Merge branch 'jc/push-default-explicit' into maint
[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 #include "mru.h"
27
28 static const char *pack_usage[] = {
29         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
30         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
31         NULL
32 };
33
34 /*
35  * Objects we are going to pack are collected in the `to_pack` structure.
36  * It contains an array (dynamically expanded) of the object data, and a map
37  * that can resolve SHA1s to their position in the array.
38  */
39 static struct packing_data to_pack;
40
41 static struct pack_idx_entry **written_list;
42 static uint32_t nr_result, nr_written;
43
44 static int non_empty;
45 static int reuse_delta = 1, reuse_object = 1;
46 static int keep_unreachable, unpack_unreachable, include_tag;
47 static unsigned long unpack_unreachable_expiration;
48 static int pack_loose_unreachable;
49 static int local;
50 static int have_non_local_packs;
51 static int incremental;
52 static int ignore_packed_keep;
53 static int allow_ofs_delta;
54 static struct pack_idx_option pack_idx_opts;
55 static const char *base_name;
56 static int progress = 1;
57 static int window = 10;
58 static unsigned long pack_size_limit;
59 static int depth = 50;
60 static int delta_search_threads;
61 static int pack_to_stdout;
62 static int num_preferred_base;
63 static struct progress *progress_state;
64 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
65 static int pack_compression_seen;
66
67 static struct packed_git *reuse_packfile;
68 static uint32_t reuse_packfile_objects;
69 static off_t reuse_packfile_offset;
70
71 static int use_bitmap_index_default = 1;
72 static int use_bitmap_index = -1;
73 static int write_bitmap_index;
74 static uint16_t write_bitmap_options;
75
76 static unsigned long delta_cache_size = 0;
77 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
78 static unsigned long cache_max_small_delta_size = 1000;
79
80 static unsigned long window_memory_limit = 0;
81
82 /*
83  * stats
84  */
85 static uint32_t written, written_delta;
86 static uint32_t reused, reused_delta;
87
88 /*
89  * Indexed commits
90  */
91 static struct commit **indexed_commits;
92 static unsigned int indexed_commits_nr;
93 static unsigned int indexed_commits_alloc;
94
95 static void index_commit_for_bitmap(struct commit *commit)
96 {
97         if (indexed_commits_nr >= indexed_commits_alloc) {
98                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
99                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
100         }
101
102         indexed_commits[indexed_commits_nr++] = commit;
103 }
104
105 static void *get_delta(struct object_entry *entry)
106 {
107         unsigned long size, base_size, delta_size;
108         void *buf, *base_buf, *delta_buf;
109         enum object_type type;
110
111         buf = read_sha1_file(entry->idx.sha1, &type, &size);
112         if (!buf)
113                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
114         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
115         if (!base_buf)
116                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
117         delta_buf = diff_delta(base_buf, base_size,
118                                buf, size, &delta_size, 0);
119         if (!delta_buf || delta_size != entry->delta_size)
120                 die("delta size changed");
121         free(buf);
122         free(base_buf);
123         return delta_buf;
124 }
125
126 static unsigned long do_compress(void **pptr, unsigned long size)
127 {
128         git_zstream stream;
129         void *in, *out;
130         unsigned long maxsize;
131
132         git_deflate_init(&stream, pack_compression_level);
133         maxsize = git_deflate_bound(&stream, size);
134
135         in = *pptr;
136         out = xmalloc(maxsize);
137         *pptr = out;
138
139         stream.next_in = in;
140         stream.avail_in = size;
141         stream.next_out = out;
142         stream.avail_out = maxsize;
143         while (git_deflate(&stream, Z_FINISH) == Z_OK)
144                 ; /* nothing */
145         git_deflate_end(&stream);
146
147         free(in);
148         return stream.total_out;
149 }
150
151 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
152                                            const unsigned char *sha1)
153 {
154         git_zstream stream;
155         unsigned char ibuf[1024 * 16];
156         unsigned char obuf[1024 * 16];
157         unsigned long olen = 0;
158
159         git_deflate_init(&stream, pack_compression_level);
160
161         for (;;) {
162                 ssize_t readlen;
163                 int zret = Z_OK;
164                 readlen = read_istream(st, ibuf, sizeof(ibuf));
165                 if (readlen == -1)
166                         die(_("unable to read %s"), sha1_to_hex(sha1));
167
168                 stream.next_in = ibuf;
169                 stream.avail_in = readlen;
170                 while ((stream.avail_in || readlen == 0) &&
171                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
172                         stream.next_out = obuf;
173                         stream.avail_out = sizeof(obuf);
174                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
175                         sha1write(f, obuf, stream.next_out - obuf);
176                         olen += stream.next_out - obuf;
177                 }
178                 if (stream.avail_in)
179                         die(_("deflate error (%d)"), zret);
180                 if (readlen == 0) {
181                         if (zret != Z_STREAM_END)
182                                 die(_("deflate error (%d)"), zret);
183                         break;
184                 }
185         }
186         git_deflate_end(&stream);
187         return olen;
188 }
189
190 /*
191  * we are going to reuse the existing object data as is.  make
192  * sure it is not corrupt.
193  */
194 static int check_pack_inflate(struct packed_git *p,
195                 struct pack_window **w_curs,
196                 off_t offset,
197                 off_t len,
198                 unsigned long expect)
199 {
200         git_zstream stream;
201         unsigned char fakebuf[4096], *in;
202         int st;
203
204         memset(&stream, 0, sizeof(stream));
205         git_inflate_init(&stream);
206         do {
207                 in = use_pack(p, w_curs, offset, &stream.avail_in);
208                 stream.next_in = in;
209                 stream.next_out = fakebuf;
210                 stream.avail_out = sizeof(fakebuf);
211                 st = git_inflate(&stream, Z_FINISH);
212                 offset += stream.next_in - in;
213         } while (st == Z_OK || st == Z_BUF_ERROR);
214         git_inflate_end(&stream);
215         return (st == Z_STREAM_END &&
216                 stream.total_out == expect &&
217                 stream.total_in == len) ? 0 : -1;
218 }
219
220 static void copy_pack_data(struct sha1file *f,
221                 struct packed_git *p,
222                 struct pack_window **w_curs,
223                 off_t offset,
224                 off_t len)
225 {
226         unsigned char *in;
227         unsigned long avail;
228
229         while (len) {
230                 in = use_pack(p, w_curs, offset, &avail);
231                 if (avail > len)
232                         avail = (unsigned long)len;
233                 sha1write(f, in, avail);
234                 offset += avail;
235                 len -= avail;
236         }
237 }
238
239 /* Return 0 if we will bust the pack-size limit */
240 static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
241                                            unsigned long limit, int usable_delta)
242 {
243         unsigned long size, datalen;
244         unsigned char header[10], dheader[10];
245         unsigned hdrlen;
246         enum object_type type;
247         void *buf;
248         struct git_istream *st = NULL;
249
250         if (!usable_delta) {
251                 if (entry->type == OBJ_BLOB &&
252                     entry->size > big_file_threshold &&
253                     (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
254                         buf = NULL;
255                 else {
256                         buf = read_sha1_file(entry->idx.sha1, &type, &size);
257                         if (!buf)
258                                 die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
259                 }
260                 /*
261                  * make sure no cached delta data remains from a
262                  * previous attempt before a pack split occurred.
263                  */
264                 free(entry->delta_data);
265                 entry->delta_data = NULL;
266                 entry->z_delta_size = 0;
267         } else if (entry->delta_data) {
268                 size = entry->delta_size;
269                 buf = entry->delta_data;
270                 entry->delta_data = NULL;
271                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
272                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
273         } else {
274                 buf = get_delta(entry);
275                 size = entry->delta_size;
276                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
277                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
278         }
279
280         if (st) /* large blob case, just assume we don't compress well */
281                 datalen = size;
282         else if (entry->z_delta_size)
283                 datalen = entry->z_delta_size;
284         else
285                 datalen = do_compress(&buf, size);
286
287         /*
288          * The object header is a byte of 'type' followed by zero or
289          * more bytes of length.
290          */
291         hdrlen = encode_in_pack_object_header(type, size, header);
292
293         if (type == OBJ_OFS_DELTA) {
294                 /*
295                  * Deltas with relative base contain an additional
296                  * encoding of the relative offset for the delta
297                  * base from this object's position in the pack.
298                  */
299                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
300                 unsigned pos = sizeof(dheader) - 1;
301                 dheader[pos] = ofs & 127;
302                 while (ofs >>= 7)
303                         dheader[--pos] = 128 | (--ofs & 127);
304                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
305                         if (st)
306                                 close_istream(st);
307                         free(buf);
308                         return 0;
309                 }
310                 sha1write(f, header, hdrlen);
311                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
312                 hdrlen += sizeof(dheader) - pos;
313         } else if (type == OBJ_REF_DELTA) {
314                 /*
315                  * Deltas with a base reference contain
316                  * an additional 20 bytes for the base sha1.
317                  */
318                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
319                         if (st)
320                                 close_istream(st);
321                         free(buf);
322                         return 0;
323                 }
324                 sha1write(f, header, hdrlen);
325                 sha1write(f, entry->delta->idx.sha1, 20);
326                 hdrlen += 20;
327         } else {
328                 if (limit && hdrlen + datalen + 20 >= limit) {
329                         if (st)
330                                 close_istream(st);
331                         free(buf);
332                         return 0;
333                 }
334                 sha1write(f, header, hdrlen);
335         }
336         if (st) {
337                 datalen = write_large_blob_data(st, f, entry->idx.sha1);
338                 close_istream(st);
339         } else {
340                 sha1write(f, buf, datalen);
341                 free(buf);
342         }
343
344         return hdrlen + datalen;
345 }
346
347 /* Return 0 if we will bust the pack-size limit */
348 static off_t write_reuse_object(struct sha1file *f, struct object_entry *entry,
349                                 unsigned long limit, int usable_delta)
350 {
351         struct packed_git *p = entry->in_pack;
352         struct pack_window *w_curs = NULL;
353         struct revindex_entry *revidx;
354         off_t offset;
355         enum object_type type = entry->type;
356         off_t datalen;
357         unsigned char header[10], dheader[10];
358         unsigned hdrlen;
359
360         if (entry->delta)
361                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
362                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
363         hdrlen = encode_in_pack_object_header(type, entry->size, header);
364
365         offset = entry->in_pack_offset;
366         revidx = find_pack_revindex(p, offset);
367         datalen = revidx[1].offset - offset;
368         if (!pack_to_stdout && p->index_version > 1 &&
369             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
370                 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
371                 unuse_pack(&w_curs);
372                 return write_no_reuse_object(f, entry, limit, usable_delta);
373         }
374
375         offset += entry->in_pack_header_size;
376         datalen -= entry->in_pack_header_size;
377
378         if (!pack_to_stdout && p->index_version == 1 &&
379             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
380                 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
381                 unuse_pack(&w_curs);
382                 return write_no_reuse_object(f, entry, limit, usable_delta);
383         }
384
385         if (type == OBJ_OFS_DELTA) {
386                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
387                 unsigned pos = sizeof(dheader) - 1;
388                 dheader[pos] = ofs & 127;
389                 while (ofs >>= 7)
390                         dheader[--pos] = 128 | (--ofs & 127);
391                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
392                         unuse_pack(&w_curs);
393                         return 0;
394                 }
395                 sha1write(f, header, hdrlen);
396                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
397                 hdrlen += sizeof(dheader) - pos;
398                 reused_delta++;
399         } else if (type == OBJ_REF_DELTA) {
400                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
401                         unuse_pack(&w_curs);
402                         return 0;
403                 }
404                 sha1write(f, header, hdrlen);
405                 sha1write(f, entry->delta->idx.sha1, 20);
406                 hdrlen += 20;
407                 reused_delta++;
408         } else {
409                 if (limit && hdrlen + datalen + 20 >= limit) {
410                         unuse_pack(&w_curs);
411                         return 0;
412                 }
413                 sha1write(f, header, hdrlen);
414         }
415         copy_pack_data(f, p, &w_curs, offset, datalen);
416         unuse_pack(&w_curs);
417         reused++;
418         return hdrlen + datalen;
419 }
420
421 /* Return 0 if we will bust the pack-size limit */
422 static off_t write_object(struct sha1file *f,
423                           struct object_entry *entry,
424                           off_t write_offset)
425 {
426         unsigned long limit;
427         off_t len;
428         int usable_delta, to_reuse;
429
430         if (!pack_to_stdout)
431                 crc32_begin(f);
432
433         /* apply size limit if limited packsize and not first object */
434         if (!pack_size_limit || !nr_written)
435                 limit = 0;
436         else if (pack_size_limit <= write_offset)
437                 /*
438                  * the earlier object did not fit the limit; avoid
439                  * mistaking this with unlimited (i.e. limit = 0).
