Merge branch 'dd/use-alloc-grow'
[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
24 static const char *pack_usage[] = {
25         N_("git pack-objects --stdout [options...] [< ref-list | < object-list]"),
26         N_("git pack-objects [options...] base-name [< ref-list | < object-list]"),
27         NULL
28 };
29
30 /*
31  * Objects we are going to pack are collected in the `to_pack` structure.
32  * It contains an array (dynamically expanded) of the object data, and a map
33  * that can resolve SHA1s to their position in the array.
34  */
35 static struct packing_data to_pack;
36
37 static struct pack_idx_entry **written_list;
38 static uint32_t nr_result, nr_written;
39
40 static int non_empty;
41 static int reuse_delta = 1, reuse_object = 1;
42 static int keep_unreachable, unpack_unreachable, include_tag;
43 static unsigned long unpack_unreachable_expiration;
44 static int local;
45 static int incremental;
46 static int ignore_packed_keep;
47 static int allow_ofs_delta;
48 static struct pack_idx_option pack_idx_opts;
49 static const char *base_name;
50 static int progress = 1;
51 static int window = 10;
52 static unsigned long pack_size_limit;
53 static int depth = 50;
54 static int delta_search_threads;
55 static int pack_to_stdout;
56 static int num_preferred_base;
57 static struct progress *progress_state;
58 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
59 static int pack_compression_seen;
60
61 static struct packed_git *reuse_packfile;
62 static uint32_t reuse_packfile_objects;
63 static off_t reuse_packfile_offset;
64
65 static int use_bitmap_index = 1;
66 static int write_bitmap_index;
67 static uint16_t write_bitmap_options;
68
69 static unsigned long delta_cache_size = 0;
70 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
71 static unsigned long cache_max_small_delta_size = 1000;
72
73 static unsigned long window_memory_limit = 0;
74
75 /*
76  * stats
77  */
78 static uint32_t written, written_delta;
79 static uint32_t reused, reused_delta;
80
81 /*
82  * Indexed commits
83  */
84 static struct commit **indexed_commits;
85 static unsigned int indexed_commits_nr;
86 static unsigned int indexed_commits_alloc;
87
88 static void index_commit_for_bitmap(struct commit *commit)
89 {
90         if (indexed_commits_nr >= indexed_commits_alloc) {
91                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
92                 indexed_commits = xrealloc(indexed_commits,
93                         indexed_commits_alloc * sizeof(struct commit *));
94         }
95
96         indexed_commits[indexed_commits_nr++] = commit;
97 }
98
99 static void *get_delta(struct object_entry *entry)
100 {
101         unsigned long size, base_size, delta_size;
102         void *buf, *base_buf, *delta_buf;
103         enum object_type type;
104
105         buf = read_sha1_file(entry->idx.sha1, &type, &size);
106         if (!buf)
107                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
108         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
109         if (!base_buf)
110                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
111         delta_buf = diff_delta(base_buf, base_size,
112                                buf, size, &delta_size, 0);
113         if (!delta_buf || delta_size != entry->delta_size)
114                 die("delta size changed");
115         free(buf);
116         free(base_buf);
117         return delta_buf;
118 }
119
120 static unsigned long do_compress(void **pptr, unsigned long size)
121 {
122         git_zstream stream;
123         void *in, *out;
124         unsigned long maxsize;
125
126         memset(&stream, 0, sizeof(stream));
127         git_deflate_init(&stream, pack_compression_level);
128         maxsize = git_deflate_bound(&stream, size);
129
130         in = *pptr;
131         out = xmalloc(maxsize);
132         *pptr = out;
133
134         stream.next_in = in;
135         stream.avail_in = size;
136         stream.next_out = out;
137         stream.avail_out = maxsize;
138         while (git_deflate(&stream, Z_FINISH) == Z_OK)
139                 ; /* nothing */
140         git_deflate_end(&stream);
141
142         free(in);
143         return stream.total_out;
144 }
145
146 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
147                                            const unsigned char *sha1)
148 {
149         git_zstream stream;
150         unsigned char ibuf[1024 * 16];
151         unsigned char obuf[1024 * 16];
152         unsigned long olen = 0;
153
154         memset(&stream, 0, sizeof(stream));
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 unsigned char *sha1, int flag,
544                        void *cb_data)
545 {
546         unsigned char peeled[20];
547         struct object_entry *entry = packlist_find(&to_pack, sha1, 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;
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         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         close(fd);
744         written += reuse_packfile_objects;
745         return reuse_packfile_offset - sizeof(struct pack_header);
746 }
747
748 static void write_pack_file(void)
749 {
750         uint32_t i = 0, j;
751         struct sha1file *f;
752         off_t offset;
753         uint32_t nr_remaining = nr_result;
754         time_t last_mtime = 0;
755         struct object_entry **write_order;
756
757         if (progress > pack_to_stdout)
758                 progress_state = start_progress(_("Writing objects"), nr_result);
759         written_list = xmalloc(to_pack.nr_objects * sizeof(*written_list));
760         write_order = compute_write_order();
761
762         do {
763                 unsigned char sha1[20];
764                 char *pack_tmp_name = NULL;
765
766                 if (pack_to_stdout)
767                         f = sha1fd_throughput(1, "<stdout>", progress_state);
768                 else
769                         f = create_tmp_packfile(&pack_tmp_name);
770
771                 offset = write_pack_header(f, nr_remaining);
772
773                 if (reuse_packfile) {
774                         off_t packfile_size;
775                         assert(pack_to_stdout);
776
777                         packfile_size = write_reused_pack(f);
778                         offset += packfile_size;
779                 }
780
781                 nr_written = 0;
782                 for (; i < to_pack.nr_objects; i++) {
783                         struct object_entry *e = write_order[i];
784                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
785                                 break;
786                         display_progress(progress_state, written);
787                 }
788
789                 /*
790                  * Did we write the wrong # entries in the header?
791                  * If so, rewrite it like in fast-import
792                  */
793                 if (pack_to_stdout) {
794                         sha1close(f, sha1, CSUM_CLOSE);
795                 } else if (nr_written == nr_remaining) {
796                         sha1close(f, sha1, CSUM_FSYNC);
797                 } else {
798                         int fd = sha1close(f, sha1, 0);
799                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
800                                                  nr_written, sha1, offset);
801                         close(fd);
802                 }
803
804                 if (!pack_to_stdout) {
805                         struct stat st;
806                         char tmpname[PATH_MAX];
807
808                         /*
809                          * Packs are runtime accessed in their mtime
810                          * order since newer packs are more likely to contain
811                          * younger objects.  So if we are creating multiple
812                          * packs then we should modify the mtime of later ones
813                          * to preserve this property.
