Merge branch 'jc/submodule-anchor-git-dir'
[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 have_non_local_packs;
50 static int incremental;
51 static int ignore_packed_keep;
52 static int allow_ofs_delta;
53 static struct pack_idx_option pack_idx_opts;
54 static const char *base_name;
55 static int progress = 1;
56 static int window = 10;
57 static unsigned long pack_size_limit;
58 static int depth = 50;
59 static int delta_search_threads;
60 static int pack_to_stdout;
61 static int num_preferred_base;
62 static struct progress *progress_state;
63 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
64 static int pack_compression_seen;
65
66 static struct packed_git *reuse_packfile;
67 static uint32_t reuse_packfile_objects;
68 static off_t reuse_packfile_offset;
69
70 static int use_bitmap_index = 1;
71 static int write_bitmap_index;
72 static uint16_t write_bitmap_options;
73
74 static unsigned long delta_cache_size = 0;
75 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
76 static unsigned long cache_max_small_delta_size = 1000;
77
78 static unsigned long window_memory_limit = 0;
79
80 /*
81  * stats
82  */
83 static uint32_t written, written_delta;
84 static uint32_t reused, reused_delta;
85
86 /*
87  * Indexed commits
88  */
89 static struct commit **indexed_commits;
90 static unsigned int indexed_commits_nr;
91 static unsigned int indexed_commits_alloc;
92
93 static void index_commit_for_bitmap(struct commit *commit)
94 {
95         if (indexed_commits_nr >= indexed_commits_alloc) {
96                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
97                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
98         }
99
100         indexed_commits[indexed_commits_nr++] = commit;
101 }
102
103 static void *get_delta(struct object_entry *entry)
104 {
105         unsigned long size, base_size, delta_size;
106         void *buf, *base_buf, *delta_buf;
107         enum object_type type;
108
109         buf = read_sha1_file(entry->idx.sha1, &type, &size);
110         if (!buf)
111                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
112         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
113         if (!base_buf)
114                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
115         delta_buf = diff_delta(base_buf, base_size,
116                                buf, size, &delta_size, 0);
117         if (!delta_buf || delta_size != entry->delta_size)
118                 die("delta size changed");
119         free(buf);
120         free(base_buf);
121         return delta_buf;
122 }
123
124 static unsigned long do_compress(void **pptr, unsigned long size)
125 {
126         git_zstream stream;
127         void *in, *out;
128         unsigned long maxsize;
129
130         git_deflate_init(&stream, pack_compression_level);
131         maxsize = git_deflate_bound(&stream, size);
132
133         in = *pptr;
134         out = xmalloc(maxsize);
135         *pptr = out;
136
137         stream.next_in = in;
138         stream.avail_in = size;
139         stream.next_out = out;
140         stream.avail_out = maxsize;
141         while (git_deflate(&stream, Z_FINISH) == Z_OK)
142                 ; /* nothing */
143         git_deflate_end(&stream);
144
145         free(in);
146         return stream.total_out;
147 }
148
149 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
150                                            const unsigned char *sha1)
151 {
152         git_zstream stream;
153         unsigned char ibuf[1024 * 16];
154         unsigned char obuf[1024 * 16];
155         unsigned long olen = 0;
156
157         git_deflate_init(&stream, pack_compression_level);
158
159         for (;;) {
160                 ssize_t readlen;
161                 int zret = Z_OK;
162                 readlen = read_istream(st, ibuf, sizeof(ibuf));
163                 if (readlen == -1)
164                         die(_("unable to read %s"), sha1_to_hex(sha1));
165
166                 stream.next_in = ibuf;
167                 stream.avail_in = readlen;
168                 while ((stream.avail_in || readlen == 0) &&
169                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
170                         stream.next_out = obuf;
171                         stream.avail_out = sizeof(obuf);
172                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
173                         sha1write(f, obuf, stream.next_out - obuf);
174                         olen += stream.next_out - obuf;
175                 }
176                 if (stream.avail_in)
177                         die(_("deflate error (%d)"), zret);
178                 if (readlen == 0) {
179                         if (zret != Z_STREAM_END)
180                                 die(_("deflate error (%d)"), zret);
181                         break;
182                 }
183         }
184         git_deflate_end(&stream);
185         return olen;
186 }
187
188 /*
189  * we are going to reuse the existing object data as is.  make
190  * sure it is not corrupt.
191  */
192 static int check_pack_inflate(struct packed_git *p,
193                 struct pack_window **w_curs,
194                 off_t offset,
195                 off_t len,
196                 unsigned long expect)
197 {
198         git_zstream stream;
199         unsigned char fakebuf[4096], *in;
200         int st;
201
202         memset(&stream, 0, sizeof(stream));
203         git_inflate_init(&stream);
204         do {
205                 in = use_pack(p, w_curs, offset, &stream.avail_in);
206                 stream.next_in = in;
207                 stream.next_out = fakebuf;
208                 stream.avail_out = sizeof(fakebuf);
209                 st = git_inflate(&stream, Z_FINISH);
210                 offset += stream.next_in - in;
211         } while (st == Z_OK || st == Z_BUF_ERROR);
212         git_inflate_end(&stream);
213         return (st == Z_STREAM_END &&
214                 stream.total_out == expect &&
215                 stream.total_in == len) ? 0 : -1;
216 }
217
218 static void copy_pack_data(struct sha1file *f,
219                 struct packed_git *p,
220                 struct pack_window **w_curs,
221                 off_t offset,
222                 off_t len)
223 {
224         unsigned char *in;
225         unsigned long avail;
226
227         while (len) {
228                 in = use_pack(p, w_curs, offset, &avail);
229                 if (avail > len)
230                         avail = (unsigned long)len;
231                 sha1write(f, in, avail);
232                 offset += avail;
233                 len -= avail;
234         }
235 }
236
237 /* Return 0 if we will bust the pack-size limit */
238 static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
239                                            unsigned long limit, int usable_delta)
240 {
241         unsigned long size, datalen;
242         unsigned char header[10], dheader[10];
243         unsigned hdrlen;
244         enum object_type type;
245         void *buf;
246         struct git_istream *st = NULL;
247
248         if (!usable_delta) {
249                 if (entry->type == OBJ_BLOB &&
250                     entry->size > big_file_threshold &&
251                     (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
252                         buf = NULL;
253                 else {
254                         buf = read_sha1_file(entry->idx.sha1, &type, &size);
255                         if (!buf)
256                                 die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
257                 }
258                 /*
259                  * make sure no cached delta data remains from a
260                  * previous attempt before a pack split occurred.
261                  */
262                 free(entry->delta_data);
263                 entry->delta_data = NULL;
264                 entry->z_delta_size = 0;
265         } else if (entry->delta_data) {
266                 size = entry->delta_size;
267                 buf = entry->delta_data;
268                 entry->delta_data = NULL;
269                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
270                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
271         } else {
272                 buf = get_delta(entry);
273                 size = entry->delta_size;
274                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
275                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
276         }
277
278         if (st) /* large blob case, just assume we don't compress well */
279                 datalen = size;
280         else if (entry->z_delta_size)
281                 datalen = entry->z_delta_size;
282         else
283                 datalen = do_compress(&buf, size);
284
285         /*
286          * The object header is a byte of 'type' followed by zero or
287          * more bytes of length.
288          */
289         hdrlen = encode_in_pack_object_header(type, size, header);
290
291         if (type == OBJ_OFS_DELTA) {
292                 /*
293                  * Deltas with relative base contain an additional
294                  * encoding of the relative offset for the delta
295                  * base from this object's position in the pack.
296                  */
297                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
298                 unsigned pos = sizeof(dheader) - 1;
299                 dheader[pos] = ofs & 127;
300                 while (ofs >>= 7)
301                         dheader[--pos] = 128 | (--ofs & 127);
302                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
303                         if (st)
304                                 close_istream(st);
305                         free(buf);
306                         return 0;
307                 }
308                 sha1write(f, header, hdrlen);
309                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
310                 hdrlen += sizeof(dheader) - pos;
311         } else if (type == OBJ_REF_DELTA) {
312                 /*
313                  * Deltas with a base reference contain
314                  * an additional 20 bytes for the base sha1.
315                  */
316                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
317                         if (st)
318                                 close_istream(st);
319                         free(buf);
320                         return 0;
321                 }
322                 sha1write(f, header, hdrlen);
323                 sha1write(f, entry->delta->idx.sha1, 20);
324                 hdrlen += 20;
325         } else {
326                 if (limit && hdrlen + datalen + 20 >= limit) {
327                         if (st)
328                                 close_istream(st);
329                         free(buf);
330                         return 0;
331                 }
332                 sha1write(f, header, hdrlen);
333         }
334         if (st) {
335                 datalen = write_large_blob_data(st, f, entry->idx.sha1);
336                 close_istream(st);
337         } else {
338                 sha1write(f, buf, datalen);
339                 free(buf);
340         }
341
342         return hdrlen + datalen;
343 }
344
345 /* Return 0 if we will bust the pack-size limit */
346 static off_t write_reuse_object(struct sha1file *f, struct object_entry *entry,
347                                 unsigned long limit, int usable_delta)
348 {
349         struct packed_git *p = entry->in_pack;
350         struct pack_window *w_curs = NULL;
351         struct revindex_entry *revidx;
352         off_t offset;
353         enum object_type type = entry->type;
354         off_t datalen;
355         unsigned char header[10], dheader[10];
356         unsigned hdrlen;
357
358         if (entry->delta)
359                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
360                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
361         hdrlen = encode_in_pack_object_header(type, entry->size, header);
362
363         offset = entry->in_pack_offset;
364         revidx = find_pack_revindex(p, offset);
365         datalen = revidx[1].offset - offset;
366         if (!pack_to_stdout && p->index_version > 1 &&
367             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
368                 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
369                 unuse_pack(&w_curs);
370                 return write_no_reuse_object(f, entry, limit, usable_delta);
371         }
372
373         offset += entry->in_pack_header_size;
374         datalen -= entry->in_pack_header_size;
375
376         if (!pack_to_stdout && p->index_version == 1 &&
377             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
378                 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
379                 unuse_pack(&w_curs);
380                 return write_no_reuse_object(f, entry, limit, usable_delta);
381         }
382
383         if (type == OBJ_OFS_DELTA) {
384                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
385                 unsigned pos = sizeof(dheader) - 1;
386                 dheader[pos] = ofs & 127;
387                 while (ofs >>= 7)
388                         dheader[--pos] = 128 | (--ofs & 127);
389                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
390                         unuse_pack(&w_curs);
391                         return 0;
392                 }
393                 sha1write(f, header, hdrlen);
394                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
395                 hdrlen += sizeof(dheader) - pos;
396                 reused_delta++;
397         } else if (type == OBJ_REF_DELTA) {
398                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
399                         unuse_pack(&w_curs);
400                         return 0;
401                 }
402                 sha1write(f, header, hdrlen);
403                 sha1write(f, entry->delta->idx.sha1, 20);
404                 hdrlen += 20;
405                 reused_delta++;
406         } else {
407                 if (limit && hdrlen + datalen + 20 >= limit) {
408                         unuse_pack(&w_curs);
409                         return 0;
410                 }
411                 sha1write(f, header, hdrlen);
412         }
413         copy_pack_data(f, p, &w_curs, offset, datalen);
414         unuse_pack(&w_curs);
415         reused++;
416         return hdrlen + datalen;
417 }
418
419 /* Return 0 if we will bust the pack-size limit */
420 static off_t write_object(struct sha1file *f,
421                           struct object_entry *entry,
422                           off_t write_offset)
423 {
424         unsigned long limit;
425         off_t len;
426         int usable_delta, to_reuse;
427
428         if (!pack_to_stdout)
429                 crc32_begin(f);
430
431         /* apply size limit if limited packsize and not first object */
432         if (!pack_size_limit || !nr_written)
433                 limit = 0;
434         else if (pack_size_limit <= write_offset)
435                 /*
436                  * the earlier object did not fit the limit; avoid
437                  * mistaking this with unlimited (i.e. limit = 0).
