pack-objects: compute local/ignore_pack_keep early
[git] / builtin / pack-objects.c
1 #include "builtin.h"
2 #include "cache.h"
3 #include "attr.h"
4 #include "object.h"
5 #include "blob.h"
6 #include "commit.h"
7 #include "tag.h"
8 #include "tree.h"
9 #include "delta.h"
10 #include "pack.h"
11 #include "pack-revindex.h"
12 #include "csum-file.h"
13 #include "tree-walk.h"
14 #include "diff.h"
15 #include "revision.h"
16 #include "list-objects.h"
17 #include "pack-objects.h"
18 #include "progress.h"
19 #include "refs.h"
20 #include "streaming.h"
21 #include "thread-utils.h"
22 #include "pack-bitmap.h"
23 #include "reachable.h"
24 #include "sha1-array.h"
25 #include "argv-array.h"
26
27 static const char *pack_usage[] = {
28         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
29         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
30         NULL
31 };
32
33 /*
34  * Objects we are going to pack are collected in the `to_pack` structure.
35  * It contains an array (dynamically expanded) of the object data, and a map
36  * that can resolve SHA1s to their position in the array.
37  */
38 static struct packing_data to_pack;
39
40 static struct pack_idx_entry **written_list;
41 static uint32_t nr_result, nr_written;
42
43 static int non_empty;
44 static int reuse_delta = 1, reuse_object = 1;
45 static int keep_unreachable, unpack_unreachable, include_tag;
46 static unsigned long unpack_unreachable_expiration;
47 static int pack_loose_unreachable;
48 static int local;
49 static int 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 unsigned long 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         unsigned long 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 unsigned long write_object(struct sha1file *f,
421                                   struct object_entry *entry,
422                                   off_t write_offset)
423 {
424         unsigned long limit, len;
425         int usable_delta, to_reuse;
426
427         if (!pack_to_stdout)
428                 crc32_begin(f);
429
430         /* apply size limit if limited packsize and not first object */
431         if (!pack_size_limit || !nr_written)
432                 limit = 0;
433         else if (pack_size_limit <= write_offset)
434                 /*
435                  * the earlier object did not fit the limit; avoid
436                  * mistaking this with unlimited (i.e. limit = 0).
437                  */
438                 limit = 1;
439         else
440                 limit = pack_size_limit - write_offset;
441
442         if (!entry->delta)
443                 usable_delta = 0;       /* no delta */
444         else if (!pack_size_limit)
445                usable_delta = 1;        /* unlimited packfile */
446         else if (entry->delta->idx.offset == (off_t)-1)
447                 usable_delta = 0;       /* base was written to another pack */
448         else if (entry->delta->idx.offset)
449                 usable_delta = 1;       /* base already exists in this pack */
450         else
451                 usable_delta = 0;       /* base could end up in another pack */
452
453         if (!reuse_object)
454                 to_reuse = 0;   /* explicit */
455         else if (!entry->in_pack)
456                 to_reuse = 0;   /* can't reuse what we don't have */
457         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
458                                 /* check_object() decided it for us ... */
459                 to_reuse = usable_delta;
460                                 /* ... but pack split may override that */
461         else if (entry->type != entry->in_pack_type)
462                 to_reuse = 0;   /* pack has delta which is unusable */
463         else if (entry->delta)
464                 to_reuse = 0;   /* we want to pack afresh */
465         else
466                 to_reuse = 1;   /* we have it in-pack undeltified,
467                                  * and we do not need to deltify it.
468                                  */
469
470         if (!to_reuse)
471                 len = write_no_reuse_object(f, entry, limit, usable_delta);
472         else
473                 len = write_reuse_object(f, entry, limit, usable_delta);
474         if (!len)
475                 return 0;
476
477         if (usable_delta)
478                 written_delta++;
479         written++;
480         if (!pack_to_stdout)
481                 entry->idx.crc32 = crc32_end(f);
482         return len;
483 }
484
485 enum write_one_status {
486         WRITE_ONE_SKIP = -1, /* already written */
487         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
488         WRITE_ONE_WRITTEN = 1, /* normal */
489         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
490 };
491
492 static enum write_one_status write_one(struct sha1file *f,
493                                        struct object_entry *e,
494                                        off_t *offset)
495 {
496         unsigned long size;
497         int recursing;
498
499         /*
500          * we set offset to 1 (which is an impossible value) to mark
501          * the fact that this object is involved in "write its base
502          * first before writing a deltified object" recursion.
503          */
504         recursing = (e->idx.offset == 1);
505         if (recursing) {
506                 warning("recursive delta detected for object %s",
507                         sha1_to_hex(e->idx.sha1));
508                 return WRITE_ONE_RECURSIVE;
509         } else if (e->idx.offset || e->preferred_base) {
510                 /* offset is non zero if object is written already. */
511                 return WRITE_ONE_SKIP;
512         }
513
514         /* if we are deltified, write out base object first. */
515         if (e->delta) {
516                 e->idx.offset = 1; /* now recurse */
517                 switch (write_one(f, e->delta, offset)) {
518                 case WRITE_ONE_RECURSIVE:
519                         /* we cannot depend on this one */
520                         e->delta = NULL;
521                         break;
522                 default:
523                         break;
524                 case WRITE_ONE_BREAK:
525                         e->idx.offset = recursing;
526                         return WRITE_ONE_BREAK;
527                 }
528         }
529
530         e->idx.offset = *offset;
531         size = write_object(f, e, *offset);
532         if (!size) {
533                 e->idx.offset = recursing;
534                 return WRITE_ONE_BREAK;
535         }
536         written_list[nr_written++] = &e->idx;
537
538         /* make sure off_t is sufficiently large not to wrap */
539         if (signed_add_overflows(*offset, size))
540                 die("pack too large for current definition of off_t");
541         *offset += size;
542         return WRITE_ONE_WRITTEN;
543 }
544
545 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
546                        void *cb_data)
547 {
548         unsigned char peeled[20];
549         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
550
551         if (entry)
552                 entry->tagged = 1;
553         if (!peel_ref(path, peeled)) {
554                 entry = packlist_find(&to_pack, peeled, NULL);
555                 if (entry)
556                         entry->tagged = 1;
557         }
558         return 0;
559 }
560
561 static inline void add_to_write_order(struct object_entry **wo,
562                                unsigned int *endp,
563                                struct object_entry *e)
564 {
565         if (e->filled)
566                 return;
567         wo[(*endp)++] = e;
568         e->filled = 1;
569 }
570
571 static void add_descendants_to_write_order(struct object_entry **wo,
572                                            unsigned int *endp,
573                                            struct object_entry *e)
574 {
575         int add_to_order = 1;
576         while (e) {
577                 if (add_to_order) {
578                         struct object_entry *s;
579                         /* add this node... */
580                         add_to_write_order(wo, endp, e);
581                         /* all its siblings... */
582                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
583                                 add_to_write_order(wo, endp, s);
584                         }
585                 }
586                 /* drop down a level to add left subtree nodes if possible */
587                 if (e->delta_child) {
588                         add_to_order = 1;
589                         e = e->delta_child;
590                 } else {
591                         add_to_order = 0;
592                         /* our sibling might have some children, it is next */
593                         if (e->delta_sibling) {
594                                 e = e->delta_sibling;
595                                 continue;
596                         }
597                         /* go back to our parent node */
598                         e = e->delta;
599                         while (e && !e->delta_sibling) {
600                                 /* we're on the right side of a subtree, keep
601                                  * going up until we can go right again */
602                                 e = e->delta;
603                         }
604                         if (!e) {
605                                 /* done- we hit our original root node */
606                                 return;
607                         }
608                         /* pass it off to sibling at this level */
609                         e = e->delta_sibling;
610                 }
611         };
612 }
613
614 static void add_family_to_write_order(struct object_entry **wo,
615                                       unsigned int *endp,
616                                       struct object_entry *e)
617 {
618         struct object_entry *root;
619
620         for (root = e; root->delta; root = root->delta)
621                 ; /* nothing */
622         add_descendants_to_write_order(wo, endp, root);
623 }
624
625 static struct object_entry **compute_write_order(void)
626 {
627         unsigned int i, wo_end, last_untagged;
628
629         struct object_entry **wo;
630         struct object_entry *objects = to_pack.objects;
631
632         for (i = 0; i < to_pack.nr_objects; i++) {
633                 objects[i].tagged = 0;
634                 objects[i].filled = 0;
635                 objects[i].delta_child = NULL;
636                 objects[i].delta_sibling = NULL;
637         }
638
639         /*
640          * Fully connect delta_child/delta_sibling network.
641          * Make sure delta_sibling is sorted in the original
642          * recency order.
643          */
644         for (i = to_pack.nr_objects; i > 0;) {
645                 struct object_entry *e = &objects[--i];
646                 if (!e->delta)
647                         continue;
648                 /* Mark me as the first child */
649                 e->delta_sibling = e->delta->delta_child;
650                 e->delta->delta_child = e;
651         }
652
653         /*
654          * Mark objects that are at the tip of tags.
655          */
656         for_each_tag_ref(mark_tagged, NULL);
657
658         /*
659          * Give the objects in the original recency order until
660          * we see a tagged tip.
661          */
662         ALLOC_ARRAY(wo, to_pack.nr_objects);
663         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
664                 if (objects[i].tagged)
665                         break;
666                 add_to_write_order(wo, &wo_end, &objects[i]);
667         }
668         last_untagged = i;
669
670         /*
671          * Then fill all the tagged tips.
672          */
673         for (; i < to_pack.nr_objects; i++) {
674                 if (objects[i].tagged)
675                         add_to_write_order(wo, &wo_end, &objects[i]);
676         }
677
678         /*
679          * And then all remaining commits and tags.
680          */
681         for (i = last_untagged; i < to_pack.nr_objects; i++) {
682                 if (objects[i].type != OBJ_COMMIT &&
683                     objects[i].type != OBJ_TAG)
684                         continue;
685                 add_to_write_order(wo, &wo_end, &objects[i]);
686         }
687
688         /*
689          * And then all the trees.