440                  */
441                 limit = 1;
442         else
443                 limit = pack_size_limit - write_offset;
444
445         if (!entry->delta)
446                 usable_delta = 0;       /* no delta */
447         else if (!pack_size_limit)
448                usable_delta = 1;        /* unlimited packfile */
449         else if (entry->delta->idx.offset == (off_t)-1)
450                 usable_delta = 0;       /* base was written to another pack */
451         else if (entry->delta->idx.offset)
452                 usable_delta = 1;       /* base already exists in this pack */
453         else
454                 usable_delta = 0;       /* base could end up in another pack */
455
456         if (!reuse_object)
457                 to_reuse = 0;   /* explicit */
458         else if (!entry->in_pack)
459                 to_reuse = 0;   /* can't reuse what we don't have */
460         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
461                                 /* check_object() decided it for us ... */
462                 to_reuse = usable_delta;
463                                 /* ... but pack split may override that */
464         else if (entry->type != entry->in_pack_type)
465                 to_reuse = 0;   /* pack has delta which is unusable */
466         else if (entry->delta)
467                 to_reuse = 0;   /* we want to pack afresh */
468         else
469                 to_reuse = 1;   /* we have it in-pack undeltified,
470                                  * and we do not need to deltify it.
471                                  */
472
473         if (!to_reuse)
474                 len = write_no_reuse_object(f, entry, limit, usable_delta);
475         else
476                 len = write_reuse_object(f, entry, limit, usable_delta);
477         if (!len)
478                 return 0;
479
480         if (usable_delta)
481                 written_delta++;
482         written++;
483         if (!pack_to_stdout)
484                 entry->idx.crc32 = crc32_end(f);
485         return len;
486 }
487
488 enum write_one_status {
489         WRITE_ONE_SKIP = -1, /* already written */
490         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
491         WRITE_ONE_WRITTEN = 1, /* normal */
492         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
493 };
494
495 static enum write_one_status write_one(struct sha1file *f,
496                                        struct object_entry *e,
497                                        off_t *offset)
498 {
499         off_t size;
500         int recursing;
501
502         /*
503          * we set offset to 1 (which is an impossible value) to mark
504          * the fact that this object is involved in "write its base
505          * first before writing a deltified object" recursion.
506          */
507         recursing = (e->idx.offset == 1);
508         if (recursing) {
509                 warning("recursive delta detected for object %s",
510                         sha1_to_hex(e->idx.sha1));
511                 return WRITE_ONE_RECURSIVE;
512         } else if (e->idx.offset || e->preferred_base) {
513                 /* offset is non zero if object is written already. */
514                 return WRITE_ONE_SKIP;
515         }
516
517         /* if we are deltified, write out base object first. */
518         if (e->delta) {
519                 e->idx.offset = 1; /* now recurse */
520                 switch (write_one(f, e->delta, offset)) {
521                 case WRITE_ONE_RECURSIVE:
522                         /* we cannot depend on this one */
523                         e->delta = NULL;
524                         break;
525                 default:
526                         break;
527                 case WRITE_ONE_BREAK:
528                         e->idx.offset = recursing;
529                         return WRITE_ONE_BREAK;
530                 }
531         }
532
533         e->idx.offset = *offset;
534         size = write_object(f, e, *offset);
535         if (!size) {
536                 e->idx.offset = recursing;
537                 return WRITE_ONE_BREAK;
538         }
539         written_list[nr_written++] = &e->idx;
540
541         /* make sure off_t is sufficiently large not to wrap */
542         if (signed_add_overflows(*offset, size))
543                 die("pack too large for current definition of off_t");
544         *offset += size;
545         return WRITE_ONE_WRITTEN;
546 }
547
548 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
549                        void *cb_data)
550 {
551         unsigned char peeled[20];
552         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
553
554         if (entry)
555                 entry->tagged = 1;
556         if (!peel_ref(path, peeled)) {
557                 entry = packlist_find(&to_pack, peeled, NULL);
558                 if (entry)
559                         entry->tagged = 1;
560         }
561         return 0;
562 }
563
564 static inline void add_to_write_order(struct object_entry **wo,
565                                unsigned int *endp,
566                                struct object_entry *e)
567 {
568         if (e->filled)
569                 return;
570         wo[(*endp)++] = e;
571         e->filled = 1;
572 }
573
574 static void add_descendants_to_write_order(struct object_entry **wo,
575                                            unsigned int *endp,
576                                            struct object_entry *e)
577 {
578         int add_to_order = 1;
579         while (e) {
580                 if (add_to_order) {
581                         struct object_entry *s;
582                         /* add this node... */
583                         add_to_write_order(wo, endp, e);
584                         /* all its siblings... */
585                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
586                                 add_to_write_order(wo, endp, s);
587                         }
588                 }
589                 /* drop down a level to add left subtree nodes if possible */
590                 if (e->delta_child) {
591                         add_to_order = 1;
592                         e = e->delta_child;
593                 } else {
594                         add_to_order = 0;
595                         /* our sibling might have some children, it is next */
596                         if (e->delta_sibling) {
597                                 e = e->delta_sibling;
598                                 continue;
599                         }
600                         /* go back to our parent node */
601                         e = e->delta;
602                         while (e && !e->delta_sibling) {
603                                 /* we're on the right side of a subtree, keep
604                                  * going up until we can go right again */
605                                 e = e->delta;
606                         }
607                         if (!e) {
608                                 /* done- we hit our original root node */
609                                 return;
610                         }
611                         /* pass it off to sibling at this level */
612                         e = e->delta_sibling;
613                 }
614         };
615 }
616
617 static void add_family_to_write_order(struct object_entry **wo,
618                                       unsigned int *endp,
619                                       struct object_entry *e)
620 {
621         struct object_entry *root;
622
623         for (root = e; root->delta; root = root->delta)
624                 ; /* nothing */
625         add_descendants_to_write_order(wo, endp, root);
626 }
627
628 static struct object_entry **compute_write_order(void)
629 {
630         unsigned int i, wo_end, last_untagged;
631
632         struct object_entry **wo;
633         struct object_entry *objects = to_pack.objects;
634
635         for (i = 0; i < to_pack.nr_objects; i++) {
636                 objects[i].tagged = 0;
637                 objects[i].filled = 0;
638                 objects[i].delta_child = NULL;
639                 objects[i].delta_sibling = NULL;
640         }
641
642         /*
643          * Fully connect delta_child/delta_sibling network.
644          * Make sure delta_sibling is sorted in the original
645          * recency order.
646          */
647         for (i = to_pack.nr_objects; i > 0;) {
648                 struct object_entry *e = &objects[--i];
649                 if (!e->delta)
650                         continue;
651                 /* Mark me as the first child */
652                 e->delta_sibling = e->delta->delta_child;
653                 e->delta->delta_child = e;
654         }
655
656         /*
657          * Mark objects that are at the tip of tags.
658          */
659         for_each_tag_ref(mark_tagged, NULL);
660
661         /*
662          * Give the objects in the original recency order until
663          * we see a tagged tip.
664          */
665         ALLOC_ARRAY(wo, to_pack.nr_objects);
666         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
667                 if (objects[i].tagged)
668                         break;
669                 add_to_write_order(wo, &wo_end, &objects[i]);
670         }
671         last_untagged = i;
672
673         /*
674          * Then fill all the tagged tips.
675          */
676         for (; i < to_pack.nr_objects; i++) {
677                 if (objects[i].tagged)
678                         add_to_write_order(wo, &wo_end, &objects[i]);
679         }
680
681         /*
682          * And then all remaining commits and tags.
683          */
684         for (i = last_untagged; i < to_pack.nr_objects; i++) {
685                 if (objects[i].type != OBJ_COMMIT &&
686                     objects[i].type != OBJ_TAG)
687                         continue;
688                 add_to_write_order(wo, &wo_end, &objects[i]);
689         }
690
691         /*
692          * And then all the trees.
693          */
694         for (i = last_untagged; i < to_pack.nr_objects; i++) {
695                 if (objects[i].type != OBJ_TREE)
696                         continue;
697                 add_to_write_order(wo, &wo_end, &objects[i]);
698         }
699
700         /*
701          * Finally all the rest in really tight order
702          */
703         for (i = last_untagged; i < to_pack.nr_objects; i++) {
704                 if (!objects[i].filled)
705                         add_family_to_write_order(wo, &wo_end, &objects[i]);
706         }
707
708         if (wo_end != to_pack.nr_objects)
709                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
710
711         return wo;
712 }
713
714 static off_t write_reused_pack(struct sha1file *f)
715 {
716         unsigned char buffer[8192];
717         off_t to_write, total;
718         int fd;
719
720         if (!is_pack_valid(reuse_packfile))
721                 die("packfile is invalid: %s", reuse_packfile->pack_name);
722
723         fd = git_open(reuse_packfile->pack_name);
724         if (fd < 0)
725                 die_errno("unable to open packfile for reuse: %s",
726                           reuse_packfile->pack_name);
727
728         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
729                 die_errno("unable to seek in reused packfile");
730
731         if (reuse_packfile_offset < 0)
732                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
733
734         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
735
736         while (to_write) {
737                 int read_pack = xread(fd, buffer, sizeof(buffer));
738
739                 if (read_pack <= 0)
740                         die_errno("unable to read from reused packfile");
741
742                 if (read_pack > to_write)
743                         read_pack = to_write;
744
745                 sha1write(f, buffer, read_pack);
746                 to_write -= read_pack;
747
748                 /*
749                  * We don't know the actual number of objects written,
750                  * only how many bytes written, how many bytes total, and
751                  * how many objects total. So we can fake it by pretending all
752                  * objects we are writing are the same size. This gives us a
753                  * smooth progress meter, and at the end it matches the true
754                  * answer.
755                  */
756                 written = reuse_packfile_objects *
757                                 (((double)(total - to_write)) / total);
758                 display_progress(progress_state, written);
759         }
760
761         close(fd);
762         written = reuse_packfile_objects;
763         display_progress(progress_state, written);
764         return reuse_packfile_offset - sizeof(struct pack_header);
765 }
766
767 static const char no_split_warning[] = N_(
768 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
769 );
770
771 static void write_pack_file(void)
772 {
773         uint32_t i = 0, j;
774         struct sha1file *f;
775         off_t offset;
776         uint32_t nr_remaining = nr_result;
777         time_t last_mtime = 0;
778         struct object_entry **write_order;
779
780         if (progress > pack_to_stdout)
781                 progress_state = start_progress(_("Writing objects"), nr_result);
782         ALLOC_ARRAY(written_list, to_pack.nr_objects);
783         write_order = compute_write_order();
784
785         do {
786                 unsigned char sha1[20];
787                 char *pack_tmp_name = NULL;
788
789                 if (pack_to_stdout)
790                         f = sha1fd_throughput(1, "<stdout>", progress_state);
791                 else
792                         f = create_tmp_packfile(&pack_tmp_name);
793
794                 offset = write_pack_header(f, nr_remaining);
795
796                 if (reuse_packfile) {
797                         off_t packfile_size;
798                         assert(pack_to_stdout);
799
800                         packfile_size = write_reused_pack(f);
801                         offset += packfile_size;
802                 }
803
804                 nr_written = 0;
805                 for (; i < to_pack.nr_objects; i++) {
806                         struct object_entry *e = write_order[i];
807                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
808                                 break;
809                         display_progress(progress_state, written);
810                 }
811
812                 /*
813                  * Did we write the wrong # entries in the header?