438                  */
439                 limit = 1;
440         else
441                 limit = pack_size_limit - write_offset;
442
443         if (!entry->delta)
444                 usable_delta = 0;       /* no delta */
445         else if (!pack_size_limit)
446                usable_delta = 1;        /* unlimited packfile */
447         else if (entry->delta->idx.offset == (off_t)-1)
448                 usable_delta = 0;       /* base was written to another pack */
449         else if (entry->delta->idx.offset)
450                 usable_delta = 1;       /* base already exists in this pack */
451         else
452                 usable_delta = 0;       /* base could end up in another pack */
453
454         if (!reuse_object)
455                 to_reuse = 0;   /* explicit */
456         else if (!entry->in_pack)
457                 to_reuse = 0;   /* can't reuse what we don't have */
458         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
459                                 /* check_object() decided it for us ... */
460                 to_reuse = usable_delta;
461                                 /* ... but pack split may override that */
462         else if (entry->type != entry->in_pack_type)
463                 to_reuse = 0;   /* pack has delta which is unusable */
464         else if (entry->delta)
465                 to_reuse = 0;   /* we want to pack afresh */
466         else
467                 to_reuse = 1;   /* we have it in-pack undeltified,
468                                  * and we do not need to deltify it.
469                                  */
470
471         if (!to_reuse)
472                 len = write_no_reuse_object(f, entry, limit, usable_delta);
473         else
474                 len = write_reuse_object(f, entry, limit, usable_delta);
475         if (!len)
476                 return 0;
477
478         if (usable_delta)
479                 written_delta++;
480         written++;
481         if (!pack_to_stdout)
482                 entry->idx.crc32 = crc32_end(f);
483         return len;
484 }
485
486 enum write_one_status {
487         WRITE_ONE_SKIP = -1, /* already written */
488         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
489         WRITE_ONE_WRITTEN = 1, /* normal */
490         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
491 };
492
493 static enum write_one_status write_one(struct sha1file *f,
494                                        struct object_entry *e,
495                                        off_t *offset)
496 {
497         off_t size;
498         int recursing;
499
500         /*
501          * we set offset to 1 (which is an impossible value) to mark
502          * the fact that this object is involved in "write its base
503          * first before writing a deltified object" recursion.
504          */
505         recursing = (e->idx.offset == 1);
506         if (recursing) {
507                 warning("recursive delta detected for object %s",
508                         sha1_to_hex(e->idx.sha1));
509                 return WRITE_ONE_RECURSIVE;
510         } else if (e->idx.offset || e->preferred_base) {
511                 /* offset is non zero if object is written already. */
512                 return WRITE_ONE_SKIP;
513         }
514
515         /* if we are deltified, write out base object first. */
516         if (e->delta) {
517                 e->idx.offset = 1; /* now recurse */
518                 switch (write_one(f, e->delta, offset)) {
519                 case WRITE_ONE_RECURSIVE:
520                         /* we cannot depend on this one */
521                         e->delta = NULL;
522                         break;
523                 default:
524                         break;
525                 case WRITE_ONE_BREAK:
526                         e->idx.offset = recursing;
527                         return WRITE_ONE_BREAK;
528                 }
529         }
530
531         e->idx.offset = *offset;
532         size = write_object(f, e, *offset);
533         if (!size) {
534                 e->idx.offset = recursing;
535                 return WRITE_ONE_BREAK;
536         }
537         written_list[nr_written++] = &e->idx;
538
539         /* make sure off_t is sufficiently large not to wrap */
540         if (signed_add_overflows(*offset, size))
541                 die("pack too large for current definition of off_t");
542         *offset += size;
543         return WRITE_ONE_WRITTEN;
544 }
545
546 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
547                        void *cb_data)
548 {
549         unsigned char peeled[20];
550         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
551
552         if (entry)
553                 entry->tagged = 1;
554         if (!peel_ref(path, peeled)) {
555                 entry = packlist_find(&to_pack, peeled, NULL);
556                 if (entry)
557                         entry->tagged = 1;
558         }
559         return 0;
560 }
561
562 static inline void add_to_write_order(struct object_entry **wo,
563                                unsigned int *endp,
564                                struct object_entry *e)
565 {
566         if (e->filled)
567                 return;
568         wo[(*endp)++] = e;
569         e->filled = 1;
570 }
571
572 static void add_descendants_to_write_order(struct object_entry **wo,
573                                            unsigned int *endp,
574                                            struct object_entry *e)
575 {
576         int add_to_order = 1;
577         while (e) {
578                 if (add_to_order) {
579                         struct object_entry *s;
580                         /* add this node... */
581                         add_to_write_order(wo, endp, e);
582                         /* all its siblings... */
583                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
584                                 add_to_write_order(wo, endp, s);
585                         }
586                 }
587                 /* drop down a level to add left subtree nodes if possible */
588                 if (e->delta_child) {
589                         add_to_order = 1;
590                         e = e->delta_child;
591                 } else {
592                         add_to_order = 0;
593                         /* our sibling might have some children, it is next */
594                         if (e->delta_sibling) {
595                                 e = e->delta_sibling;
596                                 continue;
597                         }
598                         /* go back to our parent node */
599                         e = e->delta;
600                         while (e && !e->delta_sibling) {
601                                 /* we're on the right side of a subtree, keep
602                                  * going up until we can go right again */
603                                 e = e->delta;
604                         }
605                         if (!e) {
606                                 /* done- we hit our original root node */
607                                 return;
608                         }
609                         /* pass it off to sibling at this level */
610                         e = e->delta_sibling;
611                 }
612         };
613 }
614
615 static void add_family_to_write_order(struct object_entry **wo,
616                                       unsigned int *endp,
617                                       struct object_entry *e)
618 {
619         struct object_entry *root;
620
621         for (root = e; root->delta; root = root->delta)
622                 ; /* nothing */
623         add_descendants_to_write_order(wo, endp, root);
624 }
625
626 static struct object_entry **compute_write_order(void)
627 {
628         unsigned int i, wo_end, last_untagged;
629
630         struct object_entry **wo;
631         struct object_entry *objects = to_pack.objects;
632
633         for (i = 0; i < to_pack.nr_objects; i++) {
634                 objects[i].tagged = 0;
635                 objects[i].filled = 0;
636                 objects[i].delta_child = NULL;
637                 objects[i].delta_sibling = NULL;
638         }
639
640         /*
641          * Fully connect delta_child/delta_sibling network.
642          * Make sure delta_sibling is sorted in the original
643          * recency order.
644          */
645         for (i = to_pack.nr_objects; i > 0;) {
646                 struct object_entry *e = &objects[--i];
647                 if (!e->delta)
648                         continue;
649                 /* Mark me as the first child */
650                 e->delta_sibling = e->delta->delta_child;
651                 e->delta->delta_child = e;
652         }
653
654         /*
655          * Mark objects that are at the tip of tags.
656          */
657         for_each_tag_ref(mark_tagged, NULL);
658
659         /*
660          * Give the objects in the original recency order until
661          * we see a tagged tip.
662          */
663         ALLOC_ARRAY(wo, to_pack.nr_objects);
664         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
665                 if (objects[i].tagged)
666                         break;
667                 add_to_write_order(wo, &wo_end, &objects[i]);
668         }
669         last_untagged = i;
670
671         /*
672          * Then fill all the tagged tips.
673          */
674         for (; i < to_pack.nr_objects; i++) {
675                 if (objects[i].tagged)
676                         add_to_write_order(wo, &wo_end, &objects[i]);
677         }
678
679         /*
680          * And then all remaining commits and tags.
681          */
682         for (i = last_untagged; i < to_pack.nr_objects; i++) {
683                 if (objects[i].type != OBJ_COMMIT &&
684                     objects[i].type != OBJ_TAG)
685                         continue;
686                 add_to_write_order(wo, &wo_end, &objects[i]);
687         }
688
689         /*
690          * And then all the trees.
691          */
692         for (i = last_untagged; i < to_pack.nr_objects; i++) {
693                 if (objects[i].type != OBJ_TREE)
694                         continue;
695                 add_to_write_order(wo, &wo_end, &objects[i]);
696         }
697
698         /*
699          * Finally all the rest in really tight order
700          */
701         for (i = last_untagged; i < to_pack.nr_objects; i++) {
702                 if (!objects[i].filled)
703                         add_family_to_write_order(wo, &wo_end, &objects[i]);
704         }
705
706         if (wo_end != to_pack.nr_objects)
707                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
708
709         return wo;
710 }
711
712 static off_t write_reused_pack(struct sha1file *f)
713 {
714         unsigned char buffer[8192];
715         off_t to_write, total;
716         int fd;
717
718         if (!is_pack_valid(reuse_packfile))
719                 die("packfile is invalid: %s", reuse_packfile->pack_name);
720
721         fd = git_open_noatime(reuse_packfile->pack_name);
722         if (fd < 0)
723                 die_errno("unable to open packfile for reuse: %s",
724                           reuse_packfile->pack_name);
725
726         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
727                 die_errno("unable to seek in reused packfile");
728
729         if (reuse_packfile_offset < 0)
730                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
731
732         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
733
734         while (to_write) {
735                 int read_pack = xread(fd, buffer, sizeof(buffer));
736
737                 if (read_pack <= 0)
738                         die_errno("unable to read from reused packfile");
739
740                 if (read_pack > to_write)
741                         read_pack = to_write;
742
743                 sha1write(f, buffer, read_pack);
744                 to_write -= read_pack;
745
746                 /*
747                  * We don't know the actual number of objects written,
748                  * only how many bytes written, how many bytes total, and
749                  * how many objects total. So we can fake it by pretending all
750                  * objects we are writing are the same size. This gives us a
751                  * smooth progress meter, and at the end it matches the true
752                  * answer.