690          */
691         for (i = last_untagged; i < to_pack.nr_objects; i++) {
692                 if (objects[i].type != OBJ_TREE)
693                         continue;
694                 add_to_write_order(wo, &wo_end, &objects[i]);
695         }
696
697         /*
698          * Finally all the rest in really tight order
699          */
700         for (i = last_untagged; i < to_pack.nr_objects; i++) {
701                 if (!objects[i].filled)
702                         add_family_to_write_order(wo, &wo_end, &objects[i]);
703         }
704
705         if (wo_end != to_pack.nr_objects)
706                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
707
708         return wo;
709 }
710
711 static off_t write_reused_pack(struct sha1file *f)
712 {
713         unsigned char buffer[8192];
714         off_t to_write, total;
715         int fd;
716
717         if (!is_pack_valid(reuse_packfile))
718                 die("packfile is invalid: %s", reuse_packfile->pack_name);
719
720         fd = git_open_noatime(reuse_packfile->pack_name);
721         if (fd < 0)
722                 die_errno("unable to open packfile for reuse: %s",
723                           reuse_packfile->pack_name);
724
725         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
726                 die_errno("unable to seek in reused packfile");
727
728         if (reuse_packfile_offset < 0)
729                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
730
731         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
732
733         while (to_write) {
734                 int read_pack = xread(fd, buffer, sizeof(buffer));
735
736                 if (read_pack <= 0)
737                         die_errno("unable to read from reused packfile");
738
739                 if (read_pack > to_write)
740                         read_pack = to_write;
741
742                 sha1write(f, buffer, read_pack);
743                 to_write -= read_pack;
744
745                 /*
746                  * We don't know the actual number of objects written,
747                  * only how many bytes written, how many bytes total, and
748                  * how many objects total. So we can fake it by pretending all
749                  * objects we are writing are the same size. This gives us a
750                  * smooth progress meter, and at the end it matches the true
751                  * answer.
752                  */
753                 written = reuse_packfile_objects *
754                                 (((double)(total - to_write)) / total);
755                 display_progress(progress_state, written);
756         }
757
758         close(fd);
759         written = reuse_packfile_objects;
760         display_progress(progress_state, written);
761         return reuse_packfile_offset - sizeof(struct pack_header);
762 }
763
764 static const char no_split_warning[] = N_(
765 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
766 );
767
768 static void write_pack_file(void)
769 {
770         uint32_t i = 0, j;
771         struct sha1file *f;
772         off_t offset;
773         uint32_t nr_remaining = nr_result;
774         time_t last_mtime = 0;
775         struct object_entry **write_order;
776
777         if (progress > pack_to_stdout)
778                 progress_state = start_progress(_("Writing objects"), nr_result);
779         ALLOC_ARRAY(written_list, to_pack.nr_objects);
780         write_order = compute_write_order();
781
782         do {
783                 unsigned char sha1[20];
784                 char *pack_tmp_name = NULL;
785
786                 if (pack_to_stdout)
787                         f = sha1fd_throughput(1, "<stdout>", progress_state);
788                 else
789                         f = create_tmp_packfile(&pack_tmp_name);
790
791                 offset = write_pack_header(f, nr_remaining);
792
793                 if (reuse_packfile) {
794                         off_t packfile_size;
795                         assert(pack_to_stdout);
796
797                         packfile_size = write_reused_pack(f);
798                         offset += packfile_size;
799                 }
800
801                 nr_written = 0;
802                 for (; i < to_pack.nr_objects; i++) {
803                         struct object_entry *e = write_order[i];
804                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
805                                 break;
806                         display_progress(progress_state, written);
807                 }
808
809                 /*
810                  * Did we write the wrong # entries in the header?
811                  * If so, rewrite it like in fast-import
812                  */
813                 if (pack_to_stdout) {
814                         sha1close(f, sha1, CSUM_CLOSE);
815                 } else if (nr_written == nr_remaining) {
816                         sha1close(f, sha1, CSUM_FSYNC);
817                 } else {
818                         int fd = sha1close(f, sha1, 0);
819                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
820                                                  nr_written, sha1, offset);
821                         close(fd);
822                         if (write_bitmap_index) {
823                                 warning(_(no_split_warning));
824                                 write_bitmap_index = 0;
825                         }
826                 }
827
828                 if (!pack_to_stdout) {
829                         struct stat st;
830                         struct strbuf tmpname = STRBUF_INIT;
831
832                         /*
833                          * Packs are runtime accessed in their mtime
834                          * order since newer packs are more likely to contain
835                          * younger objects.  So if we are creating multiple
836                          * packs then we should modify the mtime of later ones
837                          * to preserve this property.
838                          */
839                         if (stat(pack_tmp_name, &st) < 0) {
840                                 warning_errno("failed to stat %s", pack_tmp_name);
841                         } else if (!last_mtime) {
842                                 last_mtime = st.st_mtime;
843                         } else {
844                                 struct utimbuf utb;
845                                 utb.actime = st.st_atime;
846                                 utb.modtime = --last_mtime;
847                                 if (utime(pack_tmp_name, &utb) < 0)
848                                         warning_errno("failed utime() on %s", pack_tmp_name);
849                         }
850
851                         strbuf_addf(&tmpname, "%s-", base_name);
852
853                         if (write_bitmap_index) {
854                                 bitmap_writer_set_checksum(sha1);
855                                 bitmap_writer_build_type_index(written_list, nr_written);
856                         }
857
858                         finish_tmp_packfile(&tmpname, pack_tmp_name,
859                                             written_list, nr_written,
860                                             &pack_idx_opts, sha1);
861
862                         if (write_bitmap_index) {
863                                 strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
864
865                                 stop_progress(&progress_state);
866
867                                 bitmap_writer_show_progress(progress);
868                                 bitmap_writer_reuse_bitmaps(&to_pack);
869                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
870                                 bitmap_writer_build(&to_pack);
871                                 bitmap_writer_finish(written_list, nr_written,
872                                                      tmpname.buf, write_bitmap_options);
873                                 write_bitmap_index = 0;
874                         }
875
876                         strbuf_release(&tmpname);
877                         free(pack_tmp_name);
878                         puts(sha1_to_hex(sha1));
879                 }
880
881                 /* mark written objects as written to previous pack */
882                 for (j = 0; j < nr_written; j++) {
883                         written_list[j]->offset = (off_t)-1;
884                 }
885                 nr_remaining -= nr_written;
886         } while (nr_remaining && i < to_pack.nr_objects);
887
888         free(written_list);
889         free(write_order);
890         stop_progress(&progress_state);
891         if (written != nr_result)
892                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
893                         written, nr_result);
894 }
895
896 static void setup_delta_attr_check(struct git_attr_check *check)
897 {
898         static struct git_attr *attr_delta;
899
900         if (!attr_delta)
901                 attr_delta = git_attr("delta");
902
903         check[0].attr = attr_delta;
904 }
905
906 static int no_try_delta(const char *path)
907 {
908         struct git_attr_check check[1];
909
910         setup_delta_attr_check(check);
911         if (git_check_attr(path, ARRAY_SIZE(check), check))
912                 return 0;
913         if (ATTR_FALSE(check->value))
914                 return 1;
915         return 0;
916 }
917
918 /*
919  * When adding an object, check whether we have already added it
920  * to our packing list. If so, we can skip. However, if we are
921  * being asked to excludei t, but the previous mention was to include
922  * it, make sure to adjust its flags and tweak our numbers accordingly.
923  *
924  * As an optimization, we pass out the index position where we would have
925  * found the item, since that saves us from having to look it up again a
926  * few lines later when we want to add the new entry.
927  */
928 static int have_duplicate_entry(const unsigned char *sha1,
929                                 int exclude,
930                                 uint32_t *index_pos)
931 {
932         struct object_entry *entry;
933
934         entry = packlist_find(&to_pack, sha1, index_pos);
935         if (!entry)
936                 return 0;
937
938         if (exclude) {
939                 if (!entry->preferred_base)
940                         nr_result--;
941                 entry->preferred_base = 1;
942         }
943
944         return 1;
945 }
946
947 /*
948  * Check whether we want the object in the pack (e.g., we do not want
949  * objects found in non-local stores if the "--local" option was used).
950  *
951  * As a side effect of this check, we will find the packed version of this
952  * object, if any. We therefore pass out the pack information to avoid having
953  * to look it up again later.
954  */
955 static int want_object_in_pack(const unsigned char *sha1,
956                                int exclude,
957                                struct packed_git **found_pack,
958                                off_t *found_offset)
959 {
960         struct packed_git *p;
961
962         if (!exclude && local && has_loose_object_nonlocal(sha1))
963                 return 0;
964
965         *found_pack = NULL;
966         *found_offset = 0;
967
968         for (p = packed_git; p; p = p->next) {
969                 off_t offset = find_pack_entry_one(sha1, p);
970                 if (offset) {
971                         if (!*found_pack) {
972                                 if (!is_pack_valid(p))
973                                         continue;
974                                 *found_offset = offset;
975                                 *found_pack = p;
976                         }
977                         if (exclude)
978                                 return 1;
979                         if (incremental)
980                                 return 0;
981
982                         /*
983                          * When asked to do --local (do not include an
984                          * object that appears in a pack we borrow
985                          * from elsewhere) or --honor-pack-keep (do not
986                          * include an object that appears in a pack marked
987                          * with .keep), we need to make sure no copy of this
988                          * object come from in _any_ pack that causes us to
989                          * omit it, and need to complete this loop.  When
990                          * neither option is in effect, we know the object
991                          * we just found is going to be packed, so break
992                          * out of the loop to return 1 now.
