Merge branch 'rs/http-push-cleanup' 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
65 static struct packed_git *reuse_packfile;
66 static uint32_t reuse_packfile_objects;
67 static off_t reuse_packfile_offset;
68
69 static int use_bitmap_index_default = 1;
70 static int use_bitmap_index = -1;
71 static int write_bitmap_index;
72 static uint16_t write_bitmap_options;
73
74 static unsigned long delta_cache_size = 0;
75 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
76 static unsigned long cache_max_small_delta_size = 1000;
77
78 static unsigned long window_memory_limit = 0;
79
80 /*
81  * stats
82  */
83 static uint32_t written, written_delta;
84 static uint32_t reused, reused_delta;
85
86 /*
87  * Indexed commits
88  */
89 static struct commit **indexed_commits;
90 static unsigned int indexed_commits_nr;
91 static unsigned int indexed_commits_alloc;
92
93 static void index_commit_for_bitmap(struct commit *commit)
94 {
95         if (indexed_commits_nr >= indexed_commits_alloc) {
96                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
97                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
98         }
99
100         indexed_commits[indexed_commits_nr++] = commit;
101 }
102
103 static void *get_delta(struct object_entry *entry)
104 {
105         unsigned long size, base_size, delta_size;
106         void *buf, *base_buf, *delta_buf;
107         enum object_type type;
108
109         buf = read_sha1_file(entry->idx.sha1, &type, &size);
110         if (!buf)
111                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
112         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
113         if (!base_buf)
114                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
115         delta_buf = diff_delta(base_buf, base_size,
116                                buf, size, &delta_size, 0);
117         if (!delta_buf || delta_size != entry->delta_size)
118                 die("delta size changed");
119         free(buf);
120         free(base_buf);
121         return delta_buf;
122 }
123
124 static unsigned long do_compress(void **pptr, unsigned long size)
125 {
126         git_zstream stream;
127         void *in, *out;
128         unsigned long maxsize;
129
130         git_deflate_init(&stream, pack_compression_level);
131         maxsize = git_deflate_bound(&stream, size);
132
133         in = *pptr;
134         out = xmalloc(maxsize);
135         *pptr = out;
136
137         stream.next_in = in;
138         stream.avail_in = size;
139         stream.next_out = out;
140         stream.avail_out = maxsize;
141         while (git_deflate(&stream, Z_FINISH) == Z_OK)
142                 ; /* nothing */
143         git_deflate_end(&stream);
144
145         free(in);
146         return stream.total_out;
147 }
148
149 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
150                                            const unsigned char *sha1)
151 {
152         git_zstream stream;
153         unsigned char ibuf[1024 * 16];
154         unsigned char obuf[1024 * 16];
155         unsigned long olen = 0;
156
157         git_deflate_init(&stream, pack_compression_level);
158
159         for (;;) {
160                 ssize_t readlen;
161                 int zret = Z_OK;
162                 readlen = read_istream(st, ibuf, sizeof(ibuf));
163                 if (readlen == -1)
164                         die(_("unable to read %s"), sha1_to_hex(sha1));
165
166                 stream.next_in = ibuf;
167                 stream.avail_in = readlen;
168                 while ((stream.avail_in || readlen == 0) &&
169                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
170                         stream.next_out = obuf;
171                         stream.avail_out = sizeof(obuf);
172                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
173                         sha1write(f, obuf, stream.next_out - obuf);
174                         olen += stream.next_out - obuf;
175                 }
176                 if (stream.avail_in)
177                         die(_("deflate error (%d)"), zret);
178                 if (readlen == 0) {
179                         if (zret != Z_STREAM_END)
180                                 die(_("deflate error (%d)"), zret);
181                         break;
182                 }
183         }
184         git_deflate_end(&stream);
185         return olen;
186 }
187
188 /*
189  * we are going to reuse the existing object data as is.  make
190  * sure it is not corrupt.
191  */
192 static int check_pack_inflate(struct packed_git *p,
193                 struct pack_window **w_curs,
194                 off_t offset,
195                 off_t len,
196                 unsigned long expect)
197 {
198         git_zstream stream;
199         unsigned char fakebuf[4096], *in;
200         int st;
201
202         memset(&stream, 0, sizeof(stream));
203         git_inflate_init(&stream);
204         do {
205                 in = use_pack(p, w_curs, offset, &stream.avail_in);
206                 stream.next_in = in;
207                 stream.next_out = fakebuf;
208                 stream.avail_out = sizeof(fakebuf);
209                 st = git_inflate(&stream, Z_FINISH);
210                 offset += stream.next_in - in;
211         } while (st == Z_OK || st == Z_BUF_ERROR);
212         git_inflate_end(&stream);
213         return (st == Z_STREAM_END &&
214                 stream.total_out == expect &&
215                 stream.total_in == len) ? 0 : -1;
216 }
217
218 static void copy_pack_data(struct sha1file *f,
219                 struct packed_git *p,
220                 struct pack_window **w_curs,
221                 off_t offset,
222                 off_t len)
223 {
224         unsigned char *in;
225         unsigned long avail;
226
227         while (len) {
228                 in = use_pack(p, w_curs, offset, &avail);
229                 if (avail > len)
230                         avail = (unsigned long)len;
231                 sha1write(f, in, avail);
232                 offset += avail;
233                 len -= avail;
234         }
235 }
236
237 /* Return 0 if we will bust the pack-size limit */
238 static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
239                                            unsigned long limit, int usable_delta)
240 {
241         unsigned long size, datalen;
242         unsigned char header[10], dheader[10];
243         unsigned hdrlen;
244         enum object_type type;
245         void *buf;
246         struct git_istream *st = NULL;
247
248         if (!usable_delta) {
249                 if (entry->type == OBJ_BLOB &&
250                     entry->size > big_file_threshold &&
251                     (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
252                         buf = NULL;
253                 else {
254                         buf = read_sha1_file(entry->idx.sha1, &type, &size);
255                         if (!buf)
256                                 die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
257                 }
258                 /*
259                  * make sure no cached delta data remains from a
260                  * previous attempt before a pack split occurred.
261                  */
262                 free(entry->delta_data);
263                 entry->delta_data = NULL;
264                 entry->z_delta_size = 0;
265         } else if (entry->delta_data) {
266                 size = entry->delta_size;
267                 buf = entry->delta_data;
268                 entry->delta_data = NULL;
269                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
270                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
271         } else {
272                 buf = get_delta(entry);
273                 size = entry->delta_size;
274                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
275                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
276         }
277
278         if (st) /* large blob case, just assume we don't compress well */
279                 datalen = size;
280         else if (entry->z_delta_size)
281                 datalen = entry->z_delta_size;
282         else
283                 datalen = do_compress(&buf, size);
284
285         /*
286          * The object header is a byte of 'type' followed by zero or
287          * more bytes of length.
288          */
289         hdrlen = encode_in_pack_object_header(type, size, header);
290
291         if (type == OBJ_OFS_DELTA) {
292                 /*
293                  * Deltas with relative base contain an additional
294                  * encoding of the relative offset for the delta
295                  * base from this object's position in the pack.
296                  */
297                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
298                 unsigned pos = sizeof(dheader) - 1;
299                 dheader[pos] = ofs & 127;
300                 while (ofs >>= 7)
301                         dheader[--pos] = 128 | (--ofs & 127);
302                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
303                         if (st)
304                                 close_istream(st);
305                         free(buf);
306                         return 0;
307                 }
308                 sha1write(f, header, hdrlen);
309                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
310                 hdrlen += sizeof(dheader) - pos;
311         } else if (type == OBJ_REF_DELTA) {
312                 /*
313                  * Deltas with a base reference contain
314                  * an additional 20 bytes for the base sha1.
315                  */
316                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
317                         if (st)
318                                 close_istream(st);
319                         free(buf);
320                         return 0;
321                 }
322                 sha1write(f, header, hdrlen);
323                 sha1write(f, entry->delta->idx.sha1, 20);
324                 hdrlen += 20;
325         } else {
326                 if (limit && hdrlen + datalen + 20 >= limit) {
327                         if (st)
328                                 close_istream(st);
329                         free(buf);
330                         return 0;
331                 }
332                 sha1write(f, header, hdrlen);
333         }
334         if (st) {
335                 datalen = write_large_blob_data(st, f, entry->idx.sha1);
336                 close_istream(st);
337         } else {
338                 sha1write(f, buf, datalen);
339                 free(buf);
340         }
341
342         return hdrlen + datalen;
343 }
344
345 /* Return 0 if we will bust the pack-size limit */
346 static off_t write_reuse_object(struct sha1file *f, struct object_entry *entry,
347                                 unsigned long limit, int usable_delta)
348 {
349         struct packed_git *p = entry->in_pack;
350         struct pack_window *w_curs = NULL;
351         struct revindex_entry *revidx;
352         off_t offset;
353         enum object_type type = entry->type;
354         off_t datalen;
355         unsigned char header[10], dheader[10];
356         unsigned hdrlen;
357
358         if (entry->delta)
359                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
360                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
361         hdrlen = encode_in_pack_object_header(type, entry->size, header);
362
363         offset = entry->in_pack_offset;
364         revidx = find_pack_revindex(p, offset);
365         datalen = revidx[1].offset - offset;
366         if (!pack_to_stdout && p->index_version > 1 &&
367             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
368                 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
369                 unuse_pack(&w_curs);
370                 return write_no_reuse_object(f, entry, limit, usable_delta);
371         }
372
373         offset += entry->in_pack_header_size;
374         datalen -= entry->in_pack_header_size;
375
376         if (!pack_to_stdout && p->index_version == 1 &&
377             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
378                 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
379                 unuse_pack(&w_curs);
380                 return write_no_reuse_object(f, entry, limit, usable_delta);
381         }
382
383         if (type == OBJ_OFS_DELTA) {
384                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
385                 unsigned pos = sizeof(dheader) - 1;
386                 dheader[pos] = ofs & 127;
387                 while (ofs >>= 7)
388                         dheader[--pos] = 128 | (--ofs & 127);
389                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
390                         unuse_pack(&w_curs);
391                         return 0;
392                 }
393                 sha1write(f, header, hdrlen);
394                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
395                 hdrlen += sizeof(dheader) - pos;
396                 reused_delta++;
397         } else if (type == OBJ_REF_DELTA) {
398                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
399                         unuse_pack(&w_curs);
400                         return 0;
401                 }
402                 sha1write(f, header, hdrlen);
403                 sha1write(f, entry->delta->idx.sha1, 20);
404                 hdrlen += 20;
405                 reused_delta++;
406         } else {
407                 if (limit && hdrlen + datalen + 20 >= limit) {
408                         unuse_pack(&w_curs);
409                         return 0;
410                 }
411                 sha1write(f, header, hdrlen);
412         }
413         copy_pack_data(f, p, &w_curs, offset, datalen);
414         unuse_pack(&w_curs);
415         reused++;
416         return hdrlen + datalen;
417 }
418
419 /* Return 0 if we will bust the pack-size limit */
420 static off_t write_object(struct sha1file *f,
421                           struct object_entry *entry,
422                           off_t write_offset)
423 {
424         unsigned long limit;
425         off_t len;
426         int usable_delta, to_reuse;
427
428         if (!pack_to_stdout)
429                 crc32_begin(f);
430
431         /* apply size limit if limited packsize and not first object */
432         if (!pack_size_limit || !nr_written)
433                 limit = 0;
434         else if (pack_size_limit <= write_offset)
435                 /*
436                  * the earlier object did not fit the limit; avoid
437                  * mistaking this with unlimited (i.e. limit = 0).
438                  */
439                 limit = 1;
440         else
441                 limit = pack_size_limit - write_offset;
442
443         if (!entry->delta)
444                 usable_delta = 0;       /* no delta */
445         else if (!pack_size_limit)
446                usable_delta = 1;        /* unlimited packfile */
447         else if (entry->delta->idx.offset == (off_t)-1)
448                 usable_delta = 0;       /* base was written to another pack */
449         else if (entry->delta->idx.offset)
450                 usable_delta = 1;       /* base already exists in this pack */
451         else
452                 usable_delta = 0;       /* base could end up in another pack */
453
454         if (!reuse_object)
455                 to_reuse = 0;   /* explicit */
456         else if (!entry->in_pack)
457                 to_reuse = 0;   /* can't reuse what we don't have */
458         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
459                                 /* check_object() decided it for us ... */
460                 to_reuse = usable_delta;
461                                 /* ... but pack split may override that */
462         else if (entry->type != entry->in_pack_type)
463                 to_reuse = 0;   /* pack has delta which is unusable */
464         else if (entry->delta)
465                 to_reuse = 0;   /* we want to pack afresh */
466         else
467                 to_reuse = 1;   /* we have it in-pack undeltified,
468                                  * and we do not need to deltify it.