753                  */
754                 written = reuse_packfile_objects *
755                                 (((double)(total - to_write)) / total);
756                 display_progress(progress_state, written);
757         }
758
759         close(fd);
760         written = reuse_packfile_objects;
761         display_progress(progress_state, written);
762         return reuse_packfile_offset - sizeof(struct pack_header);
763 }
764
765 static const char no_split_warning[] = N_(
766 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
767 );
768
769 static void write_pack_file(void)
770 {
771         uint32_t i = 0, j;
772         struct sha1file *f;
773         off_t offset;
774         uint32_t nr_remaining = nr_result;
775         time_t last_mtime = 0;
776         struct object_entry **write_order;
777
778         if (progress > pack_to_stdout)
779                 progress_state = start_progress(_("Writing objects"), nr_result);
780         ALLOC_ARRAY(written_list, to_pack.nr_objects);
781         write_order = compute_write_order();
782
783         do {
784                 unsigned char sha1[20];
785                 char *pack_tmp_name = NULL;
786
787                 if (pack_to_stdout)
788                         f = sha1fd_throughput(1, "<stdout>", progress_state);
789                 else
790                         f = create_tmp_packfile(&pack_tmp_name);
791
792                 offset = write_pack_header(f, nr_remaining);
793
794                 if (reuse_packfile) {
795                         off_t packfile_size;
796                         assert(pack_to_stdout);
797
798                         packfile_size = write_reused_pack(f);
799                         offset += packfile_size;
800                 }
801
802                 nr_written = 0;
803                 for (; i < to_pack.nr_objects; i++) {
804                         struct object_entry *e = write_order[i];
805                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
806                                 break;
807                         display_progress(progress_state, written);
808                 }
809
810                 /*
811                  * Did we write the wrong # entries in the header?
812                  * If so, rewrite it like in fast-import
813                  */
814                 if (pack_to_stdout) {
815                         sha1close(f, sha1, CSUM_CLOSE);
816                 } else if (nr_written == nr_remaining) {
817                         sha1close(f, sha1, CSUM_FSYNC);
818                 } else {
819                         int fd = sha1close(f, sha1, 0);
820                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
821                                                  nr_written, sha1, offset);
822                         close(fd);
823                         if (write_bitmap_index) {
824                                 warning(_(no_split_warning));
825                                 write_bitmap_index = 0;
826                         }
827                 }
828
829                 if (!pack_to_stdout) {
830                         struct stat st;
831                         struct strbuf tmpname = STRBUF_INIT;
832
833                         /*
834                          * Packs are runtime accessed in their mtime
835                          * order since newer packs are more likely to contain
836                          * younger objects.  So if we are creating multiple
837                          * packs then we should modify the mtime of later ones
838                          * to preserve this property.
839                          */
840                         if (stat(pack_tmp_name, &st) < 0) {
841                                 warning_errno("failed to stat %s", pack_tmp_name);
842                         } else if (!last_mtime) {
843                                 last_mtime = st.st_mtime;
844                         } else {
845                                 struct utimbuf utb;
846                                 utb.actime = st.st_atime;
847                                 utb.modtime = --last_mtime;
848                                 if (utime(pack_tmp_name, &utb) < 0)
849                                         warning_errno("failed utime() on %s", pack_tmp_name);
850                         }
851
852                         strbuf_addf(&tmpname, "%s-", base_name);
853
854                         if (write_bitmap_index) {
855                                 bitmap_writer_set_checksum(sha1);
856                                 bitmap_writer_build_type_index(written_list, nr_written);
857                         }
858
859                         finish_tmp_packfile(&tmpname, pack_tmp_name,
860                                             written_list, nr_written,
861                                             &pack_idx_opts, sha1);
862
863                         if (write_bitmap_index) {
864                                 strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
865
866                                 stop_progress(&progress_state);
867
868                                 bitmap_writer_show_progress(progress);
869                                 bitmap_writer_reuse_bitmaps(&to_pack);
870                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
871                                 bitmap_writer_build(&to_pack);
872                                 bitmap_writer_finish(written_list, nr_written,
873                                                      tmpname.buf, write_bitmap_options);
874                                 write_bitmap_index = 0;
875                         }
876
877                         strbuf_release(&tmpname);
878                         free(pack_tmp_name);
879                         puts(sha1_to_hex(sha1));
880                 }
881
882                 /* mark written objects as written to previous pack */
883                 for (j = 0; j < nr_written; j++) {
884                         written_list[j]->offset = (off_t)-1;
885                 }
886                 nr_remaining -= nr_written;
887         } while (nr_remaining && i < to_pack.nr_objects);
888
889         free(written_list);
890         free(write_order);
891         stop_progress(&progress_state);
892         if (written != nr_result)
893                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
894                         written, nr_result);
895 }
896
897 static void setup_delta_attr_check(struct git_attr_check *check)
898 {
899         static struct git_attr *attr_delta;
900
901         if (!attr_delta)
902                 attr_delta = git_attr("delta");
903
904         check[0].attr = attr_delta;
905 }
906
907 static int no_try_delta(const char *path)
908 {
909         struct git_attr_check check[1];
910
911         setup_delta_attr_check(check);
912         if (git_check_attr(path, ARRAY_SIZE(check), check))
913                 return 0;
914         if (ATTR_FALSE(check->value))
915                 return 1;
916         return 0;
917 }
918
919 /*
920  * When adding an object, check whether we have already added it
921  * to our packing list. If so, we can skip. However, if we are
922  * being asked to excludei t, but the previous mention was to include
923  * it, make sure to adjust its flags and tweak our numbers accordingly.
924  *
925  * As an optimization, we pass out the index position where we would have
926  * found the item, since that saves us from having to look it up again a
927  * few lines later when we want to add the new entry.
928  */
929 static int have_duplicate_entry(const unsigned char *sha1,
930                                 int exclude,
931                                 uint32_t *index_pos)
932 {
933         struct object_entry *entry;
934
935         entry = packlist_find(&to_pack, sha1, index_pos);
936         if (!entry)
937                 return 0;
938
939         if (exclude) {
940                 if (!entry->preferred_base)
941                         nr_result--;
942                 entry->preferred_base = 1;
943         }
944
945         return 1;
946 }
947
948 /*
949  * Check whether we want the object in the pack (e.g., we do not want
950  * objects found in non-local stores if the "--local" option was used).
951  *
952  * As a side effect of this check, we will find the packed version of this
953  * object, if any. We therefore pass out the pack information to avoid having
954  * to look it up again later.
955  */
956 static int want_object_in_pack(const unsigned char *sha1,
957                                int exclude,
958                                struct packed_git **found_pack,
959                                off_t *found_offset)
960 {
961         struct packed_git *p;
962
963         if (!exclude && local && has_loose_object_nonlocal(sha1))
964                 return 0;
965
966         *found_pack = NULL;
967         *found_offset = 0;
968
969         for (p = packed_git; p; p = p->next) {
970                 off_t offset = find_pack_entry_one(sha1, p);
971                 if (offset) {
972                         if (!*found_pack) {
973                                 if (!is_pack_valid(p))
974                                         continue;
975                                 *found_offset = offset;
976                                 *found_pack = p;
977                         }
978                         if (exclude)
979                                 return 1;
980                         if (incremental)
981                                 return 0;
982
983                         /*
984                          * When asked to do --local (do not include an
985                          * object that appears in a pack we borrow
986                          * from elsewhere) or --honor-pack-keep (do not
987                          * include an object that appears in a pack marked
988                          * with .keep), we need to make sure no copy of this
989                          * object come from in _any_ pack that causes us to
990                          * omit it, and need to complete this loop.  When
991                          * neither option is in effect, we know the object
992                          * we just found is going to be packed, so break
993                          * out of the loop to return 1 now.
994                          */
995                         if (!ignore_packed_keep &&
996                             (!local || !have_non_local_packs))
997                                 break;
998
999                         if (local && !p->pack_local)
1000                                 return 0;
1001                         if (ignore_packed_keep && p->pack_local && p->pack_keep)
1002                                 return 0;
1003                 }
1004         }
1005
1006         return 1;
1007 }
1008
1009 static void create_object_entry(const unsigned char *sha1,
1010                                 enum object_type type,
1011                                 uint32_t hash,
1012                                 int exclude,
1013                                 int no_try_delta,
1014                                 uint32_t index_pos,
1015                                 struct packed_git *found_pack,
1016                                 off_t found_offset)
1017 {
1018         struct object_entry *entry;
1019
1020         entry = packlist_alloc(&to_pack, sha1, index_pos);
1021         entry->hash = hash;
1022         if (type)
1023                 entry->type = type;
1024         if (exclude)
1025                 entry->preferred_base = 1;
1026         else
1027                 nr_result++;
1028         if (found_pack) {
1029                 entry->in_pack = found_pack;
1030                 entry->in_pack_offset = found_offset;
1031         }
1032
1033         entry->no_try_delta = no_try_delta;
1034 }
1035
1036 static const char no_closure_warning[] = N_(
1037 "disabling bitmap writing, as some objects are not being packed"
1038 );
1039
1040 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1041                             const char *name, int exclude)
1042 {
1043         struct packed_git *found_pack;
1044         off_t found_offset;
1045         uint32_t index_pos;
1046
1047         if (have_duplicate_entry(sha1, exclude, &index_pos))
1048                 return 0;
1049
1050         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1051                 /* The pack is missing an object, so it will not have closure */
1052                 if (write_bitmap_index) {
1053                         warning(_(no_closure_warning));
1054                         write_bitmap_index = 0;
1055                 }
1056                 return 0;
1057         }
1058
1059         create_object_entry(sha1, type, pack_name_hash(name),
1060                             exclude, name && no_try_delta(name),
1061                             index_pos, found_pack, found_offset);
1062
1063         display_progress(progress_state, nr_result);
1064         return 1;
1065 }
1066
1067 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1068                                         enum object_type type,
1069                                         int flags, uint32_t name_hash,
1070                                         struct packed_git *pack, off_t offset)
1071 {
1072         uint32_t index_pos;
1073
1074         if (have_duplicate_entry(sha1, 0, &index_pos))
1075                 return 0;
1076
1077         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1078
1079         display_progress(progress_state, nr_result);
1080         return 1;
1081 }
1082
1083 struct pbase_tree_cache {
1084         unsigned char sha1[20];
1085         int ref;
1086         int temporary;
1087         void *tree_data;
1088         unsigned long tree_size;
1089 };
1090
1091 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1092 static int pbase_tree_cache_ix(const unsigned char *sha1)
1093 {
1094         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1095 }
1096 static int pbase_tree_cache_ix_incr(int ix)
1097 {
1098         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1099 }
1100
1101 static struct pbase_tree {
1102         struct pbase_tree *next;
1103         /* This is a phony "cache" entry; we are not
1104          * going to evict it or find it through _get()
1105          * mechanism -- this is for the toplevel node that
1106          * would almost always change with any commit.
1107          */
1108         struct pbase_tree_cache pcache;
1109 } *pbase_tree;
1110
1111 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1112 {
1113         struct pbase_tree_cache *ent, *nent;
1114         void *data;
1115         unsigned long size;
1116         enum object_type type;
1117         int neigh;
1118         int my_ix = pbase_tree_cache_ix(sha1);
1119         int available_ix = -1;
1120
1121         /* pbase-tree-cache acts as a limited hashtable.
1122          * your object will be found at your index or within a few
1123          * slots after that slot if it is cached.
1124          */
1125         for (neigh = 0; neigh < 8; neigh++) {
1126                 ent = pbase_tree_cache[my_ix];
1127                 if (ent && !hashcmp(ent->sha1, sha1)) {
1128                         ent->ref++;
1129                         return ent;
1130                 }
1131                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1132                          ((0 <= available_ix) &&
1133                           (!ent && pbase_tree_cache[available_ix])))
1134                         available_ix = my_ix;
1135                 if (!ent)
1136                         break;
1137                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1138         }
1139
1140         /* Did not find one.  Either we got a bogus request or
1141          * we need to read and perhaps cache.