814                          */
815                         if (stat(pack_tmp_name, &st) < 0) {
816                                 warning("failed to stat %s: %s",
817                                         pack_tmp_name, strerror(errno));
818                         } else if (!last_mtime) {
819                                 last_mtime = st.st_mtime;
820                         } else {
821                                 struct utimbuf utb;
822                                 utb.actime = st.st_atime;
823                                 utb.modtime = --last_mtime;
824                                 if (utime(pack_tmp_name, &utb) < 0)
825                                         warning("failed utime() on %s: %s",
826                                                 pack_tmp_name, strerror(errno));
827                         }
828
829                         /* Enough space for "-<sha-1>.pack"? */
830                         if (sizeof(tmpname) <= strlen(base_name) + 50)
831                                 die("pack base name '%s' too long", base_name);
832                         snprintf(tmpname, sizeof(tmpname), "%s-", base_name);
833
834                         if (write_bitmap_index) {
835                                 bitmap_writer_set_checksum(sha1);
836                                 bitmap_writer_build_type_index(written_list, nr_written);
837                         }
838
839                         finish_tmp_packfile(tmpname, pack_tmp_name,
840                                             written_list, nr_written,
841                                             &pack_idx_opts, sha1);
842
843                         if (write_bitmap_index) {
844                                 char *end_of_name_prefix = strrchr(tmpname, 0);
845                                 sprintf(end_of_name_prefix, "%s.bitmap", sha1_to_hex(sha1));
846
847                                 stop_progress(&progress_state);
848
849                                 bitmap_writer_show_progress(progress);
850                                 bitmap_writer_reuse_bitmaps(&to_pack);
851                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
852                                 bitmap_writer_build(&to_pack);
853                                 bitmap_writer_finish(written_list, nr_written,
854                                                      tmpname, write_bitmap_options);
855                                 write_bitmap_index = 0;
856                         }
857
858                         free(pack_tmp_name);
859                         puts(sha1_to_hex(sha1));
860                 }
861
862                 /* mark written objects as written to previous pack */
863                 for (j = 0; j < nr_written; j++) {
864                         written_list[j]->offset = (off_t)-1;
865                 }
866                 nr_remaining -= nr_written;
867         } while (nr_remaining && i < to_pack.nr_objects);
868
869         free(written_list);
870         free(write_order);
871         stop_progress(&progress_state);
872         if (written != nr_result)
873                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
874                         written, nr_result);
875 }
876
877 static void setup_delta_attr_check(struct git_attr_check *check)
878 {
879         static struct git_attr *attr_delta;
880
881         if (!attr_delta)
882                 attr_delta = git_attr("delta");
883
884         check[0].attr = attr_delta;
885 }
886
887 static int no_try_delta(const char *path)
888 {
889         struct git_attr_check check[1];
890
891         setup_delta_attr_check(check);
892         if (git_check_attr(path, ARRAY_SIZE(check), check))
893                 return 0;
894         if (ATTR_FALSE(check->value))
895                 return 1;
896         return 0;
897 }
898
899 /*
900  * When adding an object, check whether we have already added it
901  * to our packing list. If so, we can skip. However, if we are
902  * being asked to excludei t, but the previous mention was to include
903  * it, make sure to adjust its flags and tweak our numbers accordingly.
904  *
905  * As an optimization, we pass out the index position where we would have
906  * found the item, since that saves us from having to look it up again a
907  * few lines later when we want to add the new entry.
908  */
909 static int have_duplicate_entry(const unsigned char *sha1,
910                                 int exclude,
911                                 uint32_t *index_pos)
912 {
913         struct object_entry *entry;
914
915         entry = packlist_find(&to_pack, sha1, index_pos);
916         if (!entry)
917                 return 0;
918
919         if (exclude) {
920                 if (!entry->preferred_base)
921                         nr_result--;
922                 entry->preferred_base = 1;
923         }
924
925         return 1;
926 }
927
928 /*
929  * Check whether we want the object in the pack (e.g., we do not want
930  * objects found in non-local stores if the "--local" option was used).
931  *
932  * As a side effect of this check, we will find the packed version of this
933  * object, if any. We therefore pass out the pack information to avoid having
934  * to look it up again later.
935  */
936 static int want_object_in_pack(const unsigned char *sha1,
937                                int exclude,
938                                struct packed_git **found_pack,
939                                off_t *found_offset)
940 {
941         struct packed_git *p;
942
943         if (!exclude && local && has_loose_object_nonlocal(sha1))
944                 return 0;
945
946         *found_pack = NULL;
947         *found_offset = 0;
948
949         for (p = packed_git; p; p = p->next) {
950                 off_t offset = find_pack_entry_one(sha1, p);
951                 if (offset) {
952                         if (!*found_pack) {
953                                 if (!is_pack_valid(p)) {
954                                         warning("packfile %s cannot be accessed", p->pack_name);
955                                         continue;
956                                 }
957                                 *found_offset = offset;
958                                 *found_pack = p;
959                         }
960                         if (exclude)
961                                 return 1;
962                         if (incremental)
963                                 return 0;
964                         if (local && !p->pack_local)
965                                 return 0;
966                         if (ignore_packed_keep && p->pack_local && p->pack_keep)
967                                 return 0;
968                 }
969         }
970
971         return 1;
972 }
973
974 static void create_object_entry(const unsigned char *sha1,
975                                 enum object_type type,
976                                 uint32_t hash,
977                                 int exclude,
978                                 int no_try_delta,
979                                 uint32_t index_pos,
980                                 struct packed_git *found_pack,
981                                 off_t found_offset)
982 {
983         struct object_entry *entry;
984
985         entry = packlist_alloc(&to_pack, sha1, index_pos);
986         entry->hash = hash;
987         if (type)
988                 entry->type = type;
989         if (exclude)
990                 entry->preferred_base = 1;
991         else
992                 nr_result++;
993         if (found_pack) {
994                 entry->in_pack = found_pack;
995                 entry->in_pack_offset = found_offset;
996         }
997
998         entry->no_try_delta = no_try_delta;
999 }
1000
1001 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1002                             const char *name, int exclude)
1003 {
1004         struct packed_git *found_pack;
1005         off_t found_offset;
1006         uint32_t index_pos;
1007
1008         if (have_duplicate_entry(sha1, exclude, &index_pos))
1009                 return 0;
1010
1011         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset))
1012                 return 0;
1013
1014         create_object_entry(sha1, type, pack_name_hash(name),
1015                             exclude, name && no_try_delta(name),
1016                             index_pos, found_pack, found_offset);
1017
1018         display_progress(progress_state, to_pack.nr_objects);
1019         return 1;
1020 }
1021
1022 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1023                                         enum object_type type,
1024                                         int flags, uint32_t name_hash,
1025                                         struct packed_git *pack, off_t offset)
1026 {
1027         uint32_t index_pos;
1028
1029         if (have_duplicate_entry(sha1, 0, &index_pos))
1030                 return 0;
1031
1032         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1033
1034         display_progress(progress_state, to_pack.nr_objects);
1035         return 1;
1036 }
1037
1038 struct pbase_tree_cache {
1039         unsigned char sha1[20];
1040         int ref;
1041         int temporary;
1042         void *tree_data;
1043         unsigned long tree_size;
1044 };
1045
1046 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1047 static int pbase_tree_cache_ix(const unsigned char *sha1)
1048 {
1049         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1050 }
1051 static int pbase_tree_cache_ix_incr(int ix)
1052 {
1053         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1054 }
1055
1056 static struct pbase_tree {
1057         struct pbase_tree *next;
1058         /* This is a phony "cache" entry; we are not
1059          * going to evict it nor find it through _get()
1060          * mechanism -- this is for the toplevel node that
1061          * would almost always change with any commit.
1062          */
1063         struct pbase_tree_cache pcache;
1064 } *pbase_tree;
1065
1066 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1067 {
1068         struct pbase_tree_cache *ent, *nent;
1069         void *data;
1070         unsigned long size;
1071         enum object_type type;
1072         int neigh;
1073         int my_ix = pbase_tree_cache_ix(sha1);
1074         int available_ix = -1;
1075
1076         /* pbase-tree-cache acts as a limited hashtable.
1077          * your object will be found at your index or within a few
1078          * slots after that slot if it is cached.
1079          */
1080         for (neigh = 0; neigh < 8; neigh++) {
1081                 ent = pbase_tree_cache[my_ix];
1082                 if (ent && !hashcmp(ent->sha1, sha1)) {
1083                         ent->ref++;
1084                         return ent;
1085                 }
1086                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1087                          ((0 <= available_ix) &&
1088                           (!ent && pbase_tree_cache[available_ix])))
1089                         available_ix = my_ix;
1090                 if (!ent)
1091                         break;
1092                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1093         }
1094
1095         /* Did not find one.  Either we got a bogus request or
1096          * we need to read and perhaps cache.
