)) {
1125                         ent->ref++;
1126                         return ent;
1127                 }
1128                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1129                          ((0 <= available_ix) &&
1130                           (!ent && pbase_tree_cache[available_ix])))
1131                         available_ix = my_ix;
1132                 if (!ent)
1133                         break;
1134                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1135         }
1136
1137         /* Did not find one.  Either we got a bogus request or
1138          * we need to read and perhaps cache.
1139          */
1140         data = read_sha1_file(sha1, &type, &size);
1141         if (!data)
1142                 return NULL;
1143         if (type != OBJ_TREE) {
1144                 free(data);
1145                 return NULL;
1146         }
1147
1148         /* We need to either cache or return a throwaway copy */
1149
1150         if (available_ix < 0)
1151                 ent = NULL;
1152         else {
1153                 ent = pbase_tree_cache[available_ix];
1154                 my_ix = available_ix;
1155         }
1156
1157         if (!ent) {
1158                 nent = xmalloc(sizeof(*nent));
1159                 nent->temporary = (available_ix < 0);
1160         }
1161         else {
1162                 /* evict and reuse */
1163                 free(ent->tree_data);
1164                 nent = ent;
1165         }
1166         hashcpy(nent->sha1, sha1);
1167         nent->tree_data = data;
1168         nent->tree_size = size;
1169         nent->ref = 1;
1170         if (!nent->temporary)
1171                 pbase_tree_cache[my_ix] = nent;
1172         return nent;
1173 }
1174
1175 static void pbase_tree_put(struct pbase_tree_cache *cache)
1176 {
1177         if (!cache->temporary) {
1178                 cache->ref--;
1179                 return;
1180         }
1181         free(cache->tree_data);
1182         free(cache);
1183 }
1184
1185 static int name_cmp_len(const char *name)
1186 {
1187         int i;
1188         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1189                 ;
1190         return i;
1191 }
1192
1193 static void add_pbase_object(struct tree_desc *tree,
1194                              const char *name,
1195                              int cmplen,
1196                              const char *fullname)
1197 {
1198         struct name_entry entry;
1199         int cmp;
1200
1201         while (tree_entry(tree,&entry)) {
1202                 if (S_ISGITLINK(entry.mode))
1203                         continue;
1204                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1205                       memcmp(name, entry.path, cmplen);
1206                 if (cmp > 0)
1207                         continue;
1208                 if (cmp < 0)
1209                         return;
1210                 if (name[cmplen] != '/') {
1211                         add_object_entry(entry.oid->hash,
1212                                          object_type(entry.mode),
1213                                          fullname, 1);
1214                         return;
1215                 }
1216                 if (S_ISDIR(entry.mode)) {
1217                         struct tree_desc sub;
1218                         struct pbase_tree_cache *tree;
1219                         const char *down = name+cmplen+1;
1220                         int downlen = name_cmp_len(down);
1221
1222                         tree = pbase_tree_get(entry.oid->hash);
1223                         if (!tree)
1224                                 return;
1225                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1226
1227                         add_pbase_object(&sub, down, downlen, fullname);
1228                         pbase_tree_put(tree);
1229                 }
1230         }
1231 }
1232
1233 static unsigned *done_pbase_paths;
1234 static int done_pbase_paths_num;
1235 static int done_pbase_paths_alloc;
1236 static int done_pbase_path_pos(unsigned hash)
1237 {
1238         int lo = 0;
1239         int hi = done_pbase_paths_num;
1240         while (lo < hi) {
1241                 int mi = (hi + lo) / 2;
1242                 if (done_pbase_paths[mi] == hash)
1243                         return mi;
1244                 if (done_pbase_paths[mi] < hash)
1245                         hi = mi;
1246                 else
1247                         lo = mi + 1;
1248         }
1249         return -lo-1;
1250 }
1251
1252 static int check_pbase_path(unsigned hash)
1253 {
1254         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1255         if (0 <= pos)
1256                 return 1;
1257         pos = -pos - 1;
1258         ALLOC_GROW(done_pbase_paths,
1259                    done_pbase_paths_num + 1,
1260                    done_pbase_paths_alloc);
1261         done_pbase_paths_num++;
1262         if (pos < done_pbase_paths_num)
1263                 memmove(done_pbase_paths + pos + 1,
1264                         done_pbase_paths + pos,
1265                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1266         done_pbase_paths[pos] = hash;
1267         return 0;
1268 }
1269
1270 static void add_preferred_base_object(const char *name)
1271 {
1272         struct pbase_tree *it;
1273         int cmplen;
1274         unsigned hash = pack_name_hash(name);
1275
1276         if (!num_preferred_base || check_pbase_path(hash))
1277                 return;
1278
1279         cmplen = name_cmp_len(name);
1280         for (it = pbase_tree; it; it = it->next) {
1281                 if (cmplen == 0) {
1282                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1283                 }
1284                 else {
1285                         struct tree_desc tree;
1286                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1287                         add_pbase_object(&tree, name, cmplen, name);
1288                 }
1289         }
1290 }
1291
1292 static void add_preferred_base(unsigned char *sha1)
1293 {
1294         struct pbase_tree *it;
1295         void *data;
1296         unsigned long size;
1297         unsigned char tree_sha1[20];
1298
1299         if (window <= num_preferred_base++)
1300                 return;
1301
1302         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1303         if (!data)
1304                 return;
1305
1306         for (it = pbase_tree; it; it = it->next) {
1307                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1308                         free(data);
1309                         return;
1310                 }
1311         }
1312
1313         it = xcalloc(1, sizeof(*it));
1314         it->next = pbase_tree;
1315         pbase_tree = it;
1316
1317         hashcpy(it->pcache.sha1, tree_sha1);
1318         it->pcache.tree_data = data;
1319         it->pcache.tree_size = size;
1320 }
1321
1322 static void cleanup_preferred_base(void)
1323 {
1324         struct pbase_tree *it;
1325         unsigned i;
1326
1327         it = pbase_tree;
1328         pbase_tree = NULL;
1329         while (it) {
1330                 struct pbase_tree *this = it;
1331                 it = this->next;
1332                 free(this->pcache.tree_data);
1333                 free(this);
1334         }
1335
1336         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1337                 if (!pbase_tree_cache[i])
1338                         continue;
1339                 free(pbase_tree_cache[i]->tree_data);
1340                 free(pbase_tree_cache[i]);
1341                 pbase_tree_cache[i] = NULL;
1342         }
1343
1344         free(done_pbase_paths);
1345         done_pbase_paths = NULL;
1346         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1347 }
1348
1349 static void check_object(struct object_entry *entry)
1350 {
1351         if (entry->in_pack) {
1352                 struct packed_git *p = entry->in_pack;
1353                 struct pack_window *w_curs = NULL;
1354                 const unsigned char *base_ref = NULL;
1355                 struct object_entry *base_entry;
1356                 unsigned long used, used_0;
1357                 unsigned long avail;
1358                 off_t ofs;
1359                 unsigned char *buf, c;
1360
1361                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1362
1363                 /*
1364                  * We want in_pack_type even if we do not reuse delta
1365                  * since non-delta representations could still be reused.