814                  * If so, rewrite it like in fast-import
815                  */
816                 if (pack_to_stdout) {
817                         sha1close(f, sha1, CSUM_CLOSE);
818                 } else if (nr_written == nr_remaining) {
819                         sha1close(f, sha1, CSUM_FSYNC);
820                 } else {
821                         int fd = sha1close(f, sha1, 0);
822                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
823                                                  nr_written, sha1, offset);
824                         close(fd);
825                         if (write_bitmap_index) {
826                                 warning(_(no_split_warning));
827                                 write_bitmap_index = 0;
828                         }
829                 }
830
831                 if (!pack_to_stdout) {
832                         struct stat st;
833                         struct strbuf tmpname = STRBUF_INIT;
834
835                         /*
836                          * Packs are runtime accessed in their mtime
837                          * order since newer packs are more likely to contain
838                          * younger objects.  So if we are creating multiple
839                          * packs then we should modify the mtime of later ones
840                          * to preserve this property.
841                          */
842                         if (stat(pack_tmp_name, &st) < 0) {
843                                 warning_errno("failed to stat %s", pack_tmp_name);
844                         } else if (!last_mtime) {
845                                 last_mtime = st.st_mtime;
846                         } else {
847                                 struct utimbuf utb;
848                                 utb.actime = st.st_atime;
849                                 utb.modtime = --last_mtime;
850                                 if (utime(pack_tmp_name, &utb) < 0)
851                                         warning_errno("failed utime() on %s", pack_tmp_name);
852                         }
853
854                         strbuf_addf(&tmpname, "%s-", base_name);
855
856                         if (write_bitmap_index) {
857                                 bitmap_writer_set_checksum(sha1);
858                                 bitmap_writer_build_type_index(written_list, nr_written);
859                         }
860
861                         finish_tmp_packfile(&tmpname, pack_tmp_name,
862                                             written_list, nr_written,
863                                             &pack_idx_opts, sha1);
864
865                         if (write_bitmap_index) {
866                                 strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
867
868                                 stop_progress(&progress_state);
869
870                                 bitmap_writer_show_progress(progress);
871                                 bitmap_writer_reuse_bitmaps(&to_pack);
872                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
873                                 bitmap_writer_build(&to_pack);
874                                 bitmap_writer_finish(written_list, nr_written,
875                                                      tmpname.buf, write_bitmap_options);
876                                 write_bitmap_index = 0;
877                         }
878
879                         strbuf_release(&tmpname);
880                         free(pack_tmp_name);
881                         puts(sha1_to_hex(sha1));
882                 }
883
884                 /* mark written objects as written to previous pack */
885                 for (j = 0; j < nr_written; j++) {
886                         written_list[j]->offset = (off_t)-1;
887                 }
888                 nr_remaining -= nr_written;
889         } while (nr_remaining && i < to_pack.nr_objects);
890
891         free(written_list);
892         free(write_order);
893         stop_progress(&progress_state);
894         if (written != nr_result)
895                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
896                         written, nr_result);
897 }
898
899 static void setup_delta_attr_check(struct git_attr_check *check)
900 {
901         static struct git_attr *attr_delta;
902
903         if (!attr_delta)
904                 attr_delta = git_attr("delta");
905
906         check[0].attr = attr_delta;
907 }
908
909 static int no_try_delta(const char *path)
910 {
911         struct git_attr_check check[1];
912
913         setup_delta_attr_check(check);
914         if (git_check_attr(path, ARRAY_SIZE(check), check))
915                 return 0;
916         if (ATTR_FALSE(check->value))
917                 return 1;
918         return 0;
919 }
920
921 /*
922  * When adding an object, check whether we have already added it
923  * to our packing list. If so, we can skip. However, if we are
924  * being asked to excludei t, but the previous mention was to include
925  * it, make sure to adjust its flags and tweak our numbers accordingly.
926  *
927  * As an optimization, we pass out the index position where we would have
928  * found the item, since that saves us from having to look it up again a
929  * few lines later when we want to add the new entry.
930  */
931 static int have_duplicate_entry(const unsigned char *sha1,
932                                 int exclude,
933                                 uint32_t *index_pos)
934 {
935         struct object_entry *entry;
936
937         entry = packlist_find(&to_pack, sha1, index_pos);
938         if (!entry)
939                 return 0;
940
941         if (exclude) {
942                 if (!entry->preferred_base)
943                         nr_result--;
944                 entry->preferred_base = 1;
945         }
946
947         return 1;
948 }
949
950 static int want_found_object(int exclude, struct packed_git *p)
951 {
952         if (exclude)
953                 return 1;
954         if (incremental)
955                 return 0;
956
957         /*
958          * When asked to do --local (do not include an object that appears in a
959          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
960          * an object that appears in a pack marked with .keep), finding a pack
961          * that matches the criteria is sufficient for us to decide to omit it.
962          * However, even if this pack does not satisfy the criteria, we need to
963          * make sure no copy of this object appears in _any_ pack that makes us
964          * to omit the object, so we need to check all the packs.
965          *
966          * We can however first check whether these options can possible matter;
967          * if they do not matter we know we want the object in generated pack.
968          * Otherwise, we signal "-1" at the end to tell the caller that we do
969          * not know either way, and it needs to check more packs.
970          */
971         if (!ignore_packed_keep &&
972             (!local || !have_non_local_packs))
973                 return 1;
974
975         if (local && !p->pack_local)
976                 return 0;
977         if (ignore_packed_keep && p->pack_local && p->pack_keep)
978                 return 0;
979
980         /* we don't know yet; keep looking for more packs */
981         return -1;
982 }
983
984 /*
985  * Check whether we want the object in the pack (e.g., we do not want
986  * objects found in non-local stores if the "--local" option was used).
987  *
988  * If the caller already knows an existing pack it wants to take the object
989  * from, that is passed in *found_pack and *found_offset; otherwise this
990  * function finds if there is any pack that has the object and returns the pack
991  * and its offset in these variables.
992  */
993 static int want_object_in_pack(const unsigned char *sha1,
994                                int exclude,
995                                struct packed_git **found_pack,
996                                off_t *found_offset)
997 {
998         struct mru_entry *entry;
999         int want;
1000
1001         if (!exclude && local && has_loose_object_nonlocal(sha1))
1002                 return 0;
1003
1004         /*
1005          * If we already know the pack object lives in, start checks from that
1006          * pack - in the usual case when neither --local was given nor .keep files
1007          * are present we will determine the answer right now.
1008          */
1009         if (*found_pack) {
1010                 want = want_found_object(exclude, *found_pack);
1011                 if (want != -1)
1012                         return want;
1013         }
1014
1015         for (entry = packed_git_mru->head; entry; entry = entry->next) {
1016                 struct packed_git *p = entry->item;
1017                 off_t offset;
1018
1019                 if (p == *found_pack)
1020                         offset = *found_offset;
1021                 else
1022                         offset = find_pack_entry_one(sha1, p);
1023
1024                 if (offset) {
1025                         if (!*found_pack) {
1026                                 if (!is_pack_valid(p))
1027                                         continue;
1028                                 *found_offset = offset;
1029                                 *found_pack = p;
1030                         }
1031                         want = want_found_object(exclude, p);
1032                         if (!exclude && want > 0)
1033                                 mru_mark(packed_git_mru, entry);
1034                         if (want != -1)
1035                                 return want;
1036                 }
1037         }
1038
1039         return 1;
1040 }
1041
1042 static void create_object_entry(const unsigned char *sha1,
1043                                 enum object_type type,
1044                                 uint32_t hash,
1045                                 int exclude,
1046                                 int no_try_delta,
1047                                 uint32_t index_pos,
1048                                 struct packed_git *found_pack,
1049                                 off_t found_offset)
1050 {
1051         struct object_entry *entry;
1052
1053         entry = packlist_alloc(&to_pack, sha1, index_pos);
1054         entry->hash = hash;
1055         if (type)
1056                 entry->type = type;
1057         if (exclude)
1058                 entry->preferred_base = 1;
1059         else
1060                 nr_result++;
1061         if (found_pack) {
1062                 entry->in_pack = found_pack;
1063                 entry->in_pack_offset = found_offset;
1064         }
1065
1066         entry->no_try_delta = no_try_delta;
1067 }
1068
1069 static const char no_closure_warning[] = N_(
1070 "disabling bitmap writing, as some objects are not being packed"
1071 );
1072
1073 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1074                             const char *name, int exclude)
1075 {
1076         struct packed_git *found_pack = NULL;
1077         off_t found_offset = 0;
1078         uint32_t index_pos;
1079
1080         if (have_duplicate_entry(sha1, exclude, &index_pos))
1081                 return 0;
1082
1083         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1084                 /* The pack is missing an object, so it will not have closure */
1085                 if (write_bitmap_index) {
1086                         warning(_(no_closure_warning));
1087                         write_bitmap_index = 0;
1088                 }
1089                 return 0;
1090         }
1091
1092         create_object_entry(sha1, type, pack_name_hash(name),
1093                             exclude, name && no_try_delta(name),
1094                             index_pos, found_pack, found_offset);
1095
1096         display_progress(progress_state, nr_result);
1097         return 1;
1098 }
1099
1100 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1101                                         enum object_type type,
1102                                         int flags, uint32_t name_hash,
1103                                         struct packed_git *pack, off_t offset)
1104 {
1105         uint32_t index_pos;
1106
1107         if (have_duplicate_entry(sha1, 0, &index_pos))
1108                 return 0;
1109
1110         if (!want_object_in_pack(sha1, 0, &pack, &offset))
1111                 return 0;
1112
1113         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1114
1115         display_progress(progress_state, nr_result);
1116         return 1;
1117 }
1118
1119 struct pbase_tree_cache {
1120         unsigned char sha1[20];
1121         int ref;
1122         int temporary;
1123         void *tree_data;
1124         unsigned long tree_size;
1125 };
1126
1127 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1128 static int pbase_tree_cache_ix(const unsigned char *sha1)
1129 {
1130         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1131 }
1132 static int pbase_tree_cache_ix_incr(int ix)
1133 {
1134         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1135 }
1136
1137 static struct pbase_tree {
1138         struct pbase_tree *next;
1139         /* This is a phony "cache" entry; we are not
1140          * going to evict it or find it through _get()
1141          * mechanism -- this is for the toplevel node that
1142          * would almost always change with any commit.
1143          */
1144         struct pbase_tree_cache pcache;
1145 } *pbase_tree;
1146
1147 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1148 {
1149         struct pbase_tree_cache *ent, *nent;
1150         void *data;
1151         unsigned long size;
1152         enum object_type type;
1153         int neigh;
1154         int my_ix = pbase_tree_cache_ix(sha1);
1155         int available_ix = -1;
1156
1157         /* pbase-tree-cache acts as a limited hashtable.
1158          * your object will be found at your index or within a few
1159          * slots after that slot if it is cached.
1160          */
1161         for (neigh = 0; neigh < 8; neigh++) {
1162                 ent = pbase_tree_cache[my_ix];
1163                 if (ent && !hashcmp(ent->sha1, sha1)) {
1164                         ent->ref++;
1165                         return ent;
1166                 }
1167                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1168                          ((0 <= available_ix) &&
1169                           (!ent && pbase_tree_cache[available_ix])))
1170                         available_ix = my_ix;
1171                 if (!ent)
1172                         break;
1173                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1174         }
1175
1176         /* Did not find one.  Either we got a bogus request or
1177          * we need to read and perhaps cache.