1097          */
1098         data = read_sha1_file(sha1, &type, &size);
1099         if (!data)
1100                 return NULL;
1101         if (type != OBJ_TREE) {
1102                 free(data);
1103                 return NULL;
1104         }
1105
1106         /* We need to either cache or return a throwaway copy */
1107
1108         if (available_ix < 0)
1109                 ent = NULL;
1110         else {
1111                 ent = pbase_tree_cache[available_ix];
1112                 my_ix = available_ix;
1113         }
1114
1115         if (!ent) {
1116                 nent = xmalloc(sizeof(*nent));
1117                 nent->temporary = (available_ix < 0);
1118         }
1119         else {
1120                 /* evict and reuse */
1121                 free(ent->tree_data);
1122                 nent = ent;
1123         }
1124         hashcpy(nent->sha1, sha1);
1125         nent->tree_data = data;
1126         nent->tree_size = size;
1127         nent->ref = 1;
1128         if (!nent->temporary)
1129                 pbase_tree_cache[my_ix] = nent;
1130         return nent;
1131 }
1132
1133 static void pbase_tree_put(struct pbase_tree_cache *cache)
1134 {
1135         if (!cache->temporary) {
1136                 cache->ref--;
1137                 return;
1138         }
1139         free(cache->tree_data);
1140         free(cache);
1141 }
1142
1143 static int name_cmp_len(const char *name)
1144 {
1145         int i;
1146         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1147                 ;
1148         return i;
1149 }
1150
1151 static void add_pbase_object(struct tree_desc *tree,
1152                              const char *name,
1153                              int cmplen,
1154                              const char *fullname)
1155 {
1156         struct name_entry entry;
1157         int cmp;
1158
1159         while (tree_entry(tree,&entry)) {
1160                 if (S_ISGITLINK(entry.mode))
1161                         continue;
1162                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1163                       memcmp(name, entry.path, cmplen);
1164                 if (cmp > 0)
1165                         continue;
1166                 if (cmp < 0)
1167                         return;
1168                 if (name[cmplen] != '/') {
1169                         add_object_entry(entry.sha1,
1170                                          object_type(entry.mode),
1171                                          fullname, 1);
1172                         return;
1173                 }
1174                 if (S_ISDIR(entry.mode)) {
1175                         struct tree_desc sub;
1176                         struct pbase_tree_cache *tree;
1177                         const char *down = name+cmplen+1;
1178                         int downlen = name_cmp_len(down);
1179
1180                         tree = pbase_tree_get(entry.sha1);
1181                         if (!tree)
1182                                 return;
1183                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1184
1185                         add_pbase_object(&sub, down, downlen, fullname);
1186                         pbase_tree_put(tree);
1187                 }
1188         }
1189 }
1190
1191 static unsigned *done_pbase_paths;
1192 static int done_pbase_paths_num;
1193 static int done_pbase_paths_alloc;
1194 static int done_pbase_path_pos(unsigned hash)
1195 {
1196         int lo = 0;
1197         int hi = done_pbase_paths_num;
1198         while (lo < hi) {
1199                 int mi = (hi + lo) / 2;
1200                 if (done_pbase_paths[mi] == hash)
1201                         return mi;
1202                 if (done_pbase_paths[mi] < hash)
1203                         hi = mi;
1204                 else
1205                         lo = mi + 1;
1206         }
1207         return -lo-1;
1208 }
1209
1210 static int check_pbase_path(unsigned hash)
1211 {
1212         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1213         if (0 <= pos)
1214                 return 1;
1215         pos = -pos - 1;
1216         ALLOC_GROW(done_pbase_paths,
1217                    done_pbase_paths_num + 1,
1218                    done_pbase_paths_alloc);
1219         done_pbase_paths_num++;
1220         if (pos < done_pbase_paths_num)
1221                 memmove(done_pbase_paths + pos + 1,
1222                         done_pbase_paths + pos,
1223                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1224         done_pbase_paths[pos] = hash;
1225         return 0;
1226 }
1227
1228 static void add_preferred_base_object(const char *name)
1229 {
1230         struct pbase_tree *it;
1231         int cmplen;
1232         unsigned hash = pack_name_hash(name);
1233
1234         if (!num_preferred_base || check_pbase_path(hash))
1235                 return;
1236
1237         cmplen = name_cmp_len(name);
1238         for (it = pbase_tree; it; it = it->next) {
1239                 if (cmplen == 0) {
1240                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1241                 }
1242                 else {
1243                         struct tree_desc tree;
1244                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1245                         add_pbase_object(&tree, name, cmplen, name);
1246                 }
1247         }
1248 }
1249
1250 static void add_preferred_base(unsigned char *sha1)
1251 {
1252         struct pbase_tree *it;
1253         void *data;
1254         unsigned long size;
1255         unsigned char tree_sha1[20];
1256
1257         if (window <= num_preferred_base++)
1258                 return;
1259
1260         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1261         if (!data)
1262                 return;
1263
1264         for (it = pbase_tree; it; it = it->next) {
1265                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1266                         free(data);
1267                         return;
1268                 }
1269         }
1270
1271         it = xcalloc(1, sizeof(*it));
1272         it->next = pbase_tree;
1273         pbase_tree = it;
1274
1275         hashcpy(it->pcache.sha1, tree_sha1);
1276         it->pcache.tree_data = data;
1277         it->pcache.tree_size = size;
1278 }
1279
1280 static void cleanup_preferred_base(void)
1281 {
1282         struct pbase_tree *it;
1283         unsigned i;
1284
1285         it = pbase_tree;
1286         pbase_tree = NULL;
1287         while (it) {
1288                 struct pbase_tree *this = it;
1289                 it = this->next;
1290                 free(this->pcache.tree_data);
1291                 free(this);
1292         }
1293
1294         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1295                 if (!pbase_tree_cache[i])
1296                         continue;
1297                 free(pbase_tree_cache[i]->tree_data);
1298                 free(pbase_tree_cache[i]);
1299                 pbase_tree_cache[i] = NULL;
1300         }
1301
1302         free(done_pbase_paths);
1303         done_pbase_paths = NULL;
1304         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1305 }
1306
1307 static void check_object(struct object_entry *entry)
1308 {
1309         if (entry->in_pack) {
1310                 struct packed_git *p = entry->in_pack;
1311                 struct pack_window *w_curs = NULL;
1312                 const unsigned char *base_ref = NULL;
1313                 struct object_entry *base_entry;
1314                 unsigned long used, used_0;
1315                 unsigned long avail;
1316                 off_t ofs;
1317                 unsigned char *buf, c;
1318
1319                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1320
1321                 /*
1322                  * We want in_pack_type even if we do not reuse delta
1323                  * since non-delta representations could still be reused.
1324                  */
1325                 used = unpack_object_header_buffer(buf, avail,
1326                                                    &entry->in_pack_type,
1327                                                    &entry->size);
1328                 if (used == 0)
1329                         goto give_up;
1330
1331                 /*
1332                  * Determine if this is a delta and if so whether we can
1333                  * reuse it or not.  Otherwise let's find out as cheaply as
1334                  * possible what the actual type and size for this object is.
1335                  */
1336                 switch (entry->in_pack_type) {
1337                 default:
1338                         /* Not a delta hence we've already got all we need. */
1339                         entry->type = entry->in_pack_type;
1340                         entry->in_pack_header_size = used;
1341                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1342                                 goto give_up;
1343                         unuse_pack(&w_curs);
1344                         return;
1345                 case OBJ_REF_DELTA:
1346                         if (reuse_delta && !entry->preferred_base)
1347                                 base_ref = use_pack(p, &w_curs,
1348                                                 entry->in_pack_offset + used, NULL);
1349                         entry->in_pack_header_size = used + 20;
1350                         break;
1351                 case OBJ_OFS_DELTA:
1352                         buf = use_pack(p, &w_curs,
1353                                        entry->in_pack_offset + used, NULL);
1354                         used_0 = 0;
1355                         c = buf[used_0++];
1356                         ofs = c & 127;
1357                         while (c & 128) {
1358                                 ofs += 1;
1359                                 if (!ofs || MSB(ofs, 7)) {
1360                                         error("delta base offset overflow in pack for %s",
1361                                               sha1_to_hex(entry->idx.sha1));
1362                                         goto give_up;
1363                                 }
1364                                 c = buf[used_0++];
1365                                 ofs = (ofs << 7) + (c & 127);
1366                         }
1367                         ofs = entry->in_pack_offset - ofs;
1368                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1369                                 error("delta base offset out of bound for %s",
1370                                       sha1_to_hex(entry->idx.sha1));
1371                                 goto give_up;
1372                         }
1373                         if (reuse_delta && !entry->preferred_base) {
1374                                 struct revindex_entry *revidx;
1375                                 revidx = find_pack_revindex(p, ofs);
1376                                 if (!revidx)
1377                                         goto give_up;
1378                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1379                         }
1380                         entry->in_pack_header_size = used + used_0;
1381                         break;
1382                 }
1383
1384                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1385                         /*
1386                          * If base_ref was set above that means we wish to
1387                          * reuse delta data, and we even found that base
1388                          * in the list of objects we want to pack. Goodie!