469                                  */
470
471         if (!to_reuse)
472                 len = write_no_reuse_object(f, entry, limit, usable_delta);
473         else
474                 len = write_reuse_object(f, entry, limit, usable_delta);
475         if (!len)
476                 return 0;
477
478         if (usable_delta)
479                 written_delta++;
480         written++;
481         if (!pack_to_stdout)
482                 entry->idx.crc32 = crc32_end(f);
483         return len;
484 }
485
486 enum write_one_status {
487         WRITE_ONE_SKIP = -1, /* already written */
488         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
489         WRITE_ONE_WRITTEN = 1, /* normal */
490         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
491 };
492
493 static enum write_one_status write_one(struct sha1file *f,
494                                        struct object_entry *e,
495                                        off_t *offset)
496 {
497         off_t size;
498         int recursing;
499
500         /*
501          * we set offset to 1 (which is an impossible value) to mark
502          * the fact that this object is involved in "write its base
503          * first before writing a deltified object" recursion.
504          */
505         recursing = (e->idx.offset == 1);
506         if (recursing) {
507                 warning("recursive delta detected for object %s",
508                         sha1_to_hex(e->idx.sha1));
509                 return WRITE_ONE_RECURSIVE;
510         } else if (e->idx.offset || e->preferred_base) {
511                 /* offset is non zero if object is written already. */
512                 return WRITE_ONE_SKIP;
513         }
514
515         /* if we are deltified, write out base object first. */
516         if (e->delta) {
517                 e->idx.offset = 1; /* now recurse */
518                 switch (write_one(f, e->delta, offset)) {
519                 case WRITE_ONE_RECURSIVE:
520                         /* we cannot depend on this one */
521                         e->delta = NULL;
522                         break;
523                 default:
524                         break;
525                 case WRITE_ONE_BREAK:
526                         e->idx.offset = recursing;
527                         return WRITE_ONE_BREAK;
528                 }
529         }
530
531         e->idx.offset = *offset;
532         size = write_object(f, e, *offset);
533         if (!size) {
534                 e->idx.offset = recursing;
535                 return WRITE_ONE_BREAK;
536         }
537         written_list[nr_written++] = &e->idx;
538
539         /* make sure off_t is sufficiently large not to wrap */
540         if (signed_add_overflows(*offset, size))
541                 die("pack too large for current definition of off_t");
542         *offset += size;
543         return WRITE_ONE_WRITTEN;
544 }
545
546 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
547                        void *cb_data)
548 {
549         unsigned char peeled[20];
550         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
551
552         if (entry)
553                 entry->tagged = 1;
554         if (!peel_ref(path, peeled)) {
555                 entry = packlist_find(&to_pack, peeled, NULL);
556                 if (entry)
557                         entry->tagged = 1;
558         }
559         return 0;
560 }
561
562 static inline void add_to_write_order(struct object_entry **wo,
563                                unsigned int *endp,
564                                struct object_entry *e)
565 {
566         if (e->filled)
567                 return;
568         wo[(*endp)++] = e;
569         e->filled = 1;
570 }
571
572 static void add_descendants_to_write_order(struct object_entry **wo,
573                                            unsigned int *endp,
574                                            struct object_entry *e)
575 {
576         int add_to_order = 1;
577         while (e) {
578                 if (add_to_order) {
579                         struct object_entry *s;
580                         /* add this node... */
581                         add_to_write_order(wo, endp, e);
582                         /* all its siblings... */
583                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
584                                 add_to_write_order(wo, endp, s);
585                         }
586                 }
587                 /* drop down a level to add left subtree nodes if possible */
588                 if (e->delta_child) {
589                         add_to_order = 1;
590                         e = e->delta_child;
591                 } else {
592                         add_to_order = 0;
593                         /* our sibling might have some children, it is next */
594                         if (e->delta_sibling) {
595                                 e = e->delta_sibling;
596                                 continue;
597                         }
598                         /* go back to our parent node */
599                         e = e->delta;
600                         while (e && !e->delta_sibling) {
601                                 /* we're on the right side of a subtree, keep
602                                  * going up until we can go right again */
603                                 e = e->delta;
604                         }
605                         if (!e) {
606                                 /* done- we hit our original root node */
607                                 return;
608                         }
609                         /* pass it off to sibling at this level */
610                         e = e->delta_sibling;
611                 }
612         };
613 }
614
615 static void add_family_to_write_order(struct object_entry **wo,
616                                       unsigned int *endp,
617                                       struct object_entry *e)
618 {
619         struct object_entry *root;
620
621         for (root = e; root->delta; root = root->delta)
622                 ; /* nothing */
623         add_descendants_to_write_order(wo, endp, root);
624 }
625
626 static struct object_entry **compute_write_order(void)
627 {
628         unsigned int i, wo_end, last_untagged;
629
630         struct object_entry **wo;
631         struct object_entry *objects = to_pack.objects;
632
633         for (i = 0; i < to_pack.nr_objects; i++) {
634                 objects[i].tagged = 0;
635                 objects[i].filled = 0;
636                 objects[i].delta_child = NULL;
637                 objects[i].delta_sibling = NULL;
638         }
639
640         /*
641          * Fully connect delta_child/delta_sibling network.
642          * Make sure delta_sibling is sorted in the original
643          * recency order.
644          */
645         for (i = to_pack.nr_objects; i > 0;) {
646                 struct object_entry *e = &objects[--i];
647                 if (!e->delta)
648                         continue;
649                 /* Mark me as the first child */
650                 e->delta_sibling = e->delta->delta_child;
651                 e->delta->delta_child = e;
652         }
653
654         /*
655          * Mark objects that are at the tip of tags.
656          */
657         for_each_tag_ref(mark_tagged, NULL);
658
659         /*
660          * Give the objects in the original recency order until
661          * we see a tagged tip.
662          */
663         ALLOC_ARRAY(wo, to_pack.nr_objects);
664         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
665                 if (objects[i].tagged)
666                         break;
667                 add_to_write_order(wo, &wo_end, &objects[i]);
668         }
669         last_untagged = i;
670
671         /*
672          * Then fill all the tagged tips.
673          */
674         for (; i < to_pack.nr_objects; i++) {
675                 if (objects[i].tagged)
676                         add_to_write_order(wo, &wo_end, &objects[i]);
677         }
678
679         /*
680          * And then all remaining commits and tags.
681          */
682         for (i = last_untagged; i < to_pack.nr_objects; i++) {
683                 if (objects[i].type != OBJ_COMMIT &&
684                     objects[i].type != OBJ_TAG)
685                         continue;
686                 add_to_write_order(wo, &wo_end, &objects[i]);
687         }
688
689         /*
690          * And then all the trees.
691          */
692         for (i = last_untagged; i < to_pack.nr_objects; i++) {
693                 if (objects[i].type != OBJ_TREE)
694                         continue;
695                 add_to_write_order(wo, &wo_end, &objects[i]);
696         }
697
698         /*
699          * Finally all the rest in really tight order
700          */
701         for (i = last_untagged; i < to_pack.nr_objects; i++) {
702                 if (!objects[i].filled)
703                         add_family_to_write_order(wo, &wo_end, &objects[i]);
704         }
705
706         if (wo_end != to_pack.nr_objects)
707                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
708
709         return wo;
710 }
711
712 static off_t write_reused_pack(struct sha1file *f)
713 {
714         unsigned char buffer[8192];
715         off_t to_write, total;
716         int fd;
717
718         if (!is_pack_valid(reuse_packfile))
719                 die("packfile is invalid: %s", reuse_packfile->pack_name);
720
721         fd = git_open(reuse_packfile->pack_name);
722         if (fd < 0)
723                 die_errno("unable to open packfile for reuse: %s",
724                           reuse_packfile->pack_name);
725
726         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
727                 die_errno("unable to seek in reused packfile");
728
729         if (reuse_packfile_offset < 0)
730                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
731
732         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
733
734         while (to_write) {
735                 int read_pack = xread(fd, buffer, sizeof(buffer));
736
737                 if (read_pack <= 0)
738                         die_errno("unable to read from reused packfile");
739
740                 if (read_pack > to_write)
741                         read_pack = to_write;
742
743                 sha1write(f, buffer, read_pack);
744                 to_write -= read_pack;
745
746                 /*
747                  * We don't know the actual number of objects written,
748                  * only how many bytes written, how many bytes total, and
749                  * how many objects total. So we can fake it by pretending all
750                  * objects we are writing are the same size. This gives us a
751                  * smooth progress meter, and at the end it matches the true
752                  * answer.
753                  */
754                 written = reuse_packfile_objects *
755                                 (((double)(total - to_write)) / total);
756                 display_progress(progress_state, written);
757         }
758
759         close(fd);
760         written = reuse_packfile_objects;
761         display_progress(progress_state, written);
762         return reuse_packfile_offset - sizeof(struct pack_header);
763 }
764
765 static const char no_split_warning[] = N_(
766 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
767 );
768
769 static void write_pack_file(void)
770 {
771         uint32_t i = 0, j;
772         struct sha1file *f;
773         off_t offset;
774         uint32_t nr_remaining = nr_result;
775         time_t last_mtime = 0;
776         struct object_entry **write_order;
777
778         if (progress > pack_to_stdout)
779                 progress_state = start_progress(_("Writing objects"), nr_result);
780         ALLOC_ARRAY(written_list, to_pack.nr_objects);
781         write_order = compute_write_order();
782
783         do {
784                 unsigned char sha1[20];
785                 char *pack_tmp_name = NULL;
786
787                 if (pack_to_stdout)
788                         f = sha1fd_throughput(1, "<stdout>", progress_state);
789                 else
790                         f = create_tmp_packfile(&pack_tmp_name);
791
792                 offset = write_pack_header(f, nr_remaining);
793
794                 if (reuse_packfile) {
795                         off_t packfile_size;
796                         assert(pack_to_stdout);
797
798                         packfile_size = write_reused_pack(f);
799                         offset += packfile_size;
800                 }
801
802                 nr_written = 0;
803                 for (; i < to_pack.nr_objects; i++) {
804                         struct object_entry *e = write_order[i];
805                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
806                                 break;
807                         display_progress(progress_state, written);
808                 }
809
810                 /*
811                  * Did we write the wrong # entries in the header?
812                  * If so, rewrite it like in fast-import
813                  */
814                 if (pack_to_stdout) {
815                         sha1close(f, sha1, CSUM_CLOSE);
816                 } else if (nr_written == nr_remaining) {
817                         sha1close(f, sha1, CSUM_FSYNC);
818                 } else {
819                         int fd = sha1close(f, sha1, 0);
820                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
821                                                  nr_written, sha1, offset);
822                         close(fd);
823                         if (write_bitmap_index) {
824                                 warning(_(no_split_warning));
825                                 write_bitmap_index = 0;
826                         }
827                 }
828
829                 if (!pack_to_stdout) {
830                         struct stat st;
831                         struct strbuf tmpname = STRBUF_INIT;
832
833                         /*
834                          * Packs are runtime accessed in their mtime
835                          * order since newer packs are more likely to contain
836                          * younger objects.  So if we are creating multiple
837                          * packs then we should modify the mtime of later ones
838                          * to preserve this property.