1142          */
1143         data = read_sha1_file(sha1, &type, &size);
1144         if (!data)
1145                 return NULL;
1146         if (type != OBJ_TREE) {
1147                 free(data);
1148                 return NULL;
1149         }
1150
1151         /* We need to either cache or return a throwaway copy */
1152
1153         if (available_ix < 0)
1154                 ent = NULL;
1155         else {
1156                 ent = pbase_tree_cache[available_ix];
1157                 my_ix = available_ix;
1158         }
1159
1160         if (!ent) {
1161                 nent = xmalloc(sizeof(*nent));
1162                 nent->temporary = (available_ix < 0);
1163         }
1164         else {
1165                 /* evict and reuse */
1166                 free(ent->tree_data);
1167                 nent = ent;
1168         }
1169         hashcpy(nent->sha1, sha1);
1170         nent->tree_data = data;
1171         nent->tree_size = size;
1172         nent->ref = 1;
1173         if (!nent->temporary)
1174                 pbase_tree_cache[my_ix] = nent;
1175         return nent;
1176 }
1177
1178 static void pbase_tree_put(struct pbase_tree_cache *cache)
1179 {
1180         if (!cache->temporary) {
1181                 cache->ref--;
1182                 return;
1183         }
1184         free(cache->tree_data);
1185         free(cache);
1186 }
1187
1188 static int name_cmp_len(const char *name)
1189 {
1190         int i;
1191         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1192                 ;
1193         return i;
1194 }
1195
1196 static void add_pbase_object(struct tree_desc *tree,
1197                              const char *name,
1198                              int cmplen,
1199                              const char *fullname)
1200 {
1201         struct name_entry entry;
1202         int cmp;
1203
1204         while (tree_entry(tree,&entry)) {
1205                 if (S_ISGITLINK(entry.mode))
1206                         continue;
1207                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1208                       memcmp(name, entry.path, cmplen);
1209                 if (cmp > 0)
1210                         continue;
1211                 if (cmp < 0)
1212                         return;
1213                 if (name[cmplen] != '/') {
1214                         add_object_entry(entry.oid->hash,
1215                                          object_type(entry.mode),
1216                                          fullname, 1);
1217                         return;
1218                 }
1219                 if (S_ISDIR(entry.mode)) {
1220                         struct tree_desc sub;
1221                         struct pbase_tree_cache *tree;
1222                         const char *down = name+cmplen+1;
1223                         int downlen = name_cmp_len(down);
1224
1225                         tree = pbase_tree_get(entry.oid->hash);
1226                         if (!tree)
1227                                 return;
1228                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1229
1230                         add_pbase_object(&sub, down, downlen, fullname);
1231                         pbase_tree_put(tree);
1232                 }
1233         }
1234 }
1235
1236 static unsigned *done_pbase_paths;
1237 static int done_pbase_paths_num;
1238 static int done_pbase_paths_alloc;
1239 static int done_pbase_path_pos(unsigned hash)
1240 {
1241         int lo = 0;
1242         int hi = done_pbase_paths_num;
1243         while (lo < hi) {
1244                 int mi = (hi + lo) / 2;
1245                 if (done_pbase_paths[mi] == hash)
1246                         return mi;
1247                 if (done_pbase_paths[mi] < hash)
1248                         hi = mi;
1249                 else
1250                         lo = mi + 1;
1251         }
1252         return -lo-1;
1253 }
1254
1255 static int check_pbase_path(unsigned hash)
1256 {
1257         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1258         if (0 <= pos)
1259                 return 1;
1260         pos = -pos - 1;
1261         ALLOC_GROW(done_pbase_paths,
1262                    done_pbase_paths_num + 1,
1263                    done_pbase_paths_alloc);
1264         done_pbase_paths_num++;
1265         if (pos < done_pbase_paths_num)
1266                 memmove(done_pbase_paths + pos + 1,
1267                         done_pbase_paths + pos,
1268                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1269         done_pbase_paths[pos] = hash;
1270         return 0;
1271 }
1272
1273 static void add_preferred_base_object(const char *name)
1274 {
1275         struct pbase_tree *it;
1276         int cmplen;
1277         unsigned hash = pack_name_hash(name);
1278
1279         if (!num_preferred_base || check_pbase_path(hash))
1280                 return;
1281
1282         cmplen = name_cmp_len(name);
1283         for (it = pbase_tree; it; it = it->next) {
1284                 if (cmplen == 0) {
1285                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1286                 }
1287                 else {
1288                         struct tree_desc tree;
1289                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1290                         add_pbase_object(&tree, name, cmplen, name);
1291                 }
1292         }
1293 }
1294
1295 static void add_preferred_base(unsigned char *sha1)
1296 {
1297         struct pbase_tree *it;
1298         void *data;
1299         unsigned long size;
1300         unsigned char tree_sha1[20];
1301
1302         if (window <= num_preferred_base++)
1303                 return;
1304
1305         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1306         if (!data)
1307                 return;
1308
1309         for (it = pbase_tree; it; it = it->next) {
1310                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1311                         free(data);
1312                         return;
1313                 }
1314         }
1315
1316         it = xcalloc(1, sizeof(*it));
1317         it->next = pbase_tree;
1318         pbase_tree = it;
1319
1320         hashcpy(it->pcache.sha1, tree_sha1);
1321         it->pcache.tree_data = data;
1322         it->pcache.tree_size = size;
1323 }
1324
1325 static void cleanup_preferred_base(void)
1326 {
1327         struct pbase_tree *it;
1328         unsigned i;
1329
1330         it = pbase_tree;
1331         pbase_tree = NULL;
1332         while (it) {
1333                 struct pbase_tree *this = it;
1334                 it = this->next;
1335                 free(this->pcache.tree_data);
1336                 free(this);
1337         }
1338
1339         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1340                 if (!pbase_tree_cache[i])
1341                         continue;
1342                 free(pbase_tree_cache[i]->tree_data);
1343                 free(pbase_tree_cache[i]);
1344                 pbase_tree_cache[i] = NULL;
1345         }
1346
1347         free(done_pbase_paths);
1348         done_pbase_paths = NULL;
1349         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1350 }
1351
1352 static void check_object(struct object_entry *entry)
1353 {
1354         if (entry->in_pack) {
1355                 struct packed_git *p = entry->in_pack;
1356                 struct pack_window *w_curs = NULL;
1357                 const unsigned char *base_ref = NULL;
1358                 struct object_entry *base_entry;
1359                 unsigned long used, used_0;
1360                 unsigned long avail;
1361                 off_t ofs;
1362                 unsigned char *buf, c;
1363
1364                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1365
1366                 /*
1367                  * We want in_pack_type even if we do not reuse delta
1368                  * since non-delta representations could still be reused.
1369                  */
1370                 used = unpack_object_header_buffer(buf, avail,
1371                                                    &entry->in_pack_type,
1372                                                    &entry->size);
1373                 if (used == 0)
1374                         goto give_up;
1375
1376                 /*
1377                  * Determine if this is a delta and if so whether we can
1378                  * reuse it or not.  Otherwise let's find out as cheaply as
1379                  * possible what the actual type and size for this object is.
1380                  */
1381                 switch (entry->in_pack_type) {
1382                 default:
1383                         /* Not a delta hence we've already got all we need. */
1384                         entry->type = entry->in_pack_type;
1385                         entry->in_pack_header_size = used;
1386                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1387                                 goto give_up;
1388                         unuse_pack(&w_curs);
1389                         return;
1390                 case OBJ_REF_DELTA:
1391                         if (reuse_delta && !entry->preferred_base)
1392                                 base_ref = use_pack(p, &w_curs,
1393                                                 entry->in_pack_offset + used, NULL);
1394                         entry->in_pack_header_size = used + 20;
1395                         break;
1396                 case OBJ_OFS_DELTA:
1397                         buf = use_pack(p, &w_curs,
1398                                        entry->in_pack_offset + used, NULL);
1399                         used_0 = 0;
1400                         c = buf[used_0++];
1401                         ofs = c & 127;
1402                         while (c & 128) {
1403                                 ofs += 1;
1404                                 if (!ofs || MSB(ofs, 7)) {
1405                                         error("delta base offset overflow in pack for %s",
1406                                               sha1_to_hex(entry->idx.sha1));
1407                                         goto give_up;
1408                                 }
1409                                 c = buf[used_0++];
1410                                 ofs = (ofs << 7) + (c & 127);
1411                         }
1412                         ofs = entry->in_pack_offset - ofs;
1413                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1414                                 error("delta base offset out of bound for %s",
1415                                       sha1_to_hex(entry->idx.sha1));
1416                                 goto give_up;
1417                         }
1418                         if (reuse_delta && !entry->preferred_base) {
1419                                 struct revindex_entry *revidx;
1420                                 revidx = find_pack_revindex(p, ofs);
1421                                 if (!revidx)
1422                                         goto give_up;
1423                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1424                         }
1425                         entry->in_pack_header_size = used + used_0;
1426                         break;
1427                 }
1428
1429                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1430                         /*
1431                          * If base_ref was set above that means we wish to
1432                          * reuse delta data, and we even found that base
1433                          * in the list of objects we want to pack. Goodie!
1434                          *
1435                          * Depth value does not matter - find_deltas() will
1436                          * never consider reused delta as the base object to
1437                          * deltify other objects against, in order to avoid
1438                          * circular deltas.
1439                          */
1440                         entry->type = entry->in_pack_type;
1441                         entry->delta = base_entry;
1442                         entry->delta_size = entry->size;
1443                         entry->delta_sibling = base_entry->delta_child;
1444                         base_entry->delta_child = entry;
1445                         unuse_pack(&w_curs);
1446                         return;
1447                 }
1448
1449                 if (entry->type) {
1450                         /*
1451                          * This must be a delta and we already know what the
1452                          * final object type is.  Let's extract the actual
1453                          * object size from the delta header.
1454                          */
1455                         entry->size = get_size_from_delta(p, &w_curs,
1456                                         entry->in_pack_offset + entry->in_pack_header_size);
1457                         if (entry->size == 0)
1458                                 goto give_up;
1459                         unuse_pack(&w_curs);
1460                         return;
1461                 }
1462
1463                 /*
1464                  * No choice but to fall back to the recursive delta walk
1465                  * with sha1_object_info() to find about the object type
1466                  * at this point...
1467                  */
1468                 give_up:
1469                 unuse_pack(&w_curs);
1470         }
1471
1472         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1473         /*
1474          * The error condition is checked in prepare_pack().  This is
1475          * to permit a missing preferred base object to be ignored
1476          * as a preferred base.  Doing so can result in a larger
1477          * pack file, but the transfer will still take place.
1478          */
1479 }
1480
1481 static int pack_offset_sort(const void *_a, const void *_b)
1482 {
1483         const struct object_entry *a = *(struct object_entry **)_a;
1484         const struct object_entry *b = *(struct object_entry **)_b;
1485
1486         /* avoid filesystem trashing with loose objects */
1487         if (!a->in_pack && !b->in_pack)
1488                 return hashcmp(a->idx.sha1, b->idx.sha1);
1489
1490         if (a->in_pack < b->in_pack)
1491                 return -1;
1492         if (a->in_pack > b->in_pack)
1493                 return 1;
1494         return a->in_pack_offset < b->in_pack_offset ? -1 :
1495                         (a->in_pack_offset > b->in_pack_offset);
1496 }
1497
1498 static void get_object_details(void)
1499 {
1500         uint32_t i;
1501         struct object_entry **sorted_by_offset;
1502
1503         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1504         for (i = 0; i < to_pack.nr_objects; i++)
1505                 sorted_by_offset[i] = to_pack.objects + i;
1506         qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1507
1508         for (i = 0; i < to_pack.nr_objects; i++) {
1509                 struct object_entry *entry = sorted_by_offset[i];
1510                 check_object(entry);
1511                 if (big_file_threshold < entry->size)
1512                         entry->no_try_delta = 1;
1513         }
1514
1515         free(sorted_by_offset);
1516 }
1517
1518 /*
1519  * We search for deltas in a list sorted by type, by filename hash, and then
1520  * by size, so that we see progressively smaller and smaller files.