1389                          *
1390                          * Depth value does not matter - find_deltas() will
1391                          * never consider reused delta as the base object to
1392                          * deltify other objects against, in order to avoid
1393                          * circular deltas.
1394                          */
1395                         entry->type = entry->in_pack_type;
1396                         entry->delta = base_entry;
1397                         entry->delta_size = entry->size;
1398                         entry->delta_sibling = base_entry->delta_child;
1399                         base_entry->delta_child = entry;
1400                         unuse_pack(&w_curs);
1401                         return;
1402                 }
1403
1404                 if (entry->type) {
1405                         /*
1406                          * This must be a delta and we already know what the
1407                          * final object type is.  Let's extract the actual
1408                          * object size from the delta header.
1409                          */
1410                         entry->size = get_size_from_delta(p, &w_curs,
1411                                         entry->in_pack_offset + entry->in_pack_header_size);
1412                         if (entry->size == 0)
1413                                 goto give_up;
1414                         unuse_pack(&w_curs);
1415                         return;
1416                 }
1417
1418                 /*
1419                  * No choice but to fall back to the recursive delta walk
1420                  * with sha1_object_info() to find about the object type
1421                  * at this point...
1422                  */
1423                 give_up:
1424                 unuse_pack(&w_curs);
1425         }
1426
1427         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1428         /*
1429          * The error condition is checked in prepare_pack().  This is
1430          * to permit a missing preferred base object to be ignored
1431          * as a preferred base.  Doing so can result in a larger
1432          * pack file, but the transfer will still take place.
1433          */
1434 }
1435
1436 static int pack_offset_sort(const void *_a, const void *_b)
1437 {
1438         const struct object_entry *a = *(struct object_entry **)_a;
1439         const struct object_entry *b = *(struct object_entry **)_b;
1440
1441         /* avoid filesystem trashing with loose objects */
1442         if (!a->in_pack && !b->in_pack)
1443                 return hashcmp(a->idx.sha1, b->idx.sha1);
1444
1445         if (a->in_pack < b->in_pack)
1446                 return -1;
1447         if (a->in_pack > b->in_pack)
1448                 return 1;
1449         return a->in_pack_offset < b->in_pack_offset ? -1 :
1450                         (a->in_pack_offset > b->in_pack_offset);
1451 }
1452
1453 static void get_object_details(void)
1454 {
1455         uint32_t i;
1456         struct object_entry **sorted_by_offset;
1457
1458         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1459         for (i = 0; i < to_pack.nr_objects; i++)
1460                 sorted_by_offset[i] = to_pack.objects + i;
1461         qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1462
1463         for (i = 0; i < to_pack.nr_objects; i++) {
1464                 struct object_entry *entry = sorted_by_offset[i];
1465                 check_object(entry);
1466                 if (big_file_threshold < entry->size)
1467                         entry->no_try_delta = 1;
1468         }
1469
1470         free(sorted_by_offset);
1471 }
1472
1473 /*
1474  * We search for deltas in a list sorted by type, by filename hash, and then
1475  * by size, so that we see progressively smaller and smaller files.
1476  * That's because we prefer deltas to be from the bigger file
1477  * to the smaller -- deletes are potentially cheaper, but perhaps
1478  * more importantly, the bigger file is likely the more recent
1479  * one.  The deepest deltas are therefore the oldest objects which are
1480  * less susceptible to be accessed often.
1481  */
1482 static int type_size_sort(const void *_a, const void *_b)
1483 {
1484         const struct object_entry *a = *(struct object_entry **)_a;
1485         const struct object_entry *b = *(struct object_entry **)_b;
1486
1487         if (a->type > b->type)
1488                 return -1;
1489         if (a->type < b->type)
1490                 return 1;
1491         if (a->hash > b->hash)
1492                 return -1;
1493         if (a->hash < b->hash)
1494                 return 1;
1495         if (a->preferred_base > b->preferred_base)
1496                 return -1;
1497         if (a->preferred_base < b->preferred_base)
1498                 return 1;
1499         if (a->size > b->size)
1500                 return -1;
1501         if (a->size < b->size)
1502                 return 1;
1503         return a < b ? -1 : (a > b);  /* newest first */
1504 }
1505
1506 struct unpacked {
1507         struct object_entry *entry;
1508         void *data;
1509         struct delta_index *index;
1510         unsigned depth;
1511 };
1512
1513 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1514                            unsigned long delta_size)
1515 {
1516         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1517                 return 0;
1518
1519         if (delta_size < cache_max_small_delta_size)
1520                 return 1;
1521
1522         /* cache delta, if objects are large enough compared to delta size */
1523         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1524                 return 1;
1525
1526         return 0;
1527 }
1528
1529 #ifndef NO_PTHREADS
1530
1531 static pthread_mutex_t read_mutex;
1532 #define read_lock()             pthread_mutex_lock(&read_mutex)
1533 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1534
1535 static pthread_mutex_t cache_mutex;
1536 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1537 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1538
1539 static pthread_mutex_t progress_mutex;
1540 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1541 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1542
1543 #else
1544
1545 #define read_lock()             (void)0
1546 #define read_unlock()           (void)0
1547 #define cache_lock()            (void)0
1548 #define cache_unlock()          (void)0
1549 #define progress_lock()         (void)0
1550 #define progress_unlock()       (void)0
1551
1552 #endif
1553
1554 static int try_delta(struct unpacked *trg, struct unpacked *src,
1555                      unsigned max_depth, unsigned long *mem_usage)
1556 {
1557         struct object_entry *trg_entry = trg->entry;
1558         struct object_entry *src_entry = src->entry;
1559         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1560         unsigned ref_depth;
1561         enum object_type type;
1562         void *delta_buf;
1563
1564         /* Don't bother doing diffs between different types */
1565         if (trg_entry->type != src_entry->type)
1566                 return -1;
1567
1568         /*
1569          * We do not bother to try a delta that we discarded on an
1570          * earlier try, but only when reusing delta data.  Note that
1571          * src_entry that is marked as the preferred_base should always
1572          * be considered, as even if we produce a suboptimal delta against
1573          * it, we will still save the transfer cost, as we already know
1574          * the other side has it and we won't send src_entry at all.
1575          */
1576         if (reuse_delta && trg_entry->in_pack &&
1577             trg_entry->in_pack == src_entry->in_pack &&
1578             !src_entry->preferred_base &&
1579             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1580             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1581                 return 0;
1582
1583         /* Let's not bust the allowed depth. */
1584         if (src->depth >= max_depth)
1585                 return 0;
1586
1587         /* Now some size filtering heuristics. */
1588         trg_size = trg_entry->size;
1589         if (!trg_entry->delta) {
1590                 max_size = trg_size/2 - 20;
1591                 ref_depth = 1;
1592         } else {
1593                 max_size = trg_entry->delta_size;
1594                 ref_depth = trg->depth;
1595         }
1596         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1597                                                 (max_depth - ref_depth + 1);
1598         if (max_size == 0)
1599                 return 0;
1600         src_size = src_entry->size;
1601         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1602         if (sizediff >= max_size)
1603                 return 0;
1604         if (trg_size < src_size / 32)
1605                 return 0;
1606
1607         /* Load data if not already done */
1608         if (!trg->data) {
1609                 read_lock();
1610                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1611                 read_unlock();
1612                 if (!trg->data)
1613                         die("object %s cannot be read",
1614                             sha1_to_hex(trg_entry->idx.sha1));
1615                 if (sz != trg_size)
1616                         die("object %s inconsistent object length (%lu vs %lu)",
1617                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1618                 *mem_usage += sz;
1619         }
1620         if (!src->data) {
1621                 read_lock();
1622                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1623                 read_unlock();
1624                 if (!src->data) {
1625                         if (src_entry->preferred_base) {
1626                                 static int warned = 0;
1627                                 if (!warned++)
1628                                         warning("object %s cannot be read",
1629                                                 sha1_to_hex(src_entry->idx.sha1));
1630                                 /*
1631                                  * Those objects are not included in the
1632                                  * resulting pack.  Be resilient and ignore
1633                                  * them if they can't be read, in case the
1634                                  * pack could be created nevertheless.