839                          */
840                         if (stat(pack_tmp_name, &st) < 0) {
841                                 warning_errno("failed to stat %s", pack_tmp_name);
842                         } else if (!last_mtime) {
843                                 last_mtime = st.st_mtime;
844                         } else {
845                                 struct utimbuf utb;
846                                 utb.actime = st.st_atime;
847                                 utb.modtime = --last_mtime;
848                                 if (utime(pack_tmp_name, &utb) < 0)
849                                         warning_errno("failed utime() on %s", pack_tmp_name);
850                         }
851
852                         strbuf_addf(&tmpname, "%s-", base_name);
853
854                         if (write_bitmap_index) {
855                                 bitmap_writer_set_checksum(sha1);
856                                 bitmap_writer_build_type_index(written_list, nr_written);
857                         }
858
859                         finish_tmp_packfile(&tmpname, pack_tmp_name,
860                                             written_list, nr_written,
861                                             &pack_idx_opts, sha1);
862
863                         if (write_bitmap_index) {
864                                 strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
865
866                                 stop_progress(&progress_state);
867
868                                 bitmap_writer_show_progress(progress);
869                                 bitmap_writer_reuse_bitmaps(&to_pack);
870                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
871                                 bitmap_writer_build(&to_pack);
872                                 bitmap_writer_finish(written_list, nr_written,
873                                                      tmpname.buf, write_bitmap_options);
874                                 write_bitmap_index = 0;
875                         }
876
877                         strbuf_release(&tmpname);
878                         free(pack_tmp_name);
879                         puts(sha1_to_hex(sha1));
880                 }
881
882                 /* mark written objects as written to previous pack */
883                 for (j = 0; j < nr_written; j++) {
884                         written_list[j]->offset = (off_t)-1;
885                 }
886                 nr_remaining -= nr_written;
887         } while (nr_remaining && i < to_pack.nr_objects);
888
889         free(written_list);
890         free(write_order);
891         stop_progress(&progress_state);
892         if (written != nr_result)
893                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
894                         written, nr_result);
895 }
896
897 static void setup_delta_attr_check(struct git_attr_check *check)
898 {
899         static struct git_attr *attr_delta;
900
901         if (!attr_delta)
902                 attr_delta = git_attr("delta");
903
904         check[0].attr = attr_delta;
905 }
906
907 static int no_try_delta(const char *path)
908 {
909         struct git_attr_check check[1];
910
911         setup_delta_attr_check(check);
912         if (git_check_attr(path, ARRAY_SIZE(check), check))
913                 return 0;
914         if (ATTR_FALSE(check->value))
915                 return 1;
916         return 0;
917 }
918
919 /*
920  * When adding an object, check whether we have already added it
921  * to our packing list. If so, we can skip. However, if we are
922  * being asked to excludei t, but the previous mention was to include
923  * it, make sure to adjust its flags and tweak our numbers accordingly.
924  *
925  * As an optimization, we pass out the index position where we would have
926  * found the item, since that saves us from having to look it up again a
927  * few lines later when we want to add the new entry.
928  */
929 static int have_duplicate_entry(const unsigned char *sha1,
930                                 int exclude,
931                                 uint32_t *index_pos)
932 {
933         struct object_entry *entry;
934
935         entry = packlist_find(&to_pack, sha1, index_pos);
936         if (!entry)
937                 return 0;
938
939         if (exclude) {
940                 if (!entry->preferred_base)
941                         nr_result--;
942                 entry->preferred_base = 1;
943         }
944
945         return 1;
946 }
947
948 static int want_found_object(int exclude, struct packed_git *p)
949 {
950         if (exclude)
951                 return 1;
952         if (incremental)
953                 return 0;
954
955         /*
956          * When asked to do --local (do not include an object that appears in a
957          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
958          * an object that appears in a pack marked with .keep), finding a pack
959          * that matches the criteria is sufficient for us to decide to omit it.
960          * However, even if this pack does not satisfy the criteria, we need to
961          * make sure no copy of this object appears in _any_ pack that makes us
962          * to omit the object, so we need to check all the packs.
963          *
964          * We can however first check whether these options can possible matter;
965          * if they do not matter we know we want the object in generated pack.
966          * Otherwise, we signal "-1" at the end to tell the caller that we do
967          * not know either way, and it needs to check more packs.
968          */
969         if (!ignore_packed_keep &&
970             (!local || !have_non_local_packs))
971                 return 1;
972
973         if (local && !p->pack_local)
974                 return 0;
975         if (ignore_packed_keep && p->pack_local && p->pack_keep)
976                 return 0;
977
978         /* we don't know yet; keep looking for more packs */
979         return -1;
980 }
981
982 /*
983  * Check whether we want the object in the pack (e.g., we do not want
984  * objects found in non-local stores if the "--local" option was used).
985  *
986  * If the caller already knows an existing pack it wants to take the object
987  * from, that is passed in *found_pack and *found_offset; otherwise this
988  * function finds if there is any pack that has the object and returns the pack
989  * and its offset in these variables.
990  */
991 static int want_object_in_pack(const unsigned char *sha1,
992                                int exclude,
993                                struct packed_git **found_pack,
994                                off_t *found_offset)
995 {
996         struct mru_entry *entry;
997         int want;
998
999         if (!exclude && local && has_loose_object_nonlocal(sha1))
1000                 return 0;
1001
1002         /*
1003          * If we already know the pack object lives in, start checks from that
1004          * pack - in the usual case when neither --local was given nor .keep files
1005          * are present we will determine the answer right now.
1006          */
1007         if (*found_pack) {
1008                 want = want_found_object(exclude, *found_pack);
1009                 if (want != -1)
1010                         return want;
1011         }
1012
1013         for (entry = packed_git_mru->head; entry; entry = entry->next) {
1014                 struct packed_git *p = entry->item;
1015                 off_t offset;
1016
1017                 if (p == *found_pack)
1018                         offset = *found_offset;
1019                 else
1020                         offset = find_pack_entry_one(sha1, p);
1021
1022                 if (offset) {
1023                         if (!*found_pack) {
1024                                 if (!is_pack_valid(p))
1025                                         continue;
1026                                 *found_offset = offset;
1027                                 *found_pack = p;
1028                         }
1029                         want = want_found_object(exclude, p);
1030                         if (!exclude && want > 0)
1031                                 mru_mark(packed_git_mru, entry);
1032                         if (want != -1)
1033                                 return want;
1034                 }
1035         }
1036
1037         return 1;
1038 }
1039
1040 static void create_object_entry(const unsigned char *sha1,
1041                                 enum object_type type,
1042                                 uint32_t hash,
1043                                 int exclude,
1044                                 int no_try_delta,
1045                                 uint32_t index_pos,
1046                                 struct packed_git *found_pack,
1047                                 off_t found_offset)
1048 {
1049         struct object_entry *entry;
1050
1051         entry = packlist_alloc(&to_pack, sha1, index_pos);
1052         entry->hash = hash;
1053         if (type)
1054                 entry->type = type;
1055         if (exclude)
1056                 entry->preferred_base = 1;
1057         else
1058                 nr_result++;
1059         if (found_pack) {
1060                 entry->in_pack = found_pack;
1061                 entry->in_pack_offset = found_offset;
1062         }
1063
1064         entry->no_try_delta = no_try_delta;
1065 }
1066
1067 static const char no_closure_warning[] = N_(
1068 "disabling bitmap writing, as some objects are not being packed"
1069 );
1070
1071 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1072                             const char *name, int exclude)
1073 {
1074         struct packed_git *found_pack = NULL;
1075         off_t found_offset = 0;
1076         uint32_t index_pos;
1077
1078         if (have_duplicate_entry(sha1, exclude, &index_pos))
1079                 return 0;
1080
1081         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1082                 /* The pack is missing an object, so it will not have closure */
1083                 if (write_bitmap_index) {
1084                         warning(_(no_closure_warning));
1085                         write_bitmap_index = 0;
1086                 }
1087                 return 0;
1088         }
1089
1090         create_object_entry(sha1, type, pack_name_hash(name),
1091                             exclude, name && no_try_delta(name),
1092                             index_pos, found_pack, found_offset);
1093
1094         display_progress(progress_state, nr_result);
1095         return 1;
1096 }
1097
1098 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1099                                         enum object_type type,
1100                                         int flags, uint32_t name_hash,
1101                                         struct packed_git *pack, off_t offset)
1102 {
1103         uint32_t index_pos;
1104
1105         if (have_duplicate_entry(sha1, 0, &index_pos))
1106                 return 0;
1107
1108         if (!want_object_in_pack(sha1, 0, &pack, &offset))
1109                 return 0;
1110
1111         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1112
1113         display_progress(progress_state, nr_result);
1114         return 1;
1115 }
1116
1117 struct pbase_tree_cache {
1118         unsigned char sha1[20];
1119         int ref;
1120         int temporary;
1121         void *tree_data;
1122         unsigned long tree_size;
1123 };
1124
1125 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1126 static int pbase_tree_cache_ix(const unsigned char *sha1)
1127 {
1128         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1129 }
1130 static int pbase_tree_cache_ix_incr(int ix)
1131 {
1132         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1133 }
1134
1135 static struct pbase_tree {
1136         struct pbase_tree *next;
1137         /* This is a phony "cache" entry; we are not
1138          * going to evict it or find it through _get()
1139          * mechanism -- this is for the toplevel node that
1140          * would almost always change with any commit.
1141          */
1142         struct pbase_tree_cache pcache;
1143 } *pbase_tree;
1144
1145 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1146 {
1147         struct pbase_tree_cache *ent, *nent;
1148         void *data;
1149         unsigned long size;
1150         enum object_type type;
1151         int neigh;
1152         int my_ix = pbase_tree_cache_ix(sha1);
1153         int available_ix = -1;
1154
1155         /* pbase-tree-cache acts as a limited hashtable.
1156          * your object will be found at your index or within a few
1157          * slots after that slot if it is cached.
1158          */
1159         for (neigh = 0; neigh < 8; neigh++) {
1160                 ent = pbase_tree_cache[my_ix];
1161                 if (ent && !hashcmp(ent->sha1, sha1)) {
1162                         ent->ref++;
1163                         return ent;
1164                 }
1165                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1166                          ((0 <= available_ix) &&
1167                           (!ent && pbase_tree_cache[available_ix])))
1168                         available_ix = my_ix;
1169                 if (!ent)
1170                         break;
1171                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1172         }
1173
1174         /* Did not find one.  Either we got a bogus request or
1175          * we need to read and perhaps cache.