1366                  */
1367                 used = unpack_object_header_buffer(buf, avail,
1368                                                    &entry->in_pack_type,
1369                                                    &entry->size);
1370                 if (used == 0)
1371                         goto give_up;
1372
1373                 /*
1374                  * Determine if this is a delta and if so whether we can
1375                  * reuse it or not.  Otherwise let's find out as cheaply as
1376                  * possible what the actual type and size for this object is.
1377                  */
1378                 switch (entry->in_pack_type) {
1379                 default:
1380                         /* Not a delta hence we've already got all we need. */
1381                         entry->type = entry->in_pack_type;
1382                         entry->in_pack_header_size = used;
1383                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1384                                 goto give_up;
1385                         unuse_pack(&w_curs);
1386                         return;
1387                 case OBJ_REF_DELTA:
1388                         if (reuse_delta && !entry->preferred_base)
1389                                 base_ref = use_pack(p, &w_curs,
1390                                                 entry->in_pack_offset + used, NULL);
1391                         entry->in_pack_header_size = used + 20;
1392                         break;
1393                 case OBJ_OFS_DELTA:
1394                         buf = use_pack(p, &w_curs,
1395                                        entry->in_pack_offset + used, NULL);
1396                         used_0 = 0;
1397                         c = buf[used_0++];
1398                         ofs = c & 127;
1399                         while (c & 128) {
1400                                 ofs += 1;
1401                                 if (!ofs || MSB(ofs, 7)) {
1402                                         error("delta base offset overflow in pack for %s",
1403                                               sha1_to_hex(entry->idx.sha1));
1404                                         goto give_up;
1405                                 }
1406                                 c = buf[used_0++];
1407                                 ofs = (ofs << 7) + (c & 127);
1408                         }
1409                         ofs = entry->in_pack_offset - ofs;
1410                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1411                                 error("delta base offset out of bound for %s",
1412                                       sha1_to_hex(entry->idx.sha1));
1413                                 goto give_up;
1414                         }
1415                         if (reuse_delta && !entry->preferred_base) {
1416                                 struct revindex_entry *revidx;
1417                                 revidx = find_pack_revindex(p, ofs);
1418                                 if (!revidx)
1419                                         goto give_up;
1420                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1421                         }
1422                         entry->in_pack_header_size = used + used_0;
1423                         break;
1424                 }
1425
1426                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1427                         /*
1428                          * If base_ref was set above that means we wish to
1429                          * reuse delta data, and we even found that base
1430                          * in the list of objects we want to pack. Goodie!
1431                          *
1432                          * Depth value does not matter - find_deltas() will
1433                          * never consider reused delta as the base object to
1434                          * deltify other objects against, in order to avoid
1435                          * circular deltas.
1436                          */
1437                         entry->type = entry->in_pack_type;
1438                         entry->delta = base_entry;
1439                         entry->delta_size = entry->size;
1440                         entry->delta_sibling = base_entry->delta_child;
1441                         base_entry->delta_child = entry;
1442                         unuse_pack(&w_curs);
1443                         return;
1444                 }
1445
1446                 if (entry->type) {
1447                         /*
1448                          * This must be a delta and we already know what the
1449                          * final object type is.  Let's extract the actual
1450                          * object size from the delta header.
1451                          */
1452                         entry->size = get_size_from_delta(p, &w_curs,
1453                                         entry->in_pack_offset + entry->in_pack_header_size);
1454                         if (entry->size == 0)
1455                                 goto give_up;
1456                         unuse_pack(&w_curs);
1457                         return;
1458                 }
1459
1460                 /*
1461                  * No choice but to fall back to the recursive delta walk
1462                  * with sha1_object_info() to find about the object type
1463                  * at this point...
1464                  */
1465                 give_up:
1466                 unuse_pack(&w_curs);
1467         }
1468
1469         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1470         /*
1471          * The error condition is checked in prepare_pack().  This is
1472          * to permit a missing preferred base object to be ignored
1473          * as a preferred base.  Doing so can result in a larger
1474          * pack file, but the transfer will still take place.
1475          */
1476 }
1477
1478 static int pack_offset_sort(const void *_a, const void *_b)
1479 {
1480         const struct object_entry *a = *(struct object_entry **)_a;
1481         const struct object_entry *b = *(struct object_entry **)_b;
1482
1483         /* avoid filesystem trashing with loose objects */
1484         if (!a->in_pack && !b->in_pack)
1485                 return hashcmp(a->idx.sha1, b->idx.sha1);
1486
1487         if (a->in_pack < b->in_pack)
1488                 return -1;
1489         if (a->in_pack > b->in_pack)
1490                 return 1;
1491         return a->in_pack_offset < b->in_pack_offset ? -1 :
1492                         (a->in_pack_offset > b->in_pack_offset);
1493 }
1494
1495 static void get_object_details(void)
1496 {
1497         uint32_t i;
1498         struct object_entry **sorted_by_offset;
1499
1500         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1501         for (i = 0; i < to_pack.nr_objects; i++)
1502                 sorted_by_offset[i] = to_pack.objects + i;
1503         qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1504
1505         for (i = 0; i < to_pack.nr_objects; i++) {
1506                 struct object_entry *entry = sorted_by_offset[i];
1507                 check_object(entry);
1508                 if (big_file_threshold < entry->size)
1509                         entry->no_try_delta = 1;
1510         }
1511
1512         free(sorted_by_offset);
1513 }
1514
1515 /*
1516  * We search for deltas in a list sorted by type, by filename hash, and then
1517  * by size, so that we see progressively smaller and smaller files.
1518  * That's because we prefer deltas to be from the bigger file
1519  * to the smaller -- deletes are potentially cheaper, but perhaps
1520  * more importantly, the bigger file is likely the more recent
1521  * one.  The deepest deltas are therefore the oldest objects which are
1522  * less susceptible to be accessed often.
