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