1176          */
1177         data = read_sha1_file(sha1, &type, &size);
1178         if (!data)
1179                 return NULL;
1180         if (type != OBJ_TREE) {
1181                 free(data);
1182                 return NULL;
1183         }
1184
1185         /* We need to either cache or return a throwaway copy */
1186
1187         if (available_ix < 0)
1188                 ent = NULL;
1189         else {
1190                 ent = pbase_tree_cache[available_ix];
1191                 my_ix = available_ix;
1192         }
1193
1194         if (!ent) {
1195                 nent = xmalloc(sizeof(*nent));
1196                 nent->temporary = (available_ix < 0);
1197         }
1198         else {
1199                 /* evict and reuse */
1200                 free(ent->tree_data);
1201                 nent = ent;
1202         }
1203         hashcpy(nent->sha1, sha1);
1204         nent->tree_data = data;
1205         nent->tree_size = size;
1206         nent->ref = 1;
1207         if (!nent->temporary)
1208                 pbase_tree_cache[my_ix] = nent;
1209         return nent;
1210 }
1211
1212 static void pbase_tree_put(struct pbase_tree_cache *cache)
1213 {
1214         if (!cache->temporary) {
1215                 cache->ref--;
1216                 return;
1217         }
1218         free(cache->tree_data);
1219         free(cache);
1220 }
1221
1222 static int name_cmp_len(const char *name)
1223 {
1224         int i;
1225         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1226                 ;
1227         return i;
1228 }
1229
1230 static void add_pbase_object(struct tree_desc *tree,
1231                              const char *name,
1232                              int cmplen,
1233                              const char *fullname)
1234 {
1235         struct name_entry entry;
1236         int cmp;
1237
1238         while (tree_entry(tree,&entry)) {
1239                 if (S_ISGITLINK(entry.mode))
1240                         continue;
1241                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1242                       memcmp(name, entry.path, cmplen);
1243                 if (cmp > 0)
1244                         continue;
1245                 if (cmp < 0)
1246                         return;
1247                 if (name[cmplen] != '/') {
1248                         add_object_entry(entry.oid->hash,
1249                                          object_type(entry.mode),
1250                                          fullname, 1);
1251                         return;
1252                 }
1253                 if (S_ISDIR(entry.mode)) {
1254                         struct tree_desc sub;
1255                         struct pbase_tree_cache *tree;
1256                         const char *down = name+cmplen+1;
1257                         int downlen = name_cmp_len(down);
1258
1259                         tree = pbase_tree_get(entry.oid->hash);
1260                         if (!tree)
1261                                 return;
1262                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1263
1264                         add_pbase_object(&sub, down, downlen, fullname);
1265                         pbase_tree_put(tree);
1266                 }
1267         }
1268 }
1269
1270 static unsigned *done_pbase_paths;
1271 static int done_pbase_paths_num;
1272 static int done_pbase_paths_alloc;
1273 static int done_pbase_path_pos(unsigned hash)
1274 {
1275         int lo = 0;
1276         int hi = done_pbase_paths_num;
1277         while (lo < hi) {
1278                 int mi = (hi + lo) / 2;
1279                 if (done_pbase_paths[mi] == hash)
1280                         return mi;
1281                 if (done_pbase_paths[mi] < hash)
1282                         hi = mi;
1283                 else
1284                         lo = mi + 1;
1285         }
1286         return -lo-1;
1287 }
1288
1289 static int check_pbase_path(unsigned hash)
1290 {
1291         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1292         if (0 <= pos)
1293                 return 1;
1294         pos = -pos - 1;
1295         ALLOC_GROW(done_pbase_paths,
1296                    done_pbase_paths_num + 1,
1297                    done_pbase_paths_alloc);
1298         done_pbase_paths_num++;
1299         if (pos < done_pbase_paths_num)
1300                 memmove(done_pbase_paths + pos + 1,
1301                         done_pbase_paths + pos,
1302                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1303         done_pbase_paths[pos] = hash;
1304         return 0;
1305 }
1306
1307 static void add_preferred_base_object(const char *name)
1308 {
1309         struct pbase_tree *it;
1310         int cmplen;
1311         unsigned hash = pack_name_hash(name);
1312
1313         if (!num_preferred_base || check_pbase_path(hash))
1314                 return;
1315
1316         cmplen = name_cmp_len(name);
1317         for (it = pbase_tree; it; it = it->next) {
1318                 if (cmplen == 0) {
1319                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1320                 }
1321                 else {
1322                         struct tree_desc tree;
1323                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1324                         add_pbase_object(&tree, name, cmplen, name);
1325                 }
1326         }
1327 }
1328
1329 static void add_preferred_base(unsigned char *sha1)
1330 {
1331         struct pbase_tree *it;
1332         void *data;
1333         unsigned long size;
1334         unsigned char tree_sha1[20];
1335
1336         if (window <= num_preferred_base++)
1337                 return;
1338
1339         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1340         if (!data)
1341                 return;
1342
1343         for (it = pbase_tree; it; it = it->next) {
1344                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1345                         free(data);
1346                         return;
1347                 }
1348         }
1349
1350         it = xcalloc(1, sizeof(*it));
1351         it->next = pbase_tree;
1352         pbase_tree = it;
1353
1354         hashcpy(it->pcache.sha1, tree_sha1);
1355         it->pcache.tree_data = data;
1356         it->pcache.tree_size = size;
1357 }
1358
1359 static void cleanup_preferred_base(void)
1360 {
1361         struct pbase_tree *it;
1362         unsigned i;
1363
1364         it = pbase_tree;
1365         pbase_tree = NULL;
1366         while (it) {
1367                 struct pbase_tree *this = it;
1368                 it = this->next;
1369                 free(this->pcache.tree_data);
1370                 free(this);
1371         }
1372
1373         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1374                 if (!pbase_tree_cache[i])
1375                         continue;
1376                 free(pbase_tree_cache[i]->tree_data);
1377                 free(pbase_tree_cache[i]);
1378                 pbase_tree_cache[i] = NULL;
1379         }
1380
1381         free(done_pbase_paths);
1382         done_pbase_paths = NULL;
1383         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1384 }
1385
1386 static void check_object(struct object_entry *entry)
1387 {
1388         if (entry->in_pack) {
1389                 struct packed_git *p = entry->in_pack;
1390                 struct pack_window *w_curs = NULL;
1391                 const unsigned char *base_ref = NULL;
1392                 struct object_entry *base_entry;
1393                 unsigned long used, used_0;
1394                 unsigned long avail;
1395                 off_t ofs;
1396                 unsigned char *buf, c;
1397
1398                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1399
1400                 /*
1401                  * We want in_pack_type even if we do not reuse delta
1402                  * since non-delta representations could still be reused.
1403                  */
1404                 used = unpack_object_header_buffer(buf, avail,
1405                                                    &entry->in_pack_type,
1406                                                    &entry->size);
1407                 if (used == 0)
1408                         goto give_up;
1409
1410                 /*
1411                  * Determine if this is a delta and if so whether we can
1412                  * reuse it or not.  Otherwise let's find out as cheaply as
1413                  * possible what the actual type and size for this object is.
1414                  */
1415                 switch (entry->in_pack_type) {
1416                 default:
1417                         /* Not a delta hence we've already got all we need. */
1418                         entry->type = entry->in_pack_type;
1419                         entry->in_pack_header_size = used;
1420                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1421                                 goto give_up;
1422                         unuse_pack(&w_curs);
1423                         return;
1424                 case OBJ_REF_DELTA:
1425                         if (reuse_delta && !entry->preferred_base)
1426                                 base_ref = use_pack(p, &w_curs,
1427                                                 entry->in_pack_offset + used, NULL);
1428                         entry->in_pack_header_size = used + 20;
1429                         break;
1430                 case OBJ_OFS_DELTA:
1431                         buf = use_pack(p, &w_curs,
1432                                        entry->in_pack_offset + used, NULL);
1433                         used_0 = 0;
1434                         c = buf[used_0++];
1435                         ofs = c & 127;
1436                         while (c & 128) {
1437                                 ofs += 1;
1438                                 if (!ofs || MSB(ofs, 7)) {
1439                                         error("delta base offset overflow in pack for %s",
1440                                               sha1_to_hex(entry->idx.sha1));
1441                                         goto give_up;
1442                                 }
1443                                 c = buf[used_0++];
1444                                 ofs = (ofs << 7) + (c & 127);
1445                         }
1446                         ofs = entry->in_pack_offset - ofs;
1447                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1448                                 error("delta base offset out of bound for %s",
1449                                       sha1_to_hex(entry->idx.sha1));
1450                                 goto give_up;
1451                         }
1452                         if (reuse_delta && !entry->preferred_base) {
1453                                 struct revindex_entry *revidx;
1454                                 revidx = find_pack_revindex(p, ofs);
1455                                 if (!revidx)
1456                                         goto give_up;
1457                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1458                         }
1459                         entry->in_pack_header_size = used + used_0;
1460                         break;
1461                 }
1462
1463                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1464                         /*
1465                          * If base_ref was set above that means we wish to
1466                          * reuse delta data, and we even found that base
1467                          * in the list of objects we want to pack. Goodie!
1468                          *
1469                          * Depth value does not matter - find_deltas() will
1470                          * never consider reused delta as the base object to
1471                          * deltify other objects against, in order to avoid
1472                          * circular deltas.
1473                          */
1474                         entry->type = entry->in_pack_type;
1475                         entry->delta = base_entry;
1476                         entry->delta_size = entry->size;
1477                         entry->delta_sibling = base_entry->delta_child;
1478                         base_entry->delta_child = entry;
1479                         unuse_pack(&w_curs);
1480                         return;
1481                 }
1482
1483                 if (entry->type) {
1484                         /*
1485                          * This must be a delta and we already know what the
1486                          * final object type is.  Let's extract the actual
1487                          * object size from the delta header.
1488                          */
1489                         entry->size = get_size_from_delta(p, &w_curs,
1490                                         entry->in_pack_offset + entry->in_pack_header_size);
1491                         if (entry->size == 0)
1492                                 goto give_up;
1493                         unuse_pack(&w_curs);
1494                         return;
1495                 }
1496
1497                 /*
1498                  * No choice but to fall back to the recursive delta walk
1499                  * with sha1_object_info() to find about the object type
1500                  * at this point...
1501                  */
1502                 give_up:
1503                 unuse_pack(&w_curs);
1504         }
1505
1506         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1507         /*
1508          * The error condition is checked in prepare_pack().  This is
1509          * to permit a missing preferred base object to be ignored
1510          * as a preferred base.  Doing so can result in a larger
1511          * pack file, but the transfer will still take place.
1512          */
1513 }
1514
1515 static int pack_offset_sort(const void *_a, const void *_b)
1516 {
1517         const struct object_entry *a = *(struct object_entry **)_a;
1518         const struct object_entry *b = *(struct object_entry **)_b;
1519
1520         /* avoid filesystem trashing with loose objects */
1521         if (!a->in_pack && !b->in_pack)
1522                 return hashcmp(a->idx.sha1, b->idx.sha1);
1523
1524         if (a->in_pack < b->in_pack)
1525                 return -1;
1526         if (a->in_pack > b->in_pack)
1527                 return 1;
1528         return a->in_pack_offset < b->in_pack_offset ? -1 :
1529                         (a->in_pack_offset > b->in_pack_offset);
1530 }
1531
1532 /*
1533  * Drop an on-disk delta we were planning to reuse. Naively, this would
1534  * just involve blanking out the "delta" field, but we have to deal
1535  * with some extra book-keeping:
1536  *
1537  *   1. Removing ourselves from the delta_sibling linked list.
1538  *
1539  *   2. Updating our size/type to the non-delta representation. These were
1540  *      either not recorded initially (size) or overwritten with the delta type
1541  *      (type) when check_object() decided to reuse the delta.
1542  *
1543  *   3. Resetting our delta depth, as we are now a base object.
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         entry->depth = 0;
1558
1559         oi.sizep = &entry->size;
1560         oi.typep = &entry->type;
1561         if (packed_object_info(entry->in_pack, entry->in_pack_offset, &oi) < 0) {
1562                 /*
1563                  * We failed to get the info from this pack for some reason;
1564                  * fall back to sha1_object_info, which may find another copy.
1565                  * And if that fails, the error will be recorded in entry->type
1566                  * and dealt with in prepare_pack().
1567                  */
1568                 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1569         }
1570 }
1571
1572 /*
1573  * Follow the chain of deltas from this entry onward, throwing away any links
1574  * that cause us to hit a cycle (as determined by the DFS state flags in
1575  * the entries).
1576  *
1577  * We also detect too-long reused chains that would violate our --depth
1578  * limit.
1579  */
1580 static void break_delta_chains(struct object_entry *entry)
1581 {
1582         /*
1583          * The actual depth of each object we will write is stored as an int,
1584          * as it cannot exceed our int "depth" limit. But before we break
1585          * changes based no that limit, we may potentially go as deep as the
1586          * number of objects, which is elsewhere bounded to a uint32_t.
1587          */
1588         uint32_t total_depth;
1589         struct object_entry *cur, *next;
1590
1591         for (cur = entry, total_depth = 0;
1592              cur;
1593              cur = cur->delta, total_depth++) {
1594                 if (cur->dfs_state == DFS_DONE) {
1595                         /*
1596                          * We've already seen this object and know it isn't
1597                          * part of a cycle. We do need to append its depth
1598                          * to our count.
1599                          */
1600                         total_depth += cur->depth;
1601                         break;
1602                 }
1603
1604                 /*
1605                  * We break cycles before looping, so an ACTIVE state (or any
1606                  * other cruft which made its way into the state variable)
1607                  * is a bug.
1608                  */
1609                 if (cur->dfs_state != DFS_NONE)
1610                         die("BUG: confusing delta dfs state in first pass: %d",
1611                             cur->dfs_state);
1612
1613                 /*
1614                  * Now we know this is the first time we've seen the object. If
1615                  * it's not a delta, we're done traversing, but we'll mark it
1616                  * done to save time on future traversals.
1617                  */
1618                 if (!cur->delta) {
1619                         cur->dfs_state = DFS_DONE;
1620                         break;
1621                 }
1622
1623                 /*
1624                  * Mark ourselves as active and see if the next step causes
1625                  * us to cycle to another active object. It's important to do
1626                  * this _before_ we loop, because it impacts where we make the
1627                  * cut, and thus how our total_depth counter works.
1628                  * E.g., We may see a partial loop like:
1629                  *
1630                  *   A -> B -> C -> D -> B
1631                  *
1632                  * Cutting B->C breaks the cycle. But now the depth of A is
1633                  * only 1, and our total_depth counter is at 3. The size of the
1634                  * error is always one less than the size of the cycle we
1635                  * broke. Commits C and D were "lost" from A's chain.