1521  * That's because we prefer deltas to be from the bigger file
1522  * to the smaller -- deletes are potentially cheaper, but perhaps
1523  * more importantly, the bigger file is likely the more recent
1524  * one.  The deepest deltas are therefore the oldest objects which are
1525  * less susceptible to be accessed often.
1526  */
1527 static int type_size_sort(const void *_a, const void *_b)
1528 {
1529         const struct object_entry *a = *(struct object_entry **)_a;
1530         const struct object_entry *b = *(struct object_entry **)_b;
1531
1532         if (a->type > b->type)
1533                 return -1;
1534         if (a->type < b->type)
1535                 return 1;
1536         if (a->hash > b->hash)
1537                 return -1;
1538         if (a->hash < b->hash)
1539                 return 1;
1540         if (a->preferred_base > b->preferred_base)
1541                 return -1;
1542         if (a->preferred_base < b->preferred_base)
1543                 return 1;
1544         if (a->size > b->size)
1545                 return -1;
1546         if (a->size < b->size)
1547                 return 1;
1548         return a < b ? -1 : (a > b);  /* newest first */
1549 }
1550
1551 struct unpacked {
1552         struct object_entry *entry;
1553         void *data;
1554         struct delta_index *index;
1555         unsigned depth;
1556 };
1557
1558 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1559                            unsigned long delta_size)
1560 {
1561         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1562                 return 0;
1563
1564         if (delta_size < cache_max_small_delta_size)
1565                 return 1;
1566
1567         /* cache delta, if objects are large enough compared to delta size */
1568         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1569                 return 1;
1570
1571         return 0;
1572 }
1573
1574 #ifndef NO_PTHREADS
1575
1576 static pthread_mutex_t read_mutex;
1577 #define read_lock()             pthread_mutex_lock(&read_mutex)
1578 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1579
1580 static pthread_mutex_t cache_mutex;
1581 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1582 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1583
1584 static pthread_mutex_t progress_mutex;
1585 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1586 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1587
1588 #else
1589
1590 #define read_lock()             (void)0
1591 #define read_unlock()           (void)0
1592 #define cache_lock()            (void)0
1593 #define cache_unlock()          (void)0
1594 #define progress_lock()         (void)0
1595 #define progress_unlock()       (void)0
1596
1597 #endif
1598
1599 static int try_delta(struct unpacked *trg, struct unpacked *src,
1600                      unsigned max_depth, unsigned long *mem_usage)
1601 {
1602         struct object_entry *trg_entry = trg->entry;
1603         struct object_entry *src_entry = src->entry;
1604         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1605         unsigned ref_depth;
1606         enum object_type type;
1607         void *delta_buf;
1608
1609         /* Don't bother doing diffs between different types */
1610         if (trg_entry->type != src_entry->type)
1611                 return -1;
1612
1613         /*
1614          * We do not bother to try a delta that we discarded on an
1615          * earlier try, but only when reusing delta data.  Note that
1616          * src_entry that is marked as the preferred_base should always
1617          * be considered, as even if we produce a suboptimal delta against
1618          * it, we will still save the transfer cost, as we already know
1619          * the other side has it and we won't send src_entry at all.
1620          */
1621         if (reuse_delta && trg_entry->in_pack &&
1622             trg_entry->in_pack == src_entry->in_pack &&
1623             !src_entry->preferred_base &&
1624             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1625             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1626                 return 0;
1627
1628         /* Let's not bust the allowed depth. */
1629         if (src->depth >= max_depth)
1630                 return 0;
1631
1632         /* Now some size filtering heuristics. */
1633         trg_size = trg_entry->size;
1634         if (!trg_entry->delta) {
1635                 max_size = trg_size/2 - 20;
1636                 ref_depth = 1;
1637         } else {
1638                 max_size = trg_entry->delta_size;
1639                 ref_depth = trg->depth;
1640         }
1641         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1642                                                 (max_depth - ref_depth + 1);
1643         if (max_size == 0)
1644                 return 0;
1645         src_size = src_entry->size;
1646         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1647         if (sizediff >= max_size)
1648                 return 0;
1649         if (trg_size < src_size / 32)
1650                 return 0;
1651
1652         /* Load data if not already done */
1653         if (!trg->data) {
1654                 read_lock();
1655                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1656                 read_unlock();
1657                 if (!trg->data)
1658                         die("object %s cannot be read",
1659                             sha1_to_hex(trg_entry->idx.sha1));
1660                 if (sz != trg_size)
1661                         die("object %s inconsistent object length (%lu vs %lu)",
1662                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1663                 *mem_usage += sz;
1664         }
1665         if (!src->data) {
1666                 read_lock();
1667                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1668                 read_unlock();
1669                 if (!src->data) {
1670                         if (src_entry->preferred_base) {
1671                                 static int warned = 0;
1672                                 if (!warned++)
1673                                         warning("object %s cannot be read",
1674                                                 sha1_to_hex(src_entry->idx.sha1));
1675                                 /*
1676                                  * Those objects are not included in the
1677                                  * resulting pack.  Be resilient and ignore
1678                                  * them if they can't be read, in case the
1679                                  * pack could be created nevertheless.
1680                                  */
1681                                 return 0;
1682                         }
1683                         die("object %s cannot be read",
1684                             sha1_to_hex(src_entry->idx.sha1));
1685                 }
1686                 if (sz != src_size)
1687                         die("object %s inconsistent object length (%lu vs %lu)",
1688                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1689                 *mem_usage += sz;
1690         }
1691         if (!src->index) {
1692                 src->index = create_delta_index(src->data, src_size);
1693                 if (!src->index) {
1694                         static int warned = 0;
1695                         if (!warned++)
1696                                 warning("suboptimal pack - out of memory");
1697                         return 0;
1698                 }
1699                 *mem_usage += sizeof_delta_index(src->index);
1700         }
1701
1702         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1703         if (!delta_buf)
1704                 return 0;
1705
1706         if (trg_entry->delta) {
1707                 /* Prefer only shallower same-sized deltas. */
1708                 if (delta_size == trg_entry->delta_size &&
1709                     src->depth + 1 >= trg->depth) {
1710                         free(delta_buf);
1711                         return 0;
1712                 }
1713         }
1714
1715         /*
1716          * Handle memory allocation outside of the cache
1717          * accounting lock.  Compiler will optimize the strangeness
1718          * away when NO_PTHREADS is defined.
1719          */
1720         free(trg_entry->delta_data);
1721         cache_lock();
1722         if (trg_entry->delta_data) {
1723                 delta_cache_size -= trg_entry->delta_size;
1724                 trg_entry->delta_data = NULL;
1725         }
1726         if (delta_cacheable(src_size, trg_size, delta_size)) {
1727                 delta_cache_size += delta_size;
1728                 cache_unlock();
1729                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1730         } else {
1731                 cache_unlock();
1732                 free(delta_buf);
1733         }
1734
1735         trg_entry->delta = src_entry;
1736         trg_entry->delta_size = delta_size;
1737         trg->depth = src->depth + 1;
1738
1739         return 1;
1740 }
1741
1742 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1743 {
1744         struct object_entry *child = me->delta_child;
1745         unsigned int m = n;
1746         while (child) {
1747                 unsigned int c = check_delta_limit(child, n + 1);
1748                 if (m < c)
1749                         m = c;
1750                 child = child->delta_sibling;
1751         }
1752         return m;
1753 }
1754
1755 static unsigned long free_unpacked(struct unpacked *n)
1756 {
1757         unsigned long freed_mem = sizeof_delta_index(n->index);
1758         free_delta_index(n->index);
1759         n->index = NULL;
1760         if (n->data) {
1761                 freed_mem += n->entry->size;
1762                 free(n->data);
1763                 n->data = NULL;
1764         }
1765         n->entry = NULL;
1766         n->depth = 0;
1767         return freed_mem;
1768 }
1769
1770 static void find_deltas(struct object_entry **list, unsigned *list_size,
1771                         int window, int depth, unsigned *processed)
1772 {
1773         uint32_t i, idx = 0, count = 0;
1774         struct unpacked *array;
1775         unsigned long mem_usage = 0;
1776
1777         array = xcalloc(window, sizeof(struct unpacked));
1778
1779         for (;;) {
1780                 struct object_entry *entry;
1781                 struct unpacked *n = array + idx;
1782                 int j, max_depth, best_base = -1;
1783
1784                 progress_lock();
1785                 if (!*list_size) {
1786                         progress_unlock();
1787                         break;
1788                 }
1789                 entry = *list++;
1790                 (*list_size)--;
1791                 if (!entry->preferred_base) {
1792                         (*processed)++;
1793                         display_progress(progress_state, *processed);
1794                 }
1795                 progress_unlock();
1796
1797                 mem_usage -= free_unpacked(n);
1798                 n->entry = entry;
1799
1800                 while (window_memory_limit &&
1801                        mem_usage > window_memory_limit &&
1802                        count > 1) {
1803                         uint32_t tail = (idx + window - count) % window;
1804                         mem_usage -= free_unpacked(array + tail);
1805                         count--;
1806                 }
1807
1808                 /* We do not compute delta to *create* objects we are not
1809                  * going to pack.
1810                  */
1811                 if (entry->preferred_base)
1812                         goto next;
1813
1814                 /*
1815                  * If the current object is at pack edge, take the depth the
1816                  * objects that depend on the current object into account
1817                  * otherwise they would become too deep.
1818                  */
1819                 max_depth = depth;
1820                 if (entry->delta_child) {
1821                         max_depth -= check_delta_limit(entry, 0);
1822                         if (max_depth <= 0)
1823                                 goto next;
1824                 }
1825
1826                 j = window;
1827                 while (--j > 0) {
1828                         int ret;
1829                         uint32_t other_idx = idx + j;
1830                         struct unpacked *m;
1831                         if (other_idx >= window)
1832                                 other_idx -= window;
1833                         m = array + other_idx;
1834                         if (!m->entry)
1835                                 break;
1836                         ret = try_delta(n, m, max_depth, &mem_usage);
1837                         if (ret < 0)
1838                                 break;
1839                         else if (ret > 0)
1840                                 best_base = other_idx;
1841                 }
1842
1843                 /*
1844                  * If we decided to cache the delta data, then it is best
1845                  * to compress it right away.  First because we have to do
1846                  * it anyway, and doing it here while we're threaded will
1847                  * save a lot of time in the non threaded write phase,
1848                  * as well as allow for caching more deltas within
1849                  * the same cache size limit.
1850                  * ...
1851                  * But only if not writing to stdout, since in that case
1852                  * the network is most likely throttling writes anyway,
1853                  * and therefore it is best to go to the write phase ASAP
1854                  * instead, as we can afford spending more time compressing
1855                  * between writes at that moment.