1523  */
1524 static int type_size_sort(const void *_a, const void *_b)
1525 {
1526         const struct object_entry *a = *(struct object_entry **)_a;
1527         const struct object_entry *b = *(struct object_entry **)_b;
1528
1529         if (a->type > b->type)
1530                 return -1;
1531         if (a->type < b->type)
1532                 return 1;
1533         if (a->hash > b->hash)
1534                 return -1;
1535         if (a->hash < b->hash)
1536                 return 1;
1537         if (a->preferred_base > b->preferred_base)
1538                 return -1;
1539         if (a->preferred_base < b->preferred_base)
1540                 return 1;
1541         if (a->size > b->size)
1542                 return -1;
1543         if (a->size < b->size)
1544                 return 1;
1545         return a < b ? -1 : (a > b);  /* newest first */
1546 }
1547
1548 struct unpacked {
1549         struct object_entry *entry;
1550         void *data;
1551         struct delta_index *index;
1552         unsigned depth;
1553 };
1554
1555 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1556                            unsigned long delta_size)
1557 {
1558         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1559                 return 0;
1560
1561         if (delta_size < cache_max_small_delta_size)
1562                 return 1;
1563
1564         /* cache delta, if objects are large enough compared to delta size */
1565         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1566                 return 1;
1567
1568         return 0;
1569 }
1570
1571 #ifndef NO_PTHREADS
1572
1573 static pthread_mutex_t read_mutex;
1574 #define read_lock()             pthread_mutex_lock(&read_mutex)
1575 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1576
1577 static pthread_mutex_t cache_mutex;
1578 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1579 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1580
1581 static pthread_mutex_t progress_mutex;
1582 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1583 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1584
1585 #else
1586
1587 #define read_lock()             (void)0
1588 #define read_unlock()           (void)0
1589 #define cache_lock()            (void)0
1590 #define cache_unlock()          (void)0
1591 #define progress_lock()         (void)0
1592 #define progress_unlock()       (void)0
1593
1594 #endif
1595
1596 static int try_delta(struct unpacked *trg, struct unpacked *src,
1597                      unsigned max_depth, unsigned long *mem_usage)
1598 {
1599         struct object_entry *trg_entry = trg->entry;
1600         struct object_entry *src_entry = src->entry;
1601         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1602         unsigned ref_depth;
1603         enum object_type type;
1604         void *delta_buf;
1605
1606         /* Don't bother doing diffs between different types */
1607         if (trg_entry->type != src_entry->type)
1608                 return -1;
1609
1610         /*
1611          * We do not bother to try a delta that we discarded on an
1612          * earlier try, but only when reusing delta data.  Note that
1613          * src_entry that is marked as the preferred_base should always
1614          * be considered, as even if we produce a suboptimal delta against
1615          * it, we will still save the transfer cost, as we already know
1616          * the other side has it and we won't send src_entry at all.
1617          */
1618         if (reuse_delta && trg_entry->in_pack &&
1619             trg_entry->in_pack == src_entry->in_pack &&
1620             !src_entry->preferred_base &&
1621             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1622             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1623                 return 0;
1624
1625         /* Let's not bust the allowed depth. */
1626         if (src->depth >= max_depth)
1627                 return 0;
1628
1629         /* Now some size filtering heuristics. */
1630         trg_size = trg_entry->size;
1631         if (!trg_entry->delta) {
1632                 max_size = trg_size/2 - 20;
1633                 ref_depth = 1;
1634         } else {
1635                 max_size = trg_entry->delta_size;
1636                 ref_depth = trg->depth;
1637         }
1638         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1639                                                 (max_depth - ref_depth + 1);
1640         if (max_size == 0)
1641                 return 0;
1642         src_size = src_entry->size;
1643         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1644         if (sizediff >= max_size)
1645                 return 0;
1646         if (trg_size < src_size / 32)
1647                 return 0;
1648
1649         /* Load data if not already done */
1650         if (!trg->data) {
1651                 read_lock();
1652                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1653                 read_unlock();
1654                 if (!trg->data)
1655                         die("object %s cannot be read",
1656                             sha1_to_hex(trg_entry->idx.sha1));
1657                 if (sz != trg_size)
1658                         die("object %s inconsistent object length (%lu vs %lu)",
1659                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1660                 *mem_usage += sz;
1661         }
1662         if (!src->data) {
1663                 read_lock();
1664                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1665                 read_unlock();
1666                 if (!src->data) {
1667                         if (src_entry->preferred_base) {
1668                                 static int warned = 0;
1669                                 if (!warned++)
1670                                         warning("object %s cannot be read",
1671                                                 sha1_to_hex(src_entry->idx.sha1));
1672                                 /*
1673                                  * Those objects are not included in the
1674                                  * resulting pack.  Be resilient and ignore
1675                                  * them if they can't be read, in case the
1676                                  * pack could be created nevertheless.
1677                                  */
1678                                 return 0;
1679                         }
1680                         die("object %s cannot be read",
1681                             sha1_to_hex(src_entry->idx.sha1));
1682                 }
1683                 if (sz != src_size)
1684                         die("object %s inconsistent object length (%lu vs %lu)",
1685                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1686                 *mem_usage += sz;
1687         }
1688         if (!src->index) {
1689                 src->index = create_delta_index(src->data, src_size);
1690                 if (!src->index) {
1691                         static int warned = 0;
1692                         if (!warned++)
1693                                 warning("suboptimal pack - out of memory");
1694                         return 0;
1695                 }
1696                 *mem_usage += sizeof_delta_index(src->index);
1697         }
1698
1699         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1700         if (!delta_buf)
1701                 return 0;
1702
1703         if (trg_entry->delta) {
1704                 /* Prefer only shallower same-sized deltas. */
1705                 if (delta_size == trg_entry->delta_size &&
1706                     src->depth + 1 >= trg->depth) {
1707                         free(delta_buf);
1708                         return 0;
1709                 }
1710         }
1711
1712         /*
1713          * Handle memory allocation outside of the cache
1714          * accounting lock.  Compiler will optimize the strangeness
1715          * away when NO_PTHREADS is defined.
1716          */
1717         free(trg_entry->delta_data);
1718         cache_lock();
1719         if (trg_entry->delta_data) {
1720                 delta_cache_size -= trg_entry->delta_size;
1721                 trg_entry->delta_data = NULL;
1722         }
1723         if (delta_cacheable(src_size, trg_size, delta_size)) {
1724                 delta_cache_size += delta_size;
1725                 cache_unlock();
1726                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1727         } else {
1728                 cache_unlock();
1729                 free(delta_buf);
1730         }
1731
1732         trg_entry->delta = src_entry;
1733         trg_entry->delta_size = delta_size;
1734         trg->depth = src->depth + 1;
1735
1736         return 1;
1737 }
1738
1739 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1740 {
1741         struct object_entry *child = me->delta_child;
1742         unsigned int m = n;
1743         while (child) {
1744                 unsigned int c = check_delta_limit(child, n + 1);
1745                 if (m < c)
1746                         m = c;
1747                 child = child->delta_sibling;
1748         }
1749         return m;
1750 }
1751
1752 static unsigned long free_unpacked(struct unpacked *n)
1753 {
1754         unsigned long freed_mem = sizeof_delta_index(n->index);
1755         free_delta_index(n->index);
1756         n->index = NULL;
1757         if (n->data) {
1758                 freed_mem += n->entry->size;
1759                 free(n->data);
1760                 n->data = NULL;
1761         }
1762         n->entry = NULL;
1763         n->depth = 0;
1764         return freed_mem;
1765 }
1766
1767 static void find_deltas(struct object_entry **list, unsigned *list_size,
1768                         int window, int depth, unsigned *processed)
1769 {
1770         uint32_t i, idx = 0, count = 0;
1771         struct unpacked *array;
1772         unsigned long mem_usage = 0;
1773
1774         array = xcalloc(window, sizeof(struct unpacked));
1775
1776         for (;;) {
1777                 struct object_entry *entry;
1778                 struct unpacked *n = array + idx;
1779                 int j, max_depth, best_base = -1;
1780
1781                 progress_lock();
1782                 if (!*list_size) {
1783                         progress_unlock();
1784                         break;
1785                 }
1786                 entry = *list++;
1787                 (*list_size)--;
1788                 if (!entry->preferred_base) {
1789                         (*processed)++;
1790                         display_progress(progress_state, *processed);
1791                 }
1792                 progress_unlock();
1793
1794                 mem_usage -= free_unpacked(n);
1795                 n->entry = entry;
1796
1797                 while (window_memory_limit &&
1798                        mem_usage > window_memory_limit &&
1799                        count > 1) {
1800                         uint32_t tail = (idx + window - count) % window;
1801                         mem_usage -= free_unpacked(array + tail);
1802                         count--;
1803                 }
1804
1805                 /* We do not compute delta to *create* objects we are not
1806                  * going to pack.