1636                  *
1637                  * If we instead cut D->B, then the depth of A is correct at 3.
1638                  * We keep all commits in the chain that we examined.
1639                  */
1640                 cur->dfs_state = DFS_ACTIVE;
1641                 if (cur->delta->dfs_state == DFS_ACTIVE) {
1642                         drop_reused_delta(cur);
1643                         cur->dfs_state = DFS_DONE;
1644                         break;
1645                 }
1646         }
1647
1648         /*
1649          * And now that we've gone all the way to the bottom of the chain, we
1650          * need to clear the active flags and set the depth fields as
1651          * appropriate. Unlike the loop above, which can quit when it drops a
1652          * delta, we need to keep going to look for more depth cuts. So we need
1653          * an extra "next" pointer to keep going after we reset cur->delta.
1654          */
1655         for (cur = entry; cur; cur = next) {
1656                 next = cur->delta;
1657
1658                 /*
1659                  * We should have a chain of zero or more ACTIVE states down to
1660                  * a final DONE. We can quit after the DONE, because either it
1661                  * has no bases, or we've already handled them in a previous
1662                  * call.
1663                  */
1664                 if (cur->dfs_state == DFS_DONE)
1665                         break;
1666                 else if (cur->dfs_state != DFS_ACTIVE)
1667                         die("BUG: confusing delta dfs state in second pass: %d",
1668                             cur->dfs_state);
1669
1670                 /*
1671                  * If the total_depth is more than depth, then we need to snip
1672                  * the chain into two or more smaller chains that don't exceed
1673                  * the maximum depth. Most of the resulting chains will contain
1674                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1675                  * the last chain (i.e., the one containing entry) will contain
1676                  * whatever entries are left over, namely
1677                  * (total_depth % (depth + 1)) of them.
1678                  *
1679                  * Since we are iterating towards decreasing depth, we need to
1680                  * decrement total_depth as we go, and we need to write to the
1681                  * entry what its final depth will be after all of the
1682                  * snipping. Since we're snipping into chains of length (depth
1683                  * + 1) entries, the final depth of an entry will be its
1684                  * original depth modulo (depth + 1). Any time we encounter an
1685                  * entry whose final depth is supposed to be zero, we snip it
1686                  * from its delta base, thereby making it so.
1687                  */
1688                 cur->depth = (total_depth--) % (depth + 1);
1689                 if (!cur->depth)
1690                         drop_reused_delta(cur);
1691
1692                 cur->dfs_state = DFS_DONE;
1693         }
1694 }
1695
1696 static void get_object_details(void)
1697 {
1698         uint32_t i;
1699         struct object_entry **sorted_by_offset;
1700
1701         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1702         for (i = 0; i < to_pack.nr_objects; i++)
1703                 sorted_by_offset[i] = to_pack.objects + i;
1704         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1705
1706         for (i = 0; i < to_pack.nr_objects; i++) {
1707                 struct object_entry *entry = sorted_by_offset[i];
1708                 check_object(entry);
1709                 if (big_file_threshold < entry->size)
1710                         entry->no_try_delta = 1;
1711         }
1712
1713         /*
1714          * This must happen in a second pass, since we rely on the delta
1715          * information for the whole list being completed.
1716          */
1717         for (i = 0; i < to_pack.nr_objects; i++)
1718                 break_delta_chains(&to_pack.objects[i]);
1719
1720         free(sorted_by_offset);
1721 }
1722
1723 /*
1724  * We search for deltas in a list sorted by type, by filename hash, and then
1725  * by size, so that we see progressively smaller and smaller files.
1726  * That's because we prefer deltas to be from the bigger file
1727  * to the smaller -- deletes are potentially cheaper, but perhaps
1728  * more importantly, the bigger file is likely the more recent
1729  * one.  The deepest deltas are therefore the oldest objects which are
1730  * less susceptible to be accessed often.
1731  */
1732 static int type_size_sort(const void *_a, const void *_b)
1733 {
1734         const struct object_entry *a = *(struct object_entry **)_a;
1735         const struct object_entry *b = *(struct object_entry **)_b;
1736
1737         if (a->type > b->type)
1738                 return -1;
1739         if (a->type < b->type)
1740                 return 1;
1741         if (a->hash > b->hash)
1742                 return -1;
1743         if (a->hash < b->hash)
1744                 return 1;
1745         if (a->preferred_base > b->preferred_base)
1746                 return -1;
1747         if (a->preferred_base < b->preferred_base)
1748                 return 1;
1749         if (a->size > b->size)
1750                 return -1;
1751         if (a->size < b->size)
1752                 return 1;
1753         return a < b ? -1 : (a > b);  /* newest first */
1754 }
1755
1756 struct unpacked {
1757         struct object_entry *entry;
1758         void *data;
1759         struct delta_index *index;
1760         unsigned depth;
1761 };
1762
1763 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1764                            unsigned long delta_size)
1765 {
1766         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1767                 return 0;
1768
1769         if (delta_size < cache_max_small_delta_size)
1770                 return 1;
1771
1772         /* cache delta, if objects are large enough compared to delta size */
1773         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1774                 return 1;
1775
1776         return 0;
1777 }
1778
1779 #ifndef NO_PTHREADS
1780
1781 static pthread_mutex_t read_mutex;
1782 #define read_lock()             pthread_mutex_lock(&read_mutex)
1783 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1784
1785 static pthread_mutex_t cache_mutex;
1786 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1787 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1788
1789 static pthread_mutex_t progress_mutex;
1790 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1791 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1792
1793 #else
1794
1795 #define read_lock()             (void)0
1796 #define read_unlock()           (void)0
1797 #define cache_lock()            (void)0
1798 #define cache_unlock()          (void)0
1799 #define progress_lock()         (void)0
1800 #define progress_unlock()       (void)0
1801
1802 #endif
1803
1804 static int try_delta(struct unpacked *trg, struct unpacked *src,
1805                      unsigned max_depth, unsigned long *mem_usage)
1806 {
1807         struct object_entry *trg_entry = trg->entry;
1808         struct object_entry *src_entry = src->entry;
1809         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1810         unsigned ref_depth;
1811         enum object_type type;
1812         void *delta_buf;
1813
1814         /* Don't bother doing diffs between different types */
1815         if (trg_entry->type != src_entry->type)
1816                 return -1;
1817
1818         /*
1819          * We do not bother to try a delta that we discarded on an
1820          * earlier try, but only when reusing delta data.  Note that
1821          * src_entry that is marked as the preferred_base should always
1822          * be considered, as even if we produce a suboptimal delta against
1823          * it, we will still save the transfer cost, as we already know
1824          * the other side has it and we won't send src_entry at all.
1825          */
1826         if (reuse_delta && trg_entry->in_pack &&
1827             trg_entry->in_pack == src_entry->in_pack &&
1828             !src_entry->preferred_base &&
1829             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1830             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1831                 return 0;
1832
1833         /* Let's not bust the allowed depth. */
1834         if (src->depth >= max_depth)
1835                 return 0;
1836
1837         /* Now some size filtering heuristics. */
1838         trg_size = trg_entry->size;
1839         if (!trg_entry->delta) {
1840                 max_size = trg_size/2 - 20;
1841                 ref_depth = 1;
1842         } else {
1843                 max_size = trg_entry->delta_size;
1844                 ref_depth = trg->depth;
1845         }
1846         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1847                                                 (max_depth - ref_depth + 1);
1848         if (max_size == 0)
1849                 return 0;
1850         src_size = src_entry->size;
1851         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1852         if (sizediff >= max_size)
1853                 return 0;
1854         if (trg_size < src_size / 32)
1855                 return 0;
1856
1857         /* Load data if not already done */
1858         if (!trg->data) {
1859                 read_lock();
1860                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1861                 read_unlock();
1862                 if (!trg->data)
1863                         die("object %s cannot be read",
1864                             sha1_to_hex(trg_entry->idx.sha1));
1865                 if (sz != trg_size)
1866                         die("object %s inconsistent object length (%lu vs %lu)",
1867                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1868                 *mem_usage += sz;
1869         }
1870         if (!src->data) {
1871                 read_lock();
1872                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1873                 read_unlock();
1874                 if (!src->data) {
1875                         if (src_entry->preferred_base) {
1876                                 static int warned = 0;
1877                                 if (!warned++)
1878                                         warning("object %s cannot be read",
1879                                                 sha1_to_hex(src_entry->idx.sha1));
1880                                 /*
1881                                  * Those objects are not included in the
1882                                  * resulting pack.  Be resilient and ignore
1883                                  * them if they can't be read, in case the
1884                                  * pack could be created nevertheless.
1885                                  */
1886                                 return 0;
1887                         }
1888                         die("object %s cannot be read",
1889                             sha1_to_hex(src_entry->idx.sha1));
1890                 }
1891                 if (sz != src_size)
1892                         die("object %s inconsistent object length (%lu vs %lu)",
1893                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1894                 *mem_usage += sz;
1895         }
1896         if (!src->index) {
1897                 src->index = create_delta_index(src->data, src_size);
1898                 if (!src->index) {
1899                         static int warned = 0;
1900                         if (!warned++)
1901                                 warning("suboptimal pack - out of memory");
1902                         return 0;
1903                 }
1904                 *mem_usage += sizeof_delta_index(src->index);
1905         }
1906
1907         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1908         if (!delta_buf)
1909                 return 0;
1910
1911         if (trg_entry->delta) {
1912                 /* Prefer only shallower same-sized deltas. */
1913                 if (delta_size == trg_entry->delta_size &&
1914                     src->depth + 1 >= trg->depth) {
1915                         free(delta_buf);
1916                         return 0;
1917                 }
1918         }
1919
1920         /*
1921          * Handle memory allocation outside of the cache
1922          * accounting lock.  Compiler will optimize the strangeness
1923          * away when NO_PTHREADS is defined.
1924          */
1925         free(trg_entry->delta_data);
1926         cache_lock();
1927         if (trg_entry->delta_data) {
1928                 delta_cache_size -= trg_entry->delta_size;
1929                 trg_entry->delta_data = NULL;
1930         }
1931         if (delta_cacheable(src_size, trg_size, delta_size)) {
1932                 delta_cache_size += delta_size;
1933                 cache_unlock();
1934                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1935         } else {
1936                 cache_unlock();
1937                 free(delta_buf);
1938         }
1939
1940         trg_entry->delta = src_entry;
1941         trg_entry->delta_size = delta_size;
1942         trg->depth = src->depth + 1;
1943
1944         return 1;
1945 }
1946
1947 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1948 {
1949         struct object_entry *child = me->delta_child;
1950         unsigned int m = n;
1951         while (child) {
1952                 unsigned int c = check_delta_limit(child, n + 1);
1953                 if (m < c)
1954                         m = c;
1955                 child = child->delta_sibling;
1956         }
1957         return m;
1958 }
1959
1960 static unsigned long free_unpacked(struct unpacked *n)
1961 {
1962         unsigned long freed_mem = sizeof_delta_index(n->index);
1963         free_delta_index(n->index);
1964         n->index = NULL;
1965         if (n->data) {
1966                 freed_mem += n->entry->size;
1967                 free(n->data);
1968                 n->data = NULL;
1969         }
1970         n->entry = NULL;
1971         n->depth = 0;
1972         return freed_mem;
1973 }
1974
1975 static void find_deltas(struct object_entry **list, unsigned *list_size,
1976                         int window, int depth, unsigned *processed)
1977 {
1978         uint32_t i, idx = 0, count = 0;
1979         struct unpacked *array;
1980         unsigned long mem_usage = 0;
1981
1982         array = xcalloc(window, sizeof(struct unpacked));
1983
1984         for (;;) {
1985                 struct object_entry *entry;
1986                 struct unpacked *n = array + idx;
1987                 int j, max_depth, best_base = -1;
1988
1989                 progress_lock();
1990                 if (!*list_size) {
1991                         progress_unlock();
1992                         break;
1993                 }
1994                 entry = *list++;
1995                 (*list_size)--;
1996                 if (!entry->preferred_base) {
1997                         (*processed)++;
1998                         display_progress(progress_state, *processed);
1999                 }
2000                 progress_unlock();
2001
2002                 mem_usage -= free_unpacked(n);
2003                 n->entry = entry;
2004
2005                 while (window_memory_limit &&
2006                        mem_usage > window_memory_limit &&
2007                        count > 1) {
2008                         uint32_t tail = (idx + window - count) % window;
2009                         mem_usage -= free_unpacked(array + tail);
2010                         count--;
2011                 }
2012
2013                 /* We do not compute delta to *create* objects we are not
2014                  * going to pack.