1178          */
1179         data = read_sha1_file(sha1, &type, &size);
1180         if (!data)
1181                 return NULL;
1182         if (type != OBJ_TREE) {
1183                 free(data);
1184                 return NULL;
1185         }
1186
1187         /* We need to either cache or return a throwaway copy */
1188
1189         if (available_ix < 0)
1190                 ent = NULL;
1191         else {
1192                 ent = pbase_tree_cache[available_ix];
1193                 my_ix = available_ix;
1194         }
1195
1196         if (!ent) {
1197                 nent = xmalloc(sizeof(*nent));
1198                 nent->temporary = (available_ix < 0);
1199         }
1200         else {
1201                 /* evict and reuse */
1202                 free(ent->tree_data);
1203                 nent = ent;
1204         }
1205         hashcpy(nent->sha1, sha1);
1206         nent->tree_data = data;
1207         nent->tree_size = size;
1208         nent->ref = 1;
1209         if (!nent->temporary)
1210                 pbase_tree_cache[my_ix] = nent;
1211         return nent;
1212 }
1213
1214 static void pbase_tree_put(struct pbase_tree_cache *cache)
1215 {
1216         if (!cache->temporary) {
1217                 cache->ref--;
1218                 return;
1219         }
1220         free(cache->tree_data);
1221         free(cache);
1222 }
1223
1224 static int name_cmp_len(const char *name)
1225 {
1226         int i;
1227         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1228                 ;
1229         return i;
1230 }
1231
1232 static void add_pbase_object(struct tree_desc *tree,
1233                              const char *name,
1234                              int cmplen,
1235                              const char *fullname)
1236 {
1237         struct name_entry entry;
1238         int cmp;
1239
1240         while (tree_entry(tree,&entry)) {
1241                 if (S_ISGITLINK(entry.mode))
1242                         continue;
1243                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1244                       memcmp(name, entry.path, cmplen);
1245                 if (cmp > 0)
1246                         continue;
1247                 if (cmp < 0)
1248                         return;
1249                 if (name[cmplen] != '/') {
1250                         add_object_entry(entry.oid->hash,
1251                                          object_type(entry.mode),
1252                                          fullname, 1);
1253                         return;
1254                 }
1255                 if (S_ISDIR(entry.mode)) {
1256                         struct tree_desc sub;
1257                         struct pbase_tree_cache *tree;
1258                         const char *down = name+cmplen+1;
1259                         int downlen = name_cmp_len(down);
1260
1261                         tree = pbase_tree_get(entry.oid->hash);
1262                         if (!tree)
1263                                 return;
1264                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1265
1266                         add_pbase_object(&sub, down, downlen, fullname);
1267                         pbase_tree_put(tree);
1268                 }
1269         }
1270 }
1271
1272 static unsigned *done_pbase_paths;
1273 static int done_pbase_paths_num;
1274 static int done_pbase_paths_alloc;
1275 static int done_pbase_path_pos(unsigned hash)
1276 {
1277         int lo = 0;
1278         int hi = done_pbase_paths_num;
1279         while (lo < hi) {
1280                 int mi = (hi + lo) / 2;
1281                 if (done_pbase_paths[mi] == hash)
1282                         return mi;
1283                 if (done_pbase_paths[mi] < hash)
1284                         hi = mi;
1285                 else
1286                         lo = mi + 1;
1287         }
1288         return -lo-1;
1289 }
1290
1291 static int check_pbase_path(unsigned hash)
1292 {
1293         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1294         if (0 <= pos)
1295                 return 1;
1296         pos = -pos - 1;
1297         ALLOC_GROW(done_pbase_paths,
1298                    done_pbase_paths_num + 1,
1299                    done_pbase_paths_alloc);
1300         done_pbase_paths_num++;
1301         if (pos < done_pbase_paths_num)
1302                 memmove(done_pbase_paths + pos + 1,
1303                         done_pbase_paths + pos,
1304                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1305         done_pbase_paths[pos] = hash;
1306         return 0;
1307 }
1308
1309 static void add_preferred_base_object(const char *name)
1310 {
1311         struct pbase_tree *it;
1312         int cmplen;
1313         unsigned hash = pack_name_hash(name);
1314
1315         if (!num_preferred_base || check_pbase_path(hash))
1316                 return;
1317
1318         cmplen = name_cmp_len(name);
1319         for (it = pbase_tree; it; it = it->next) {
1320                 if (cmplen == 0) {
1321                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1322                 }
1323                 else {
1324                         struct tree_desc tree;
1325                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1326                         add_pbase_object(&tree, name, cmplen, name);
1327                 }
1328         }
1329 }
1330
1331 static void add_preferred_base(unsigned char *sha1)
1332 {
1333         struct pbase_tree *it;
1334         void *data;
1335         unsigned long size;
1336         unsigned char tree_sha1[20];
1337
1338         if (window <= num_preferred_base++)
1339                 return;
1340
1341         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1342         if (!data)
1343                 return;
1344
1345         for (it = pbase_tree; it; it = it->next) {
1346                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1347                         free(data);
1348                         return;
1349                 }
1350         }
1351
1352         it = xcalloc(1, sizeof(*it));
1353         it->next = pbase_tree;
1354         pbase_tree = it;
1355
1356         hashcpy(it->pcache.sha1, tree_sha1);
1357         it->pcache.tree_data = data;
1358         it->pcache.tree_size = size;
1359 }
1360
1361 static void cleanup_preferred_base(void)
1362 {
1363         struct pbase_tree *it;
1364         unsigned i;
1365
1366         it = pbase_tree;
1367         pbase_tree = NULL;
1368         while (it) {
1369                 struct pbase_tree *this = it;
1370                 it = this->next;
1371                 free(this->pcache.tree_data);
1372                 free(this);
1373         }
1374
1375         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1376                 if (!pbase_tree_cache[i])
1377                         continue;
1378                 free(pbase_tree_cache[i]->tree_data);
1379                 free(pbase_tree_cache[i]);
1380                 pbase_tree_cache[i] = NULL;
1381         }
1382
1383         free(done_pbase_paths);
1384         done_pbase_paths = NULL;
1385         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1386 }
1387
1388 static void check_object(struct object_entry *entry)
1389 {
1390         if (entry->in_pack) {
1391                 struct packed_git *p = entry->in_pack;
1392                 struct pack_window *w_curs = NULL;
1393                 const unsigned char *base_ref = NULL;
1394                 struct object_entry *base_entry;
1395                 unsigned long used, used_0;
1396                 unsigned long avail;
1397                 off_t ofs;
1398                 unsigned char *buf, c;
1399
1400                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1401
1402                 /*
1403                  * We want in_pack_type even if we do not reuse delta
1404                  * since non-delta representations could still be reused.
1405                  */
1406                 used = unpack_object_header_buffer(buf, avail,
1407                                                    &entry->in_pack_type,
1408                                                    &entry->size);
1409                 if (used == 0)
1410                         goto give_up;
1411
1412                 /*
1413                  * Determine if this is a delta and if so whether we can
1414                  * reuse it or not.  Otherwise let's find out as cheaply as
1415                  * possible what the actual type and size for this object is.
1416                  */
1417                 switch (entry->in_pack_type) {
1418                 default:
1419                         /* Not a delta hence we've already got all we need. */
1420                         entry->type = entry->in_pack_type;
1421                         entry->in_pack_header_size = used;
1422                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1423                                 goto give_up;
1424                         unuse_pack(&w_curs);
1425                         return;
1426                 case OBJ_REF_DELTA:
1427                         if (reuse_delta && !entry->preferred_base)
1428                                 base_ref = use_pack(p, &w_curs,
1429                                                 entry->in_pack_offset + used, NULL);
1430                         entry->in_pack_header_size = used + 20;
1431                         break;
1432                 case OBJ_OFS_DELTA:
1433                         buf = use_pack(p, &w_curs,
1434                                        entry->in_pack_offset + used, NULL);
1435                         used_0 = 0;
1436                         c = buf[used_0++];
1437                         ofs = c & 127;
1438                         while (c & 128) {
1439                                 ofs += 1;
1440                                 if (!ofs || MSB(ofs, 7)) {
1441                                         error("delta base offset overflow in pack for %s",
1442                                               sha1_to_hex(entry->idx.sha1));
1443                                         goto give_up;
1444                                 }
1445                                 c = buf[used_0++];
1446                                 ofs = (ofs << 7) + (c & 127);
1447                         }
1448                         ofs = entry->in_pack_offset - ofs;
1449                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1450                                 error("delta base offset out of bound for %s",
1451                                       sha1_to_hex(entry->idx.sha1));
1452                                 goto give_up;
1453                         }
1454                         if (reuse_delta && !entry->preferred_base) {
1455                                 struct revindex_entry *revidx;
1456                                 revidx = find_pack_revindex(p, ofs);
1457                                 if (!revidx)
1458                                         goto give_up;
1459                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1460                         }
1461                         entry->in_pack_header_size = used + used_0;
1462                         break;
1463                 }
1464
1465                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1466                         /*
1467                          * If base_ref was set above that means we wish to
1468                          * reuse delta data, and we even found that base
1469                          * in the list of objects we want to pack. Goodie!
1470                          *
1471                          * Depth value does not matter - find_deltas() will
1472                          * never consider reused delta as the base object to
1473                          * deltify other objects against, in order to avoid
1474                          * circular deltas.
1475                          */
1476                         entry->type = entry->in_pack_type;
1477                         entry->delta = base_entry;
1478                         entry->delta_size = entry->size;
1479                         entry->delta_sibling = base_entry->delta_child;
1480                         base_entry->delta_child = entry;
1481                         unuse_pack(&w_curs);
1482                         return;
1483                 }
1484
1485                 if (entry->type) {
1486                         /*
1487                          * This must be a delta and we already know what the
1488                          * final object type is.  Let's extract the actual
1489                          * object size from the delta header.
1490                          */
1491                         entry->size = get_size_from_delta(p, &w_curs,
1492                                         entry->in_pack_offset + entry->in_pack_header_size);
1493                         if (entry->size == 0)
1494                                 goto give_up;
1495                         unuse_pack(&w_curs);
1496                         return;
1497                 }
1498
1499                 /*
1500                  * No choice but to fall back to the recursive delta walk
1501                  * with sha1_object_info() to find about the object type
1502                  * at this point...
1503                  */
1504                 give_up:
1505                 unuse_pack(&w_curs);
1506         }
1507
1508         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1509         /*
1510          * The error condition is checked in prepare_pack().  This is
1511          * to permit a missing preferred base object to be ignored
1512          * as a preferred base.  Doing so can result in a larger
1513          * pack file, but the transfer will still take place.
1514          */
1515 }
1516
1517 static int pack_offset_sort(const void *_a, const void *_b)
1518 {
1519         const struct object_entry *a = *(struct object_entry **)_a;
1520         const struct object_entry *b = *(struct object_entry **)_b;
1521
1522         /* avoid filesystem trashing with loose objects */
1523         if (!a->in_pack && !b->in_pack)
1524                 return hashcmp(a->idx.sha1, b->idx.sha1);
1525
1526         if (a->in_pack < b->in_pack)
1527                 return -1;
1528         if (a->in_pack > b->in_pack)
1529                 return 1;
1530         return a->in_pack_offset < b->in_pack_offset ? -1 :
1531                         (a->in_pack_offset > b->in_pack_offset);
1532 }
1533
1534 /*
1535  * Drop an on-disk delta we were planning to reuse. Naively, this would
1536  * just involve blanking out the "delta" field, but we have to deal
1537  * with some extra book-keeping:
1538  *
1539  *   1. Removing ourselves from the delta_sibling linked list.
1540  *
1541  *   2. Updating our size/type to the non-delta representation. These were
1542  *      either not recorded initially (size) or overwritten with the delta type
1543  *      (type) when check_object() decided to reuse the delta.
1544  */
1545 static void drop_reused_delta(struct object_entry *entry)
1546 {
1547         struct object_entry **p = &entry->delta->delta_child;
1548         struct object_info oi = OBJECT_INFO_INIT;
1549
1550         while (*p) {
1551                 if (*p == entry)
1552                         *p = (*p)->delta_sibling;
1553                 else
1554                         p = &(*p)->delta_sibling;
1555         }
1556         entry->delta = NULL;
1557
1558         oi.sizep = &entry->size;
1559         oi.typep = &entry->type;
1560         if (packed_object_info(entry->in_pack, entry->in_pack_offset, &oi) < 0) {
1561                 /*
1562                  * We failed to get the info from this pack for some reason;
1563                  * fall back to sha1_object_info, which may find another copy.
1564                  * And if that fails, the error will be recorded in entry->type
1565                  * and dealt with in prepare_pack().