1856                  */
1857                 if (entry->delta_data && !pack_to_stdout) {
1858                         entry->z_delta_size = do_compress(&entry->delta_data,
1859                                                           entry->delta_size);
1860                         cache_lock();
1861                         delta_cache_size -= entry->delta_size;
1862                         delta_cache_size += entry->z_delta_size;
1863                         cache_unlock();
1864                 }
1865
1866                 /* if we made n a delta, and if n is already at max
1867                  * depth, leaving it in the window is pointless.  we
1868                  * should evict it first.
1869                  */
1870                 if (entry->delta && max_depth <= n->depth)
1871                         continue;
1872
1873                 /*
1874                  * Move the best delta base up in the window, after the
1875                  * currently deltified object, to keep it longer.  It will
1876                  * be the first base object to be attempted next.
1877                  */
1878                 if (entry->delta) {
1879                         struct unpacked swap = array[best_base];
1880                         int dist = (window + idx - best_base) % window;
1881                         int dst = best_base;
1882                         while (dist--) {
1883                                 int src = (dst + 1) % window;
1884                                 array[dst] = array[src];
1885                                 dst = src;
1886                         }
1887                         array[dst] = swap;
1888                 }
1889
1890                 next:
1891                 idx++;
1892                 if (count + 1 < window)
1893                         count++;
1894                 if (idx >= window)
1895                         idx = 0;
1896         }
1897
1898         for (i = 0; i < window; ++i) {
1899                 free_delta_index(array[i].index);
1900                 free(array[i].data);
1901         }
1902         free(array);
1903 }
1904
1905 #ifndef NO_PTHREADS
1906
1907 static void try_to_free_from_threads(size_t size)
1908 {
1909         read_lock();
1910         release_pack_memory(size);
1911         read_unlock();
1912 }
1913
1914 static try_to_free_t old_try_to_free_routine;
1915
1916 /*
1917  * The main thread waits on the condition that (at least) one of the workers
1918  * has stopped working (which is indicated in the .working member of
1919  * struct thread_params).
1920  * When a work thread has completed its work, it sets .working to 0 and
1921  * signals the main thread and waits on the condition that .data_ready
1922  * becomes 1.
1923  */
1924
1925 struct thread_params {
1926         pthread_t thread;
1927         struct object_entry **list;
1928         unsigned list_size;
1929         unsigned remaining;
1930         int window;
1931         int depth;
1932         int working;
1933         int data_ready;
1934         pthread_mutex_t mutex;
1935         pthread_cond_t cond;
1936         unsigned *processed;
1937 };
1938
1939 static pthread_cond_t progress_cond;
1940
1941 /*
1942  * Mutex and conditional variable can't be statically-initialized on Windows.
1943  */
1944 static void init_threaded_search(void)
1945 {
1946         init_recursive_mutex(&read_mutex);
1947         pthread_mutex_init(&cache_mutex, NULL);
1948         pthread_mutex_init(&progress_mutex, NULL);
1949         pthread_cond_init(&progress_cond, NULL);
1950         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1951 }
1952
1953 static void cleanup_threaded_search(void)
1954 {
1955         set_try_to_free_routine(old_try_to_free_routine);
1956         pthread_cond_destroy(&progress_cond);
1957         pthread_mutex_destroy(&read_mutex);
1958         pthread_mutex_destroy(&cache_mutex);
1959         pthread_mutex_destroy(&progress_mutex);
1960 }
1961
1962 static void *threaded_find_deltas(void *arg)
1963 {
1964         struct thread_params *me = arg;
1965
1966         while (me->remaining) {
1967                 find_deltas(me->list, &me->remaining,
1968                             me->window, me->depth, me->processed);
1969
1970                 progress_lock();
1971                 me->working = 0;
1972                 pthread_cond_signal(&progress_cond);
1973                 progress_unlock();
1974
1975                 /*
1976                  * We must not set ->data_ready before we wait on the
1977                  * condition because the main thread may have set it to 1
1978                  * before we get here. In order to be sure that new
1979                  * work is available if we see 1 in ->data_ready, it
1980                  * was initialized to 0 before this thread was spawned
1981                  * and we reset it to 0 right away.
1982                  */
1983                 pthread_mutex_lock(&me->mutex);
1984                 while (!me->data_ready)
1985                         pthread_cond_wait(&me->cond, &me->mutex);
1986                 me->data_ready = 0;
1987                 pthread_mutex_unlock(&me->mutex);
1988         }
1989         /* leave ->working 1 so that this doesn't get more work assigned */
1990         return NULL;
1991 }
1992
1993 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1994                            int window, int depth, unsigned *processed)
1995 {
1996         struct thread_params *p;
1997         int i, ret, active_threads = 0;
1998
1999         init_threaded_search();
2000
2001         if (delta_search_threads <= 1) {
2002                 find_deltas(list, &list_size, window, depth, processed);
2003                 cleanup_threaded_search();
2004                 return;
2005         }
2006         if (progress > pack_to_stdout)
2007                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2008                                 delta_search_threads);
2009         p = xcalloc(delta_search_threads, sizeof(*p));
2010
2011         /* Partition the work amongst work threads. */
2012         for (i = 0; i < delta_search_threads; i++) {
2013                 unsigned sub_size = list_size / (delta_search_threads - i);
2014
2015                 /* don't use too small segments or no deltas will be found */
2016                 if (sub_size < 2*window && i+1 < delta_search_threads)
2017                         sub_size = 0;
2018
2019                 p[i].window = window;
2020                 p[i].depth = depth;
2021                 p[i].processed = processed;
2022                 p[i].working = 1;
2023                 p[i].data_ready = 0;
2024
2025                 /* try to split chunks on "path" boundaries */
2026                 while (sub_size && sub_size < list_size &&
2027                        list[sub_size]->hash &&
2028                        list[sub_size]->hash == list[sub_size-1]->hash)
2029                         sub_size++;
2030
2031                 p[i].list = list;
2032                 p[i].list_size = sub_size;
2033                 p[i].remaining = sub_size;
2034
2035                 list += sub_size;
2036                 list_size -= sub_size;
2037         }
2038
2039         /* Start work threads. */
2040         for (i = 0; i < delta_search_threads; i++) {
2041                 if (!p[i].list_size)
2042                         continue;
2043                 pthread_mutex_init(&p[i].mutex, NULL);
2044                 pthread_cond_init(&p[i].cond, NULL);
2045                 ret = pthread_create(&p[i].thread, NULL,
2046                                      threaded_find_deltas, &p[i]);
2047                 if (ret)
2048                         die("unable to create thread: %s", strerror(ret));
2049                 active_threads++;
2050         }
2051
2052         /*
2053          * Now let's wait for work completion.  Each time a thread is done
2054          * with its work, we steal half of the remaining work from the
2055          * thread with the largest number of unprocessed objects and give
2056          * it to that newly idle thread.  This ensure good load balancing
2057          * until the remaining object list segments are simply too short
2058          * to be worth splitting anymore.
2059          */
2060         while (active_threads) {
2061                 struct thread_params *target = NULL;
2062                 struct thread_params *victim = NULL;
2063                 unsigned sub_size = 0;
2064
2065                 progress_lock();
2066                 for (;;) {
2067                         for (i = 0; !target && i < delta_search_threads; i++)
2068                                 if (!p[i].working)
2069                                         target = &p[i];
2070                         if (target)
2071                                 break;
2072                         pthread_cond_wait(&progress_cond, &progress_mutex);
2073                 }
2074
2075                 for (i = 0; i < delta_search_threads; i++)
2076                         if (p[i].remaining > 2*window &&
2077                             (!victim || victim->remaining < p[i].remaining))
2078                                 victim = &p[i];
2079                 if (victim) {
2080                         sub_size = victim->remaining / 2;
2081                         list = victim->list + victim->list_size - sub_size;
2082                         while (sub_size && list[0]->hash &&
2083                                list[0]->hash == list[-1]->hash) {
2084                                 list++;
2085                                 sub_size--;
2086                         }
2087                         if (!sub_size) {
2088                                 /*
2089                                  * It is possible for some "paths" to have
2090                                  * so many objects that no hash boundary
2091                                  * might be found.  Let's just steal the
2092                                  * exact half in that case.
2093                                  */
2094                                 sub_size = victim->remaining / 2;
2095                                 list -= sub_size;
2096                         }
2097                         target->list = list;
2098                         victim->list_size -= sub_size;
2099                         victim->remaining -= sub_size;
2100                 }
2101                 target->list_size = sub_size;
2102                 target->remaining = sub_size;
2103                 target->working = 1;
2104                 progress_unlock();
2105
2106                 pthread_mutex_lock(&target->mutex);
2107                 target->data_ready = 1;
2108                 pthread_cond_signal(&target->cond);
2109                 pthread_mutex_unlock(&target->mutex);
2110
2111                 if (!sub_size) {
2112                         pthread_join(target->thread, NULL);
2113                         pthread_cond_destroy(&target->cond);
2114                         pthread_mutex_destroy(&target->mutex);
2115                         active_threads--;
2116                 }
2117         }
2118         cleanup_threaded_search();
2119         free(p);
2120 }
2121
2122 #else
2123 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2124 #endif
2125
2126 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2127 {
2128         struct object_id peeled;
2129
2130         if (starts_with(path, "refs/tags/") && /* is a tag? */
2131             !peel_ref(path, peeled.hash)    && /* peelable? */
2132             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2133                 add_object_entry(oid->hash, OBJ_TAG, NULL, 0);
2134         return 0;
2135 }
2136
2137 static void prepare_pack(int window, int depth)
2138 {
2139         struct object_entry **delta_list;
2140         uint32_t i, nr_deltas;
2141         unsigned n;
2142
2143         get_object_details();
2144
2145         /*
2146          * If we're locally repacking then we need to be doubly careful
2147          * from now on in order to make sure no stealth corruption gets
2148          * propagated to the new pack.  Clients receiving streamed packs
2149          * should validate everything they get anyway so no need to incur
2150          * the additional cost here in that case.
2151          */
2152         if (!pack_to_stdout)
2153                 do_check_packed_object_crc = 1;
2154
2155         if (!to_pack.nr_objects || !window || !depth)
2156                 return;
2157
2158         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2159         nr_deltas = n = 0;
2160
2161         for (i = 0; i < to_pack.nr_objects; i++) {
2162                 struct object_entry *entry = to_pack.objects + i;
2163
2164                 if (entry->delta)
2165                         /* This happens if we decided to reuse existing
2166                          * delta from a pack.  "reuse_delta &&" is implied.
2167                          */
2168                         continue;
2169
2170                 if (entry->size < 50)
2171                         continue;
2172
2173                 if (entry->no_try_delta)
2174                         continue;
2175
2176                 if (!entry->preferred_base) {
2177                         nr_deltas++;
2178                         if (entry->type < 0)
2179                                 die("unable to get type of object %s",
2180                                     sha1_to_hex(entry->idx.sha1));
2181                 } else {
2182                         if (entry->type < 0) {
2183                                 /*
2184                                  * This object is not found, but we
2185                                  * don't have to include it anyway.