2015                  */
2016                 if (entry->preferred_base)
2017                         goto next;
2018
2019                 /*
2020                  * If the current object is at pack edge, take the depth the
2021                  * objects that depend on the current object into account
2022                  * otherwise they would become too deep.
2023                  */
2024                 max_depth = depth;
2025                 if (entry->delta_child) {
2026                         max_depth -= check_delta_limit(entry, 0);
2027                         if (max_depth <= 0)
2028                                 goto next;
2029                 }
2030
2031                 j = window;
2032                 while (--j > 0) {
2033                         int ret;
2034                         uint32_t other_idx = idx + j;
2035                         struct unpacked *m;
2036                         if (other_idx >= window)
2037                                 other_idx -= window;
2038                         m = array + other_idx;
2039                         if (!m->entry)
2040                                 break;
2041                         ret = try_delta(n, m, max_depth, &mem_usage);
2042                         if (ret < 0)
2043                                 break;
2044                         else if (ret > 0)
2045                                 best_base = other_idx;
2046                 }
2047
2048                 /*
2049                  * If we decided to cache the delta data, then it is best
2050                  * to compress it right away.  First because we have to do
2051                  * it anyway, and doing it here while we're threaded will
2052                  * save a lot of time in the non threaded write phase,
2053                  * as well as allow for caching more deltas within
2054                  * the same cache size limit.
2055                  * ...
2056                  * But only if not writing to stdout, since in that case
2057                  * the network is most likely throttling writes anyway,
2058                  * and therefore it is best to go to the write phase ASAP
2059                  * instead, as we can afford spending more time compressing
2060                  * between writes at that moment.
2061                  */
2062                 if (entry->delta_data && !pack_to_stdout) {
2063                         entry->z_delta_size = do_compress(&entry->delta_data,
2064                                                           entry->delta_size);
2065                         cache_lock();
2066                         delta_cache_size -= entry->delta_size;
2067                         delta_cache_size += entry->z_delta_size;
2068                         cache_unlock();
2069                 }
2070
2071                 /* if we made n a delta, and if n is already at max
2072                  * depth, leaving it in the window is pointless.  we
2073                  * should evict it first.
2074                  */
2075                 if (entry->delta && max_depth <= n->depth)
2076                         continue;
2077
2078                 /*
2079                  * Move the best delta base up in the window, after the
2080                  * currently deltified object, to keep it longer.  It will
2081                  * be the first base object to be attempted next.
2082                  */
2083                 if (entry->delta) {
2084                         struct unpacked swap = array[best_base];
2085                         int dist = (window + idx - best_base) % window;
2086                         int dst = best_base;
2087                         while (dist--) {
2088                                 int src = (dst + 1) % window;
2089                                 array[dst] = array[src];
2090                                 dst = src;
2091                         }
2092                         array[dst] = swap;
2093                 }
2094
2095                 next:
2096                 idx++;
2097                 if (count + 1 < window)
2098                         count++;
2099                 if (idx >= window)
2100                         idx = 0;
2101         }
2102
2103         for (i = 0; i < window; ++i) {
2104                 free_delta_index(array[i].index);
2105                 free(array[i].data);
2106         }
2107         free(array);
2108 }
2109
2110 #ifndef NO_PTHREADS
2111
2112 static void try_to_free_from_threads(size_t size)
2113 {
2114         read_lock();
2115         release_pack_memory(size);
2116         read_unlock();
2117 }
2118
2119 static try_to_free_t old_try_to_free_routine;
2120
2121 /*
2122  * The main thread waits on the condition that (at least) one of the workers
2123  * has stopped working (which is indicated in the .working member of
2124  * struct thread_params).
2125  * When a work thread has completed its work, it sets .working to 0 and
2126  * signals the main thread and waits on the condition that .data_ready
2127  * becomes 1.
2128  */
2129
2130 struct thread_params {
2131         pthread_t thread;
2132         struct object_entry **list;
2133         unsigned list_size;
2134         unsigned remaining;
2135         int window;
2136         int depth;
2137         int working;
2138         int data_ready;
2139         pthread_mutex_t mutex;
2140         pthread_cond_t cond;
2141         unsigned *processed;
2142 };
2143
2144 static pthread_cond_t progress_cond;
2145
2146 /*
2147  * Mutex and conditional variable can't be statically-initialized on Windows.
2148  */
2149 static void init_threaded_search(void)
2150 {
2151         init_recursive_mutex(&read_mutex);
2152         pthread_mutex_init(&cache_mutex, NULL);
2153         pthread_mutex_init(&progress_mutex, NULL);
2154         pthread_cond_init(&progress_cond, NULL);
2155         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2156 }
2157
2158 static void cleanup_threaded_search(void)
2159 {
2160         set_try_to_free_routine(old_try_to_free_routine);
2161         pthread_cond_destroy(&progress_cond);
2162         pthread_mutex_destroy(&read_mutex);
2163         pthread_mutex_destroy(&cache_mutex);
2164         pthread_mutex_destroy(&progress_mutex);
2165 }
2166
2167 static void *threaded_find_deltas(void *arg)
2168 {
2169         struct thread_params *me = arg;
2170
2171         while (me->remaining) {
2172                 find_deltas(me->list, &me->remaining,
2173                             me->window, me->depth, me->processed);
2174
2175                 progress_lock();
2176                 me->working = 0;
2177                 pthread_cond_signal(&progress_cond);
2178                 progress_unlock();
2179
2180                 /*
2181                  * We must not set ->data_ready before we wait on the
2182                  * condition because the main thread may have set it to 1
2183                  * before we get here. In order to be sure that new
2184                  * work is available if we see 1 in ->data_ready, it
2185                  * was initialized to 0 before this thread was spawned
2186                  * and we reset it to 0 right away.
2187                  */
2188                 pthread_mutex_lock(&me->mutex);
2189                 while (!me->data_ready)
2190                         pthread_cond_wait(&me->cond, &me->mutex);
2191                 me->data_ready = 0;
2192                 pthread_mutex_unlock(&me->mutex);
2193         }
2194         /* leave ->working 1 so that this doesn't get more work assigned */
2195         return NULL;
2196 }
2197
2198 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2199                            int window, int depth, unsigned *processed)
2200 {
2201         struct thread_params *p;
2202         int i, ret, active_threads = 0;
2203
2204         init_threaded_search();
2205
2206         if (delta_search_threads <= 1) {
2207                 find_deltas(list, &list_size, window, depth, processed);
2208                 cleanup_threaded_search();
2209                 return;
2210         }
2211         if (progress > pack_to_stdout)
2212                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2213                                 delta_search_threads);
2214         p = xcalloc(delta_search_threads, sizeof(*p));
2215
2216         /* Partition the work amongst work threads. */
2217         for (i = 0; i < delta_search_threads; i++) {
2218                 unsigned sub_size = list_size / (delta_search_threads - i);
2219
2220                 /* don't use too small segments or no deltas will be found */
2221                 if (sub_size < 2*window && i+1 < delta_search_threads)
2222                         sub_size = 0;
2223
2224                 p[i].window = window;
2225                 p[i].depth = depth;
2226                 p[i].processed = processed;
2227                 p[i].working = 1;
2228                 p[i].data_ready = 0;
2229
2230                 /* try to split chunks on "path" boundaries */
2231                 while (sub_size && sub_size < list_size &&
2232                        list[sub_size]->hash &&
2233                        list[sub_size]->hash == list[sub_size-1]->hash)
2234                         sub_size++;
2235
2236                 p[i].list = list;
2237                 p[i].list_size = sub_size;
2238                 p[i].remaining = sub_size;
2239
2240                 list += sub_size;
2241                 list_size -= sub_size;
2242         }
2243
2244         /* Start work threads. */
2245         for (i = 0; i < delta_search_threads; i++) {
2246                 if (!p[i].list_size)
2247                         continue;
2248                 pthread_mutex_init(&p[i].mutex, NULL);
2249                 pthread_cond_init(&p[i].cond, NULL);
2250                 ret = pthread_create(&p[i].thread, NULL,
2251                                      threaded_find_deltas, &p[i]);
2252                 if (ret)
2253                         die("unable to create thread: %s", strerror(ret));
2254                 active_threads++;
2255         }
2256
2257         /*
2258          * Now let's wait for work completion.  Each time a thread is done
2259          * with its work, we steal half of the remaining work from the
2260          * thread with the largest number of unprocessed objects and give
2261          * it to that newly idle thread.  This ensure good load balancing
2262          * until the remaining object list segments are simply too short
2263          * to be worth splitting anymore.
2264          */
2265         while (active_threads) {
2266                 struct thread_params *target = NULL;
2267                 struct thread_params *victim = NULL;
2268                 unsigned sub_size = 0;
2269
2270                 progress_lock();
2271                 for (;;) {
2272                         for (i = 0; !target && i < delta_search_threads; i++)
2273                                 if (!p[i].working)
2274                                         target = &p[i];
2275                         if (target)
2276                                 break;
2277                         pthread_cond_wait(&progress_cond, &progress_mutex);
2278                 }
2279
2280                 for (i = 0; i < delta_search_threads; i++)
2281                         if (p[i].remaining > 2*window &&
2282                             (!victim || victim->remaining < p[i].remaining))
2283                                 victim = &p[i];
2284                 if (victim) {
2285                         sub_size = victim->remaining / 2;
2286                         list = victim->list + victim->list_size - sub_size;
2287                         while (sub_size && list[0]->hash &&
2288                                list[0]->hash == list[-1]->hash) {
2289                                 list++;
2290                                 sub_size--;
2291                         }
2292                         if (!sub_size) {
2293                                 /*
2294                                  * It is possible for some "paths" to have
2295                                  * so many objects that no hash boundary
2296                                  * might be found.  Let's just steal the
2297                                  * exact half in that case.
2298                                  */
2299                                 sub_size = victim->remaining / 2;
2300                                 list -= sub_size;
2301                         }
2302                         target->list = list;
2303                         victim->list_size -= sub_size;
2304                         victim->remaining -= sub_size;
2305                 }
2306                 target->list_size = sub_size;
2307                 target->remaining = sub_size;
2308                 target->working = 1;
2309                 progress_unlock();
2310
2311                 pthread_mutex_lock(&target->mutex);
2312                 target->data_ready = 1;
2313                 pthread_cond_signal(&target->cond);
2314                 pthread_mutex_unlock(&target->mutex);
2315
2316                 if (!sub_size) {
2317                         pthread_join(target->thread, NULL);
2318                         pthread_cond_destroy(&target->cond);
2319                         pthread_mutex_destroy(&target->mutex);
2320                         active_threads--;
2321                 }
2322         }
2323         cleanup_threaded_search();
2324         free(p);
2325 }
2326
2327 #else
2328 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2329 #endif
2330
2331 static void add_tag_chain(const struct object_id *oid)
2332 {
2333         struct tag *tag;
2334
2335         /*
2336          * We catch duplicates already in add_object_entry(), but we'd
2337          * prefer to do this extra check to avoid having to parse the
2338          * tag at all if we already know that it's being packed (e.g., if
2339          * it was included via bitmaps, we would not have parsed it
2340          * previously).
2341          */
2342         if (packlist_find(&to_pack, oid->hash, NULL))
2343                 return;
2344
2345         tag = lookup_tag(oid->hash);
2346         while (1) {
2347                 if (!tag || parse_tag(tag) || !tag->tagged)
2348                         die("unable to pack objects reachable from tag %s",
2349                             oid_to_hex(oid));
2350
2351                 add_object_entry(tag->object.oid.hash, OBJ_TAG, NULL, 0);
2352
2353                 if (tag->tagged->type != OBJ_TAG)
2354                         return;
2355
2356                 tag = (struct tag *)tag->tagged;
2357         }
2358 }
2359
2360 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2361 {
2362         struct object_id peeled;
2363
2364         if (starts_with(path, "refs/tags/") && /* is a tag? */
2365             !peel_ref(path, peeled.hash)    && /* peelable? */
2366             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2367                 add_tag_chain(oid);
2368         return 0;
2369 }
2370
2371 static void prepare_pack(int window, int depth)
2372 {
2373         struct object_entry **delta_list;
2374         uint32_t i, nr_deltas;
2375         unsigned n;
2376
2377         get_object_details();
2378
2379         /*
2380          * If we're locally repacking then we need to be doubly careful
2381          * from now on in order to make sure no stealth corruption gets
2382          * propagated to the new pack.  Clients receiving streamed packs
2383          * should validate everything they get anyway so no need to incur
2384          * the additional cost here in that case.