1635                                  */
1636                                 return 0;
1637                         }
1638                         die("object %s cannot be read",
1639                             sha1_to_hex(src_entry->idx.sha1));
1640                 }
1641                 if (sz != src_size)
1642                         die("object %s inconsistent object length (%lu vs %lu)",
1643                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1644                 *mem_usage += sz;
1645         }
1646         if (!src->index) {
1647                 src->index = create_delta_index(src->data, src_size);
1648                 if (!src->index) {
1649                         static int warned = 0;
1650                         if (!warned++)
1651                                 warning("suboptimal pack - out of memory");
1652                         return 0;
1653                 }
1654                 *mem_usage += sizeof_delta_index(src->index);
1655         }
1656
1657         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1658         if (!delta_buf)
1659                 return 0;
1660
1661         if (trg_entry->delta) {
1662                 /* Prefer only shallower same-sized deltas. */
1663                 if (delta_size == trg_entry->delta_size &&
1664                     src->depth + 1 >= trg->depth) {
1665                         free(delta_buf);
1666                         return 0;
1667                 }
1668         }
1669
1670         /*
1671          * Handle memory allocation outside of the cache
1672          * accounting lock.  Compiler will optimize the strangeness
1673          * away when NO_PTHREADS is defined.
1674          */
1675         free(trg_entry->delta_data);
1676         cache_lock();
1677         if (trg_entry->delta_data) {
1678                 delta_cache_size -= trg_entry->delta_size;
1679                 trg_entry->delta_data = NULL;
1680         }
1681         if (delta_cacheable(src_size, trg_size, delta_size)) {
1682                 delta_cache_size += delta_size;
1683                 cache_unlock();
1684                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1685         } else {
1686                 cache_unlock();
1687                 free(delta_buf);
1688         }
1689
1690         trg_entry->delta = src_entry;
1691         trg_entry->delta_size = delta_size;
1692         trg->depth = src->depth + 1;
1693
1694         return 1;
1695 }
1696
1697 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1698 {
1699         struct object_entry *child = me->delta_child;
1700         unsigned int m = n;
1701         while (child) {
1702                 unsigned int c = check_delta_limit(child, n + 1);
1703                 if (m < c)
1704                         m = c;
1705                 child = child->delta_sibling;
1706         }
1707         return m;
1708 }
1709
1710 static unsigned long free_unpacked(struct unpacked *n)
1711 {
1712         unsigned long freed_mem = sizeof_delta_index(n->index);
1713         free_delta_index(n->index);
1714         n->index = NULL;
1715         if (n->data) {
1716                 freed_mem += n->entry->size;
1717                 free(n->data);
1718                 n->data = NULL;
1719         }
1720         n->entry = NULL;
1721         n->depth = 0;
1722         return freed_mem;
1723 }
1724
1725 static void find_deltas(struct object_entry **list, unsigned *list_size,
1726                         int window, int depth, unsigned *processed)
1727 {
1728         uint32_t i, idx = 0, count = 0;
1729         struct unpacked *array;
1730         unsigned long mem_usage = 0;
1731
1732         array = xcalloc(window, sizeof(struct unpacked));
1733
1734         for (;;) {
1735                 struct object_entry *entry;
1736                 struct unpacked *n = array + idx;
1737                 int j, max_depth, best_base = -1;
1738
1739                 progress_lock();
1740                 if (!*list_size) {
1741                         progress_unlock();
1742                         break;
1743                 }
1744                 entry = *list++;
1745                 (*list_size)--;
1746                 if (!entry->preferred_base) {
1747                         (*processed)++;
1748                         display_progress(progress_state, *processed);
1749                 }
1750                 progress_unlock();
1751
1752                 mem_usage -= free_unpacked(n);
1753                 n->entry = entry;
1754
1755                 while (window_memory_limit &&
1756                        mem_usage > window_memory_limit &&
1757                        count > 1) {
1758                         uint32_t tail = (idx + window - count) % window;
1759                         mem_usage -= free_unpacked(array + tail);
1760                         count--;
1761                 }
1762
1763                 /* We do not compute delta to *create* objects we are not
1764                  * going to pack.
1765                  */
1766                 if (entry->preferred_base)
1767                         goto next;
1768
1769                 /*
1770                  * If the current object is at pack edge, take the depth the
1771                  * objects that depend on the current object into account
1772                  * otherwise they would become too deep.
1773                  */
1774                 max_depth = depth;
1775                 if (entry->delta_child) {
1776                         max_depth -= check_delta_limit(entry, 0);
1777                         if (max_depth <= 0)
1778                                 goto next;
1779                 }
1780
1781                 j = window;
1782                 while (--j > 0) {
1783                         int ret;
1784                         uint32_t other_idx = idx + j;
1785                         struct unpacked *m;
1786                         if (other_idx >= window)
1787                                 other_idx -= window;
1788                         m = array + other_idx;
1789                         if (!m->entry)
1790                                 break;
1791                         ret = try_delta(n, m, max_depth, &mem_usage);
1792                         if (ret < 0)
1793                                 break;
1794                         else if (ret > 0)
1795                                 best_base = other_idx;
1796                 }
1797
1798                 /*
1799                  * If we decided to cache the delta data, then it is best
1800                  * to compress it right away.  First because we have to do
1801                  * it anyway, and doing it here while we're threaded will
1802                  * save a lot of time in the non threaded write phase,
1803                  * as well as allow for caching more deltas within
1804                  * the same cache size limit.
1805                  * ...
1806                  * But only if not writing to stdout, since in that case
1807                  * the network is most likely throttling writes anyway,
1808                  * and therefore it is best to go to the write phase ASAP
1809                  * instead, as we can afford spending more time compressing
1810                  * between writes at that moment.
1811                  */
1812                 if (entry->delta_data && !pack_to_stdout) {
1813                         entry->z_delta_size = do_compress(&entry->delta_data,
1814                                                           entry->delta_size);
1815                         cache_lock();
1816                         delta_cache_size -= entry->delta_size;
1817                         delta_cache_size += entry->z_delta_size;
1818                         cache_unlock();
1819                 }
1820
1821                 /* if we made n a delta, and if n is already at max
1822                  * depth, leaving it in the window is pointless.  we
1823                  * should evict it first.
1824                  */
1825                 if (entry->delta && max_depth <= n->depth)
1826                         continue;
1827
1828                 /*
1829                  * Move the best delta base up in the window, after the
1830                  * currently deltified object, to keep it longer.  It will
1831                  * be the first base object to be attempted next.
1832                  */
1833                 if (entry->delta) {
1834                         struct unpacked swap = array[best_base];
1835                         int dist = (window + idx - best_base) % window;
1836                         int dst = best_base;
1837                         while (dist--) {
1838                                 int src = (dst + 1) % window;
1839                                 array[dst] = array[src];
1840                                 dst = src;
1841                         }
1842                         array[dst] = swap;
1843                 }
1844
1845                 next:
1846                 idx++;
1847                 if (count + 1 < window)
1848                         count++;
1849                 if (idx >= window)
1850                         idx = 0;
1851         }
1852
1853         for (i = 0; i < window; ++i) {
1854                 free_delta_index(array[i].index);
1855                 free(array[i].data);
1856         }
1857         free(array);
1858 }
1859
1860 #ifndef NO_PTHREADS
1861
1862 static void try_to_free_from_threads(size_t size)
1863 {
1864         read_lock();
1865         release_pack_memory(size);
1866         read_unlock();
1867 }
1868
1869 static try_to_free_t old_try_to_free_routine;
1870
1871 /*
1872  * The main thread waits on the condition that (at least) one of the workers
1873  * has stopped working (which is indicated in the .working member of
1874  * struct thread_params).