1566                  */
1567                 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1568         }
1569 }
1570
1571 /*
1572  * Follow the chain of deltas from this entry onward, throwing away any links
1573  * that cause us to hit a cycle (as determined by the DFS state flags in
1574  * the entries).
1575  */
1576 static void break_delta_chains(struct object_entry *entry)
1577 {
1578         /* If it's not a delta, it can't be part of a cycle. */
1579         if (!entry->delta) {
1580                 entry->dfs_state = DFS_DONE;
1581                 return;
1582         }
1583
1584         switch (entry->dfs_state) {
1585         case DFS_NONE:
1586                 /*
1587                  * This is the first time we've seen the object. We mark it as
1588                  * part of the active potential cycle and recurse.
1589                  */
1590                 entry->dfs_state = DFS_ACTIVE;
1591                 break_delta_chains(entry->delta);
1592                 entry->dfs_state = DFS_DONE;
1593                 break;
1594
1595         case DFS_DONE:
1596                 /* object already examined, and not part of a cycle */
1597                 break;
1598
1599         case DFS_ACTIVE:
1600                 /*
1601                  * We found a cycle that needs broken. It would be correct to
1602                  * break any link in the chain, but it's convenient to
1603                  * break this one.
1604                  */
1605                 drop_reused_delta(entry);
1606                 entry->dfs_state = DFS_DONE;
1607                 break;
1608         }
1609 }
1610
1611 static void get_object_details(void)
1612 {
1613         uint32_t i;
1614         struct object_entry **sorted_by_offset;
1615
1616         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1617         for (i = 0; i < to_pack.nr_objects; i++)
1618                 sorted_by_offset[i] = to_pack.objects + i;
1619         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1620
1621         for (i = 0; i < to_pack.nr_objects; i++) {
1622                 struct object_entry *entry = sorted_by_offset[i];
1623                 check_object(entry);
1624                 if (big_file_threshold < entry->size)
1625                         entry->no_try_delta = 1;
1626         }
1627
1628         /*
1629          * This must happen in a second pass, since we rely on the delta
1630          * information for the whole list being completed.
1631          */
1632         for (i = 0; i < to_pack.nr_objects; i++)
1633                 break_delta_chains(&to_pack.objects[i]);
1634
1635         free(sorted_by_offset);
1636 }
1637
1638 /*
1639  * We search for deltas in a list sorted by type, by filename hash, and then
1640  * by size, so that we see progressively smaller and smaller files.
1641  * That's because we prefer deltas to be from the bigger file
1642  * to the smaller -- deletes are potentially cheaper, but perhaps
1643  * more importantly, the bigger file is likely the more recent
1644  * one.  The deepest deltas are therefore the oldest objects which are
1645  * less susceptible to be accessed often.
1646  */
1647 static int type_size_sort(const void *_a, const void *_b)
1648 {
1649         const struct object_entry *a = *(struct object_entry **)_a;
1650         const struct object_entry *b = *(struct object_entry **)_b;
1651
1652         if (a->type > b->type)
1653                 return -1;
1654         if (a->type < b->type)
1655                 return 1;
1656         if (a->hash > b->hash)
1657                 return -1;
1658         if (a->hash < b->hash)
1659                 return 1;
1660         if (a->preferred_base > b->preferred_base)
1661                 return -1;
1662         if (a->preferred_base < b->preferred_base)
1663                 return 1;
1664         if (a->size > b->size)
1665                 return -1;
1666         if (a->size < b->size)
1667                 return 1;
1668         return a < b ? -1 : (a > b);  /* newest first */
1669 }
1670
1671 struct unpacked {
1672         struct object_entry *entry;
1673         void *data;
1674         struct delta_index *index;
1675         unsigned depth;
1676 };
1677
1678 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1679                            unsigned long delta_size)
1680 {
1681         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1682                 return 0;
1683
1684         if (delta_size < cache_max_small_delta_size)
1685                 return 1;
1686
1687         /* cache delta, if objects are large enough compared to delta size */
1688         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1689                 return 1;
1690
1691         return 0;
1692 }
1693
1694 #ifndef NO_PTHREADS
1695
1696 static pthread_mutex_t read_mutex;
1697 #define read_lock()             pthread_mutex_lock(&read_mutex)
1698 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1699
1700 static pthread_mutex_t cache_mutex;
1701 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1702 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1703
1704 static pthread_mutex_t progress_mutex;
1705 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1706 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1707
1708 #else
1709
1710 #define read_lock()             (void)0
1711 #define read_unlock()           (void)0
1712 #define cache_lock()            (void)0
1713 #define cache_unlock()          (void)0
1714 #define progress_lock()         (void)0
1715 #define progress_unlock()       (void)0
1716
1717 #endif
1718
1719 static int try_delta(struct unpacked *trg, struct unpacked *src,
1720                      unsigned max_depth, unsigned long *mem_usage)
1721 {
1722         struct object_entry *trg_entry = trg->entry;
1723         struct object_entry *src_entry = src->entry;
1724         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1725         unsigned ref_depth;
1726         enum object_type type;
1727         void *delta_buf;
1728
1729         /* Don't bother doing diffs between different types */
1730         if (trg_entry->type != src_entry->type)
1731                 return -1;
1732
1733         /*
1734          * We do not bother to try a delta that we discarded on an
1735          * earlier try, but only when reusing delta data.  Note that
1736          * src_entry that is marked as the preferred_base should always
1737          * be considered, as even if we produce a suboptimal delta against
1738          * it, we will still save the transfer cost, as we already know
1739          * the other side has it and we won't send src_entry at all.
1740          */
1741         if (reuse_delta && trg_entry->in_pack &&
1742             trg_entry->in_pack == src_entry->in_pack &&
1743             !src_entry->preferred_base &&
1744             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1745             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1746                 return 0;
1747
1748         /* Let's not bust the allowed depth. */
1749         if (src->depth >= max_depth)
1750                 return 0;
1751
1752         /* Now some size filtering heuristics. */
1753         trg_size = trg_entry->size;
1754         if (!trg_entry->delta) {
1755                 max_size = trg_size/2 - 20;
1756                 ref_depth = 1;
1757         } else {
1758                 max_size = trg_entry->delta_size;
1759                 ref_depth = trg->depth;
1760         }
1761         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1762                                                 (max_depth - ref_depth + 1);
1763         if (max_size == 0)
1764                 return 0;
1765         src_size = src_entry->size;
1766         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1767         if (sizediff >= max_size)
1768                 return 0;
1769         if (trg_size < src_size / 32)
1770                 return 0;
1771
1772         /* Load data if not already done */
1773         if (!trg->data) {
1774                 read_lock();
1775                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1776                 read_unlock();
1777                 if (!trg->data)
1778                         die("object %s cannot be read",
1779                             sha1_to_hex(trg_entry->idx.sha1));
1780                 if (sz != trg_size)
1781                         die("object %s inconsistent object length (%lu vs %lu)",
1782                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1783                 *mem_usage += sz;
1784         }
1785         if (!src->data) {
1786                 read_lock();
1787                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1788                 read_unlock();
1789                 if (!src->data) {
1790                         if (src_entry->preferred_base) {
1791                                 static int warned = 0;
1792                                 if (!warned++)
1793                                         warning("object %s cannot be read",
1794                                                 sha1_to_hex(src_entry->idx.sha1));
1795                                 /*
1796                                  * Those objects are not included in the
1797                                  * resulting pack.  Be resilient and ignore
1798                                  * them if they can't be read, in case the
1799                                  * pack could be created nevertheless.
1800                                  */
1801                                 return 0;
1802                         }
1803                         die("object %s cannot be read",
1804                             sha1_to_hex(src_entry->idx.sha1));
1805                 }
1806                 if (sz != src_size)
1807                         die("object %s inconsistent object length (%lu vs %lu)",
1808                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1809                 *mem_usage += sz;
1810         }
1811         if (!src->index) {
1812                 src->index = create_delta_index(src->data, src_size);
1813                 if (!src->index) {
1814                         static int warned = 0;
1815                         if (!warned++)
1816                                 warning("suboptimal pack - out of memory");
1817                         return 0;
1818                 }
1819                 *mem_usage += sizeof_delta_index(src->index);
1820         }
1821
1822         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1823         if (!delta_buf)
1824                 return 0;
1825
1826         if (trg_entry->delta) {
1827                 /* Prefer only shallower same-sized deltas. */
1828                 if (delta_size == trg_entry->delta_size &&
1829                     src->depth + 1 >= trg->depth) {
1830                         free(delta_buf);
1831                         return 0;
1832                 }
1833         }
1834
1835         /*
1836          * Handle memory allocation outside of the cache
1837          * accounting lock.  Compiler will optimize the strangeness
1838          * away when NO_PTHREADS is defined.
1839          */
1840         free(trg_entry->delta_data);
1841         cache_lock();
1842         if (trg_entry->delta_data) {
1843                 delta_cache_size -= trg_entry->delta_size;
1844                 trg_entry->delta_data = NULL;
1845         }
1846         if (delta_cacheable(src_size, trg_size, delta_size)) {
1847                 delta_cache_size += delta_size;
1848                 cache_unlock();
1849                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1850         } else {
1851                 cache_unlock();
1852                 free(delta_buf);
1853         }
1854
1855         trg_entry->delta = src_entry;
1856         trg_entry->delta_size = delta_size;
1857         trg->depth = src->depth + 1;
1858
1859         return 1;
1860 }
1861
1862 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1863 {
1864         struct object_entry *child = me->delta_child;
1865         unsigned int m = n;
1866         while (child) {
1867                 unsigned int c = check_delta_limit(child, n + 1);
1868                 if (m < c)
1869                         m = c;
1870                 child = child->delta_sibling;
1871         }
1872         return m;
1873 }
1874
1875 static unsigned long free_unpacked(struct unpacked *n)
1876 {
1877         unsigned long freed_mem = sizeof_delta_index(n->index);
1878         free_delta_index(n->index);
1879         n->index = NULL;
1880         if (n->data) {
1881                 freed_mem += n->entry->size;
1882                 free(n->data);
1883                 n->data = NULL;
1884         }
1885         n->entry = NULL;
1886         n->depth = 0;
1887         return freed_mem;
1888 }
1889
1890 static void find_deltas(struct object_entry **list, unsigned *list_size,
1891                         int window, int depth, unsigned *processed)
1892 {
1893         uint32_t i, idx = 0, count = 0;
1894         struct unpacked *array;
1895         unsigned long mem_usage = 0;
1896
1897         array = xcalloc(window, sizeof(struct unpacked));
1898
1899         for (;;) {
1900                 struct object_entry *entry;
1901                 struct unpacked *n = array + idx;
1902                 int j, max_depth, best_base = -1;
1903
1904                 progress_lock();
1905                 if (!*list_size) {
1906                         progress_unlock();
1907                         break;
1908                 }
1909                 entry = *list++;
1910                 (*list_size)--;
1911                 if (!entry->preferred_base) {
1912                         (*processed)++;
1913                         display_progress(progress_state, *processed);
1914                 }
1915                 progress_unlock();
1916
1917                 mem_usage -= free_unpacked(n);
1918                 n->entry = entry;
1919
1920                 while (window_memory_limit &&
1921                        mem_usage > window_memory_limit &&
1922                        count > 1) {
1923                         uint32_t tail = (idx + window - count) % window;
1924                         mem_usage -= free_unpacked(array + tail);
1925                         count--;
1926                 }
1927
1928                 /* We do not compute delta to *create* objects we are not
1929                  * going to pack.
1930                  */
1931                 if (entry->preferred_base)
1932                         goto next;
1933
1934                 /*
1935                  * If the current object is at pack edge, take the depth the
1936                  * objects that depend on the current object into account
1937                  * otherwise they would become too deep.