993                          */
994                         if (!ignore_packed_keep &&
995                             (!local || !have_non_local_packs))
996                                 break;
997
998                         if (local && !p->pack_local)
999                                 return 0;
1000                         if (ignore_packed_keep && p->pack_local && p->pack_keep)
1001                                 return 0;
1002                 }
1003         }
1004
1005         return 1;
1006 }
1007
1008 static void create_object_entry(const unsigned char *sha1,
1009                                 enum object_type type,
1010                                 uint32_t hash,
1011                                 int exclude,
1012                                 int no_try_delta,
1013                                 uint32_t index_pos,
1014                                 struct packed_git *found_pack,
1015                                 off_t found_offset)
1016 {
1017         struct object_entry *entry;
1018
1019         entry = packlist_alloc(&to_pack, sha1, index_pos);
1020         entry->hash = hash;
1021         if (type)
1022                 entry->type = type;
1023         if (exclude)
1024                 entry->preferred_base = 1;
1025         else
1026                 nr_result++;
1027         if (found_pack) {
1028                 entry->in_pack = found_pack;
1029                 entry->in_pack_offset = found_offset;
1030         }
1031
1032         entry->no_try_delta = no_try_delta;
1033 }
1034
1035 static const char no_closure_warning[] = N_(
1036 "disabling bitmap writing, as some objects are not being packed"
1037 );
1038
1039 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1040                             const char *name, int exclude)
1041 {
1042         struct packed_git *found_pack;
1043         off_t found_offset;
1044         uint32_t index_pos;
1045
1046         if (have_duplicate_entry(sha1, exclude, &index_pos))
1047                 return 0;
1048
1049         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1050                 /* The pack is missing an object, so it will not have closure */
1051                 if (write_bitmap_index) {
1052                         warning(_(no_closure_warning));
1053                         write_bitmap_index = 0;
1054                 }
1055                 return 0;
1056         }
1057
1058         create_object_entry(sha1, type, pack_name_hash(name),
1059                             exclude, name && no_try_delta(name),
1060                             index_pos, found_pack, found_offset);
1061
1062         display_progress(progress_state, nr_result);
1063         return 1;
1064 }
1065
1066 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1067                                         enum object_type type,
1068                                         int flags, uint32_t name_hash,
1069                                         struct packed_git *pack, off_t offset)
1070 {
1071         uint32_t index_pos;
1072
1073         if (have_duplicate_entry(sha1, 0, &index_pos))
1074                 return 0;
1075
1076         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1077
1078         display_progress(progress_state, nr_result);
1079         return 1;
1080 }
1081
1082 struct pbase_tree_cache {
1083         unsigned char sha1[20];
1084         int ref;
1085         int temporary;
1086         void *tree_data;
1087         unsigned long tree_size;
1088 };
1089
1090 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1091 static int pbase_tree_cache_ix(const unsigned char *sha1)
1092 {
1093         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1094 }
1095 static int pbase_tree_cache_ix_incr(int ix)
1096 {
1097         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1098 }
1099
1100 static struct pbase_tree {
1101         struct pbase_tree *next;
1102         /* This is a phony "cache" entry; we are not
1103          * going to evict it or find it through _get()
1104          * mechanism -- this is for the toplevel node that
1105          * would almost always change with any commit.
1106          */
1107         struct pbase_tree_cache pcache;
1108 } *pbase_tree;
1109
1110 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1111 {
1112         struct pbase_tree_cache *ent, *nent;
1113         void *data;
1114         unsigned long size;
1115         enum object_type type;
1116         int neigh;
1117         int my_ix = pbase_tree_cache_ix(sha1);
1118         int available_ix = -1;
1119
1120         /* pbase-tree-cache acts as a limited hashtable.
1121          * your object will be found at your index or within a few
1122          * slots after that slot if it is cached.
1123          */
1124         for (neigh = 0; neigh < 8; neigh++) {
1125                 ent = pbase_tree_cache[my_ix];
1126                 if (ent && !hashcmp(ent->sha1, sha1)) {
1127                         ent->ref++;
1128                         return ent;
1129                 }
1130                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1131                          ((0 <= available_ix) &&
1132                           (!ent && pbase_tree_cache[available_ix])))
1133                         available_ix = my_ix;
1134                 if (!ent)
1135                         break;
1136                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1137         }
1138
1139         /* Did not find one.  Either we got a bogus request or
1140          * we need to read and perhaps cache.
1141          */
1142         data = read_sha1_file(sha1, &type, &size);
1143         if (!data)
1144                 return NULL;
1145         if (type != OBJ_TREE) {
1146                 free(data);
1147                 return NULL;
1148         }
1149
1150         /* We need to either cache or return a throwaway copy */
1151
1152         if (available_ix < 0)
1153                 ent = NULL;
1154         else {
1155                 ent = pbase_tree_cache[available_ix];
1156                 my_ix = available_ix;
1157         }
1158
1159         if (!ent) {
1160                 nent = xmalloc(sizeof(*nent));
1161                 nent->temporary = (available_ix < 0);
1162         }
1163         else {
1164                 /* evict and reuse */
1165                 free(ent->tree_data);
1166                 nent = ent;
1167         }
1168         hashcpy(nent->sha1, sha1);
1169         nent->tree_data = data;
1170         nent->tree_size = size;
1171         nent->ref = 1;
1172         if (!nent->temporary)
1173                 pbase_tree_cache[my_ix] = nent;
1174         return nent;
1175 }
1176
1177 static void pbase_tree_put(struct pbase_tree_cache *cache)
1178 {
1179         if (!cache->temporary) {
1180                 cache->ref--;
1181                 return;
1182         }
1183         free(cache->tree_data);
1184         free(cache);
1185 }
1186
1187 static int name_cmp_len(const char *name)
1188 {
1189         int i;
1190         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1191                 ;
1192         return i;
1193 }
1194
1195 static void add_pbase_object(struct tree_desc *tree,
1196                              const char *name,
1197                              int cmplen,
1198                              const char *fullname)
1199 {
1200         struct name_entry entry;
1201         int cmp;
1202
1203         while (tree_entry(tree,&entry)) {
1204                 if (S_ISGITLINK(entry.mode))
1205                         continue;
1206                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1207                       memcmp(name, entry.path, cmplen);
1208                 if (cmp > 0)
1209                         continue;
1210                 if (cmp < 0)
1211                         return;
1212                 if (name[cmplen] != '/') {
1213                         add_object_entry(entry.oid->hash,
1214                                          object_type(entry.mode),
1215                                          fullname, 1);
1216                         return;
1217                 }
1218                 if (S_ISDIR(entry.mode)) {
1219                         struct tree_desc sub;
1220                         struct pbase_tree_cache *tree;
1221                         const char *down = name+cmplen+1;
1222                         int downlen = name_cmp_len(down);
1223
1224                         tree = pbase_tree_get(entry.oid->hash);
1225                         if (!tree)
1226                                 return;
1227                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1228
1229                         add_pbase_object(&sub, down, downlen, fullname);
1230                         pbase_tree_put(tree);
1231                 }
1232         }
1233 }
1234
1235 static unsigned *done_pbase_paths;
1236 static int done_pbase_paths_num;
1237 static int done_pbase_paths_alloc;
1238 static int done_pbase_path_pos(unsigned hash)
1239 {
1240         int lo = 0;
1241         int hi = done_pbase_paths_num;
1242         while (lo < hi) {
1243                 int mi = (hi + lo) / 2;
1244                 if (done_pbase_paths[mi] == hash)
1245                         return mi;
1246                 if (done_pbase_paths[mi] < hash)
1247                         hi = mi;
1248                 else
1249                         lo = mi + 1;
1250         }
1251         return -lo-1;
1252 }
1253
1254 static int check_pbase_path(unsigned hash)
1255 {
1256         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1257         if (0 <= pos)
1258                 return 1;
1259         pos = -pos - 1;
1260         ALLOC_GROW(done_pbase_paths,
1261                    done_pbase_paths_num + 1,
1262                    done_pbase_paths_alloc);
1263         done_pbase_paths_num++;
1264         if (pos < done_pbase_paths_num)
1265                 memmove(done_pbase_paths + pos + 1,
1266                         done_pbase_paths + pos,
1267                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1268         done_pbase_paths[pos] = hash;
1269         return 0;
1270 }
1271
1272 static void add_preferred_base_object(const char *name)
1273 {
1274         struct pbase_tree *it;
1275         int cmplen;
1276         unsigned hash = pack_name_hash(name);
1277
1278         if (!num_preferred_base || check_pbase_path(hash))
1279                 return;
1280
1281         cmplen = name_cmp_len(name);
1282         for (it = pbase_tree; it; it = it->next) {
1283                 if (cmplen == 0) {
1284                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1285                 }
1286                 else {
1287                         struct tree_desc tree;
1288                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1289                         add_pbase_object(&tree, name, cmplen, name);
1290                 }
1291         }
1292 }
1293
1294 static void add_preferred_base(unsigned char *sha1)
1295 {
1296         struct pbase_tree *it;
1297         void *data;
1298         unsigned long size;
1299         unsigned char tree_sha1[20];
1300
1301         if (window <= num_preferred_base++)
1302                 return;
1303
1304         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1305         if (!data)
1306                 return;
1307
1308         for (it = pbase_tree; it; it = it->next) {
1309                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1310                         free(data);
1311                         return;
1312                 }
1313         }
1314
1315         it = xcalloc(1, sizeof(*it));
1316         it->next = pbase_tree;
1317         pbase_tree = it;
1318
1319         hashcpy(it->pcache.sha1, tree_sha1);
1320         it->pcache.tree_data = data;
1321         it->pcache.tree_size = size;
1322 }
1323
1324 static void cleanup_preferred_base(void)
1325 {
1326         struct pbase_tree *it;
1327         unsigned i;
1328
1329         it = pbase_tree;
1330         pbase_tree = NULL;
1331         while (it) {
1332                 struct pbase_tree *this = it;
1333                 it = this->next;
1334                 free(this->pcache.tree_data);
1335                 free(this);
1336         }
1337
1338         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1339                 if (!pbase_tree_cache[i])
1340                         continue;
1341                 free(pbase_tree_cache[i]->tree_data);
1342                 free(pbase_tree_cache[i]);
1343                 pbase_tree_cache[i] = NULL;
1344         }
1345
1346         free(done_pbase_paths);
1347         done_pbase_paths = NULL;
1348         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1349 }
1350
1351 static void check_object(struct object_entry *entry)
1352 {
1353         if (entry->in_pack) {
1354                 struct packed_git *p = entry->in_pack;
1355                 struct pack_window *w_curs = NULL;
1356                 const unsigned char *base_ref = NULL;
1357                 struct object_entry *base_entry;
1358                 unsigned long used, used_0;
1359                 unsigned long avail;
1360                 off_t ofs;
1361                 unsigned char *buf, c;
1362
1363                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1364
1365                 /*
1366                  * We want in_pack_type even if we do not reuse delta
1367                  * since non-delta representations could still be reused.