2186                                  */
2187                                 continue;
2188                         }
2189                 }
2190
2191                 delta_list[n++] = entry;
2192         }
2193
2194         if (nr_deltas && n > 1) {
2195                 unsigned nr_done = 0;
2196                 if (progress)
2197                         progress_state = start_progress(_("Compressing objects"),
2198                                                         nr_deltas);
2199                 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2200                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2201                 stop_progress(&progress_state);
2202                 if (nr_done != nr_deltas)
2203                         die("inconsistency with delta count");
2204         }
2205         free(delta_list);
2206 }
2207
2208 static int git_pack_config(const char *k, const char *v, void *cb)
2209 {
2210         if (!strcmp(k, "pack.window")) {
2211                 window = git_config_int(k, v);
2212                 return 0;
2213         }
2214         if (!strcmp(k, "pack.windowmemory")) {
2215                 window_memory_limit = git_config_ulong(k, v);
2216                 return 0;
2217         }
2218         if (!strcmp(k, "pack.depth")) {
2219                 depth = git_config_int(k, v);
2220                 return 0;
2221         }
2222         if (!strcmp(k, "pack.compression")) {
2223                 int level = git_config_int(k, v);
2224                 if (level == -1)
2225                         level = Z_DEFAULT_COMPRESSION;
2226                 else if (level < 0 || level > Z_BEST_COMPRESSION)
2227                         die("bad pack compression level %d", level);
2228                 pack_compression_level = level;
2229                 pack_compression_seen = 1;
2230                 return 0;
2231         }
2232         if (!strcmp(k, "pack.deltacachesize")) {
2233                 max_delta_cache_size = git_config_int(k, v);
2234                 return 0;
2235         }
2236         if (!strcmp(k, "pack.deltacachelimit")) {
2237                 cache_max_small_delta_size = git_config_int(k, v);
2238                 return 0;
2239         }
2240         if (!strcmp(k, "pack.writebitmaphashcache")) {
2241                 if (git_config_bool(k, v))
2242                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2243                 else
2244                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2245         }
2246         if (!strcmp(k, "pack.usebitmaps")) {
2247                 use_bitmap_index = git_config_bool(k, v);
2248                 return 0;
2249         }
2250         if (!strcmp(k, "pack.threads")) {
2251                 delta_search_threads = git_config_int(k, v);
2252                 if (delta_search_threads < 0)
2253                         die("invalid number of threads specified (%d)",
2254                             delta_search_threads);
2255 #ifdef NO_PTHREADS
2256                 if (delta_search_threads != 1)
2257                         warning("no threads support, ignoring %s", k);
2258 #endif
2259                 return 0;
2260         }
2261         if (!strcmp(k, "pack.indexversion")) {
2262                 pack_idx_opts.version = git_config_int(k, v);
2263                 if (pack_idx_opts.version > 2)
2264                         die("bad pack.indexversion=%"PRIu32,
2265                             pack_idx_opts.version);
2266                 return 0;
2267         }
2268         return git_default_config(k, v, cb);
2269 }
2270
2271 static void read_object_list_from_stdin(void)
2272 {
2273         char line[40 + 1 + PATH_MAX + 2];
2274         unsigned char sha1[20];
2275
2276         for (;;) {
2277                 if (!fgets(line, sizeof(line), stdin)) {
2278                         if (feof(stdin))
2279                                 break;
2280                         if (!ferror(stdin))
2281                                 die("fgets returned NULL, not EOF, not error!");
2282                         if (errno != EINTR)
2283                                 die_errno("fgets");
2284                         clearerr(stdin);
2285                         continue;
2286                 }
2287                 if (line[0] == '-') {
2288                         if (get_sha1_hex(line+1, sha1))
2289                                 die("expected edge sha1, got garbage:\n %s",
2290                                     line);
2291                         add_preferred_base(sha1);
2292                         continue;
2293                 }
2294                 if (get_sha1_hex(line, sha1))
2295                         die("expected sha1, got garbage:\n %s", line);
2296
2297                 add_preferred_base_object(line+41);
2298                 add_object_entry(sha1, 0, line+41, 0);
2299         }
2300 }
2301
2302 #define OBJECT_ADDED (1u<<20)
2303
2304 static void show_commit(struct commit *commit, void *data)
2305 {
2306         add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2307         commit->object.flags |= OBJECT_ADDED;
2308
2309         if (write_bitmap_index)
2310                 index_commit_for_bitmap(commit);
2311 }
2312
2313 static void show_object(struct object *obj, const char *name, void *data)
2314 {
2315         add_preferred_base_object(name);
2316         add_object_entry(obj->oid.hash, obj->type, name, 0);
2317         obj->flags |= OBJECT_ADDED;
2318 }
2319
2320 static void show_edge(struct commit *commit)
2321 {
2322         add_preferred_base(commit->object.oid.hash);
2323 }
2324
2325 struct in_pack_object {
2326         off_t offset;
2327         struct object *object;
2328 };
2329
2330 struct in_pack {
2331         int alloc;
2332         int nr;
2333         struct in_pack_object *array;
2334 };
2335
2336 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2337 {
2338         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2339         in_pack->array[in_pack->nr].object = object;
2340         in_pack->nr++;
2341 }
2342
2343 /*
2344  * Compare the objects in the offset order, in order to emulate the
2345  * "git rev-list --objects" output that produced the pack originally.
2346  */
2347 static int ofscmp(const void *a_, const void *b_)
2348 {
2349         struct in_pack_object *a = (struct in_pack_object *)a_;
2350         struct in_pack_object *b = (struct in_pack_object *)b_;
2351
2352         if (a->offset < b->offset)
2353                 return -1;
2354         else if (a->offset > b->offset)
2355                 return 1;
2356         else
2357                 return oidcmp(&a->object->oid, &b->object->oid);
2358 }
2359
2360 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2361 {
2362         struct packed_git *p;
2363         struct in_pack in_pack;
2364         uint32_t i;
2365
2366         memset(&in_pack, 0, sizeof(in_pack));
2367
2368         for (p = packed_git; p; p = p->next) {
2369                 const unsigned char *sha1;
2370                 struct object *o;
2371
2372                 if (!p->pack_local || p->pack_keep)
2373                         continue;
2374                 if (open_pack_index(p))
2375                         die("cannot open pack index");
2376
2377                 ALLOC_GROW(in_pack.array,
2378                            in_pack.nr + p->num_objects,
2379                            in_pack.alloc);
2380
2381                 for (i = 0; i < p->num_objects; i++) {
2382                         sha1 = nth_packed_object_sha1(p, i);
2383                         o = lookup_unknown_object(sha1);
2384                         if (!(o->flags & OBJECT_ADDED))
2385                                 mark_in_pack_object(o, p, &in_pack);
2386                         o->flags |= OBJECT_ADDED;
2387                 }
2388         }
2389
2390         if (in_pack.nr) {
2391                 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2392                       ofscmp);
2393                 for (i = 0; i < in_pack.nr; i++) {
2394                         struct object *o = in_pack.array[i].object;
2395                         add_object_entry(o->oid.hash, o->type, "", 0);
2396                 }
2397         }
2398         free(in_pack.array);
2399 }
2400
2401 static int add_loose_object(const unsigned char *sha1, const char *path,
2402                             void *data)
2403 {
2404         enum object_type type = sha1_object_info(sha1, NULL);
2405
2406         if (type < 0) {
2407                 warning("loose object at %s could not be examined", path);
2408                 return 0;
2409         }
2410
2411         add_object_entry(sha1, type, "", 0);
2412         return 0;
2413 }
2414
2415 /*
2416  * We actually don't even have to worry about reachability here.
2417  * add_object_entry will weed out duplicates, so we just add every
2418  * loose object we find.
2419  */
2420 static void add_unreachable_loose_objects(void)
2421 {
2422         for_each_loose_file_in_objdir(get_object_directory(),
2423                                       add_loose_object,
2424                                       NULL, NULL, NULL);
2425 }
2426
2427 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2428 {
2429         static struct packed_git *last_found = (void *)1;
2430         struct packed_git *p;
2431
2432         p = (last_found != (void *)1) ? last_found : packed_git;
2433
2434         while (p) {
2435                 if ((!p->pack_local || p->pack_keep) &&
2436                         find_pack_entry_one(sha1, p)) {
2437                         last_found = p;
2438                         return 1;
2439                 }
2440                 if (p == last_found)
2441                         p = packed_git;
2442                 else
2443                         p = p->next;
2444                 if (p == last_found)
2445                         p = p->next;
2446         }
2447         return 0;
2448 }
2449
2450 /*
2451  * Store a list of sha1s that are should not be discarded
2452  * because they are either written too recently, or are
2453  * reachable from another object that was.
2454  *
2455  * This is filled by get_object_list.
2456  */
2457 static struct sha1_array recent_objects;
2458
2459 static int loosened_object_can_be_discarded(const unsigned char *sha1,
2460                                             unsigned long mtime)
2461 {
2462         if (!unpack_unreachable_expiration)
2463                 return 0;
2464         if (mtime > unpack_unreachable_expiration)
2465                 return 0;
2466         if (sha1_array_lookup(&recent_objects, sha1) >= 0)
2467                 return 0;
2468         return 1;
2469 }
2470
2471 static void loosen_unused_packed_objects(struct rev_info *revs)
2472 {
2473         struct packed_git *p;
2474         uint32_t i;
2475         const unsigned char *sha1;
2476
2477         for (p = packed_git; p; p = p->next) {
2478                 if (!p->pack_local || p->pack_keep)
2479                         continue;
2480
2481                 if (open_pack_index(p))
2482                         die("cannot open pack index");
2483
2484                 for (i = 0; i < p->num_objects; i++) {
2485                         sha1 = nth_packed_object_sha1(p, i);
2486                         if (!packlist_find(&to_pack, sha1, NULL) &&
2487                             !has_sha1_pack_kept_or_nonlocal(sha1) &&
2488                             !loosened_object_can_be_discarded(sha1, p->mtime))
2489                                 if (force_object_loose(sha1, p->mtime))
2490                                         die("unable to force loose object");
2491                 }
2492         }
2493 }
2494
2495 /*
2496  * This tracks any options which a reader of the pack might
2497  * not understand, and which would therefore prevent blind reuse
2498  * of what we have on disk.