1875  * When a work thread has completed its work, it sets .working to 0 and
1876  * signals the main thread and waits on the condition that .data_ready
1877  * becomes 1.
1878  */
1879
1880 struct thread_params {
1881         pthread_t thread;
1882         struct object_entry **list;
1883         unsigned list_size;
1884         unsigned remaining;
1885         int window;
1886         int depth;
1887         int working;
1888         int data_ready;
1889         pthread_mutex_t mutex;
1890         pthread_cond_t cond;
1891         unsigned *processed;
1892 };
1893
1894 static pthread_cond_t progress_cond;
1895
1896 /*
1897  * Mutex and conditional variable can't be statically-initialized on Windows.
1898  */
1899 static void init_threaded_search(void)
1900 {
1901         init_recursive_mutex(&read_mutex);
1902         pthread_mutex_init(&cache_mutex, NULL);
1903         pthread_mutex_init(&progress_mutex, NULL);
1904         pthread_cond_init(&progress_cond, NULL);
1905         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1906 }
1907
1908 static void cleanup_threaded_search(void)
1909 {
1910         set_try_to_free_routine(old_try_to_free_routine);
1911         pthread_cond_destroy(&progress_cond);
1912         pthread_mutex_destroy(&read_mutex);
1913         pthread_mutex_destroy(&cache_mutex);
1914         pthread_mutex_destroy(&progress_mutex);
1915 }
1916
1917 static void *threaded_find_deltas(void *arg)
1918 {
1919         struct thread_params *me = arg;
1920
1921         while (me->remaining) {
1922                 find_deltas(me->list, &me->remaining,
1923                             me->window, me->depth, me->processed);
1924
1925                 progress_lock();
1926                 me->working = 0;
1927                 pthread_cond_signal(&progress_cond);
1928                 progress_unlock();
1929
1930                 /*
1931                  * We must not set ->data_ready before we wait on the
1932                  * condition because the main thread may have set it to 1
1933                  * before we get here. In order to be sure that new
1934                  * work is available if we see 1 in ->data_ready, it
1935                  * was initialized to 0 before this thread was spawned
1936                  * and we reset it to 0 right away.
1937                  */
1938                 pthread_mutex_lock(&me->mutex);
1939                 while (!me->data_ready)
1940                         pthread_cond_wait(&me->cond, &me->mutex);
1941                 me->data_ready = 0;
1942                 pthread_mutex_unlock(&me->mutex);
1943         }
1944         /* leave ->working 1 so that this doesn't get more work assigned */
1945         return NULL;
1946 }
1947
1948 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1949                            int window, int depth, unsigned *processed)
1950 {
1951         struct thread_params *p;
1952         int i, ret, active_threads = 0;
1953
1954         init_threaded_search();
1955
1956         if (!delta_search_threads)      /* --threads=0 means autodetect */
1957                 delta_search_threads = online_cpus();
1958         if (delta_search_threads <= 1) {
1959                 find_deltas(list, &list_size, window, depth, processed);
1960                 cleanup_threaded_search();
1961                 return;
1962         }
1963         if (progress > pack_to_stdout)
1964                 fprintf(stderr, "Delta compression using up to %d threads.\n",
1965                                 delta_search_threads);
1966         p = xcalloc(delta_search_threads, sizeof(*p));
1967
1968         /* Partition the work amongst work threads. */
1969         for (i = 0; i < delta_search_threads; i++) {
1970                 unsigned sub_size = list_size / (delta_search_threads - i);
1971
1972                 /* don't use too small segments or no deltas will be found */
1973                 if (sub_size < 2*window && i+1 < delta_search_threads)
1974                         sub_size = 0;
1975
1976                 p[i].window = window;
1977                 p[i].depth = depth;
1978                 p[i].processed = processed;
1979                 p[i].working = 1;
1980                 p[i].data_ready = 0;
1981
1982                 /* try to split chunks on "path" boundaries */
1983                 while (sub_size && sub_size < list_size &&
1984                        list[sub_size]->hash &&
1985                        list[sub_size]->hash == list[sub_size-1]->hash)
1986                         sub_size++;
1987
1988                 p[i].list = list;
1989                 p[i].list_size = sub_size;
1990                 p[i].remaining = sub_size;
1991
1992                 list += sub_size;
1993                 list_size -= sub_size;
1994         }
1995
1996         /* Start work threads. */
1997         for (i = 0; i < delta_search_threads; i++) {
1998                 if (!p[i].list_size)
1999                         continue;
2000                 pthread_mutex_init(&p[i].mutex, NULL);
2001                 pthread_cond_init(&p[i].cond, NULL);
2002                 ret = pthread_create(&p[i].thread, NULL,
2003                                      threaded_find_deltas, &p[i]);
2004                 if (ret)
2005                         die("unable to create thread: %s", strerror(ret));
2006                 active_threads++;
2007         }
2008
2009         /*
2010          * Now let's wait for work completion.  Each time a thread is done
2011          * with its work, we steal half of the remaining work from the
2012          * thread with the largest number of unprocessed objects and give
2013          * it to that newly idle thread.  This ensure good load balancing
2014          * until the remaining object list segments are simply too short
2015          * to be worth splitting anymore.
2016          */
2017         while (active_threads) {
2018                 struct thread_params *target = NULL;
2019                 struct thread_params *victim = NULL;
2020                 unsigned sub_size = 0;
2021
2022                 progress_lock();
2023                 for (;;) {
2024                         for (i = 0; !target && i < delta_search_threads; i++)
2025                                 if (!p[i].working)
2026                                         target = &p[i];
2027                         if (target)
2028                                 break;
2029                         pthread_cond_wait(&progress_cond, &progress_mutex);
2030                 }
2031
2032                 for (i = 0; i < delta_search_threads; i++)
2033                         if (p[i].remaining > 2*window &&
2034                             (!victim || victim->remaining < p[i].remaining))
2035                                 victim = &p[i];
2036                 if (victim) {
2037                         sub_size = victim->remaining / 2;
2038                         list = victim->list + victim->list_size - sub_size;
2039                         while (sub_size && list[0]->hash &&
2040                                list[0]->hash == list[-1]->hash) {
2041                                 list++;
2042                                 sub_size--;
2043                         }
2044                         if (!sub_size) {
2045                                 /*
2046                                  * It is possible for some "paths" to have
2047                                  * so many objects that no hash boundary
2048                                  * might be found.  Let's just steal the
2049                                  * exact half in that case.
2050                                  */
2051                                 sub_size = victim->remaining / 2;
2052                                 list -= sub_size;
2053                         }
2054                         target->list = list;
2055                         victim->list_size -= sub_size;
2056                         victim->remaining -= sub_size;
2057                 }
2058                 target->list_size = sub_size;
2059                 target->remaining = sub_size;
2060                 target->working = 1;
2061                 progress_unlock();
2062
2063                 pthread_mutex_lock(&target->mutex);
2064                 target->data_ready = 1;
2065                 pthread_cond_signal(&target->cond);
2066                 pthread_mutex_unlock(&target->mutex);
2067
2068                 if (!sub_size) {
2069                         pthread_join(target->thread, NULL);
2070                         pthread_cond_destroy(&target->cond);
2071                         pthread_mutex_destroy(&target->mutex);
2072                         active_threads--;
2073                 }
2074         }
2075         cleanup_threaded_search();
2076         free(p);
2077 }
2078
2079 #else
2080 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2081 #endif
2082
2083 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
2084 {
2085         unsigned char peeled[20];
2086
2087         if (starts_with(path, "refs/tags/") && /* is a tag? */
2088             !peel_ref(path, peeled)        && /* peelable? */
2089             packlist_find(&to_pack, peeled, NULL))      /* object packed? */
2090                 add_object_entry(sha1, OBJ_TAG, NULL, 0);
2091         return 0;
2092 }
2093
2094 static void prepare_pack(int window, int depth)
2095 {
2096         struct object_entry **delta_list;
2097         uint32_t i, nr_deltas;
2098         unsigned n;
2099
2100         get_object_details();
2101
2102         /*
2103          * If we're locally repacking then we need to be doubly careful
2104          * from now on in order to make sure no stealth corruption gets
2105          * propagated to the new pack.  Clients receiving streamed packs
2106          * should validate everything they get anyway so no need to incur
2107          * the additional cost here in that case.