1368                  */
1369                 used = unpack_object_header_buffer(buf, avail,
1370                                                    &entry->in_pack_type,
1371                                                    &entry->size);
1372                 if (used == 0)
1373                         goto give_up;
1374
1375                 /*
1376                  * Determine if this is a delta and if so whether we can
1377                  * reuse it or not.  Otherwise let's find out as cheaply as
1378                  * possible what the actual type and size for this object is.
1379                  */
1380                 switch (entry->in_pack_type) {
1381                 default:
1382                         /* Not a delta hence we've already got all we need. */
1383                         entry->type = entry->in_pack_type;
1384                         entry->in_pack_header_size = used;
1385                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1386                                 goto give_up;
1387                         unuse_pack(&w_curs);
1388                         return;
1389                 case OBJ_REF_DELTA:
1390                         if (reuse_delta && !entry->preferred_base)
1391                                 base_ref = use_pack(p, &w_curs,
1392                                                 entry->in_pack_offset + used, NULL);
1393                         entry->in_pack_header_size = used + 20;
1394                         break;
1395                 case OBJ_OFS_DELTA:
1396                         buf = use_pack(p, &w_curs,
1397                                        entry->in_pack_offset + used, NULL);
1398                         used_0 = 0;
1399                         c = buf[used_0++];
1400                         ofs = c & 127;
1401                         while (c & 128) {
1402                                 ofs += 1;
1403                                 if (!ofs || MSB(ofs, 7)) {
1404                                         error("delta base offset overflow in pack for %s",
1405                                               sha1_to_hex(entry->idx.sha1));
1406                                         goto give_up;
1407                                 }
1408                                 c = buf[used_0++];
1409                                 ofs = (ofs << 7) + (c & 127);
1410                         }
1411                         ofs = entry->in_pack_offset - ofs;
1412                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1413                                 error("delta base offset out of bound for %s",
1414                                       sha1_to_hex(entry->idx.sha1));
1415                                 goto give_up;
1416                         }
1417                         if (reuse_delta && !entry->preferred_base) {
1418                                 struct revindex_entry *revidx;
1419                                 revidx = find_pack_revindex(p, ofs);
1420                                 if (!revidx)
1421                                         goto give_up;
1422                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1423                         }
1424                         entry->in_pack_header_size = used + used_0;
1425                         break;
1426                 }
1427
1428                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1429                         /*
1430                          * If base_ref was set above that means we wish to
1431                          * reuse delta data, and we even found that base
1432                          * in the list of objects we want to pack. Goodie!
1433                          *
1434                          * Depth value does not matter - find_deltas() will
1435                          * never consider reused delta as the base object to
1436                          * deltify other objects against, in order to avoid
1437                          * circular deltas.
1438                          */
1439                         entry->type = entry->in_pack_type;
1440                         entry->delta = base_entry;
1441                         entry->delta_size = entry->size;
1442                         entry->delta_sibling = base_entry->delta_child;
1443                         base_entry->delta_child = entry;
1444                         unuse_pack(&w_curs);
1445                         return;
1446                 }
1447
1448                 if (entry->type) {
1449                         /*
1450                          * This must be a delta and we already know what the
1451                          * final object type is.  Let's extract the actual
1452                          * object size from the delta header.
1453                          */
1454                         entry->size = get_size_from_delta(p, &w_curs,
1455                                         entry->in_pack_offset + entry->in_pack_header_size);
1456                         if (entry->size == 0)
1457                                 goto give_up;
1458                         unuse_pack(&w_curs);
1459                         return;
1460                 }
1461
1462                 /*
1463                  * No choice but to fall back to the recursive delta walk
1464                  * with sha1_object_info() to find about the object type
1465                  * at this point...
1466                  */
1467                 give_up:
1468                 unuse_pack(&w_curs);
1469         }
1470
1471         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1472         /*
1473          * The error condition is checked in prepare_pack().  This is
1474          * to permit a missing preferred base object to be ignored
1475          * as a preferred base.  Doing so can result in a larger
1476          * pack file, but the transfer will still take place.
1477          */
1478 }
1479
1480 static int pack_offset_sort(const void *_a, const void *_b)
1481 {
1482         const struct object_entry *a = *(struct object_entry **)_a;
1483         const struct object_entry *b = *(struct object_entry **)_b;
1484
1485         /* avoid filesystem trashing with loose objects */
1486         if (!a->in_pack && !b->in_pack)
1487                 return hashcmp(a->idx.sha1, b->idx.sha1);
1488
1489         if (a->in_pack < b->in_pack)
1490                 return -1;
1491         if (a->in_pack > b->in_pack)
1492                 return 1;
1493         return a->in_pack_offset < b->in_pack_offset ? -1 :
1494                         (a->in_pack_offset > b->in_pack_offset);
1495 }
1496
1497 static void get_object_details(void)
1498 {
1499         uint32_t i;
1500         struct object_entry **sorted_by_offset;
1501
1502         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1503         for (i = 0; i < to_pack.nr_objects; i++)
1504                 sorted_by_offset[i] = to_pack.objects + i;
1505         qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1506
1507         for (i = 0; i < to_pack.nr_objects; i++) {
1508                 struct object_entry *entry = sorted_by_offset[i];
1509                 check_object(entry);
1510                 if (big_file_threshold < entry->size)
1511                         entry->no_try_delta = 1;
1512         }
1513
1514         free(sorted_by_offset);
1515 }
1516
1517 /*
1518  * We search for deltas in a list sorted by type, by filename hash, and then
1519  * by size, so that we see progressively smaller and smaller files.
1520  * That's because we prefer deltas to be from the bigger file
1521  * to the smaller -- deletes are potentially cheaper, but perhaps
1522  * more importantly, the bigger file is likely the more recent
1523  * one.  The deepest deltas are therefore the oldest objects which are
1524  * less susceptible to be accessed often.
1525  */
1526 static int type_size_sort(const void *_a, const void *_b)
1527 {
1528         const struct object_entry *a = *(struct object_entry **)_a;
1529         const struct object_entry *b = *(struct object_entry **)_b;
1530
1531         if (a->type > b->type)
1532                 return -1;
1533         if (a->type < b->type)
1534                 return 1;
1535         if (a->hash > b->hash)
1536                 return -1;
1537         if (a->hash < b->hash)
1538                 return 1;
1539         if (a->preferred_base > b->preferred_base)
1540                 return -1;
1541         if (a->preferred_base < b->preferred_base)
1542                 return 1;
1543         if (a->size > b->size)
1544                 return -1;
1545         if (a->size < b->size)
1546                 return 1;
1547         return a < b ? -1 : (a > b);  /* newest first */
1548 }
1549
1550 struct unpacked {
1551         struct object_entry *entry;
1552         void *data;
1553         struct delta_index *index;
1554         unsigned depth;
1555 };
1556
1557 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1558                            unsigned long delta_size)
1559 {
1560         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1561                 return 0;
1562
1563         if (delta_size < cache_max_small_delta_size)
1564                 return 1;
1565
1566         /* cache delta, if objects are large enough compared to delta size */
1567         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1568                 return 1;
1569
1570         return 0;
1571 }
1572
1573 #ifndef NO_PTHREADS
1574
1575 static pthread_mutex_t read_mutex;
1576 #define read_lock()             pthread_mutex_lock(&read_mutex)
1577 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1578
1579 static pthread_mutex_t cache_mutex;
1580 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1581 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1582
1583 static pthread_mutex_t progress_mutex;
1584 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1585 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1586
1587 #else
1588
1589 #define read_lock()             (void)0
1590 #define read_unlock()           (void)0
1591 #define cache_lock()            (void)0
1592 #define cache_unlock()          (void)0
1593 #define progress_lock()         (void)0
1594 #define progress_unlock()       (void)0
1595
1596 #endif
1597
1598 static int try_delta(struct unpacked *trg, struct unpacked *src,
1599                      unsigned max_depth, unsigned long *mem_usage)
1600 {
1601         struct object_entry *trg_entry = trg->entry;
1602         struct object_entry *src_entry = src->entry;
1603         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1604         unsigned ref_depth;
1605         enum object_type type;
1606         void *delta_buf;
1607
1608         /* Don't bother doing diffs between different types */
1609         if (trg_entry->type != src_entry->type)
1610                 return -1;
1611
1612         /*
1613          * We do not bother to try a delta that we discarded on an
1614          * earlier try, but only when reusing delta data.  Note that
1615          * src_entry that is marked as the preferred_base should always
1616          * be considered, as even if we produce a suboptimal delta against
1617          * it, we will still save the transfer cost, as we already know
1618          * the other side has it and we won't send src_entry at all.
1619          */
1620         if (reuse_delta && trg_entry->in_pack &&
1621             trg_entry->in_pack == src_entry->in_pack &&
1622             !src_entry->preferred_base &&
1623             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1624             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1625                 return 0;
1626
1627         /* Let's not bust the allowed depth. */
1628         if (src->depth >= max_depth)
1629                 return 0;
1630
1631         /* Now some size filtering heuristics. */
1632         trg_size = trg_entry->size;
1633         if (!trg_entry->delta) {
1634                 max_size = trg_size/2 - 20;
1635                 ref_depth = 1;
1636         } else {
1637                 max_size = trg_entry->delta_size;
1638                 ref_depth = trg->depth;
1639         }
1640         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1641                                                 (max_depth - ref_depth + 1);
1642         if (max_size == 0)
1643                 return 0;
1644         src_size = src_entry->size;
1645         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1646         if (sizediff >= max_size)
1647                 return 0;
1648         if (trg_size < src_size / 32)
1649                 return 0;
1650
1651         /* Load data if not already done */
1652         if (!trg->data) {
1653                 read_lock();
1654                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1655                 read_unlock();
1656                 if (!trg->data)
1657                         die("object %s cannot be read",
1658                             sha1_to_hex(trg_entry->idx.sha1));
1659                 if (sz != trg_size)
1660                         die("object %s inconsistent object length (%lu vs %lu)",
1661                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1662                 *mem_usage += sz;
1663         }
1664         if (!src->data) {
1665                 read_lock();
1666                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1667                 read_unlock();
1668                 if (!src->data) {
1669                         if (src_entry->preferred_base) {
1670                                 static int warned = 0;
1671                                 if (!warned++)
1672                                         warning("object %s cannot be read",
1673                                                 sha1_to_hex(src_entry->idx.sha1));
1674                                 /*
1675                                  * Those objects are not included in the
1676                                  * resulting pack.  Be resilient and ignore
1677                                  * them if they can't be read, in case the
1678                                  * pack could be created nevertheless.