1938                  */
1939                 max_depth = depth;
1940                 if (entry->delta_child) {
1941                         max_depth -= check_delta_limit(entry, 0);
1942                         if (max_depth <= 0)
1943                                 goto next;
1944                 }
1945
1946                 j = window;
1947                 while (--j > 0) {
1948                         int ret;
1949                         uint32_t other_idx = idx + j;
1950                         struct unpacked *m;
1951                         if (other_idx >= window)
1952                                 other_idx -= window;
1953                         m = array + other_idx;
1954                         if (!m->entry)
1955                                 break;
1956                         ret = try_delta(n, m, max_depth, &mem_usage);
1957                         if (ret < 0)
1958                                 break;
1959                         else if (ret > 0)
1960                                 best_base = other_idx;
1961                 }
1962
1963                 /*
1964                  * If we decided to cache the delta data, then it is best
1965                  * to compress it right away.  First because we have to do
1966                  * it anyway, and doing it here while we're threaded will
1967                  * save a lot of time in the non threaded write phase,
1968                  * as well as allow for caching more deltas within
1969                  * the same cache size limit.
1970                  * ...
1971                  * But only if not writing to stdout, since in that case
1972                  * the network is most likely throttling writes anyway,
1973                  * and therefore it is best to go to the write phase ASAP
1974                  * instead, as we can afford spending more time compressing
1975                  * between writes at that moment.
1976                  */
1977                 if (entry->delta_data && !pack_to_stdout) {
1978                         entry->z_delta_size = do_compress(&entry->delta_data,
1979                                                           entry->delta_size);
1980                         cache_lock();
1981                         delta_cache_size -= entry->delta_size;
1982                         delta_cache_size += entry->z_delta_size;
1983                         cache_unlock();
1984                 }
1985
1986                 /* if we made n a delta, and if n is already at max
1987                  * depth, leaving it in the window is pointless.  we
1988                  * should evict it first.
1989                  */
1990                 if (entry->delta && max_depth <= n->depth)
1991                         continue;
1992
1993                 /*
1994                  * Move the best delta base up in the window, after the
1995                  * currently deltified object, to keep it longer.  It will
1996                  * be the first base object to be attempted next.
1997                  */
1998                 if (entry->delta) {
1999                         struct unpacked swap = array[best_base];
2000                         int dist = (window + idx - best_base) % window;
2001                         int dst = best_base;
2002                         while (dist--) {
2003                                 int src = (dst + 1) % window;
2004                                 array[dst] = array[src];
2005                                 dst = src;
2006                         }
2007                         array[dst] = swap;
2008                 }
2009
2010                 next:
2011                 idx++;
2012                 if (count + 1 < window)
2013                         count++;
2014                 if (idx >= window)
2015                         idx = 0;
2016         }
2017
2018         for (i = 0; i < window; ++i) {
2019                 free_delta_index(array[i].index);
2020                 free(array[i].data);
2021         }
2022         free(array);
2023 }
2024
2025 #ifndef NO_PTHREADS
2026
2027 static void try_to_free_from_threads(size_t size)
2028 {
2029         read_lock();
2030         release_pack_memory(size);
2031         read_unlock();
2032 }
2033
2034 static try_to_free_t old_try_to_free_routine;
2035
2036 /*
2037  * The main thread waits on the condition that (at least) one of the workers
2038  * has stopped working (which is indicated in the .working member of
2039  * struct thread_params).
2040  * When a work thread has completed its work, it sets .working to 0 and
2041  * signals the main thread and waits on the condition that .data_ready
2042  * becomes 1.
2043  */
2044
2045 struct thread_params {
2046         pthread_t thread;
2047         struct object_entry **list;
2048         unsigned list_size;
2049         unsigned remaining;
2050         int window;
2051         int depth;
2052         int working;
2053         int data_ready;
2054         pthread_mutex_t mutex;
2055         pthread_cond_t cond;
2056         unsigned *processed;
2057 };
2058
2059 static pthread_cond_t progress_cond;
2060
2061 /*
2062  * Mutex and conditional variable can't be statically-initialized on Windows.
2063  */
2064 static void init_threaded_search(void)
2065 {
2066         init_recursive_mutex(&read_mutex);
2067         pthread_mutex_init(&cache_mutex, NULL);
2068         pthread_mutex_init(&progress_mutex, NULL);
2069         pthread_cond_init(&progress_cond, NULL);
2070         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2071 }
2072
2073 static void cleanup_threaded_search(void)
2074 {
2075         set_try_to_free_routine(old_try_to_free_routine);
2076         pthread_cond_destroy(&progress_cond);
2077         pthread_mutex_destroy(&read_mutex);
2078         pthread_mutex_destroy(&cache_mutex);
2079         pthread_mutex_destroy(&progress_mutex);
2080 }
2081
2082 static void *threaded_find_deltas(void *arg)
2083 {
2084         struct thread_params *me = arg;
2085
2086         while (me->remaining) {
2087                 find_deltas(me->list, &me->remaining,
2088                             me->window, me->depth, me->processed);
2089
2090                 progress_lock();
2091                 me->working = 0;
2092                 pthread_cond_signal(&progress_cond);
2093                 progress_unlock();
2094
2095                 /*
2096                  * We must not set ->data_ready before we wait on the
2097                  * condition because the main thread may have set it to 1
2098                  * before we get here. In order to be sure that new
2099                  * work is available if we see 1 in ->data_ready, it
2100                  * was initialized to 0 before this thread was spawned
2101                  * and we reset it to 0 right away.
2102                  */
2103                 pthread_mutex_lock(&me->mutex);
2104                 while (!me->data_ready)
2105                         pthread_cond_wait(&me->cond, &me->mutex);
2106                 me->data_ready = 0;
2107                 pthread_mutex_unlock(&me->mutex);
2108         }
2109         /* leave ->working 1 so that this doesn't get more work assigned */
2110         return NULL;
2111 }
2112
2113 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2114                            int window, int depth, unsigned *processed)
2115 {
2116         struct thread_params *p;
2117         int i, ret, active_threads = 0;
2118
2119         init_threaded_search();
2120
2121         if (delta_search_threads <= 1) {
2122                 find_deltas(list, &list_size, window, depth, processed);
2123                 cleanup_threaded_search();
2124                 return;
2125         }
2126         if (progress > pack_to_stdout)
2127                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2128                                 delta_search_threads);
2129         p = xcalloc(delta_search_threads, sizeof(*p));
2130
2131         /* Partition the work amongst work threads. */
2132         for (i = 0; i < delta_search_threads; i++) {
2133                 unsigned sub_size = list_size / (delta_search_threads - i);
2134
2135                 /* don't use too small segments or no deltas will be found */
2136                 if (sub_size < 2*window && i+1 < delta_search_threads)
2137                         sub_size = 0;
2138
2139                 p[i].window = window;
2140                 p[i].depth = depth;
2141                 p[i].processed = processed;
2142                 p[i].working = 1;
2143                 p[i].data_ready = 0;
2144
2145                 /* try to split chunks on "path" boundaries */
2146                 while (sub_size && sub_size < list_size &&
2147                        list[sub_size]->hash &&
2148                        list[sub_size]->hash == list[sub_size-1]->hash)
2149                         sub_size++;
2150
2151                 p[i].list = list;
2152                 p[i].list_size = sub_size;
2153                 p[i].remaining = sub_size;
2154
2155                 list += sub_size;
2156                 list_size -= sub_size;
2157         }
2158
2159         /* Start work threads. */
2160         for (i = 0; i < delta_search_threads; i++) {
2161                 if (!p[i].list_size)
2162                         continue;
2163                 pthread_mutex_init(&p[i].mutex, NULL);
2164                 pthread_cond_init(&p[i].cond, NULL);
2165                 ret = pthread_create(&p[i].thread, NULL,
2166                                      threaded_find_deltas, &p[i]);
2167                 if (ret)
2168                         die("unable to create thread: %s", strerror(ret));
2169                 active_threads++;
2170         }
2171
2172         /*
2173          * Now let's wait for work completion.  Each time a thread is done
2174          * with its work, we steal half of the remaining work from the
2175          * thread with the largest number of unprocessed objects and give
2176          * it to that newly idle thread.  This ensure good load balancing
2177          * until the remaining object list segments are simply too short
2178          * to be worth splitting anymore.
2179          */
2180         while (active_threads) {
2181                 struct thread_params *target = NULL;
2182                 struct thread_params *victim = NULL;
2183                 unsigned sub_size = 0;
2184
2185                 progress_lock();
2186                 for (;;) {
2187                         for (i = 0; !target && i < delta_search_threads; i++)
2188                                 if (!p[i].working)
2189                                         target = &p[i];
2190                         if (target)
2191                                 break;
2192                         pthread_cond_wait(&progress_cond, &progress_mutex);
2193                 }
2194
2195                 for (i = 0; i < delta_search_threads; i++)
2196                         if (p[i].remaining > 2*window &&
2197                             (!victim || victim->remaining < p[i].remaining))
2198                                 victim = &p[i];
2199                 if (victim) {
2200                         sub_size = victim->remaining / 2;
2201                         list = victim->list + victim->list_size - sub_size;
2202                         while (sub_size && list[0]->hash &&
2203                                list[0]->hash == list[-1]->hash) {
2204                                 list++;
2205                                 sub_size--;
2206                         }
2207                         if (!sub_size) {
2208                                 /*
2209                                  * It is possible for some "paths" to have
2210                                  * so many objects that no hash boundary
2211                                  * might be found.  Let's just steal the
2212                                  * exact half in that case.
2213                                  */
2214                                 sub_size = victim->remaining / 2;
2215                                 list -= sub_size;
2216                         }
2217                         target->list = list;
2218                         victim->list_size -= sub_size;
2219                         victim->remaining -= sub_size;
2220                 }
2221                 target->list_size = sub_size;
2222                 target->remaining = sub_size;
2223                 target->working = 1;
2224                 progress_unlock();
2225
2226                 pthread_mutex_lock(&target->mutex);
2227                 target->data_ready = 1;
2228                 pthread_cond_signal(&target->cond);
2229                 pthread_mutex_unlock(&target->mutex);
2230
2231                 if (!sub_size) {
2232                         pthread_join(target->thread, NULL);
2233                         pthread_cond_destroy(&target->cond);
2234                         pthread_mutex_destroy(&target->mutex);
2235                         active_threads--;
2236                 }
2237         }
2238         cleanup_threaded_search();
2239         free(p);
2240 }
2241
2242 #else
2243 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2244 #endif
2245
2246 static void add_tag_chain(const struct object_id *oid)
2247 {
2248         struct tag *tag;
2249
2250         /*
2251          * We catch duplicates already in add_object_entry(), but we'd
2252          * prefer to do this extra check to avoid having to parse the
2253          * tag at all if we already know that it's being packed (e.g., if
2254          * it was included via bitmaps, we would not have parsed it
2255          * previously).
2256          */
2257         if (packlist_find(&to_pack, oid->hash, NULL))
2258                 return;
2259
2260         tag = lookup_tag(oid->hash);
2261         while (1) {
2262                 if (!tag || parse_tag(tag) || !tag->tagged)
2263                         die("unable to pack objects reachable from tag %s",
2264                             oid_to_hex(oid));
2265
2266                 add_object_entry(tag->object.oid.hash, OBJ_TAG, NULL, 0);
2267
2268                 if (tag->tagged->type != OBJ_TAG)
2269                         return;
2270
2271                 tag = (struct tag *)tag->tagged;
2272         }
2273 }
2274
2275 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2276 {
2277         struct object_id peeled;
2278
2279         if (starts_with(path, "refs/tags/") && /* is a tag? */
2280             !peel_ref(path, peeled.hash)    && /* peelable? */
2281             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2282                 add_tag_chain(oid);
2283         return 0;
2284 }
2285
2286 static void prepare_pack(int window, int depth)
2287 {
2288         struct object_entry **delta_list;
2289         uint32_t i, nr_deltas;
2290         unsigned n;
2291
2292         get_object_details();
2293
2294         /*
2295          * If we're locally repacking then we need to be doubly careful
2296          * from now on in order to make sure no stealth corruption gets
2297          * propagated to the new pack.  Clients receiving streamed packs
2298          * should validate everything they get anyway so no need to incur
2299          * the additional cost here in that case.