1807                  */
1808                 if (entry->preferred_base)
1809                         goto next;
1810
1811                 /*
1812                  * If the current object is at pack edge, take the depth the
1813                  * objects that depend on the current object into account
1814                  * otherwise they would become too deep.
1815                  */
1816                 max_depth = depth;
1817                 if (entry->delta_child) {
1818                         max_depth -= check_delta_limit(entry, 0);
1819                         if (max_depth <= 0)
1820                                 goto next;
1821                 }
1822
1823                 j = window;
1824                 while (--j > 0) {
1825                         int ret;
1826                         uint32_t other_idx = idx + j;
1827                         struct unpacked *m;
1828                         if (other_idx >= window)
1829                                 other_idx -= window;
1830                         m = array + other_idx;
1831                         if (!m->entry)
1832                                 break;
1833                         ret = try_delta(n, m, max_depth, &mem_usage);
1834                         if (ret < 0)
1835                                 break;
1836                         else if (ret > 0)
1837                                 best_base = other_idx;
1838                 }
1839
1840                 /*
1841                  * If we decided to cache the delta data, then it is best
1842                  * to compress it right away.  First because we have to do
1843                  * it anyway, and doing it here while we're threaded will
1844                  * save a lot of time in the non threaded write phase,
1845                  * as well as allow for caching more deltas within
1846                  * the same cache size limit.
1847                  * ...
1848                  * But only if not writing to stdout, since in that case
1849                  * the network is most likely throttling writes anyway,
1850                  * and therefore it is best to go to the write phase ASAP
1851                  * instead, as we can afford spending more time compressing
1852                  * between writes at that moment.
1853                  */
1854                 if (entry->delta_data && !pack_to_stdout) {
1855                         entry->z_delta_size = do_compress(&entry->delta_data,
1856                                                           entry->delta_size);
1857                         cache_lock();
1858                         delta_cache_size -= entry->delta_size;
1859                         delta_cache_size += entry->z_delta_size;
1860                         cache_unlock();
1861                 }
1862
1863                 /* if we made n a delta, and if n is already at max
1864                  * depth, leaving it in the window is pointless.  we
1865                  * should evict it first.
1866                  */
1867                 if (entry->delta && max_depth <= n->depth)
1868                         continue;
1869
1870                 /*
1871                  * Move the best delta base up in the window, after the
1872                  * currently deltified object, to keep it longer.  It will
1873                  * be the first base object to be attempted next.
1874                  */
1875                 if (entry->delta) {
1876                         struct unpacked swap = array[best_base];
1877                         int dist = (window + idx - best_base) % window;
1878                         int dst = best_base;
1879                         while (dist--) {
1880                                 int src = (dst + 1) % window;
1881                                 array[dst] = array[src];
1882                                 dst = src;
1883                         }
1884                         array[dst] = swap;
1885                 }
1886
1887                 next:
1888                 idx++;
1889                 if (count + 1 < window)
1890                         count++;
1891                 if (idx >= window)
1892                         idx = 0;
1893         }
1894
1895         for (i = 0; i < window; ++i) {
1896                 free_delta_index(array[i].index);
1897                 free(array[i].data);
1898         }
1899         free(array);
1900 }
1901
1902 #ifndef NO_PTHREADS
1903
1904 static void try_to_free_from_threads(size_t size)
1905 {
1906         read_lock();
1907         release_pack_memory(size);
1908         read_unlock();
1909 }
1910
1911 static try_to_free_t old_try_to_free_routine;
1912
1913 /*
1914  * The main thread waits on the condition that (at least) one of the workers
1915  * has stopped working (which is indicated in the .working member of
1916  * struct thread_params).
1917  * When a work thread has completed its work, it sets .working to 0 and
1918  * signals the main thread and waits on the condition that .data_ready
1919  * becomes 1.
1920  */
1921
1922 struct thread_params {
1923         pthread_t thread;
1924         struct object_entry **list;
1925         unsigned list_size;
1926         unsigned remaining;
1927         int window;
1928         int depth;
1929         int working;
1930         int data_ready;
1931         pthread_mutex_t mutex;
1932         pthread_cond_t cond;
1933         unsigned *processed;
1934 };
1935
1936 static pthread_cond_t progress_cond;
1937
1938 /*
1939  * Mutex and conditional variable can't be statically-initialized on Windows.
1940  */
1941 static void init_threaded_search(void)
1942 {
1943         init_recursive_mutex(&read_mutex);
1944         pthread_mutex_init(&cache_mutex, NULL);
1945         pthread_mutex_init(&progress_mutex, NULL);
1946         pthread_cond_init(&progress_cond, NULL);
1947         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1948 }
1949
1950 static void cleanup_threaded_search(void)
1951 {
1952         set_try_to_free_routine(old_try_to_free_routine);
1953         pthread_cond_destroy(&progress_cond);
1954         pthread_mutex_destroy(&read_mutex);
1955         pthread_mutex_destroy(&cache_mutex);
1956         pthread_mutex_destroy(&progress_mutex);
1957 }
1958
1959 static void *threaded_find_deltas(void *arg)
1960 {
1961         struct thread_params *me = arg;
1962
1963         while (me->remaining) {
1964                 find_deltas(me->list, &me->remaining,
1965                             me->window, me->depth, me->processed);
1966
1967                 progress_lock();
1968                 me->working = 0;
1969                 pthread_cond_signal(&progress_cond);
1970                 progress_unlock();
1971
1972                 /*
1973                  * We must not set ->data_ready before we wait on the
1974                  * condition because the main thread may have set it to 1
1975                  * before we get here. In order to be sure that new
1976                  * work is available if we see 1 in ->data_ready, it
1977                  * was initialized to 0 before this thread was spawned
1978                  * and we reset it to 0 right away.