1679                                  */
1680                                 return 0;
1681                         }
1682                         die("object %s cannot be read",
1683                             sha1_to_hex(src_entry->idx.sha1));
1684                 }
1685                 if (sz != src_size)
1686                         die("object %s inconsistent object length (%lu vs %lu)",
1687                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1688                 *mem_usage += sz;
1689         }
1690         if (!src->index) {
1691                 src->index = create_delta_index(src->data, src_size);
1692                 if (!src->index) {
1693                         static int warned = 0;
1694                         if (!warned++)
1695                                 warning("suboptimal pack - out of memory");
1696                         return 0;
1697                 }
1698                 *mem_usage += sizeof_delta_index(src->index);
1699         }
1700
1701         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1702         if (!delta_buf)
1703                 return 0;
1704
1705         if (trg_entry->delta) {
1706                 /* Prefer only shallower same-sized deltas. */
1707                 if (delta_size == trg_entry->delta_size &&
1708                     src->depth + 1 >= trg->depth) {
1709                         free(delta_buf);
1710                         return 0;
1711                 }
1712         }
1713
1714         /*
1715          * Handle memory allocation outside of the cache
1716          * accounting lock.  Compiler will optimize the strangeness
1717          * away when NO_PTHREADS is defined.
1718          */
1719         free(trg_entry->delta_data);
1720         cache_lock();
1721         if (trg_entry->delta_data) {
1722                 delta_cache_size -= trg_entry->delta_size;
1723                 trg_entry->delta_data = NULL;
1724         }
1725         if (delta_cacheable(src_size, trg_size, delta_size)) {
1726                 delta_cache_size += delta_size;
1727                 cache_unlock();
1728                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1729         } else {
1730                 cache_unlock();
1731                 free(delta_buf);
1732         }
1733
1734         trg_entry->delta = src_entry;
1735         trg_entry->delta_size = delta_size;
1736         trg->depth = src->depth + 1;
1737
1738         return 1;
1739 }
1740
1741 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1742 {
1743         struct object_entry *child = me->delta_child;
1744         unsigned int m = n;
1745         while (child) {
1746                 unsigned int c = check_delta_limit(child, n + 1);
1747                 if (m < c)
1748                         m = c;
1749                 child = child->delta_sibling;
1750         }
1751         return m;
1752 }
1753
1754 static unsigned long free_unpacked(struct unpacked *n)
1755 {
1756         unsigned long freed_mem = sizeof_delta_index(n->index);
1757         free_delta_index(n->index);
1758         n->index = NULL;
1759         if (n->data) {
1760                 freed_mem += n->entry->size;
1761                 free(n->data);
1762                 n->data = NULL;
1763         }
1764         n->entry = NULL;
1765         n->depth = 0;
1766         return freed_mem;
1767 }
1768
1769 static void find_deltas(struct object_entry **list, unsigned *list_size,
1770                         int window, int depth, unsigned *processed)
1771 {
1772         uint32_t i, idx = 0, count = 0;
1773         struct unpacked *array;
1774         unsigned long mem_usage = 0;
1775
1776         array = xcalloc(window, sizeof(struct unpacked));
1777
1778         for (;;) {
1779                 struct object_entry *entry;
1780                 struct unpacked *n = array + idx;
1781                 int j, max_depth, best_base = -1;
1782
1783                 progress_lock();
1784                 if (!*list_size) {
1785                         progress_unlock();
1786                         break;
1787                 }
1788                 entry = *list++;
1789                 (*list_size)--;
1790                 if (!entry->preferred_base) {
1791                         (*processed)++;
1792                         display_progress(progress_state, *processed);
1793                 }
1794                 progress_unlock();
1795
1796                 mem_usage -= free_unpacked(n);
1797                 n->entry = entry;
1798
1799                 while (window_memory_limit &&
1800                        mem_usage > window_memory_limit &&
1801                        count > 1) {
1802                         uint32_t tail = (idx + window - count) % window;
1803                         mem_usage -= free_unpacked(array + tail);
1804                         count--;
1805                 }
1806
1807                 /* We do not compute delta to *create* objects we are not
1808                  * going to pack.
1809                  */
1810                 if (entry->preferred_base)
1811                         goto next;
1812
1813                 /*
1814                  * If the current object is at pack edge, take the depth the
1815                  * objects that depend on the current object into account
1816                  * otherwise they would become too deep.
1817                  */
1818                 max_depth = depth;
1819                 if (entry->delta_child) {
1820                         max_depth -= check_delta_limit(entry, 0);
1821                         if (max_depth <= 0)
1822                                 goto next;
1823                 }
1824
1825                 j = window;
1826                 while (--j > 0) {
1827                         int ret;
1828                         uint32_t other_idx = idx + j;
1829                         struct unpacked *m;
1830                         if (other_idx >= window)
1831                                 other_idx -= window;
1832                         m = array + other_idx;
1833                         if (!m->entry)
1834                                 break;
1835                         ret = try_delta(n, m, max_depth, &mem_usage);
1836                         if (ret < 0)
1837                                 break;
1838                         else if (ret > 0)
1839                                 best_base = other_idx;
1840                 }
1841
1842                 /*
1843                  * If we decided to cache the delta data, then it is best
1844                  * to compress it right away.  First because we have to do
1845                  * it anyway, and doing it here while we're threaded will
1846                  * save a lot of time in the non threaded write phase,
1847                  * as well as allow for caching more deltas within
1848                  * the same cache size limit.
1849                  * ...
1850                  * But only if not writing to stdout, since in that case
1851                  * the network is most likely throttling writes anyway,
1852                  * and therefore it is best to go to the write phase ASAP
1853                  * instead, as we can afford spending more time compressing
1854                  * between writes at that moment.
1855                  */
1856                 if (entry->delta_data && !pack_to_stdout) {
1857                         entry->z_delta_size = do_compress(&entry->delta_data,
1858                                                           entry->delta_size);
1859                         cache_lock();
1860                         delta_cache_size -= entry->delta_size;
1861                         delta_cache_size += entry->z_delta_size;
1862                         cache_unlock();
1863                 }
1864
1865                 /* if we made n a delta, and if n is already at max
1866                  * depth, leaving it in the window is pointless.  we
1867                  * should evict it first.
1868                  */
1869                 if (entry->delta && max_depth <= n->depth)
1870                         continue;
1871
1872                 /*
1873                  * Move the best delta base up in the window, after the
1874                  * currently deltified object, to keep it longer.  It will
1875                  * be the first base object to be attempted next.
1876                  */
1877                 if (entry->delta) {
1878                         struct unpacked swap = array[best_base];
1879                         int dist = (window + idx - best_base) % window;
1880                         int dst = best_base;
1881                         while (dist--) {
1882                                 int src = (dst + 1) % window;
1883                                 array[dst] = array[src];
1884                                 dst = src;
1885                         }
1886                         array[dst] = swap;
1887                 }
1888
1889                 next:
1890                 idx++;
1891                 if (count + 1 < window)
1892                         count++;
1893                 if (idx >= window)
1894                         idx = 0;
1895         }
1896
1897         for (i = 0; i < window; ++i) {
1898                 free_delta_index(array[i].index);
1899                 free(array[i].data);
1900         }
1901         free(array);
1902 }
1903
1904 #ifndef NO_PTHREADS
1905
1906 static void try_to_free_from_threads(size_t size)
1907 {
1908         read_lock();
1909         release_pack_memory(size);
1910         read_unlock();
1911 }
1912
1913 static try_to_free_t old_try_to_free_routine;
1914
1915 /*
1916  * The main thread waits on the condition that (at least) one of the workers
1917  * has stopped working (which is indicated in the .working member of
1918  * struct thread_params).
1919  * When a work thread has completed its work, it sets .working to 0 and
1920  * signals the main thread and waits on the condition that .data_ready
1921  * becomes 1.
1922  */
1923
1924 struct thread_params {
1925         pthread_t thread;
1926         struct object_entry **list;
1927         unsigned list_size;
1928         unsigned remaining;
1929         int window;
1930         int depth;
1931         int working;
1932         int data_ready;
1933         pthread_mutex_t mutex;
1934         pthread_cond_t cond;
1935         unsigned *processed;
1936 };
1937
1938 static pthread_cond_t progress_cond;
1939
1940 /*
1941  * Mutex and conditional variable can't be statically-initialized on Windows.
1942  */
1943 static void init_threaded_search(void)
1944 {
1945         init_recursive_mutex(&read_mutex);
1946         pthread_mutex_init(&cache_mutex, NULL);
1947         pthread_mutex_init(&progress_mutex, NULL);
1948         pthread_cond_init(&progress_cond, NULL);
1949         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1950 }
1951
1952 static void cleanup_threaded_search(void)
1953 {
1954         set_try_to_free_routine(old_try_to_free_routine);
1955         pthread_cond_destroy(&progress_cond);
1956         pthread_mutex_destroy(&read_mutex);
1957         pthread_mutex_destroy(&cache_mutex);
1958         pthread_mutex_destroy(&progress_mutex);
1959 }
1960
1961 static void *threaded_find_deltas(void *arg)
1962 {
1963         struct thread_params *me = arg;
1964
1965         while (me->remaining) {
1966                 find_deltas(me->list, &me->remaining,
1967                             me->window, me->depth, me->processed);
1968
1969                 progress_lock();
1970                 me->working = 0;
1971                 pthread_cond_signal(&progress_cond);
1972                 progress_unlock();
1973
1974                 /*
1975                  * We must not set ->data_ready before we wait on the
1976                  * condition because the main thread may have set it to 1
1977                  * before we get here. In order to be sure that new
1978                  * work is available if we see 1 in ->data_ready, it
1979                  * was initialized to 0 before this thread was spawned
1980                  * and we reset it to 0 right away.