2385          */
2386         if (!pack_to_stdout)
2387                 do_check_packed_object_crc = 1;
2388
2389         if (!to_pack.nr_objects || !window || !depth)
2390                 return;
2391
2392         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2393         nr_deltas = n = 0;
2394
2395         for (i = 0; i < to_pack.nr_objects; i++) {
2396                 struct object_entry *entry = to_pack.objects + i;
2397
2398                 if (entry->delta)
2399                         /* This happens if we decided to reuse existing
2400                          * delta from a pack.  "reuse_delta &&" is implied.
2401                          */
2402                         continue;
2403
2404                 if (entry->size < 50)
2405                         continue;
2406
2407                 if (entry->no_try_delta)
2408                         continue;
2409
2410                 if (!entry->preferred_base) {
2411                         nr_deltas++;
2412                         if (entry->type < 0)
2413                                 die("unable to get type of object %s",
2414                                     sha1_to_hex(entry->idx.sha1));
2415                 } else {
2416                         if (entry->type < 0) {
2417                                 /*
2418                                  * This object is not found, but we
2419                                  * don't have to include it anyway.
2420                                  */
2421                                 continue;
2422                         }
2423                 }
2424
2425                 delta_list[n++] = entry;
2426         }
2427
2428         if (nr_deltas && n > 1) {
2429                 unsigned nr_done = 0;
2430                 if (progress)
2431                         progress_state = start_progress(_("Compressing objects"),
2432                                                         nr_deltas);
2433                 QSORT(delta_list, n, type_size_sort);
2434                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2435                 stop_progress(&progress_state);
2436                 if (nr_done != nr_deltas)
2437                         die("inconsistency with delta count");
2438         }
2439         free(delta_list);
2440 }
2441
2442 static int git_pack_config(const char *k, const char *v, void *cb)
2443 {
2444         if (!strcmp(k, "pack.window")) {
2445                 window = git_config_int(k, v);
2446                 return 0;
2447         }
2448         if (!strcmp(k, "pack.windowmemory")) {
2449                 window_memory_limit = git_config_ulong(k, v);
2450                 return 0;
2451         }
2452         if (!strcmp(k, "pack.depth")) {
2453                 depth = git_config_int(k, v);
2454                 return 0;
2455         }
2456         if (!strcmp(k, "pack.deltacachesize")) {
2457                 max_delta_cache_size = git_config_int(k, v);
2458                 return 0;
2459         }
2460         if (!strcmp(k, "pack.deltacachelimit")) {
2461                 cache_max_small_delta_size = git_config_int(k, v);
2462                 return 0;
2463         }
2464         if (!strcmp(k, "pack.writebitmaphashcache")) {
2465                 if (git_config_bool(k, v))
2466                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2467                 else
2468                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2469         }
2470         if (!strcmp(k, "pack.usebitmaps")) {
2471                 use_bitmap_index_default = git_config_bool(k, v);
2472                 return 0;
2473         }
2474         if (!strcmp(k, "pack.threads")) {
2475                 delta_search_threads = git_config_int(k, v);
2476                 if (delta_search_threads < 0)
2477                         die("invalid number of threads specified (%d)",
2478                             delta_search_threads);
2479 #ifdef NO_PTHREADS
2480                 if (delta_search_threads != 1)
2481                         warning("no threads support, ignoring %s", k);
2482 #endif
2483                 return 0;
2484         }
2485         if (!strcmp(k, "pack.indexversion")) {
2486                 pack_idx_opts.version = git_config_int(k, v);
2487                 if (pack_idx_opts.version > 2)
2488                         die("bad pack.indexversion=%"PRIu32,
2489                             pack_idx_opts.version);
2490                 return 0;
2491         }
2492         return git_default_config(k, v, cb);
2493 }
2494
2495 static void read_object_list_from_stdin(void)
2496 {
2497         char line[40 + 1 + PATH_MAX + 2];
2498         unsigned char sha1[20];
2499
2500         for (;;) {
2501                 if (!fgets(line, sizeof(line), stdin)) {
2502                         if (feof(stdin))
2503                                 break;
2504                         if (!ferror(stdin))
2505                                 die("fgets returned NULL, not EOF, not error!");
2506                         if (errno != EINTR)
2507                                 die_errno("fgets");
2508                         clearerr(stdin);
2509                         continue;
2510                 }
2511                 if (line[0] == '-') {
2512                         if (get_sha1_hex(line+1, sha1))
2513                                 die("expected edge sha1, got garbage:\n %s",
2514                                     line);
2515                         add_preferred_base(sha1);
2516                         continue;
2517                 }
2518                 if (get_sha1_hex(line, sha1))
2519                         die("expected sha1, got garbage:\n %s", line);
2520
2521                 add_preferred_base_object(line+41);
2522                 add_object_entry(sha1, 0, line+41, 0);
2523         }
2524 }
2525
2526 #define OBJECT_ADDED (1u<<20)
2527
2528 static void show_commit(struct commit *commit, void *data)
2529 {
2530         add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2531         commit->object.flags |= OBJECT_ADDED;
2532
2533         if (write_bitmap_index)
2534                 index_commit_for_bitmap(commit);
2535 }
2536
2537 static void show_object(struct object *obj, const char *name, void *data)
2538 {
2539         add_preferred_base_object(name);
2540         add_object_entry(obj->oid.hash, obj->type, name, 0);
2541         obj->flags |= OBJECT_ADDED;
2542 }
2543
2544 static void show_edge(struct commit *commit)
2545 {
2546         add_preferred_base(commit->object.oid.hash);
2547 }
2548
2549 struct in_pack_object {
2550         off_t offset;
2551         struct object *object;
2552 };
2553
2554 struct in_pack {
2555         int alloc;
2556         int nr;
2557         struct in_pack_object *array;
2558 };
2559
2560 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2561 {
2562         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2563         in_pack->array[in_pack->nr].object = object;
2564         in_pack->nr++;
2565 }
2566
2567 /*
2568  * Compare the objects in the offset order, in order to emulate the
2569  * "git rev-list --objects" output that produced the pack originally.
2570  */
2571 static int ofscmp(const void *a_, const void *b_)
2572 {
2573         struct in_pack_object *a = (struct in_pack_object *)a_;
2574         struct in_pack_object *b = (struct in_pack_object *)b_;
2575
2576         if (a->offset < b->offset)
2577                 return -1;
2578         else if (a->offset > b->offset)
2579                 return 1;
2580         else
2581                 return oidcmp(&a->object->oid, &b->object->oid);
2582 }
2583
2584 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2585 {
2586         struct packed_git *p;
2587         struct in_pack in_pack;
2588         uint32_t i;
2589
2590         memset(&in_pack, 0, sizeof(in_pack));
2591
2592         for (p = packed_git; p; p = p->next) {
2593                 const unsigned char *sha1;
2594                 struct object *o;
2595
2596                 if (!p->pack_local || p->pack_keep)
2597                         continue;
2598                 if (open_pack_index(p))
2599                         die("cannot open pack index");
2600
2601                 ALLOC_GROW(in_pack.array,
2602                            in_pack.nr + p->num_objects,
2603                            in_pack.alloc);
2604
2605                 for (i = 0; i < p->num_objects; i++) {
2606                         sha1 = nth_packed_object_sha1(p, i);
2607                         o = lookup_unknown_object(sha1);
2608                         if (!(o->flags & OBJECT_ADDED))
2609                                 mark_in_pack_object(o, p, &in_pack);
2610                         o->flags |= OBJECT_ADDED;
2611                 }
2612         }
2613
2614         if (in_pack.nr) {
2615                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2616                 for (i = 0; i < in_pack.nr; i++) {
2617                         struct object *o = in_pack.array[i].object;
2618                         add_object_entry(o->oid.hash, o->type, "", 0);
2619                 }
2620         }
2621         free(in_pack.array);
2622 }
2623
2624 static int add_loose_object(const unsigned char *sha1, const char *path,
2625                             void *data)
2626 {
2627         enum object_type type = sha1_object_info(sha1, NULL);
2628
2629         if (type < 0) {
2630                 warning("loose object at %s could not be examined", path);
2631                 return 0;
2632         }
2633
2634         add_object_entry(sha1, type, "", 0);
2635         return 0;
2636 }
2637
2638 /*
2639  * We actually don't even have to worry about reachability here.
2640  * add_object_entry will weed out duplicates, so we just add every
2641  * loose object we find.
2642  */
2643 static void add_unreachable_loose_objects(void)
2644 {
2645         for_each_loose_file_in_objdir(get_object_directory(),
2646                                       add_loose_object,
2647                                       NULL, NULL, NULL);
2648 }
2649
2650 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2651 {
2652         static struct packed_git *last_found = (void *)1;
2653         struct packed_git *p;
2654
2655         p = (last_found != (void *)1) ? last_found : packed_git;
2656
2657         while (p) {
2658                 if ((!p->pack_local || p->pack_keep) &&
2659                         find_pack_entry_one(sha1, p)) {
2660                         last_found = p;
2661                         return 1;
2662                 }
2663                 if (p == last_found)
2664                         p = packed_git;
2665                 else
2666                         p = p->next;
2667                 if (p == last_found)
2668                         p = p->next;
2669         }
2670         return 0;
2671 }
2672
2673 /*
2674  * Store a list of sha1s that are should not be discarded
2675  * because they are either written too recently, or are
2676  * reachable from another object that was.
2677  *
2678  * This is filled by get_object_list.
2679  */
2680 static struct sha1_array recent_objects;
2681
2682 static int loosened_object_can_be_discarded(const unsigned char *sha1,
2683                                             unsigned long mtime)
2684 {
2685         if (!unpack_unreachable_expiration)
2686                 return 0;
2687         if (mtime > unpack_unreachable_expiration)
2688                 return 0;
2689         if (sha1_array_lookup(&recent_objects, sha1) >= 0)
2690                 return 0;
2691         return 1;
2692 }
2693
2694 static void loosen_unused_packed_objects(struct rev_info *revs)
2695 {
2696         struct packed_git *p;
2697         uint32_t i;
2698         const unsigned char *sha1;
2699
2700         for (p = packed_git; p; p = p->next) {
2701                 if (!p->pack_local || p->pack_keep)
2702                         continue;
2703
2704                 if (open_pack_index(p))
2705                         die("cannot open pack index");
2706
2707                 for (i = 0; i < p->num_objects; i++) {
2708                         sha1 = nth_packed_object_sha1(p, i);
2709                         if (!packlist_find(&to_pack, sha1, NULL) &&
2710                             !has_sha1_pack_kept_or_nonlocal(sha1) &&
2711                             !loosened_object_can_be_discarded(sha1, p->mtime))
2712                                 if (force_object_loose(sha1, p->mtime))
2713                                         die("unable to force loose object");
2714                 }
2715         }
2716 }
2717
2718 /*
2719  * This tracks any options which pack-reuse code expects to be on, or which a
2720  * reader of the pack might not understand, and which would therefore prevent
2721  * blind reuse of what we have on disk.