1979                  */
1980                 pthread_mutex_lock(&me->mutex);
1981                 while (!me->data_ready)
1982                         pthread_cond_wait(&me->cond, &me->mutex);
1983                 me->data_ready = 0;
1984                 pthread_mutex_unlock(&me->mutex);
1985         }
1986         /* leave ->working 1 so that this doesn't get more work assigned */
1987         return NULL;
1988 }
1989
1990 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1991                            int window, int depth, unsigned *processed)
1992 {
1993         struct thread_params *p;
1994         int i, ret, active_threads = 0;
1995
1996         init_threaded_search();
1997
1998         if (delta_search_threads <= 1) {
1999                 find_deltas(list, &list_size, window, depth, processed);
2000                 cleanup_threaded_search();
2001                 return;
2002         }
2003         if (progress > pack_to_stdout)
2004                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2005                                 delta_search_threads);
2006         p = xcalloc(delta_search_threads, sizeof(*p));
2007
2008         /* Partition the work amongst work threads. */
2009         for (i = 0; i < delta_search_threads; i++) {
2010                 unsigned sub_size = list_size / (delta_search_threads - i);
2011
2012                 /* don't use too small segments or no deltas will be found */
2013                 if (sub_size < 2*window && i+1 < delta_search_threads)
2014                         sub_size = 0;
2015
2016                 p[i].window = window;
2017                 p[i].depth = depth;
2018                 p[i].processed = processed;
2019                 p[i].working = 1;
2020                 p[i].data_ready = 0;
2021
2022                 /* try to split chunks on "path" boundaries */
2023                 while (sub_size && sub_size < list_size &&
2024                        list[sub_size]->hash &&
2025                        list[sub_size]->hash == list[sub_size-1]->hash)
2026                         sub_size++;
2027
2028                 p[i].list = list;
2029                 p[i].list_size = sub_size;
2030                 p[i].remaining = sub_size;
2031
2032                 list += sub_size;
2033                 list_size -= sub_size;
2034         }
2035
2036         /* Start work threads. */
2037         for (i = 0; i < delta_search_threads; i++) {
2038                 if (!p[i].list_size)
2039                         continue;
2040                 pthread_mutex_init(&p[i].mutex, NULL);
2041                 pthread_cond_init(&p[i].cond, NULL);
2042                 ret = pthread_create(&p[i].thread, NULL,
2043                                      threaded_find_deltas, &p[i]);
2044                 if (ret)
2045                         die("unable to create thread: %s", strerror(ret));
2046                 active_threads++;
2047         }
2048
2049         /*
2050          * Now let's wait for work completion.  Each time a thread is done
2051          * with its work, we steal half of the remaining work from the
2052          * thread with the largest number of unprocessed objects and give
2053          * it to that newly idle thread.  This ensure good load balancing
2054          * until the remaining object list segments are simply too short
2055          * to be worth splitting anymore.
2056          */
2057         while (active_threads) {
2058                 struct thread_params *target = NULL;
2059                 struct thread_params *victim = NULL;
2060                 unsigned sub_size = 0;
2061
2062                 progress_lock();
2063                 for (;;) {
2064                         for (i = 0; !target && i < delta_search_threads; i++)
2065                                 if (!p[i].working)
2066                                         target = &p[i];
2067                         if (target)
2068                                 break;
2069                         pthread_cond_wait(&progress_cond, &progress_mutex);
2070                 }
2071
2072                 for (i = 0; i < delta_search_threads; i++)
2073                         if (p[i].remaining > 2*window &&
2074                             (!victim || victim->remaining < p[i].remaining))
2075                                 victim = &p[i];
2076                 if (victim) {
2077                         sub_size = victim->remaining / 2;
2078                         list = victim->list + victim->list_size - sub_size;
2079                         while (sub_size && list[0]->hash &&
2080                                list[0]->hash == list[-1]->hash) {
2081                                 list++;
2082                                 sub_size--;
2083                         }
2084                         if (!sub_size) {
2085                                 /*
2086                                  * It is possible for some "paths" to have
2087                                  * so many objects that no hash boundary
2088                                  * might be found.  Let's just steal the
2089                                  * exact half in that case.
2090                                  */
2091                                 sub_size = victim->remaining / 2;
2092                                 list -= sub_size;
2093                         }
2094                         target->list = list;
2095                         victim->list_size -= sub_size;
2096                         victim->remaining -= sub_size;
2097                 }
2098                 target->list_size = sub_size;
2099                 target->remaining = sub_size;
2100                 target->working = 1;
2101                 progress_unlock();
2102
2103                 pthread_mutex_lock(&target->mutex);
2104                 target->data_ready = 1;
2105                 pthread_cond_signal(&target->cond);
2106                 pthread_mutex_unlock(&target->mutex);
2107
2108                 if (!sub_size) {
2109                         pthread_join(target->thread, NULL);
2110                         pthread_cond_destroy(&target->cond);
2111                         pthread_mutex_destroy(&target->mutex);
2112                         active_threads--;
2113                 }
2114         }
2115         cleanup_threaded_search();
2116         free(p);
2117 }
2118
2119 #else
2120 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2121 #endif
2122
2123 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2124 {
2125         struct object_id peeled;
2126
2127         if (starts_with(path, "refs/tags/") && /* is a tag? */
2128             !peel_ref(path, peeled.hash)    && /* peelable? */
2129             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2130                 add_object_entry(oid->hash, OBJ_TAG, NULL, 0);
2131         return 0;
2132 }
2133
2134 static void prepare_pack(int window, int depth)
2135 {
2136         struct object_entry **delta_list;
2137         uint32_t i, nr_deltas;
2138         unsigned n;
2139
2140         get_object_details();
2141
2142         /*
2143          * If we're locally repacking then we need to be doubly careful
2144          * from now on in order to make sure no stealth corruption gets
2145          * propagated to the new pack.  Clients receiving streamed packs
2146          * should validate everything they get anyway so no need to incur
2147          * the additional cost here in that case.
2148          */
2149         if (!pack_to_stdout)
2150                 do_check_packed_object_crc = 1;
2151
2152         if (!to_pack.nr_objects || !window || !depth)
2153                 return;
2154
2155         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2156         nr_deltas = n = 0;
2157
2158         for (i = 0; i < to_pack.nr_objects; i++) {
2159                 struct object_entry *entry = to_pack.objects + i;
2160
2161                 if (entry->delta)
2162                         /* This happens if we decided to reuse existing
2163                          * delta from a pack.  "reuse_delta &&" is implied.
2164                          */
2165                         continue;
2166
2167                 if (entry->size < 50)
2168                         continue;
2169
2170                 if (entry->no_try_delta)
2171                         continue;
2172
2173                 if (!entry->preferred_base) {
2174                         nr_deltas++;
2175                         if (entry->type < 0)
2176                                 die("unable to get type of object %s",
2177                                     sha1_to_hex(entry->idx.sha1));
2178                 } else {
2179                         if (entry->type < 0) {
2180                                 /*
2181                                  * This object is not found, but we
2182                                  * don't have to include it anyway.