1981                  */
1982                 pthread_mutex_lock(&me->mutex);
1983                 while (!me->data_ready)
1984                         pthread_cond_wait(&me->cond, &me->mutex);
1985                 me->data_ready = 0;
1986                 pthread_mutex_unlock(&me->mutex);
1987         }
1988         /* leave ->working 1 so that this doesn't get more work assigned */
1989         return NULL;
1990 }
1991
1992 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1993                            int window, int depth, unsigned *processed)
1994 {
1995         struct thread_params *p;
1996         int i, ret, active_threads = 0;
1997
1998         init_threaded_search();
1999
2000         if (delta_search_threads <= 1) {
2001                 find_deltas(list, &list_size, window, depth, processed);
2002                 cleanup_threaded_search();
2003                 return;
2004         }
2005         if (progress > pack_to_stdout)
2006                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2007                                 delta_search_threads);
2008         p = xcalloc(delta_search_threads, sizeof(*p));
2009
2010         /* Partition the work amongst work threads. */
2011         for (i = 0; i < delta_search_threads; i++) {
2012                 unsigned sub_size = list_size / (delta_search_threads - i);
2013
2014                 /* don't use too small segments or no deltas will be found */
2015                 if (sub_size < 2*window && i+1 < delta_search_threads)
2016                         sub_size = 0;
2017
2018                 p[i].window = window;
2019                 p[i].depth = depth;
2020                 p[i].processed = processed;
2021                 p[i].working = 1;
2022                 p[i].data_ready = 0;
2023
2024                 /* try to split chunks on "path" boundaries */
2025                 while (sub_size && sub_size < list_size &&
2026                        list[sub_size]->hash &&
2027                        list[sub_size]->hash == list[sub_size-1]->hash)
2028                         sub_size++;
2029
2030                 p[i].list = list;
2031                 p[i].list_size = sub_size;
2032                 p[i].remaining = sub_size;
2033
2034                 list += sub_size;
2035                 list_size -= sub_size;
2036         }
2037
2038         /* Start work threads. */
2039         for (i = 0; i < delta_search_threads; i++) {
2040                 if (!p[i].list_size)
2041                         continue;
2042                 pthread_mutex_init(&p[i].mutex, NULL);
2043                 pthread_cond_init(&p[i].cond, NULL);
2044                 ret = pthread_create(&p[i].thread, NULL,
2045                                      threaded_find_deltas, &p[i]);
2046                 if (ret)
2047                         die("unable to create thread: %s", strerror(ret));
2048                 active_threads++;
2049         }
2050
2051         /*
2052          * Now let's wait for work completion.  Each time a thread is done
2053          * with its work, we steal half of the remaining work from the
2054          * thread with the largest number of unprocessed objects and give
2055          * it to that newly idle thread.  This ensure good load balancing
2056          * until the remaining object list segments are simply too short
2057          * to be worth splitting anymore.
2058          */
2059         while (active_threads) {
2060                 struct thread_params *target = NULL;
2061                 struct thread_params *victim = NULL;
2062                 unsigned sub_size = 0;
2063
2064                 progress_lock();
2065                 for (;;) {
2066                         for (i = 0; !target && i < delta_search_threads; i++)
2067                                 if (!p[i].working)
2068                                         target = &p[i];
2069                         if (target)
2070                                 break;
2071                         pthread_cond_wait(&progress_cond, &progress_mutex);
2072                 }
2073
2074                 for (i = 0; i < delta_search_threads; i++)
2075                         if (p[i].remaining > 2*window &&
2076                             (!victim || victim->remaining < p[i].remaining))
2077                                 victim = &p[i];
2078                 if (victim) {
2079                         sub_size = victim->remaining / 2;
2080                         list = victim->list + victim->list_size - sub_size;
2081                         while (sub_size && list[0]->hash &&
2082                                list[0]->hash == list[-1]->hash) {
2083                                 list++;
2084                                 sub_size--;
2085                         }
2086                         if (!sub_size) {
2087                                 /*
2088                                  * It is possible for some "paths" to have
2089                                  * so many objects that no hash boundary
2090                                  * might be found.  Let's just steal the
2091                                  * exact half in that case.
2092                                  */
2093                                 sub_size = victim->remaining / 2;
2094                                 list -= sub_size;
2095                         }
2096                         target->list = list;
2097                         victim->list_size -= sub_size;
2098                         victim->remaining -= sub_size;
2099                 }
2100                 target->list_size = sub_size;
2101                 target->remaining = sub_size;
2102                 target->working = 1;
2103                 progress_unlock();
2104
2105                 pthread_mutex_lock(&target->mutex);
2106                 target->data_ready = 1;
2107                 pthread_cond_signal(&target->cond);
2108                 pthread_mutex_unlock(&target->mutex);
2109
2110                 if (!sub_size) {
2111                         pthread_join(target->thread, NULL);
2112                         pthread_cond_destroy(&target->cond);
2113                         pthread_mutex_destroy(&target->mutex);
2114                         active_threads--;
2115                 }
2116         }
2117         cleanup_threaded_search();
2118         free(p);
2119 }
2120
2121 #else
2122 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2123 #endif
2124
2125 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2126 {
2127         struct object_id peeled;
2128
2129         if (starts_with(path, "refs/tags/") && /* is a tag? */
2130             !peel_ref(path, peeled.hash)    && /* peelable? */
2131             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2132                 add_object_entry(oid->hash, OBJ_TAG, NULL, 0);
2133         return 0;
2134 }
2135
2136 static void prepare_pack(int window, int depth)
2137 {
2138         struct object_entry **delta_list;
2139         uint32_t i, nr_deltas;
2140         unsigned n;
2141
2142         get_object_details();
2143
2144         /*
2145          * If we're locally repacking then we need to be doubly careful
2146          * from now on in order to make sure no stealth corruption gets
2147          * propagated to the new pack.  Clients receiving streamed packs
2148          * should validate everything they get anyway so no need to incur
2149          * the additional cost here in that case.
2150          */
2151         if (!pack_to_stdout)
2152                 do_check_packed_object_crc = 1;
2153
2154         if (!to_pack.nr_objects || !window || !depth)
2155                 return;
2156
2157         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2158         nr_deltas = n = 0;
2159
2160         for (i = 0; i < to_pack.nr_objects; i++) {
2161                 struct object_entry *entry = to_pack.objects + i;
2162
2163                 if (entry->delta)
2164                         /* This happens if we decided to reuse existing
2165                          * delta from a pack.  "reuse_delta &&" is implied.
2166                          */
2167                         continue;
2168
2169                 if (entry->size < 50)
2170                         continue;
2171
2172                 if (entry->no_try_delta)
2173                         continue;
2174
2175                 if (!entry->preferred_base) {
2176                         nr_deltas++;
2177                         if (entry->type < 0)
2178                                 die("unable to get type of object %s",
2179                                     sha1_to_hex(entry->idx.sha1));
2180                 } else {
2181                         if (entry->type < 0) {
2182                                 /*
2183                                  * This object is not found, but we
2184                                  * don't have to include it anyway.
2185                                  */
2186                                 continue;
2187                         }
2188                 }
2189
2190                 delta_list[n++] = entry;
2191         }
2192
2193         if (nr_deltas && n > 1) {
2194                 unsigned nr_done = 0;
2195                 if (progress)
2196                         progress_state = start_progress(_("Compressing objects"),
2197                                                         nr_deltas);
2198                 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2199                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2200                 stop_progress(&progress_state);
2201                 if (nr_done != nr_deltas)
2202                         die("inconsistency with delta count");
2203         }
2204         free(delta_list);
2205 }
2206
2207 static int git_pack_config(const char *k, const char *v, void *cb)
2208 {
2209         if (!strcmp(k, "pack.window")) {
2210                 window = git_config_int(k, v);
2211                 return 0;
2212         }
2213         if (!strcmp(k, "pack.windowmemory")) {
2214                 window_memory_limit = git_config_ulong(k, v);
2215                 return 0;
2216         }
2217         if (!strcmp(k, "pack.depth")) {
2218                 depth = git_config_int(k, v);
2219                 return 0;
2220         }
2221         if (!strcmp(k, "pack.compression")) {
2222                 int level = git_config_int(k, v);
2223                 if (level == -1)
2224                         level = Z_DEFAULT_COMPRESSION;
2225                 else if (level < 0 || level > Z_BEST_COMPRESSION)
2226                         die("bad pack compression level %d", level);
2227                 pack_compression_level = level;
2228                 pack_compression_seen = 1;
2229                 return 0;
2230         }
2231         if (!strcmp(k, "pack.deltacachesize")) {
2232                 max_delta_cache_size = git_config_int(k, v);
2233                 return 0;
2234         }
2235         if (!strcmp(k, "pack.deltacachelimit")) {
2236                 cache_max_small_delta_size = git_config_int(k, v);
2237                 return 0;
2238         }
2239         if (!strcmp(k, "pack.writebitmaphashcache")) {
2240                 if (git_config_bool(k, v))
2241                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2242                 else
2243                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2244         }
2245         if (!strcmp(k, "pack.usebitmaps")) {
2246                 use_bitmap_index = git_config_bool(k, v);
2247                 return 0;
2248         }
2249         if (!strcmp(k, "pack.threads")) {
2250                 delta_search_threads = git_config_int(k, v);
2251                 if (delta_search_threads < 0)
2252                         die("invalid number of threads specified (%d)",
2253                             delta_search_threads);
2254 #ifdef NO_PTHREADS
2255                 if (delta_search_threads != 1)
2256                         warning("no threads support, ignoring %s", k);
2257 #endif
2258                 return 0;
2259         }
2260         if (!strcmp(k, "pack.indexversion")) {
2261                 pack_idx_opts.version = git_config_int(k, v);
2262                 if (pack_idx_opts.version > 2)
2263                         die("bad pack.indexversion=%"PRIu32,
2264                             pack_idx_opts.version);
2265                 return 0;
2266         }
2267         return git_default_config(k, v, cb);
2268 }
2269
2270 static void read_object_list_from_stdin(void)
2271 {
2272         char line[40 + 1 + PATH_MAX + 2];
2273         unsigned char sha1[20];
2274
2275         for (;;) {
2276                 if (!fgets(line, sizeof(line), stdin)) {
2277                         if (feof(stdin))
2278                                 break;
2279                         if (!ferror(stdin))
2280                                 die("fgets returned NULL, not EOF, not error!");
2281                         if (errno != EINTR)
2282                                 die_errno("fgets");
2283                         clearerr(stdin);
2284                         continue;
2285                 }
2286                 if (line[0] == '-') {
2287                         if (get_sha1_hex(line+1, sha1))
2288                                 die("expected edge sha1, got garbage:\n %s",
2289                                     line);
2290                         add_preferred_base(sha1);
2291                         continue;
2292                 }
2293                 if (get_sha1_hex(line, sha1))
2294                         die("expected sha1, got garbage:\n %s", line);
2295
2296                 add_preferred_base_object(line+41);
2297                 add_object_entry(sha1, 0, line+41, 0);
2298         }
2299 }
2300
2301 #define OBJECT_ADDED (1u<<20)
2302
2303 static void show_commit(struct commit *commit, void *data)
2304 {
2305         add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2306         commit->object.flags |= OBJECT_ADDED;
2307
2308         if (write_bitmap_index)
2309                 index_commit_for_bitmap(commit);
2310 }
2311
2312 static void show_object(struct object *obj, const char *name, void *data)
2313 {
2314         add_preferred_base_object(name);
2315         add_object_entry(obj->oid.hash, obj->type, name, 0);
2316         obj->flags |= OBJECT_ADDED;
2317 }
2318
2319 static void show_edge(struct commit *commit)
2320 {
2321         add_preferred_base(commit->object.oid.hash);
2322 }
2323
2324 struct in_pack_object {
2325         off_t offset;
2326         struct object *object;
2327 };
2328
2329 struct in_pack {
2330         int alloc;
2331         int nr;
2332         struct in_pack_object *array;
2333 };
2334
2335 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2336 {
2337         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2338         in_pack->array[in_pack->nr].object = object;
2339         in_pack->nr++;
2340 }
2341
2342 /*
2343  * Compare the objects in the offset order, in order to emulate the
2344  * "git rev-list --objects" output that produced the pack originally.