2722  */
2723 static int pack_options_allow_reuse(void)
2724 {
2725         return pack_to_stdout && allow_ofs_delta;
2726 }
2727
2728 static int get_object_list_from_bitmap(struct rev_info *revs)
2729 {
2730         if (prepare_bitmap_walk(revs) < 0)
2731                 return -1;
2732
2733         if (pack_options_allow_reuse() &&
2734             !reuse_partial_packfile_from_bitmap(
2735                         &reuse_packfile,
2736                         &reuse_packfile_objects,
2737                         &reuse_packfile_offset)) {
2738                 assert(reuse_packfile_objects);
2739                 nr_result += reuse_packfile_objects;
2740                 display_progress(progress_state, nr_result);
2741         }
2742
2743         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2744         return 0;
2745 }
2746
2747 static void record_recent_object(struct object *obj,
2748                                  const char *name,
2749                                  void *data)
2750 {
2751         sha1_array_append(&recent_objects, obj->oid.hash);
2752 }
2753
2754 static void record_recent_commit(struct commit *commit, void *data)
2755 {
2756         sha1_array_append(&recent_objects, commit->object.oid.hash);
2757 }
2758
2759 static void get_object_list(int ac, const char **av)
2760 {
2761         struct rev_info revs;
2762         char line[1000];
2763         int flags = 0;
2764
2765         init_revisions(&revs, NULL);
2766         save_commit_buffer = 0;
2767         setup_revisions(ac, av, &revs, NULL);
2768
2769         /* make sure shallows are read */
2770         is_repository_shallow();
2771
2772         while (fgets(line, sizeof(line), stdin) != NULL) {
2773                 int len = strlen(line);
2774                 if (len && line[len - 1] == '\n')
2775                         line[--len] = 0;
2776                 if (!len)
2777                         break;
2778                 if (*line == '-') {
2779                         if (!strcmp(line, "--not")) {
2780                                 flags ^= UNINTERESTING;
2781                                 write_bitmap_index = 0;
2782                                 continue;
2783                         }
2784                         if (starts_with(line, "--shallow ")) {
2785                                 unsigned char sha1[20];
2786                                 if (get_sha1_hex(line + 10, sha1))
2787                                         die("not an SHA-1 '%s'", line + 10);
2788                                 register_shallow(sha1);
2789                                 use_bitmap_index = 0;
2790                                 continue;
2791                         }
2792                         die("not a rev '%s'", line);
2793                 }
2794                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2795                         die("bad revision '%s'", line);
2796         }
2797
2798         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2799                 return;
2800
2801         if (prepare_revision_walk(&revs))
2802                 die("revision walk setup failed");
2803         mark_edges_uninteresting(&revs, show_edge);
2804         traverse_commit_list(&revs, show_commit, show_object, NULL);
2805
2806         if (unpack_unreachable_expiration) {
2807                 revs.ignore_missing_links = 1;
2808                 if (add_unseen_recent_objects_to_traversal(&revs,
2809                                 unpack_unreachable_expiration))
2810                         die("unable to add recent objects");
2811                 if (prepare_revision_walk(&revs))
2812                         die("revision walk setup failed");
2813                 traverse_commit_list(&revs, record_recent_commit,
2814                                      record_recent_object, NULL);
2815         }
2816
2817         if (keep_unreachable)
2818                 add_objects_in_unpacked_packs(&revs);
2819         if (pack_loose_unreachable)
2820                 add_unreachable_loose_objects();
2821         if (unpack_unreachable)
2822                 loosen_unused_packed_objects(&revs);
2823
2824         sha1_array_clear(&recent_objects);
2825 }
2826
2827 static int option_parse_index_version(const struct option *opt,
2828                                       const char *arg, int unset)
2829 {
2830         char *c;
2831         const char *val = arg;
2832         pack_idx_opts.version = strtoul(val, &c, 10);
2833         if (pack_idx_opts.version > 2)
2834                 die(_("unsupported index version %s"), val);
2835         if (*c == ',' && c[1])
2836                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2837         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2838                 die(_("bad index version '%s'"), val);
2839         return 0;
2840 }
2841
2842 static int option_parse_unpack_unreachable(const struct option *opt,
2843                                            const char *arg, int unset)
2844 {
2845         if (unset) {
2846                 unpack_unreachable = 0;
2847                 unpack_unreachable_expiration = 0;
2848         }
2849         else {
2850                 unpack_unreachable = 1;
2851                 if (arg)
2852                         unpack_unreachable_expiration = approxidate(arg);
2853         }
2854         return 0;
2855 }
2856
2857 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2858 {
2859         int use_internal_rev_list = 0;
2860         int thin = 0;
2861         int shallow = 0;
2862         int all_progress_implied = 0;
2863         struct argv_array rp = ARGV_ARRAY_INIT;
2864         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2865         int rev_list_index = 0;
2866         struct option pack_objects_options[] = {
2867                 OPT_SET_INT('q', "quiet", &progress,
2868                             N_("do not show progress meter"), 0),
2869                 OPT_SET_INT(0, "progress", &progress,
2870                             N_("show progress meter"), 1),
2871                 OPT_SET_INT(0, "all-progress", &progress,
2872                             N_("show progress meter during object writing phase"), 2),
2873                 OPT_BOOL(0, "all-progress-implied",
2874                          &all_progress_implied,
2875                          N_("similar to --all-progress when progress meter is shown")),
2876                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2877                   N_("write the pack index file in the specified idx format version"),
2878                   0, option_parse_index_version },
2879                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2880                               N_("maximum size of each output pack file")),
2881                 OPT_BOOL(0, "local", &local,
2882                          N_("ignore borrowed objects from alternate object store")),
2883                 OPT_BOOL(0, "incremental", &incremental,
2884                          N_("ignore packed objects")),
2885                 OPT_INTEGER(0, "window", &window,
2886                             N_("limit pack window by objects")),
2887                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2888                               N_("limit pack window by memory in addition to object limit")),
2889                 OPT_INTEGER(0, "depth", &depth,
2890                             N_("maximum length of delta chain allowed in the resulting pack")),
2891                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2892                          N_("reuse existing deltas")),
2893                 OPT_BOOL(0, "reuse-object", &reuse_object,
2894                          N_("reuse existing objects")),
2895                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2896                          N_("use OFS_DELTA objects")),
2897                 OPT_INTEGER(0, "threads", &delta_search_threads,
2898                             N_("use threads when searching for best delta matches")),
2899                 OPT_BOOL(0, "non-empty", &non_empty,
2900                          N_("do not create an empty pack output")),
2901                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2902                          N_("read revision arguments from standard input")),
2903                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2904                   N_("limit the objects to those that are not yet packed"),
2905                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2906                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2907                   N_("include objects reachable from any reference"),
2908                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2909                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2910                   N_("include objects referred by reflog entries"),
2911                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2912                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2913                   N_("include objects referred to by the index"),
2914                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2915                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2916                          N_("output pack to stdout")),
2917                 OPT_BOOL(0, "include-tag", &include_tag,
2918                          N_("include tag objects that refer to objects to be packed")),
2919                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2920                          N_("keep unreachable objects")),
2921                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
2922                          N_("pack loose unreachable objects")),
2923                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2924                   N_("unpack unreachable objects newer than <time>"),
2925                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2926                 OPT_BOOL(0, "thin", &thin,
2927                          N_("create thin packs")),
2928                 OPT_BOOL(0, "shallow", &shallow,
2929                          N_("create packs suitable for shallow fetches")),
2930                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2931                          N_("ignore packs that have companion .keep file")),
2932                 OPT_INTEGER(0, "compression", &pack_compression_level,
2933                             N_("pack compression level")),
2934                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2935                             N_("do not hide commits by grafts"), 0),
2936                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2937                          N_("use a bitmap index if available to speed up counting objects")),
2938                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2939                          N_("write a bitmap index together with the pack index")),
2940                 OPT_END(),
2941         };
2942
2943         check_replace_refs = 0;
2944
2945         reset_pack_idx_option(&pack_idx_opts);
2946         git_config(git_pack_config, NULL);
2947
2948         progress = isatty(2);
2949         argc = parse_options(argc, argv, prefix, pack_objects_options,
2950                              pack_usage, 0);
2951
2952         if (argc) {
2953                 base_name = argv[0];
2954                 argc--;
2955         }
2956         if (pack_to_stdout != !base_name || argc)
2957                 usage_with_options(pack_usage, pack_objects_options);
2958
2959         argv_array_push(&rp, "pack-objects");
2960         if (thin) {
2961                 use_internal_rev_list = 1;
2962                 argv_array_push(&rp, shallow
2963                                 ? "--objects-edge-aggressive"
2964                                 : "--objects-edge");
2965         } else
2966                 argv_array_push(&rp, "--objects");
2967
2968         if (rev_list_all) {
2969                 use_internal_rev_list = 1;
2970                 argv_array_push(&rp, "--all");
2971         }
2972         if (rev_list_reflog) {
2973                 use_internal_rev_list = 1;
2974                 argv_array_push(&rp, "--reflog");
2975         }
2976         if (rev_list_index) {
2977                 use_internal_rev_list = 1;
2978                 argv_array_push(&rp, "--indexed-objects");
2979         }
2980         if (rev_list_unpacked) {
2981                 use_internal_rev_list = 1;
2982                 argv_array_push(&rp, "--unpacked");
2983         }
2984
2985         if (!reuse_object)
2986                 reuse_delta = 0;
2987         if (pack_compression_level == -1)
2988                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2989         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2990                 die("bad pack compression level %d", pack_compression_level);
2991
2992         if (!delta_search_threads)      /* --threads=0 means autodetect */
2993                 delta_search_threads = online_cpus();
2994
2995 #ifdef NO_PTHREADS
2996         if (delta_search_threads != 1)
2997                 warning("no threads support, ignoring --threads");
2998 #endif
2999         if (!pack_to_stdout && !pack_size_limit)
3000                 pack_size_limit = pack_size_limit_cfg;
3001         if (pack_to_stdout && pack_size_limit)
3002                 die("--max-pack-size cannot be used to build a pack for transfer.");
3003         if (pack_size_limit && pack_size_limit < 1024*1024) {
3004                 warning("minimum pack size limit is 1 MiB");
3005                 pack_size_limit = 1024*1024;
3006         }
3007
3008         if (!pack_to_stdout && thin)
3009                 die("--thin cannot be used to build an indexable pack.");
3010
3011         if (keep_unreachable && unpack_unreachable)
3012                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
3013         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3014                 unpack_unreachable_expiration = 0;
3015
3016         /*
3017          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3018          *
3019          * - to produce good pack (with bitmap index not-yet-packed objects are
3020          *   packed in suboptimal order).
3021          *
3022          * - to use more robust pack-generation codepath (avoiding possible
3023          *   bugs in bitmap code and possible bitmap index corruption).
3024          */
3025         if (!pack_to_stdout)
3026                 use_bitmap_index_default = 0;
3027
3028         if (use_bitmap_index < 0)
3029                 use_bitmap_index = use_bitmap_index_default;
3030
3031         /* "hard" reasons not to use bitmaps; these just won't work at all */
3032         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
3033                 use_bitmap_index = 0;
3034
3035         if (pack_to_stdout || !rev_list_all)
3036                 write_bitmap_index = 0;
3037
3038         if (progress && all_progress_implied)
3039                 progress = 2;
3040
3041         prepare_packed_git();
3042         if (ignore_packed_keep) {
3043                 struct packed_git *p;
3044                 for (p = packed_git; p; p = p->next)
3045                         if (p->pack_local && p->pack_keep)
3046                                 break;
3047                 if (!p) /* no keep-able packs found */
3048                         ignore_packed_keep = 0;
3049         }
3050         if (local) {
3051                 /*
3052                  * unlike ignore_packed_keep above, we do not want to
3053                  * unset "local" based on looking at packs, as it
3054                  * also covers non-local objects
3055                  */
3056                 struct packed_git *p;
3057                 for (p = packed_git; p; p = p->next) {
3058                         if (!p->pack_local) {
3059                                 have_non_local_packs = 1;
3060                                 break;
3061                         }
3062                 }
3063         }
3064
3065         if (progress)
3066                 progress_state = start_progress(_("Counting objects"), 0);
3067         if (!use_internal_rev_list)
3068                 read_object_list_from_stdin();
3069         else {
3070                 get_object_list(rp.argc, rp.argv);
3071                 argv_array_clear(&rp);
3072         }
3073         cleanup_preferred_base();
3074         if (include_tag && nr_result)
3075                 for_each_ref(add_ref_tag, NULL);
3076         stop_progress(&progress_state);
3077
3078         if (non_empty && !nr_result)
3079                 return 0;
3080         if (nr_result)
3081                 prepare_pack(window, depth);
3082         write_pack_file();
3083         if (progress)
3084                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
3085                         " reused %"PRIu32" (delta %"PRIu32")\n",
3086                         written, written_delta, reused, reused_delta);
3087         return 0;
3088 }