2183                                  */
2184                                 continue;
2185                         }
2186                 }
2187
2188                 delta_list[n++] = entry;
2189         }
2190
2191         if (nr_deltas && n > 1) {
2192                 unsigned nr_done = 0;
2193                 if (progress)
2194                         progress_state = start_progress(_("Compressing objects"),
2195                                                         nr_deltas);
2196                 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2197                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2198                 stop_progress(&progress_state);
2199                 if (nr_done != nr_deltas)
2200                         die("inconsistency with delta count");
2201         }
2202         free(delta_list);
2203 }
2204
2205 static int git_pack_config(const char *k, const char *v, void *cb)
2206 {
2207         if (!strcmp(k, "pack.window")) {
2208                 window = git_config_int(k, v);
2209                 return 0;
2210         }
2211         if (!strcmp(k, "pack.windowmemory")) {
2212                 window_memory_limit = git_config_ulong(k, v);
2213                 return 0;
2214         }
2215         if (!strcmp(k, "pack.depth")) {
2216                 depth = git_config_int(k, v);
2217                 return 0;
2218         }
2219         if (!strcmp(k, "pack.compression")) {
2220                 int level = git_config_int(k, v);
2221                 if (level == -1)
2222                         level = Z_DEFAULT_COMPRESSION;
2223                 else if (level < 0 || level > Z_BEST_COMPRESSION)
2224                         die("bad pack compression level %d", level);
2225                 pack_compression_level = level;
2226                 pack_compression_seen = 1;
2227                 return 0;
2228         }
2229         if (!strcmp(k, "pack.deltacachesize")) {
2230                 max_delta_cache_size = git_config_int(k, v);
2231                 return 0;
2232         }
2233         if (!strcmp(k, "pack.deltacachelimit")) {
2234                 cache_max_small_delta_size = git_config_int(k, v);
2235                 return 0;
2236         }
2237         if (!strcmp(k, "pack.writebitmaphashcache")) {
2238                 if (git_config_bool(k, v))
2239                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2240                 else
2241                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2242         }
2243         if (!strcmp(k, "pack.usebitmaps")) {
2244                 use_bitmap_index = git_config_bool(k, v);
2245                 return 0;
2246         }
2247         if (!strcmp(k, "pack.threads")) {
2248                 delta_search_threads = git_config_int(k, v);
2249                 if (delta_search_threads < 0)
2250                         die("invalid number of threads specified (%d)",
2251                             delta_search_threads);
2252 #ifdef NO_PTHREADS
2253                 if (delta_search_threads != 1)
2254                         warning("no threads support, ignoring %s", k);
2255 #endif
2256                 return 0;
2257         }
2258         if (!strcmp(k, "pack.indexversion")) {
2259                 pack_idx_opts.version = git_config_int(k, v);
2260                 if (pack_idx_opts.version > 2)
2261                         die("bad pack.indexversion=%"PRIu32,
2262                             pack_idx_opts.version);
2263                 return 0;
2264         }
2265         return git_default_config(k, v, cb);
2266 }
2267
2268 static void read_object_list_from_stdin(void)
2269 {
2270         char line[40 + 1 + PATH_MAX + 2];
2271         unsigned char sha1[20];
2272
2273         for (;;) {
2274                 if (!fgets(line, sizeof(line), stdin)) {
2275                         if (feof(stdin))
2276                                 break;
2277                         if (!ferror(stdin))
2278                                 die("fgets returned NULL, not EOF, not error!");
2279                         if (errno != EINTR)
2280                                 die_errno("fgets");
2281                         clearerr(stdin);
2282                         continue;
2283                 }
2284                 if (line[0] == '-') {
2285                         if (get_sha1_hex(line+1, sha1))
2286                                 die("expected edge sha1, got garbage:\n %s",
2287                                     line);
2288                         add_preferred_base(sha1);
2289                         continue;
2290                 }
2291                 if (get_sha1_hex(line, sha1))
2292                         die("expected sha1, got garbage:\n %s", line);
2293
2294                 add_preferred_base_object(line+41);
2295                 add_object_entry(sha1, 0, line+41, 0);
2296         }
2297 }
2298
2299 #define OBJECT_ADDED (1u<<20)
2300
2301 static void show_commit(struct commit *commit, void *data)
2302 {
2303         add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2304         commit->object.flags |= OBJECT_ADDED;
2305
2306         if (write_bitmap_index)
2307                 index_commit_for_bitmap(commit);
2308 }
2309
2310 static void show_object(struct object *obj, const char *name, void *data)
2311 {
2312         add_preferred_base_object(name);
2313         add_object_entry(obj->oid.hash, obj->type, name, 0);
2314         obj->flags |= OBJECT_ADDED;
2315 }
2316
2317 static void show_edge(struct commit *commit)
2318 {
2319         add_preferred_base(commit->object.oid.hash);
2320 }
2321
2322 struct in_pack_object {
2323         off_t offset;
2324         struct object *object;
2325 };
2326
2327 struct in_pack {
2328         int alloc;
2329         int nr;
2330         struct in_pack_object *array;
2331 };
2332
2333 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2334 {
2335         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2336         in_pack->array[in_pack->nr].object = object;
2337         in_pack->nr++;
2338 }
2339
2340 /*
2341  * Compare the objects in the offset order, in order to emulate the
2342  * "git rev-list --objects" output that produced the pack originally.
2343  */
2344 static int ofscmp(const void *a_, const void *b_)
2345 {
2346         struct in_pack_object *a = (struct in_pack_object *)a_;
2347         struct in_pack_object *b = (struct in_pack_object *)b_;
2348
2349         if (a->offset < b->offset)
2350                 return -1;
2351         else if (a->offset > b->offset)
2352                 return 1;
2353         else
2354                 return oidcmp(&a->object->oid, &b->object->oid);
2355 }
2356
2357 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2358 {
2359         struct packed_git *p;
2360         struct in_pack in_pack;
2361         uint32_t i;
2362
2363         memset(&in_pack, 0, sizeof(in_pack));
2364
2365         for (p = packed_git; p; p = p->next) {
2366                 const unsigned char *sha1;
2367                 struct object *o;
2368
2369                 if (!p->pack_local || p->pack_keep)
2370                         continue;
2371                 if (open_pack_index(p))
2372                         die("cannot open pack index");
2373
2374                 ALLOC_GROW(in_pack.array,
2375                            in_pack.nr + p->num_objects,
2376                            in_pack.alloc);
2377
2378                 for (i = 0; i < p->num_objects; i++) {
2379                         sha1 = nth_packed_object_sha1(p, i);
2380                         o = lookup_unknown_object(sha1);
2381                         if (!(o->flags & OBJECT_ADDED))
2382                                 mark_in_pack_object(o, p, &in_pack);
2383                         o->flags |= OBJECT_ADDED;
2384                 }
2385         }
2386
2387         if (in_pack.nr) {
2388                 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2389                       ofscmp);
2390                 for (i = 0; i < in_pack.nr; i++) {
2391                         struct object *o = in_pack.array[i].object;
2392                         add_object_entry(o->oid.hash, o->type, "", 0);
2393                 }
2394         }
2395         free(in_pack.array);
2396 }
2397
2398 static int add_loose_object(const unsigned char *sha1, const char *path,
2399                             void *data)
2400 {
2401         enum object_type type = sha1_object_info(sha1, NULL);
2402
2403         if (type < 0) {
2404                 warning("loose object at %s could not be examined", path);
2405                 return 0;
2406         }
2407
2408         add_object_entry(sha1, type, "", 0);
2409         return 0;
2410 }
2411
2412 /*
2413  * We actually don't even have to worry about reachability here.
2414  * add_object_entry will weed out duplicates, so we just add every
2415  * loose object we find.