2499  */
2500 static int pack_options_allow_reuse(void)
2501 {
2502         return allow_ofs_delta;
2503 }
2504
2505 static int get_object_list_from_bitmap(struct rev_info *revs)
2506 {
2507         if (prepare_bitmap_walk(revs) < 0)
2508                 return -1;
2509
2510         if (pack_options_allow_reuse() &&
2511             !reuse_partial_packfile_from_bitmap(
2512                         &reuse_packfile,
2513                         &reuse_packfile_objects,
2514                         &reuse_packfile_offset)) {
2515                 assert(reuse_packfile_objects);
2516                 nr_result += reuse_packfile_objects;
2517                 display_progress(progress_state, nr_result);
2518         }
2519
2520         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2521         return 0;
2522 }
2523
2524 static void record_recent_object(struct object *obj,
2525                                  const char *name,
2526                                  void *data)
2527 {
2528         sha1_array_append(&recent_objects, obj->oid.hash);
2529 }
2530
2531 static void record_recent_commit(struct commit *commit, void *data)
2532 {
2533         sha1_array_append(&recent_objects, commit->object.oid.hash);
2534 }
2535
2536 static void get_object_list(int ac, const char **av)
2537 {
2538         struct rev_info revs;
2539         char line[1000];
2540         int flags = 0;
2541
2542         init_revisions(&revs, NULL);
2543         save_commit_buffer = 0;
2544         setup_revisions(ac, av, &revs, NULL);
2545
2546         /* make sure shallows are read */
2547         is_repository_shallow();
2548
2549         while (fgets(line, sizeof(line), stdin) != NULL) {
2550                 int len = strlen(line);
2551                 if (len && line[len - 1] == '\n')
2552                         line[--len] = 0;
2553                 if (!len)
2554                         break;
2555                 if (*line == '-') {
2556                         if (!strcmp(line, "--not")) {
2557                                 flags ^= UNINTERESTING;
2558                                 write_bitmap_index = 0;
2559                                 continue;
2560                         }
2561                         if (starts_with(line, "--shallow ")) {
2562                                 unsigned char sha1[20];
2563                                 if (get_sha1_hex(line + 10, sha1))
2564                                         die("not an SHA-1 '%s'", line + 10);
2565                                 register_shallow(sha1);
2566                                 use_bitmap_index = 0;
2567                                 continue;
2568                         }
2569                         die("not a rev '%s'", line);
2570                 }
2571                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2572                         die("bad revision '%s'", line);
2573         }
2574
2575         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2576                 return;
2577
2578         if (prepare_revision_walk(&revs))
2579                 die("revision walk setup failed");
2580         mark_edges_uninteresting(&revs, show_edge);
2581         traverse_commit_list(&revs, show_commit, show_object, NULL);
2582
2583         if (unpack_unreachable_expiration) {
2584                 revs.ignore_missing_links = 1;
2585                 if (add_unseen_recent_objects_to_traversal(&revs,
2586                                 unpack_unreachable_expiration))
2587                         die("unable to add recent objects");
2588                 if (prepare_revision_walk(&revs))
2589                         die("revision walk setup failed");
2590                 traverse_commit_list(&revs, record_recent_commit,
2591                                      record_recent_object, NULL);
2592         }
2593
2594         if (keep_unreachable)
2595                 add_objects_in_unpacked_packs(&revs);
2596         if (pack_loose_unreachable)
2597                 add_unreachable_loose_objects();
2598         if (unpack_unreachable)
2599                 loosen_unused_packed_objects(&revs);
2600
2601         sha1_array_clear(&recent_objects);
2602 }
2603
2604 static int option_parse_index_version(const struct option *opt,
2605                                       const char *arg, int unset)
2606 {
2607         char *c;
2608         const char *val = arg;
2609         pack_idx_opts.version = strtoul(val, &c, 10);
2610         if (pack_idx_opts.version > 2)
2611                 die(_("unsupported index version %s"), val);
2612         if (*c == ',' && c[1])
2613                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2614         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2615                 die(_("bad index version '%s'"), val);
2616         return 0;
2617 }
2618
2619 static int option_parse_unpack_unreachable(const struct option *opt,
2620                                            const char *arg, int unset)
2621 {
2622         if (unset) {
2623                 unpack_unreachable = 0;
2624                 unpack_unreachable_expiration = 0;
2625         }
2626         else {
2627                 unpack_unreachable = 1;
2628                 if (arg)
2629                         unpack_unreachable_expiration = approxidate(arg);
2630         }
2631         return 0;
2632 }
2633
2634 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2635 {
2636         int use_internal_rev_list = 0;
2637         int thin = 0;
2638         int shallow = 0;
2639         int all_progress_implied = 0;
2640         struct argv_array rp = ARGV_ARRAY_INIT;
2641         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2642         int rev_list_index = 0;
2643         struct option pack_objects_options[] = {
2644                 OPT_SET_INT('q', "quiet", &progress,
2645                             N_("do not show progress meter"), 0),
2646                 OPT_SET_INT(0, "progress", &progress,
2647                             N_("show progress meter"), 1),
2648                 OPT_SET_INT(0, "all-progress", &progress,
2649                             N_("show progress meter during object writing phase"), 2),
2650                 OPT_BOOL(0, "all-progress-implied",
2651                          &all_progress_implied,
2652                          N_("similar to --all-progress when progress meter is shown")),
2653                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2654                   N_("write the pack index file in the specified idx format version"),
2655                   0, option_parse_index_version },
2656                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2657                               N_("maximum size of each output pack file")),
2658                 OPT_BOOL(0, "local", &local,
2659                          N_("ignore borrowed objects from alternate object store")),
2660                 OPT_BOOL(0, "incremental", &incremental,
2661                          N_("ignore packed objects")),
2662                 OPT_INTEGER(0, "window", &window,
2663                             N_("limit pack window by objects")),
2664                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2665                               N_("limit pack window by memory in addition to object limit")),
2666                 OPT_INTEGER(0, "depth", &depth,
2667                             N_("maximum length of delta chain allowed in the resulting pack")),
2668                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2669                          N_("reuse existing deltas")),
2670                 OPT_BOOL(0, "reuse-object", &reuse_object,
2671                          N_("reuse existing objects")),
2672                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2673                          N_("use OFS_DELTA objects")),
2674                 OPT_INTEGER(0, "threads", &delta_search_threads,
2675                             N_("use threads when searching for best delta matches")),
2676                 OPT_BOOL(0, "non-empty", &non_empty,
2677                          N_("do not create an empty pack output")),
2678                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2679                          N_("read revision arguments from standard input")),
2680                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2681                   N_("limit the objects to those that are not yet packed"),
2682                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2683                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2684                   N_("include objects reachable from any reference"),
2685                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2686                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2687                   N_("include objects referred by reflog entries"),
2688                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2689                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2690                   N_("include objects referred to by the index"),
2691                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2692                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2693                          N_("output pack to stdout")),
2694                 OPT_BOOL(0, "include-tag", &include_tag,
2695                          N_("include tag objects that refer to objects to be packed")),
2696                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2697                          N_("keep unreachable objects")),
2698                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
2699                          N_("pack loose unreachable objects")),
2700                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2701                   N_("unpack unreachable objects newer than <time>"),
2702                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2703                 OPT_BOOL(0, "thin", &thin,
2704                          N_("create thin packs")),
2705                 OPT_BOOL(0, "shallow", &shallow,
2706                          N_("create packs suitable for shallow fetches")),
2707                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2708                          N_("ignore packs that have companion .keep file")),
2709                 OPT_INTEGER(0, "compression", &pack_compression_level,
2710                             N_("pack compression level")),
2711                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2712                             N_("do not hide commits by grafts"), 0),
2713                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2714                          N_("use a bitmap index if available to speed up counting objects")),
2715                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2716                          N_("write a bitmap index together with the pack index")),
2717                 OPT_END(),
2718         };
2719
2720         check_replace_refs = 0;
2721
2722         reset_pack_idx_option(&pack_idx_opts);
2723         git_config(git_pack_config, NULL);
2724         if (!pack_compression_seen && core_compression_seen)
2725                 pack_compression_level = core_compression_level;
2726
2727         progress = isatty(2);
2728         argc = parse_options(argc, argv, prefix, pack_objects_options,
2729                              pack_usage, 0);
2730
2731         if (argc) {
2732                 base_name = argv[0];
2733                 argc--;
2734         }
2735         if (pack_to_stdout != !base_name || argc)
2736                 usage_with_options(pack_usage, pack_objects_options);
2737
2738         argv_array_push(&rp, "pack-objects");
2739         if (thin) {
2740                 use_internal_rev_list = 1;
2741                 argv_array_push(&rp, shallow
2742                                 ? "--objects-edge-aggressive"
2743                                 : "--objects-edge");
2744         } else
2745                 argv_array_push(&rp, "--objects");
2746
2747         if (rev_list_all) {
2748                 use_internal_rev_list = 1;
2749                 argv_array_push(&rp, "--all");
2750         }
2751         if (rev_list_reflog) {
2752                 use_internal_rev_list = 1;
2753                 argv_array_push(&rp, "--reflog");
2754         }
2755         if (rev_list_index) {
2756                 use_internal_rev_list = 1;
2757                 argv_array_push(&rp, "--indexed-objects");
2758         }
2759         if (rev_list_unpacked) {
2760                 use_internal_rev_list = 1;
2761                 argv_array_push(&rp, "--unpacked");
2762         }
2763
2764         if (!reuse_object)
2765                 reuse_delta = 0;
2766         if (pack_compression_level == -1)
2767                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2768         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2769                 die("bad pack compression level %d", pack_compression_level);
2770
2771         if (!delta_search_threads)      /* --threads=0 means autodetect */
2772                 delta_search_threads = online_cpus();
2773
2774 #ifdef NO_PTHREADS
2775         if (delta_search_threads != 1)
2776                 warning("no threads support, ignoring --threads");
2777 #endif
2778         if (!pack_to_stdout && !pack_size_limit)
2779                 pack_size_limit = pack_size_limit_cfg;
2780         if (pack_to_stdout && pack_size_limit)
2781                 die("--max-pack-size cannot be used to build a pack for transfer.");
2782         if (pack_size_limit && pack_size_limit < 1024*1024) {
2783                 warning("minimum pack size limit is 1 MiB");
2784                 pack_size_limit = 1024*1024;
2785         }
2786
2787         if (!pack_to_stdout && thin)
2788                 die("--thin cannot be used to build an indexable pack.");
2789
2790         if (keep_unreachable && unpack_unreachable)
2791                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2792         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
2793                 unpack_unreachable_expiration = 0;
2794
2795         if (!use_internal_rev_list || !pack_to_stdout || is_repository_shallow())
2796                 use_bitmap_index = 0;
2797
2798         if (pack_to_stdout || !rev_list_all)
2799                 write_bitmap_index = 0;
2800
2801         if (progress && all_progress_implied)
2802                 progress = 2;
2803
2804         prepare_packed_git();
2805         if (ignore_packed_keep) {
2806                 struct packed_git *p;
2807                 for (p = packed_git; p; p = p->next)
2808                         if (p->pack_local && p->pack_keep)
2809                                 break;
2810                 if (!p) /* no keep-able packs found */
2811                         ignore_packed_keep = 0;
2812         }
2813         if (local) {
2814                 /*
2815                  * unlike ignore_packed_keep above, we do not want to
2816                  * unset "local" based on looking at packs, as it
2817                  * also covers non-local objects
2818                  */
2819                 struct packed_git *p;
2820                 for (p = packed_git; p; p = p->next) {
2821                         if (!p->pack_local) {
2822                                 have_non_local_packs = 1;
2823                                 break;
2824                         }
2825                 }
2826         }
2827
2828         if (progress)
2829                 progress_state = start_progress(_("Counting objects"), 0);
2830         if (!use_internal_rev_list)
2831                 read_object_list_from_stdin();
2832         else {
2833                 get_object_list(rp.argc, rp.argv);
2834                 argv_array_clear(&rp);
2835         }
2836         cleanup_preferred_base();
2837         if (include_tag && nr_result)
2838                 for_each_ref(add_ref_tag, NULL);
2839         stop_progress(&progress_state);
2840
2841         if (non_empty && !nr_result)
2842                 return 0;
2843         if (nr_result)
2844                 prepare_pack(window, depth);
2845         write_pack_file();
2846         if (progress)
2847                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2848                         " reused %"PRIu32" (delta %"PRIu32")\n",
2849                         written, written_delta, reused, reused_delta);
2850         return 0;
2851 }