2345  */
2346 static int ofscmp(const void *a_, const void *b_)
2347 {
2348         struct in_pack_object *a = (struct in_pack_object *)a_;
2349         struct in_pack_object *b = (struct in_pack_object *)b_;
2350
2351         if (a->offset < b->offset)
2352                 return -1;
2353         else if (a->offset > b->offset)
2354                 return 1;
2355         else
2356                 return oidcmp(&a->object->oid, &b->object->oid);
2357 }
2358
2359 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2360 {
2361         struct packed_git *p;
2362         struct in_pack in_pack;
2363         uint32_t i;
2364
2365         memset(&in_pack, 0, sizeof(in_pack));
2366
2367         for (p = packed_git; p; p = p->next) {
2368                 const unsigned char *sha1;
2369                 struct object *o;
2370
2371                 if (!p->pack_local || p->pack_keep)
2372                         continue;
2373                 if (open_pack_index(p))
2374                         die("cannot open pack index");
2375
2376                 ALLOC_GROW(in_pack.array,
2377                            in_pack.nr + p->num_objects,
2378                            in_pack.alloc);
2379
2380                 for (i = 0; i < p->num_objects; i++) {
2381                         sha1 = nth_packed_object_sha1(p, i);
2382                         o = lookup_unknown_object(sha1);
2383                         if (!(o->flags & OBJECT_ADDED))
2384                                 mark_in_pack_object(o, p, &in_pack);
2385                         o->flags |= OBJECT_ADDED;
2386                 }
2387         }
2388
2389         if (in_pack.nr) {
2390                 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2391                       ofscmp);
2392                 for (i = 0; i < in_pack.nr; i++) {
2393                         struct object *o = in_pack.array[i].object;
2394                         add_object_entry(o->oid.hash, o->type, "", 0);
2395                 }
2396         }
2397         free(in_pack.array);
2398 }
2399
2400 static int add_loose_object(const unsigned char *sha1, const char *path,
2401                             void *data)
2402 {
2403         enum object_type type = sha1_object_info(sha1, NULL);
2404
2405         if (type < 0) {
2406                 warning("loose object at %s could not be examined", path);
2407                 return 0;
2408         }
2409
2410         add_object_entry(sha1, type, "", 0);
2411         return 0;
2412 }
2413
2414 /*
2415  * We actually don't even have to worry about reachability here.
2416  * add_object_entry will weed out duplicates, so we just add every
2417  * loose object we find.
2418  */
2419 static void add_unreachable_loose_objects(void)
2420 {
2421         for_each_loose_file_in_objdir(get_object_directory(),
2422                                       add_loose_object,
2423                                       NULL, NULL, NULL);
2424 }
2425
2426 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2427 {
2428         static struct packed_git *last_found = (void *)1;
2429         struct packed_git *p;
2430
2431         p = (last_found != (void *)1) ? last_found : packed_git;
2432
2433         while (p) {
2434                 if ((!p->pack_local || p->pack_keep) &&
2435                         find_pack_entry_one(sha1, p)) {
2436                         last_found = p;
2437                         return 1;
2438                 }
2439                 if (p == last_found)
2440                         p = packed_git;
2441                 else
2442                         p = p->next;
2443                 if (p == last_found)
2444                         p = p->next;
2445         }
2446         return 0;
2447 }
2448
2449 /*
2450  * Store a list of sha1s that are should not be discarded
2451  * because they are either written too recently, or are
2452  * reachable from another object that was.
2453  *
2454  * This is filled by get_object_list.
2455  */
2456 static struct sha1_array recent_objects;
2457
2458 static int loosened_object_can_be_discarded(const unsigned char *sha1,
2459                                             unsigned long mtime)
2460 {
2461         if (!unpack_unreachable_expiration)
2462                 return 0;
2463         if (mtime > unpack_unreachable_expiration)
2464                 return 0;
2465         if (sha1_array_lookup(&recent_objects, sha1) >= 0)
2466                 return 0;
2467         return 1;
2468 }
2469
2470 static void loosen_unused_packed_objects(struct rev_info *revs)
2471 {
2472         struct packed_git *p;
2473         uint32_t i;
2474         const unsigned char *sha1;
2475
2476         for (p = packed_git; p; p = p->next) {
2477                 if (!p->pack_local || p->pack_keep)
2478                         continue;
2479
2480                 if (open_pack_index(p))
2481                         die("cannot open pack index");
2482
2483                 for (i = 0; i < p->num_objects; i++) {
2484                         sha1 = nth_packed_object_sha1(p, i);
2485                         if (!packlist_find(&to_pack, sha1, NULL) &&
2486                             !has_sha1_pack_kept_or_nonlocal(sha1) &&
2487                             !loosened_object_can_be_discarded(sha1, p->mtime))
2488                                 if (force_object_loose(sha1, p->mtime))
2489                                         die("unable to force loose object");
2490                 }
2491         }
2492 }
2493
2494 /*
2495  * This tracks any options which a reader of the pack might
2496  * not understand, and which would therefore prevent blind reuse
2497  * of what we have on disk.