2416  */
2417 static void add_unreachable_loose_objects(void)
2418 {
2419         for_each_loose_file_in_objdir(get_object_directory(),
2420                                       add_loose_object,
2421                                       NULL, NULL, NULL);
2422 }
2423
2424 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2425 {
2426         static struct packed_git *last_found = (void *)1;
2427         struct packed_git *p;
2428
2429         p = (last_found != (void *)1) ? last_found : packed_git;
2430
2431         while (p) {
2432                 if ((!p->pack_local || p->pack_keep) &&
2433                         find_pack_entry_one(sha1, p)) {
2434                         last_found = p;
2435                         return 1;
2436                 }
2437                 if (p == last_found)
2438                         p = packed_git;
2439                 else
2440                         p = p->next;
2441                 if (p == last_found)
2442                         p = p->next;
2443         }
2444         return 0;
2445 }
2446
2447 /*
2448  * Store a list of sha1s that are should not be discarded
2449  * because they are either written too recently, or are
2450  * reachable from another object that was.
2451  *
2452  * This is filled by get_object_list.
2453  */
2454 static struct sha1_array recent_objects;
2455
2456 static int loosened_object_can_be_discarded(const unsigned char *sha1,
2457                                             unsigned long mtime)
2458 {
2459         if (!unpack_unreachable_expiration)
2460                 return 0;
2461         if (mtime > unpack_unreachable_expiration)
2462                 return 0;
2463         if (sha1_array_lookup(&recent_objects, sha1) >= 0)
2464                 return 0;
2465         return 1;
2466 }
2467
2468 static void loosen_unused_packed_objects(struct rev_info *revs)
2469 {
2470         struct packed_git *p;
2471         uint32_t i;
2472         const unsigned char *sha1;
2473
2474         for (p = packed_git; p; p = p->next) {
2475                 if (!p->pack_local || p->pack_keep)
2476                         continue;
2477
2478                 if (open_pack_index(p))
2479                         die("cannot open pack index");
2480
2481                 for (i = 0; i < p->num_objects; i++) {
2482                         sha1 = nth_packed_object_sha1(p, i);
2483                         if (!packlist_find(&to_pack, sha1, NULL) &&
2484                             !has_sha1_pack_kept_or_nonlocal(sha1) &&
2485                             !loosened_object_can_be_discarded(sha1, p->mtime))
2486                                 if (force_object_loose(sha1, p->mtime))
2487                                         die("unable to force loose object");
2488                 }
2489         }
2490 }
2491
2492 /*
2493  * This tracks any options which a reader of the pack might
2494  * not understand, and which would therefore prevent blind reuse
2495  * of what we have on disk.
2496  */
2497 static int pack_options_allow_reuse(void)
2498 {
2499         return allow_ofs_delta;
2500 }
2501
2502 static int get_object_list_from_bitmap(struct rev_info *revs)
2503 {
2504         if (prepare_bitmap_walk(revs) < 0)
2505                 return -1;
2506
2507         if (pack_options_allow_reuse() &&
2508             !reuse_partial_packfile_from_bitmap(
2509                         &reuse_packfile,
2510                         &reuse_packfile_objects,
2511                         &reuse_packfile_offset)) {
2512                 assert(reuse_packfile_objects);
2513                 nr_result += reuse_packfile_objects;
2514                 display_progress(progress_state, nr_result);
2515         }
2516
2517         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2518         return 0;
2519 }
2520
2521 static void record_recent_object(struct object *obj,
2522                                  const char *name,
2523                                  void *data)
2524 {
2525         sha1_array_append(&recent_objects, obj->oid.hash);
2526 }
2527
2528 static void record_recent_commit(struct commit *commit, void *data)
2529 {
2530         sha1_array_append(&recent_objects, commit->object.oid.hash);
2531 }
2532
2533 static void get_object_list(int ac, const char **av)
2534 {
2535         struct rev_info revs;
2536         char line[1000];
2537         int flags = 0;
2538
2539         init_revisions(&revs, NULL);
2540         save_commit_buffer = 0;
2541         setup_revisions(ac, av, &revs, NULL);
2542
2543         /* make sure shallows are read */
2544         is_repository_shallow();
2545
2546         while (fgets(line, sizeof(line), stdin) != NULL) {
2547                 int len = strlen(line);
2548                 if (len && line[len - 1] == '\n')
2549                         line[--len] = 0;
2550                 if (!len)
2551                         break;
2552                 if (*line == '-') {
2553                         if (!strcmp(line, "--not")) {
2554                                 flags ^= UNINTERESTING;
2555                                 write_bitmap_index = 0;
2556                                 continue;
2557                         }
2558                         if (starts_with(line, "--shallow ")) {
2559                                 unsigned char sha1[20];
2560                                 if (get_sha1_hex(line + 10, sha1))
2561                                         die("not an SHA-1 '%s'", line + 10);
2562                                 register_shallow(sha1);
2563                                 use_bitmap_index = 0;
2564                                 continue;
2565                         }
2566                         die("not a rev '%s'", line);
2567                 }
2568                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2569                         die("bad revision '%s'", line);
2570         }
2571
2572         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2573                 return;
2574
2575         if (prepare_revision_walk(&revs))
2576                 die("revision walk setup failed");
2577         mark_edges_uninteresting(&revs, show_edge);
2578         traverse_commit_list(&revs, show_commit, show_object, NULL);
2579
2580         if (unpack_unreachable_expiration) {
2581                 revs.ignore_missing_links = 1;
2582                 if (add_unseen_recent_objects_to_traversal(&revs,
2583                                 unpack_unreachable_expiration))
2584                         die("unable to add recent objects");
2585                 if (prepare_revision_walk(&revs))
2586                         die("revision walk setup failed");
2587                 traverse_commit_list(&revs, record_recent_commit,
2588                                      record_recent_object, NULL);
2589         }
2590
2591         if (keep_unreachable)
2592                 add_objects_in_unpacked_packs(&revs);
2593         if (pack_loose_unreachable)
2594                 add_unreachable_loose_objects();
2595         if (unpack_unreachable)
2596                 loosen_unused_packed_objects(&revs);
2597
2598         sha1_array_clear(&recent_objects);
2599 }
2600
2601 static int option_parse_index_version(const struct option *opt,
2602                                       const char *arg, int unset)
2603 {
2604         char *c;
2605         const char *val = arg;
2606         pack_idx_opts.version = strtoul(val, &c, 10);
2607         if (pack_idx_opts.version > 2)
2608                 die(_("unsupported index version %s"), val);
2609         if (*c == ',' && c[1])
2610                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2611         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2612                 die(_("bad index version '%s'"), val);
2613         return 0;
2614 }
2615
2616 static int option_parse_unpack_unreachable(const struct option *opt,
2617                                            const char *arg, int unset)
2618 {
2619         if (unset) {
2620                 unpack_unreachable = 0;
2621                 unpack_unreachable_expiration = 0;
2622         }
2623         else {
2624                 unpack_unreachable = 1;
2625                 if (arg)
2626                         unpack_unreachable_expiration = approxidate(arg);
2627         }
2628         return 0;
2629 }
2630
2631 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2632 {
2633         int use_internal_rev_list = 0;
2634         int thin = 0;
2635         int shallow = 0;
2636         int all_progress_implied = 0;
2637         struct argv_array rp = ARGV_ARRAY_INIT;
2638         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2639         int rev_list_index = 0;
2640         struct option pack_objects_options[] = {
2641                 OPT_SET_INT('q', "quiet", &progress,
2642                             N_("do not show progress meter"), 0),
2643                 OPT_SET_INT(0, "progress", &progress,
2644                             N_("show progress meter"), 1),
2645                 OPT_SET_INT(0, "all-progress", &progress,
2646                             N_("show progress meter during object writing phase"), 2),
2647                 OPT_BOOL(0, "all-progress-implied",
2648                          &all_progress_implied,