2108          */
2109         if (!pack_to_stdout)
2110                 do_check_packed_object_crc = 1;
2111
2112         if (!to_pack.nr_objects || !window || !depth)
2113                 return;
2114
2115         delta_list = xmalloc(to_pack.nr_objects * sizeof(*delta_list));
2116         nr_deltas = n = 0;
2117
2118         for (i = 0; i < to_pack.nr_objects; i++) {
2119                 struct object_entry *entry = to_pack.objects + i;
2120
2121                 if (entry->delta)
2122                         /* This happens if we decided to reuse existing
2123                          * delta from a pack.  "reuse_delta &&" is implied.
2124                          */
2125                         continue;
2126
2127                 if (entry->size < 50)
2128                         continue;
2129
2130                 if (entry->no_try_delta)
2131                         continue;
2132
2133                 if (!entry->preferred_base) {
2134                         nr_deltas++;
2135                         if (entry->type < 0)
2136                                 die("unable to get type of object %s",
2137                                     sha1_to_hex(entry->idx.sha1));
2138                 } else {
2139                         if (entry->type < 0) {
2140                                 /*
2141                                  * This object is not found, but we
2142                                  * don't have to include it anyway.
2143                                  */
2144                                 continue;
2145                         }
2146                 }
2147
2148                 delta_list[n++] = entry;
2149         }
2150
2151         if (nr_deltas && n > 1) {
2152                 unsigned nr_done = 0;
2153                 if (progress)
2154                         progress_state = start_progress(_("Compressing objects"),
2155                                                         nr_deltas);
2156                 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2157                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2158                 stop_progress(&progress_state);
2159                 if (nr_done != nr_deltas)
2160                         die("inconsistency with delta count");
2161         }
2162         free(delta_list);
2163 }
2164
2165 static int git_pack_config(const char *k, const char *v, void *cb)
2166 {
2167         if (!strcmp(k, "pack.window")) {
2168                 window = git_config_int(k, v);
2169                 return 0;
2170         }
2171         if (!strcmp(k, "pack.windowmemory")) {
2172                 window_memory_limit = git_config_ulong(k, v);
2173                 return 0;
2174         }
2175         if (!strcmp(k, "pack.depth")) {
2176                 depth = git_config_int(k, v);
2177                 return 0;
2178         }
2179         if (!strcmp(k, "pack.compression")) {
2180                 int level = git_config_int(k, v);
2181                 if (level == -1)
2182                         level = Z_DEFAULT_COMPRESSION;
2183                 else if (level < 0 || level > Z_BEST_COMPRESSION)
2184                         die("bad pack compression level %d", level);
2185                 pack_compression_level = level;
2186                 pack_compression_seen = 1;
2187                 return 0;
2188         }
2189         if (!strcmp(k, "pack.deltacachesize")) {
2190                 max_delta_cache_size = git_config_int(k, v);
2191                 return 0;
2192         }
2193         if (!strcmp(k, "pack.deltacachelimit")) {
2194                 cache_max_small_delta_size = git_config_int(k, v);
2195                 return 0;
2196         }
2197         if (!strcmp(k, "pack.writebitmaps")) {
2198                 write_bitmap_index = git_config_bool(k, v);
2199                 return 0;
2200         }
2201         if (!strcmp(k, "pack.writebitmaphashcache")) {
2202                 if (git_config_bool(k, v))
2203                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2204                 else
2205                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2206         }
2207         if (!strcmp(k, "pack.usebitmaps")) {
2208                 use_bitmap_index = git_config_bool(k, v);
2209                 return 0;
2210         }
2211         if (!strcmp(k, "pack.threads")) {
2212                 delta_search_threads = git_config_int(k, v);
2213                 if (delta_search_threads < 0)
2214                         die("invalid number of threads specified (%d)",
2215                             delta_search_threads);
2216 #ifdef NO_PTHREADS
2217                 if (delta_search_threads != 1)
2218                         warning("no threads support, ignoring %s", k);
2219 #endif
2220                 return 0;
2221         }
2222         if (!strcmp(k, "pack.indexversion")) {
2223                 pack_idx_opts.version = git_config_int(k, v);
2224                 if (pack_idx_opts.version > 2)
2225                         die("bad pack.indexversion=%"PRIu32,
2226                             pack_idx_opts.version);
2227                 return 0;
2228         }
2229         return git_default_config(k, v, cb);
2230 }
2231
2232 static void read_object_list_from_stdin(void)
2233 {
2234         char line[40 + 1 + PATH_MAX + 2];
2235         unsigned char sha1[20];
2236
2237         for (;;) {
2238                 if (!fgets(line, sizeof(line), stdin)) {
2239                         if (feof(stdin))
2240                                 break;
2241                         if (!ferror(stdin))
2242                                 die("fgets returned NULL, not EOF, not error!");
2243                         if (errno != EINTR)
2244                                 die_errno("fgets");
2245                         clearerr(stdin);
2246                         continue;
2247                 }
2248                 if (line[0] == '-') {
2249                         if (get_sha1_hex(line+1, sha1))
2250                                 die("expected edge sha1, got garbage:\n %s",
2251                                     line);
2252                         add_preferred_base(sha1);
2253                         continue;
2254                 }
2255                 if (get_sha1_hex(line, sha1))
2256                         die("expected sha1, got garbage:\n %s", line);
2257
2258                 add_preferred_base_object(line+41);
2259                 add_object_entry(sha1, 0, line+41, 0);
2260         }
2261 }
2262
2263 #define OBJECT_ADDED (1u<<20)
2264
2265 static void show_commit(struct commit *commit, void *data)
2266 {
2267         add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
2268         commit->object.flags |= OBJECT_ADDED;
2269
2270         if (write_bitmap_index)
2271                 index_commit_for_bitmap(commit);
2272 }
2273
2274 static void show_object(struct object *obj,
2275                         const struct name_path *path, const char *last,
2276                         void *data)
2277 {
2278         char *name = path_name(path, last);
2279
2280         add_preferred_base_object(name);
2281         add_object_entry(obj->sha1, obj->type, name, 0);
2282         obj->flags |= OBJECT_ADDED;
2283
2284         /*
2285          * We will have generated the hash from the name,
2286          * but not saved a pointer to it - we can free it
2287          */
2288         free((char *)name);
2289 }
2290
2291 static void show_edge(struct commit *commit)
2292 {
2293         add_preferred_base(commit->object.sha1);
2294 }
2295
2296 struct in_pack_object {
2297         off_t offset;
2298         struct object *object;
2299 };
2300
2301 struct in_pack {
2302         int alloc;
2303         int nr;
2304         struct in_pack_object *array;
2305 };
2306
2307 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2308 {
2309         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
2310         in_pack->array[in_pack->nr].object = object;
2311         in_pack->nr++;
2312 }
2313
2314 /*
2315  * Compare the objects in the offset order, in order to emulate the
2316  * "git rev-list --objects" output that produced the pack originally.