2300          */
2301         if (!pack_to_stdout)
2302                 do_check_packed_object_crc = 1;
2303
2304         if (!to_pack.nr_objects || !window || !depth)
2305                 return;
2306
2307         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2308         nr_deltas = n = 0;
2309
2310         for (i = 0; i < to_pack.nr_objects; i++) {
2311                 struct object_entry *entry = to_pack.objects + i;
2312
2313                 if (entry->delta)
2314                         /* This happens if we decided to reuse existing
2315                          * delta from a pack.  "reuse_delta &&" is implied.
2316                          */
2317                         continue;
2318
2319                 if (entry->size < 50)
2320                         continue;
2321
2322                 if (entry->no_try_delta)
2323                         continue;
2324
2325                 if (!entry->preferred_base) {
2326                         nr_deltas++;
2327                         if (entry->type < 0)
2328                                 die("unable to get type of object %s",
2329                                     sha1_to_hex(entry->idx.sha1));
2330                 } else {
2331                         if (entry->type < 0) {
2332                                 /*
2333                                  * This object is not found, but we
2334                                  * don't have to include it anyway.
2335                                  */
2336                                 continue;
2337                         }
2338                 }
2339
2340                 delta_list[n++] = entry;
2341         }
2342
2343         if (nr_deltas && n > 1) {
2344                 unsigned nr_done = 0;
2345                 if (progress)
2346                         progress_state = start_progress(_("Compressing objects"),
2347                                                         nr_deltas);
2348                 QSORT(delta_list, n, type_size_sort);
2349                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2350                 stop_progress(&progress_state);
2351                 if (nr_done != nr_deltas)
2352                         die("inconsistency with delta count");
2353         }
2354         free(delta_list);
2355 }
2356
2357 static int git_pack_config(const char *k, const char *v, void *cb)
2358 {
2359         if (!strcmp(k, "pack.window")) {
2360                 window = git_config_int(k, v);
2361                 return 0;
2362         }
2363         if (!strcmp(k, "pack.windowmemory")) {
2364                 window_memory_limit = git_config_ulong(k, v);
2365                 return 0;
2366         }
2367         if (!strcmp(k, "pack.depth")) {
2368                 depth = git_config_int(k, v);
2369                 return 0;
2370         }
2371         if (!strcmp(k, "pack.compression")) {
2372                 int level = git_config_int(k, v);
2373                 if (level == -1)
2374                         level = Z_DEFAULT_COMPRESSION;
2375                 else if (level < 0 || level > Z_BEST_COMPRESSION)
2376                         die("bad pack compression level %d", level);
2377                 pack_compression_level = level;
2378                 pack_compression_seen = 1;
2379                 return 0;
2380         }
2381         if (!strcmp(k, "pack.deltacachesize")) {
2382                 max_delta_cache_size = git_config_int(k, v);
2383                 return 0;
2384         }
2385         if (!strcmp(k, "pack.deltacachelimit")) {
2386                 cache_max_small_delta_size = git_config_int(k, v);
2387                 return 0;
2388         }
2389         if (!strcmp(k, "pack.writebitmaphashcache")) {
2390                 if (git_config_bool(k, v))
2391                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2392                 else
2393                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2394         }
2395         if (!strcmp(k, "pack.usebitmaps")) {
2396                 use_bitmap_index_default = git_config_bool(k, v);
2397                 return 0;
2398         }
2399         if (!strcmp(k, "pack.threads")) {
2400                 delta_search_threads = git_config_int(k, v);
2401                 if (delta_search_threads < 0)
2402                         die("invalid number of threads specified (%d)",
2403                             delta_search_threads);
2404 #ifdef NO_PTHREADS
2405                 if (delta_search_threads != 1)
2406                         warning("no threads support, ignoring %s", k);
2407 #endif
2408                 return 0;
2409         }
2410         if (!strcmp(k, "pack.indexversion")) {
2411                 pack_idx_opts.version = git_config_int(k, v);
2412                 if (pack_idx_opts.version > 2)
2413                         die("bad pack.indexversion=%"PRIu32,
2414                             pack_idx_opts.version);
2415                 return 0;
2416         }
2417         return git_default_config(k, v, cb);
2418 }
2419
2420 static void read_object_list_from_stdin(void)
2421 {
2422         char line[40 + 1 + PATH_MAX + 2];
2423         unsigned char sha1[20];
2424
2425         for (;;) {
2426                 if (!fgets(line, sizeof(line), stdin)) {
2427                         if (feof(stdin))
2428                                 break;
2429                         if (!ferror(stdin))
2430                                 die("fgets returned NULL, not EOF, not error!");
2431                         if (errno != EINTR)
2432                                 die_errno("fgets");
2433                         clearerr(stdin);
2434                         continue;
2435                 }
2436                 if (line[0] == '-') {
2437                         if (get_sha1_hex(line+1, sha1))
2438                                 die("expected edge sha1, got garbage:\n %s",
2439                                     line);
2440                         add_preferred_base(sha1);
2441                         continue;
2442                 }
2443                 if (get_sha1_hex(line, sha1))
2444                         die("expected sha1, got garbage:\n %s", line);
2445
2446                 add_preferred_base_object(line+41);
2447                 add_object_entry(sha1, 0, line+41, 0);
2448         }
2449 }
2450
2451 #define OBJECT_ADDED (1u<<20)
2452
2453 static void show_commit(struct commit *commit, void *data)
2454 {
2455         add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2456         commit->object.flags |= OBJECT_ADDED;
2457
2458         if (write_bitmap_index)
2459                 index_commit_for_bitmap(commit);
2460 }
2461
2462 static void show_object(struct object *obj, const char *name, void *data)
2463 {
2464         add_preferred_base_object(name);
2465         add_object_entry(obj->oid.hash, obj->type, name, 0);
2466         obj->flags |= OBJECT_ADDED;
2467 }
2468
2469 static void show_edge(struct commit *commit)
2470 {
2471         add_preferred_base(commit->object.oid.hash);
2472 }
2473
2474 struct in_pack_object {
2475         off_t offset;
2476         struct object *object;
2477 };
2478
2479 struct in_pack {
2480         int alloc;
2481         int nr;
2482         struct in_pack_object *array;
2483 };
2484
2485 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2486 {
2487         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2488         in_pack->array[in_pack->nr].object = object;
2489         in_pack->nr++;
2490 }
2491
2492 /*
2493  * Compare the objects in the offset order, in order to emulate the
2494  * "git rev-list --objects" output that produced the pack originally.
2495  */
2496 static int ofscmp(const void *a_, const void *b_)
2497 {
2498         struct in_pack_object *a = (struct in_pack_object *)a_;
2499         struct in_pack_object *b = (struct in_pack_object *)b_;
2500
2501         if (a->offset < b->offset)
2502                 return -1;
2503         else if (a->offset > b->offset)
2504                 return 1;
2505         else
2506                 return oidcmp(&a->object->oid, &b->object->oid);
2507 }
2508
2509 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2510 {
2511         struct packed_git *p;
2512         struct in_pack in_pack;
2513         uint32_t i;
2514
2515         memset(&in_pack, 0, sizeof(in_pack));
2516
2517         for (p = packed_git; p; p = p->next) {
2518                 const unsigned char *sha1;
2519                 struct object *o;
2520
2521                 if (!p->pack_local || p->pack_keep)
2522                         continue;
2523                 if (open_pack_index(p))
2524                         die("cannot open pack index");
2525
2526                 ALLOC_GROW(in_pack.array,
2527                            in_pack.nr + p->num_objects,
2528                            in_pack.alloc);
2529
2530                 for (i = 0; i < p->num_objects; i++) {
2531                         sha1 = nth_packed_object_sha1(p, i);
2532                         o = lookup_unknown_object(sha1);
2533                         if (!(o->flags & OBJECT_ADDED))
2534                                 mark_in_pack_object(o, p, &in_pack);
2535                         o->flags |= OBJECT_ADDED;
2536                 }
2537         }
2538
2539         if (in_pack.nr) {
2540                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2541                 for (i = 0; i < in_pack.nr; i++) {
2542                         struct object *o = in_pack.array[i].object;
2543                         add_object_entry(o->oid.hash, o->type, "", 0);
2544                 }
2545         }
2546         free(in_pack.array);
2547 }
2548
2549 static int add_loose_object(const unsigned char *sha1, const char *path,
2550                             void *data)
2551 {
2552         enum object_type type = sha1_object_info(sha1, NULL);
2553
2554         if (type < 0) {
2555                 warning("loose object at %s could not be examined", path);
2556                 return 0;
2557         }
2558
2559         add_object_entry(sha1, type, "", 0);
2560         return 0;
2561 }
2562
2563 /*
2564  * We actually don't even have to worry about reachability here.
2565  * add_object_entry will weed out duplicates, so we just add every
2566  * loose object we find.
2567  */
2568 static void add_unreachable_loose_objects(void)
2569 {
2570         for_each_loose_file_in_objdir(get_object_directory(),
2571                                       add_loose_object,
2572                                       NULL, NULL, NULL);
2573 }
2574
2575 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2576 {
2577         static struct packed_git *last_found = (void *)1;
2578         struct packed_git *p;
2579
2580         p = (last_found != (void *)1) ? last_found : packed_git;
2581
2582         while (p) {
2583                 if ((!p->pack_local || p->pack_keep) &&
2584                         find_pack_entry_one(sha1, p)) {
2585                         last_found = p;
2586                         return 1;
2587                 }
2588                 if (p == last_found)
2589                         p = packed_git;
2590                 else
2591                         p = p->next;
2592                 if (p == last_found)
2593                         p = p->next;
2594         }
2595         return 0;
2596 }
2597
2598 /*
2599  * Store a list of sha1s that are should not be discarded
2600  * because they are either written too recently, or are
2601  * reachable from another object that was.
2602  *
2603  * This is filled by get_object_list.
2604  */
2605 static struct sha1_array recent_objects;
2606
2607 static int loosened_object_can_be_discarded(const unsigned char *sha1,
2608                                             unsigned long mtime)
2609 {
2610         if (!unpack_unreachable_expiration)
2611                 return 0;
2612         if (mtime > unpack_unreachable_expiration)
2613                 return 0;
2614         if (sha1_array_lookup(&recent_objects, sha1) >= 0)
2615                 return 0;
2616         return 1;
2617 }
2618
2619 static void loosen_unused_packed_objects(struct rev_info *revs)
2620 {
2621         struct packed_git *p;
2622         uint32_t i;
2623         const unsigned char *sha1;
2624
2625         for (p = packed_git; p; p = p->next) {
2626                 if (!p->pack_local || p->pack_keep)
2627                         continue;
2628
2629                 if (open_pack_index(p))
2630                         die("cannot open pack index");
2631
2632                 for (i = 0; i < p->num_objects; i++) {
2633                         sha1 = nth_packed_object_sha1(p, i);
2634                         if (!packlist_find(&to_pack, sha1, NULL) &&
2635                             !has_sha1_pack_kept_or_nonlocal(sha1) &&
2636                             !loosened_object_can_be_discarded(sha1, p->mtime))
2637                                 if (force_object_loose(sha1, p->mtime))
2638                                         die("unable to force loose object");
2639                 }
2640         }
2641 }
2642
2643 /*
2644  * This tracks any options which pack-reuse code expects to be on, or which a
2645  * reader of the pack might not understand, and which would therefore prevent
2646  * blind reuse of what we have on disk.