2649                          N_("similar to --all-progress when progress meter is shown")),
2650                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2651                   N_("write the pack index file in the specified idx format version"),
2652                   0, option_parse_index_version },
2653                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2654                               N_("maximum size of each output pack file")),
2655                 OPT_BOOL(0, "local", &local,
2656                          N_("ignore borrowed objects from alternate object store")),
2657                 OPT_BOOL(0, "incremental", &incremental,
2658                          N_("ignore packed objects")),
2659                 OPT_INTEGER(0, "window", &window,
2660                             N_("limit pack window by objects")),
2661                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2662                               N_("limit pack window by memory in addition to object limit")),
2663                 OPT_INTEGER(0, "depth", &depth,
2664                             N_("maximum length of delta chain allowed in the resulting pack")),
2665                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2666                          N_("reuse existing deltas")),
2667                 OPT_BOOL(0, "reuse-object", &reuse_object,
2668                          N_("reuse existing objects")),
2669                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2670                          N_("use OFS_DELTA objects")),
2671                 OPT_INTEGER(0, "threads", &delta_search_threads,
2672                             N_("use threads when searching for best delta matches")),
2673                 OPT_BOOL(0, "non-empty", &non_empty,
2674                          N_("do not create an empty pack output")),
2675                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2676                          N_("read revision arguments from standard input")),
2677                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2678                   N_("limit the objects to those that are not yet packed"),
2679                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2680                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2681                   N_("include objects reachable from any reference"),
2682                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2683                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2684                   N_("include objects referred by reflog entries"),
2685                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2686                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2687                   N_("include objects referred to by the index"),
2688                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2689                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2690                          N_("output pack to stdout")),
2691                 OPT_BOOL(0, "include-tag", &include_tag,
2692                          N_("include tag objects that refer to objects to be packed")),
2693                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2694                          N_("keep unreachable objects")),
2695                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
2696                          N_("pack loose unreachable objects")),
2697                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2698                   N_("unpack unreachable objects newer than <time>"),
2699                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2700                 OPT_BOOL(0, "thin", &thin,
2701                          N_("create thin packs")),
2702                 OPT_BOOL(0, "shallow", &shallow,
2703                          N_("create packs suitable for shallow fetches")),
2704                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2705                          N_("ignore packs that have companion .keep file")),
2706                 OPT_INTEGER(0, "compression", &pack_compression_level,
2707                             N_("pack compression level")),
2708                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2709                             N_("do not hide commits by grafts"), 0),
2710                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2711                          N_("use a bitmap index if available to speed up counting objects")),
2712                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2713                          N_("write a bitmap index together with the pack index")),
2714                 OPT_END(),
2715         };
2716
2717         check_replace_refs = 0;
2718
2719         reset_pack_idx_option(&pack_idx_opts);
2720         git_config(git_pack_config, NULL);
2721         if (!pack_compression_seen && core_compression_seen)
2722                 pack_compression_level = core_compression_level;
2723
2724         progress = isatty(2);
2725         argc = parse_options(argc, argv, prefix, pack_objects_options,
2726                              pack_usage, 0);
2727
2728         if (argc) {
2729                 base_name = argv[0];
2730                 argc--;
2731         }
2732         if (pack_to_stdout != !base_name || argc)
2733                 usage_with_options(pack_usage, pack_objects_options);
2734
2735         argv_array_push(&rp, "pack-objects");
2736         if (thin) {
2737                 use_internal_rev_list = 1;
2738                 argv_array_push(&rp, shallow
2739                                 ? "--objects-edge-aggressive"
2740                                 : "--objects-edge");
2741         } else
2742                 argv_array_push(&rp, "--objects");
2743
2744         if (rev_list_all) {
2745                 use_internal_rev_list = 1;
2746                 argv_array_push(&rp, "--all");
2747         }
2748         if (rev_list_reflog) {
2749                 use_internal_rev_list = 1;
2750                 argv_array_push(&rp, "--reflog");
2751         }
2752         if (rev_list_index) {
2753                 use_internal_rev_list = 1;
2754                 argv_array_push(&rp, "--indexed-objects");
2755         }
2756         if (rev_list_unpacked) {
2757                 use_internal_rev_list = 1;
2758                 argv_array_push(&rp, "--unpacked");
2759         }
2760
2761         if (!reuse_object)
2762                 reuse_delta = 0;
2763         if (pack_compression_level == -1)
2764                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2765         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2766                 die("bad pack compression level %d", pack_compression_level);
2767
2768         if (!delta_search_threads)      /* --threads=0 means autodetect */
2769                 delta_search_threads = online_cpus();
2770
2771 #ifdef NO_PTHREADS
2772         if (delta_search_threads != 1)
2773                 warning("no threads support, ignoring --threads");
2774 #endif
2775         if (!pack_to_stdout && !pack_size_limit)
2776                 pack_size_limit = pack_size_limit_cfg;
2777         if (pack_to_stdout && pack_size_limit)
2778                 die("--max-pack-size cannot be used to build a pack for transfer.");
2779         if (pack_size_limit && pack_size_limit < 1024*1024) {
2780                 warning("minimum pack size limit is 1 MiB");
2781                 pack_size_limit = 1024*1024;
2782         }
2783
2784         if (!pack_to_stdout && thin)
2785                 die("--thin cannot be used to build an indexable pack.");
2786
2787         if (keep_unreachable && unpack_unreachable)
2788                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2789         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
2790                 unpack_unreachable_expiration = 0;
2791
2792         if (!use_internal_rev_list || !pack_to_stdout || is_repository_shallow())
2793                 use_bitmap_index = 0;
2794
2795         if (pack_to_stdout || !rev_list_all)
2796                 write_bitmap_index = 0;
2797
2798         if (progress && all_progress_implied)
2799                 progress = 2;
2800
2801         prepare_packed_git();
2802
2803         if (progress)
2804                 progress_state = start_progress(_("Counting objects"), 0);
2805         if (!use_internal_rev_list)
2806                 read_object_list_from_stdin();
2807         else {
2808                 get_object_list(rp.argc, rp.argv);
2809                 argv_array_clear(&rp);
2810         }
2811         cleanup_preferred_base();
2812         if (include_tag && nr_result)
2813                 for_each_ref(add_ref_tag, NULL);
2814         stop_progress(&progress_state);
2815
2816         if (non_empty && !nr_result)
2817                 return 0;
2818         if (nr_result)
2819                 prepare_pack(window, depth);
2820         write_pack_file();
2821         if (progress)
2822                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2823                         " reused %"PRIu32" (delta %"PRIu32")\n",
2824                         written, written_delta, reused, reused_delta);
2825         return 0;
2826 }