2498  */
2499 static int pack_options_allow_reuse(void)
2500 {
2501         return allow_ofs_delta;
2502 }
2503
2504 static int get_object_list_from_bitmap(struct rev_info *revs)
2505 {
2506         if (prepare_bitmap_walk(revs) < 0)
2507                 return -1;
2508
2509         if (pack_options_allow_reuse() &&
2510             !reuse_partial_packfile_from_bitmap(
2511                         &reuse_packfile,
2512                         &reuse_packfile_objects,
2513                         &reuse_packfile_offset)) {
2514                 assert(reuse_packfile_objects);
2515                 nr_result += reuse_packfile_objects;
2516                 display_progress(progress_state, nr_result);
2517         }
2518
2519         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2520         return 0;
2521 }
2522
2523 static void record_recent_object(struct object *obj,
2524                                  const char *name,
2525                                  void *data)
2526 {
2527         sha1_array_append(&recent_objects, obj->oid.hash);
2528 }
2529
2530 static void record_recent_commit(struct commit *commit, void *data)
2531 {
2532         sha1_array_append(&recent_objects, commit->object.oid.hash);
2533 }
2534
2535 static void get_object_list(int ac, const char **av)
2536 {
2537         struct rev_info revs;
2538         char line[1000];
2539         int flags = 0;
2540
2541         init_revisions(&revs, NULL);
2542         save_commit_buffer = 0;
2543         setup_revisions(ac, av, &revs, NULL);
2544
2545         /* make sure shallows are read */
2546         is_repository_shallow();
2547
2548         while (fgets(line, sizeof(line), stdin) != NULL) {
2549                 int len = strlen(line);
2550                 if (len && line[len - 1] == '\n')
2551                         line[--len] = 0;
2552                 if (!len)
2553                         break;
2554                 if (*line == '-') {
2555                         if (!strcmp(line, "--not")) {
2556                                 flags ^= UNINTERESTING;
2557                                 write_bitmap_index = 0;
2558                                 continue;
2559                         }
2560                         if (starts_with(line, "--shallow ")) {
2561                                 unsigned char sha1[20];
2562                                 if (get_sha1_hex(line + 10, sha1))
2563                                         die("not an SHA-1 '%s'", line + 10);
2564                                 register_shallow(sha1);
2565                                 use_bitmap_index = 0;
2566                                 continue;
2567                         }
2568                         die("not a rev '%s'", line);
2569                 }
2570                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2571                         die("bad revision '%s'", line);
2572         }
2573
2574         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2575                 return;
2576
2577         if (prepare_revision_walk(&revs))
2578                 die("revision walk setup failed");
2579         mark_edges_uninteresting(&revs, show_edge);
2580         traverse_commit_list(&revs, show_commit, show_object, NULL);
2581
2582         if (unpack_unreachable_expiration) {
2583                 revs.ignore_missing_links = 1;
2584                 if (add_unseen_recent_objects_to_traversal(&revs,
2585                                 unpack_unreachable_expiration))
2586                         die("unable to add recent objects");
2587                 if (prepare_revision_walk(&revs))
2588                         die("revision walk setup failed");
2589                 traverse_commit_list(&revs, record_recent_commit,
2590                                      record_recent_object, NULL);
2591         }
2592
2593         if (keep_unreachable)
2594                 add_objects_in_unpacked_packs(&revs);
2595         if (pack_loose_unreachable)
2596                 add_unreachable_loose_objects();
2597         if (unpack_unreachable)
2598                 loosen_unused_packed_objects(&revs);
2599
2600         sha1_array_clear(&recent_objects);
2601 }
2602
2603 static int option_parse_index_version(const struct option *opt,
2604                                       const char *arg, int unset)
2605 {
2606         char *c;
2607         const char *val = arg;
2608         pack_idx_opts.version = strtoul(val, &c, 10);
2609         if (pack_idx_opts.version > 2)
2610                 die(_("unsupported index version %s"), val);
2611         if (*c == ',' && c[1])
2612                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2613         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2614                 die(_("bad index version '%s'"), val);
2615         return 0;
2616 }
2617
2618 static int option_parse_unpack_unreachable(const struct option *opt,
2619                                            const char *arg, int unset)
2620 {
2621         if (unset) {
2622                 unpack_unreachable = 0;
2623                 unpack_unreachable_expiration = 0;
2624         }
2625         else {
2626                 unpack_unreachable = 1;
2627                 if (arg)
2628                         unpack_unreachable_expiration = approxidate(arg);
2629         }
2630         return 0;
2631 }
2632
2633 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2634 {
2635         int use_internal_rev_list = 0;
2636         int thin = 0;
2637         int shallow = 0;
2638         int all_progress_implied = 0;
2639         struct argv_array rp = ARGV_ARRAY_INIT;
2640         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2641         int rev_list_index = 0;
2642         struct option pack_objects_options[] = {
2643                 OPT_SET_INT('q', "quiet", &progress,
2644                             N_("do not show progress meter"), 0),
2645                 OPT_SET_INT(0, "progress", &progress,
2646                             N_("show progress meter"), 1),
2647                 OPT_SET_INT(0, "all-progress", &progress,
2648                             N_("show progress meter during object writing phase"), 2),
2649                 OPT_BOOL(0, "all-progress-implied",
2650                          &all_progress_implied,
2651                          N_("similar to --all-progress when progress meter is shown")),
2652                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2653                   N_("write the pack index file in the specified idx format version"),
2654                   0, option_parse_index_version },
2655                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2656                               N_("maximum size of each output pack file")),
2657                 OPT_BOOL(0, "local", &local,
2658                          N_("ignore borrowed objects from alternate object store")),
2659                 OPT_BOOL(0, "incremental", &incremental,
2660                          N_("ignore packed objects")),
2661                 OPT_INTEGER(0, "window", &window,
2662                             N_("limit pack window by objects")),
2663                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2664                               N_("limit pack window by memory in addition to object limit")),
2665                 OPT_INTEGER(0, "depth", &depth,
2666                             N_("maximum length of delta chain allowed in the resulting pack")),
2667                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2668                          N_("reuse existing deltas")),
2669                 OPT_BOOL(0, "reuse-object", &reuse_object,
2670                          N_("reuse existing objects")),
2671                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2672                          N_("use OFS_DELTA objects")),
2673                 OPT_INTEGER(0, "threads", &delta_search_threads,
2674                             N_("use threads when searching for best delta matches")),
2675                 OPT_BOOL(0, "non-empty", &non_empty,
2676                          N_("do not create an empty pack output")),
2677                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2678                          N_("read revision arguments from standard input")),
2679                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2680                   N_("limit the objects to those that are not yet packed"),
2681                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2682                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2683                   N_("include objects reachable from any reference"),
2684                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2685                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2686                   N_("include objects referred by reflog entries"),
2687                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2688                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2689                   N_("include objects referred to by the index"),
2690                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2691                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2692                          N_("output pack to stdout")),
2693                 OPT_BOOL(0, "include-tag", &include_tag,
2694                          N_("include tag objects that refer to objects to be packed")),
2695                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2696                          N_("keep unreachable objects")),
2697                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
2698                          N_("pack loose unreachable objects")),
2699                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2700                   N_("unpack unreachable objects newer than <time>"),
2701                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2702                 OPT_BOOL(0, "thin", &thin,
2703                          N_("create thin packs")),
2704                 OPT_BOOL(0, "shallow", &shallow,
2705                          N_("create packs suitable for shallow fetches")),
2706                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2707                          N_("ignore packs that have companion .keep file")),
2708                 OPT_INTEGER(0, "compression", &pack_compression_level,
2709                             N_("pack compression level")),
2710                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2711                             N_("do not hide commits by grafts"), 0),
2712                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2713                          N_("use a bitmap index if available to speed up counting objects")),
2714                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2715                          N_("write a bitmap index together with the pack index")),
2716                 OPT_END(),
2717         };
2718
2719         check_replace_refs = 0;
2720
2721         reset_pack_idx_option(&pack_idx_opts);
2722         git_config(git_pack_config, NULL);
2723         if (!pack_compression_seen && core_compression_seen)
2724                 pack_compression_level = core_compression_level;
2725
2726         progress = isatty(2);
2727         argc = parse_options(argc, argv, prefix, pack_objects_options,
2728                              pack_usage, 0);
2729
2730         if (argc) {
2731                 base_name = argv[0];
2732                 argc--;
2733         }
2734         if (pack_to_stdout != !base_name || argc)
2735                 usage_with_options(pack_usage, pack_objects_options);
2736
2737         argv_array_push(&rp, "pack-objects");
2738         if (thin) {
2739                 use_internal_rev_list = 1;
2740                 argv_array_push(&rp, shallow
2741                                 ? "--objects-edge-aggressive"
2742                                 : "--objects-edge");
2743         } else
2744                 argv_array_push(&rp, "--objects");
2745
2746         if (rev_list_all) {
2747                 use_internal_rev_list = 1;
2748                 argv_array_push(&rp, "--all");
2749         }
2750         if (rev_list_reflog) {
2751                 use_internal_rev_list = 1;
2752                 argv_array_push(&rp, "--reflog");
2753         }
2754         if (rev_list_index) {
2755                 use_internal_rev_list = 1;
2756                 argv_array_push(&rp, "--indexed-objects");
2757         }
2758         if (rev_list_unpacked) {
2759                 use_internal_rev_list = 1;
2760                 argv_array_push(&rp, "--unpacked");
2761         }
2762
2763         if (!reuse_object)
2764                 reuse_delta = 0;
2765         if (pack_compression_level == -1)
2766                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2767         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2768                 die("bad pack compression level %d", pack_compression_level);
2769
2770         if (!delta_search_threads)      /* --threads=0 means autodetect */
2771                 delta_search_threads = online_cpus();
2772
2773 #ifdef NO_PTHREADS
2774         if (delta_search_threads != 1)
2775                 warning("no threads support, ignoring --threads");
2776 #endif
2777         if (!pack_to_stdout && !pack_size_limit)
2778                 pack_size_limit = pack_size_limit_cfg;
2779         if (pack_to_stdout && pack_size_limit)
2780                 die("--max-pack-size cannot be used to build a pack for transfer.");
2781         if (pack_size_limit && pack_size_limit < 1024*1024) {
2782                 warning("minimum pack size limit is 1 MiB");
2783                 pack_size_limit = 1024*1024;
2784         }
2785
2786         if (!pack_to_stdout && thin)
2787                 die("--thin cannot be used to build an indexable pack.");
2788
2789         if (keep_unreachable && unpack_unreachable)
2790                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2791         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
2792                 unpack_unreachable_expiration = 0;
2793
2794         if (!use_internal_rev_list || !pack_to_stdout || is_repository_shallow())
2795                 use_bitmap_index = 0;
2796
2797         if (pack_to_stdout || !rev_list_all)
2798                 write_bitmap_index = 0;
2799
2800         if (progress && all_progress_implied)
2801                 progress = 2;
2802
2803         prepare_packed_git();
2804         if (ignore_packed_keep) {
2805                 struct packed_git *p;
2806                 for (p = packed_git; p; p = p->next)
2807                         if (p->pack_local && p->pack_keep)
2808                                 break;
2809                 if (!p) /* no keep-able packs found */
2810                         ignore_packed_keep = 0;
2811         }
2812         if (local) {
2813                 /*
2814                  * unlike ignore_packed_keep above, we do not want to
2815                  * unset "local" based on looking at packs, as it
2816                  * also covers non-local objects
2817                  */
2818                 struct packed_git *p;
2819                 for (p = packed_git; p; p = p->next) {
2820                         if (!p->pack_local) {
2821                                 have_non_local_packs = 1;
2822                                 break;
2823                         }
2824                 }
2825         }
2826
2827         if (progress)
2828                 progress_state = start_progress(_("Counting objects"), 0);
2829         if (!use_internal_rev_list)
2830                 read_object_list_from_stdin();
2831         else {
2832                 get_object_list(rp.argc, rp.argv);
2833                 argv_array_clear(&rp);
2834         }
2835         cleanup_preferred_base();
2836         if (include_tag && nr_result)
2837                 for_each_ref(add_ref_tag, NULL);
2838         stop_progress(&progress_state);
2839
2840         if (non_empty && !nr_result)
2841                 return 0;
2842         if (nr_result)
2843                 prepare_pack(window, depth);
2844         write_pack_file();
2845         if (progress)
2846                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2847                         " reused %"PRIu32" (delta %"PRIu32")\n",
2848                         written, written_delta, reused, reused_delta);
2849         return 0;
2850 }