2647  */
2648 static int pack_options_allow_reuse(void)
2649 {
2650         return pack_to_stdout && allow_ofs_delta;
2651 }
2652
2653 static int get_object_list_from_bitmap(struct rev_info *revs)
2654 {
2655         if (prepare_bitmap_walk(revs) < 0)
2656                 return -1;
2657
2658         if (pack_options_allow_reuse() &&
2659             !reuse_partial_packfile_from_bitmap(
2660                         &reuse_packfile,
2661                         &reuse_packfile_objects,
2662                         &reuse_packfile_offset)) {
2663                 assert(reuse_packfile_objects);
2664                 nr_result += reuse_packfile_objects;
2665                 display_progress(progress_state, nr_result);
2666         }
2667
2668         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2669         return 0;
2670 }
2671
2672 static void record_recent_object(struct object *obj,
2673                                  const char *name,
2674                                  void *data)
2675 {
2676         sha1_array_append(&recent_objects, obj->oid.hash);
2677 }
2678
2679 static void record_recent_commit(struct commit *commit, void *data)
2680 {
2681         sha1_array_append(&recent_objects, commit->object.oid.hash);
2682 }
2683
2684 static void get_object_list(int ac, const char **av)
2685 {
2686         struct rev_info revs;
2687         char line[1000];
2688         int flags = 0;
2689
2690         init_revisions(&revs, NULL);
2691         save_commit_buffer = 0;
2692         setup_revisions(ac, av, &revs, NULL);
2693
2694         /* make sure shallows are read */
2695         is_repository_shallow();
2696
2697         while (fgets(line, sizeof(line), stdin) != NULL) {
2698                 int len = strlen(line);
2699                 if (len && line[len - 1] == '\n')
2700                         line[--len] = 0;
2701                 if (!len)
2702                         break;
2703                 if (*line == '-') {
2704                         if (!strcmp(line, "--not")) {
2705                                 flags ^= UNINTERESTING;
2706                                 write_bitmap_index = 0;
2707                                 continue;
2708                         }
2709                         if (starts_with(line, "--shallow ")) {
2710                                 unsigned char sha1[20];
2711                                 if (get_sha1_hex(line + 10, sha1))
2712                                         die("not an SHA-1 '%s'", line + 10);
2713                                 register_shallow(sha1);
2714                                 use_bitmap_index = 0;
2715                                 continue;
2716                         }
2717                         die("not a rev '%s'", line);
2718                 }
2719                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2720                         die("bad revision '%s'", line);
2721         }
2722
2723         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2724                 return;
2725
2726         if (prepare_revision_walk(&revs))
2727                 die("revision walk setup failed");
2728         mark_edges_uninteresting(&revs, show_edge);
2729         traverse_commit_list(&revs, show_commit, show_object, NULL);
2730
2731         if (unpack_unreachable_expiration) {
2732                 revs.ignore_missing_links = 1;
2733                 if (add_unseen_recent_objects_to_traversal(&revs,
2734                                 unpack_unreachable_expiration))
2735                         die("unable to add recent objects");
2736                 if (prepare_revision_walk(&revs))
2737                         die("revision walk setup failed");
2738                 traverse_commit_list(&revs, record_recent_commit,
2739                                      record_recent_object, NULL);
2740         }
2741
2742         if (keep_unreachable)
2743                 add_objects_in_unpacked_packs(&revs);
2744         if (pack_loose_unreachable)
2745                 add_unreachable_loose_objects();
2746         if (unpack_unreachable)
2747                 loosen_unused_packed_objects(&revs);
2748
2749         sha1_array_clear(&recent_objects);
2750 }
2751
2752 static int option_parse_index_version(const struct option *opt,
2753                                       const char *arg, int unset)
2754 {
2755         char *c;
2756         const char *val = arg;
2757         pack_idx_opts.version = strtoul(val, &c, 10);
2758         if (pack_idx_opts.version > 2)
2759                 die(_("unsupported index version %s"), val);
2760         if (*c == ',' && c[1])
2761                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2762         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2763                 die(_("bad index version '%s'"), val);
2764         return 0;
2765 }
2766
2767 static int option_parse_unpack_unreachable(const struct option *opt,
2768                                            const char *arg, int unset)
2769 {
2770         if (unset) {
2771                 unpack_unreachable = 0;
2772                 unpack_unreachable_expiration = 0;
2773         }
2774         else {
2775                 unpack_unreachable = 1;
2776                 if (arg)
2777                         unpack_unreachable_expiration = approxidate(arg);
2778         }
2779         return 0;
2780 }
2781
2782 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2783 {
2784         int use_internal_rev_list = 0;
2785         int thin = 0;
2786         int shallow = 0;
2787         int all_progress_implied = 0;
2788         struct argv_array rp = ARGV_ARRAY_INIT;
2789         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2790         int rev_list_index = 0;
2791         struct option pack_objects_options[] = {
2792                 OPT_SET_INT('q', "quiet", &progress,
2793                             N_("do not show progress meter"), 0),
2794                 OPT_SET_INT(0, "progress", &progress,
2795                             N_("show progress meter"), 1),
2796                 OPT_SET_INT(0, "all-progress", &progress,
2797                             N_("show progress meter during object writing phase"), 2),
2798                 OPT_BOOL(0, "all-progress-implied",
2799                          &all_progress_implied,
2800                          N_("similar to --all-progress when progress meter is shown")),
2801                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2802                   N_("write the pack index file in the specified idx format version"),
2803                   0, option_parse_index_version },
2804                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2805                               N_("maximum size of each output pack file")),
2806                 OPT_BOOL(0, "local", &local,
2807                          N_("ignore borrowed objects from alternate object store")),
2808                 OPT_BOOL(0, "incremental", &incremental,
2809                          N_("ignore packed objects")),
2810                 OPT_INTEGER(0, "window", &window,
2811                             N_("limit pack window by objects")),
2812                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2813                               N_("limit pack window by memory in addition to object limit")),
2814                 OPT_INTEGER(0, "depth", &depth,
2815                             N_("maximum length of delta chain allowed in the resulting pack")),
2816                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2817                          N_("reuse existing deltas")),
2818                 OPT_BOOL(0, "reuse-object", &reuse_object,
2819                          N_("reuse existing objects")),
2820                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2821                          N_("use OFS_DELTA objects")),
2822                 OPT_INTEGER(0, "threads", &delta_search_threads,
2823                             N_("use threads when searching for best delta matches")),
2824                 OPT_BOOL(0, "non-empty", &non_empty,
2825                          N_("do not create an empty pack output")),
2826                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2827                          N_("read revision arguments from standard input")),
2828                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2829                   N_("limit the objects to those that are not yet packed"),
2830                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2831                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2832                   N_("include objects reachable from any reference"),
2833                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2834                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2835                   N_("include objects referred by reflog entries"),
2836                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2837                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2838                   N_("include objects referred to by the index"),
2839                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2840                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2841                          N_("output pack to stdout")),
2842                 OPT_BOOL(0, "include-tag", &include_tag,
2843                          N_("include tag objects that refer to objects to be packed")),
2844                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2845                          N_("keep unreachable objects")),
2846                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
2847                          N_("pack loose unreachable objects")),
2848                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2849                   N_("unpack unreachable objects newer than <time>"),
2850                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2851                 OPT_BOOL(0, "thin", &thin,
2852                          N_("create thin packs")),
2853                 OPT_BOOL(0, "shallow", &shallow,
2854                          N_("create packs suitable for shallow fetches")),
2855                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2856                          N_("ignore packs that have companion .keep file")),
2857                 OPT_INTEGER(0, "compression", &pack_compression_level,
2858                             N_("pack compression level")),
2859                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2860                             N_("do not hide commits by grafts"), 0),
2861                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2862                          N_("use a bitmap index if available to speed up counting objects")),
2863                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2864                          N_("write a bitmap index together with the pack index")),
2865                 OPT_END(),
2866         };
2867
2868         check_replace_refs = 0;
2869
2870         reset_pack_idx_option(&pack_idx_opts);
2871         git_config(git_pack_config, NULL);
2872         if (!pack_compression_seen && core_compression_seen)
2873                 pack_compression_level = core_compression_level;
2874
2875         progress = isatty(2);
2876         argc = parse_options(argc, argv, prefix, pack_objects_options,
2877                              pack_usage, 0);
2878
2879         if (argc) {
2880                 base_name = argv[0];
2881                 argc--;
2882         }
2883         if (pack_to_stdout != !base_name || argc)
2884                 usage_with_options(pack_usage, pack_objects_options);
2885
2886         argv_array_push(&rp, "pack-objects");
2887         if (thin) {
2888                 use_internal_rev_list = 1;
2889                 argv_array_push(&rp, shallow
2890                                 ? "--objects-edge-aggressive"
2891                                 : "--objects-edge");
2892         } else
2893                 argv_array_push(&rp, "--objects");
2894
2895         if (rev_list_all) {
2896                 use_internal_rev_list = 1;
2897                 argv_array_push(&rp, "--all");
2898         }
2899         if (rev_list_reflog) {
2900                 use_internal_rev_list = 1;
2901                 argv_array_push(&rp, "--reflog");
2902         }
2903         if (rev_list_index) {
2904                 use_internal_rev_list = 1;
2905                 argv_array_push(&rp, "--indexed-objects");
2906         }
2907         if (rev_list_unpacked) {
2908                 use_internal_rev_list = 1;
2909                 argv_array_push(&rp, "--unpacked");
2910         }
2911
2912         if (!reuse_object)
2913                 reuse_delta = 0;
2914         if (pack_compression_level == -1)
2915                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2916         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2917                 die("bad pack compression level %d", pack_compression_level);
2918
2919         if (!delta_search_threads)      /* --threads=0 means autodetect */
2920                 delta_search_threads = online_cpus();
2921
2922 #ifdef NO_PTHREADS
2923         if (delta_search_threads != 1)
2924                 warning("no threads support, ignoring --threads");
2925 #endif
2926         if (!pack_to_stdout && !pack_size_limit)
2927                 pack_size_limit = pack_size_limit_cfg;
2928         if (pack_to_stdout && pack_size_limit)
2929                 die("--max-pack-size cannot be used to build a pack for transfer.");
2930         if (pack_size_limit && pack_size_limit < 1024*1024) {
2931                 warning("minimum pack size limit is 1 MiB");
2932                 pack_size_limit = 1024*1024;
2933         }
2934
2935         if (!pack_to_stdout && thin)
2936                 die("--thin cannot be used to build an indexable pack.");
2937
2938         if (keep_unreachable && unpack_unreachable)
2939                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2940         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
2941                 unpack_unreachable_expiration = 0;
2942
2943         /*
2944          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
2945          *
2946          * - to produce good pack (with bitmap index not-yet-packed objects are
2947          *   packed in suboptimal order).
2948          *
2949          * - to use more robust pack-generation codepath (avoiding possible
2950          *   bugs in bitmap code and possible bitmap index corruption).
2951          */
2952         if (!pack_to_stdout)
2953                 use_bitmap_index_default = 0;
2954
2955         if (use_bitmap_index < 0)
2956                 use_bitmap_index = use_bitmap_index_default;
2957
2958         /* "hard" reasons not to use bitmaps; these just won't work at all */
2959         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
2960                 use_bitmap_index = 0;
2961
2962         if (pack_to_stdout || !rev_list_all)
2963                 write_bitmap_index = 0;
2964
2965         if (progress && all_progress_implied)
2966                 progress = 2;
2967
2968         prepare_packed_git();
2969         if (ignore_packed_keep) {
2970                 struct packed_git *p;
2971                 for (p = packed_git; p; p = p->next)
2972                         if (p->pack_local && p->pack_keep)
2973                                 break;
2974                 if (!p) /* no keep-able packs found */
2975                         ignore_packed_keep = 0;
2976         }
2977         if (local) {
2978                 /*
2979                  * unlike ignore_packed_keep above, we do not want to
2980                  * unset "local" based on looking at packs, as it
2981                  * also covers non-local objects
2982                  */
2983                 struct packed_git *p;
2984                 for (p = packed_git; p; p = p->next) {
2985                         if (!p->pack_local) {
2986                                 have_non_local_packs = 1;
2987                                 break;
2988                         }
2989                 }
2990         }
2991
2992         if (progress)
2993                 progress_state = start_progress(_("Counting objects"), 0);
2994         if (!use_internal_rev_list)
2995                 read_object_list_from_stdin();
2996         else {
2997                 get_object_list(rp.argc, rp.argv);
2998                 argv_array_clear(&rp);
2999         }
3000         cleanup_preferred_base();
3001         if (include_tag && nr_result)
3002                 for_each_ref(add_ref_tag, NULL);
3003         stop_progress(&progress_state);
3004
3005         if (non_empty && !nr_result)
3006                 return 0;
3007         if (nr_result)
3008                 prepare_pack(window, depth);
3009         write_pack_file();
3010         if (progress)
3011                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
3012                         " reused %"PRIu32" (delta %"PRIu32")\n",
3013                         written, written_delta, reused, reused_delta);
3014         return 0;
3015 }