2317  */
2318 static int ofscmp(const void *a_, const void *b_)
2319 {
2320         struct in_pack_object *a = (struct in_pack_object *)a_;
2321         struct in_pack_object *b = (struct in_pack_object *)b_;
2322
2323         if (a->offset < b->offset)
2324                 return -1;
2325         else if (a->offset > b->offset)
2326                 return 1;
2327         else
2328                 return hashcmp(a->object->sha1, b->object->sha1);
2329 }
2330
2331 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2332 {
2333         struct packed_git *p;
2334         struct in_pack in_pack;
2335         uint32_t i;
2336
2337         memset(&in_pack, 0, sizeof(in_pack));
2338
2339         for (p = packed_git; p; p = p->next) {
2340                 const unsigned char *sha1;
2341                 struct object *o;
2342
2343                 if (!p->pack_local || p->pack_keep)
2344                         continue;
2345                 if (open_pack_index(p))
2346                         die("cannot open pack index");
2347
2348                 ALLOC_GROW(in_pack.array,
2349                            in_pack.nr + p->num_objects,
2350                            in_pack.alloc);
2351
2352                 for (i = 0; i < p->num_objects; i++) {
2353                         sha1 = nth_packed_object_sha1(p, i);
2354                         o = lookup_unknown_object(sha1);
2355                         if (!(o->flags & OBJECT_ADDED))
2356                                 mark_in_pack_object(o, p, &in_pack);
2357                         o->flags |= OBJECT_ADDED;
2358                 }
2359         }
2360
2361         if (in_pack.nr) {
2362                 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2363                       ofscmp);
2364                 for (i = 0; i < in_pack.nr; i++) {
2365                         struct object *o = in_pack.array[i].object;
2366                         add_object_entry(o->sha1, o->type, "", 0);
2367                 }
2368         }
2369         free(in_pack.array);
2370 }
2371
2372 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2373 {
2374         static struct packed_git *last_found = (void *)1;
2375         struct packed_git *p;
2376
2377         p = (last_found != (void *)1) ? last_found : packed_git;
2378
2379         while (p) {
2380                 if ((!p->pack_local || p->pack_keep) &&
2381                         find_pack_entry_one(sha1, p)) {
2382                         last_found = p;
2383                         return 1;
2384                 }
2385                 if (p == last_found)
2386                         p = packed_git;
2387                 else
2388                         p = p->next;
2389                 if (p == last_found)
2390                         p = p->next;
2391         }
2392         return 0;
2393 }
2394
2395 static void loosen_unused_packed_objects(struct rev_info *revs)
2396 {
2397         struct packed_git *p;
2398         uint32_t i;
2399         const unsigned char *sha1;
2400
2401         for (p = packed_git; p; p = p->next) {
2402                 if (!p->pack_local || p->pack_keep)
2403                         continue;
2404
2405                 if (unpack_unreachable_expiration &&
2406                     p->mtime < unpack_unreachable_expiration)
2407                         continue;
2408
2409                 if (open_pack_index(p))
2410                         die("cannot open pack index");
2411
2412                 for (i = 0; i < p->num_objects; i++) {
2413                         sha1 = nth_packed_object_sha1(p, i);
2414                         if (!packlist_find(&to_pack, sha1, NULL) &&
2415                                 !has_sha1_pack_kept_or_nonlocal(sha1))
2416                                 if (force_object_loose(sha1, p->mtime))
2417                                         die("unable to force loose object");
2418                 }
2419         }
2420 }
2421
2422 static int get_object_list_from_bitmap(struct rev_info *revs)
2423 {
2424         if (prepare_bitmap_walk(revs) < 0)
2425                 return -1;
2426
2427         if (!reuse_partial_packfile_from_bitmap(
2428                         &reuse_packfile,
2429                         &reuse_packfile_objects,
2430                         &reuse_packfile_offset)) {
2431                 assert(reuse_packfile_objects);
2432                 nr_result += reuse_packfile_objects;
2433
2434                 if (progress) {
2435                         fprintf(stderr, "Reusing existing pack: %d, done.\n",
2436                                 reuse_packfile_objects);
2437                         fflush(stderr);
2438                 }
2439         }
2440
2441         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2442         return 0;
2443 }
2444
2445 static void get_object_list(int ac, const char **av)
2446 {
2447         struct rev_info revs;
2448         char line[1000];
2449         int flags = 0;
2450
2451         init_revisions(&revs, NULL);
2452         save_commit_buffer = 0;
2453         setup_revisions(ac, av, &revs, NULL);
2454
2455         while (fgets(line, sizeof(line), stdin) != NULL) {
2456                 int len = strlen(line);
2457                 if (len && line[len - 1] == '\n')
2458                         line[--len] = 0;
2459                 if (!len)
2460                         break;
2461                 if (*line == '-') {
2462                         if (!strcmp(line, "--not")) {
2463                                 flags ^= UNINTERESTING;
2464                                 write_bitmap_index = 0;
2465                                 continue;
2466                         }
2467                         die("not a rev '%s'", line);
2468                 }
2469                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2470                         die("bad revision '%s'", line);
2471         }
2472
2473         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2474                 return;
2475
2476         if (prepare_revision_walk(&revs))
2477                 die("revision walk setup failed");
2478         mark_edges_uninteresting(&revs, show_edge);
2479         traverse_commit_list(&revs, show_commit, show_object, NULL);
2480
2481         if (keep_unreachable)
2482                 add_objects_in_unpacked_packs(&revs);
2483         if (unpack_unreachable)
2484                 loosen_unused_packed_objects(&revs);
2485 }
2486
2487 static int option_parse_index_version(const struct option *opt,
2488                                       const char *arg, int unset)
2489 {
2490         char *c;
2491         const char *val = arg;
2492         pack_idx_opts.version = strtoul(val, &c, 10);
2493         if (pack_idx_opts.version > 2)
2494                 die(_("unsupported index version %s"), val);
2495         if (*c == ',' && c[1])
2496                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2497         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2498                 die(_("bad index version '%s'"), val);
2499         return 0;
2500 }
2501
2502 static int option_parse_unpack_unreachable(const struct option *opt,
2503                                            const char *arg, int unset)
2504 {
2505         if (unset) {
2506                 unpack_unreachable = 0;
2507                 unpack_unreachable_expiration = 0;
2508         }
2509         else {
2510                 unpack_unreachable = 1;
2511                 if (arg)
2512                         unpack_unreachable_expiration = approxidate(arg);
2513         }
2514         return 0;
2515 }
2516
2517 static int option_parse_ulong(const struct option *opt,
2518                               const char *arg, int unset)
2519 {
2520         if (unset)
2521                 die(_("option %s does not accept negative form"),
2522                     opt->long_name);
2523
2524         if (!git_parse_ulong(arg, opt->value))
2525                 die(_("unable to parse value '%s' for option %s"),
2526                     arg, opt->long_name);
2527         return 0;
2528 }
2529
2530 #define OPT_ULONG(s, l, v, h) \
2531         { OPTION_CALLBACK, (s), (l), (v), "n", (h),     \
2532           PARSE_OPT_NONEG, option_parse_ulong }
2533
2534 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2535 {
2536         int use_internal_rev_list = 0;
2537         int thin = 0;
2538         int all_progress_implied = 0;
2539         const char *rp_av[6];
2540         int rp_ac = 0;
2541         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2542         struct option pack_objects_options[] = {
2543                 OPT_SET_INT('q', "quiet", &progress,
2544                             N_("do not show progress meter"), 0),
2545                 OPT_SET_INT(0, "progress", &progress,
2546                             N_("show progress meter"), 1),
2547                 OPT_SET_INT(0, "all-progress", &progress,
2548                             N_("show progress meter during object writing phase"), 2),
2549                 OPT_BOOL(0, "all-progress-implied",
2550                          &all_progress_implied,
2551                          N_("similar to --all-progress when progress meter is shown")),
2552                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2553                   N_("write the pack index file in the specified idx format version"),
2554                   0, option_parse_index_version },
2555                 OPT_ULONG(0, "max-pack-size", &pack_size_limit,
2556                           N_("maximum size of each output pack file")),
2557                 OPT_BOOL(0, "local", &local,
2558                          N_("ignore borrowed objects from alternate object store")),
2559                 OPT_BOOL(0, "incremental", &incremental,
2560                          N_("ignore packed objects")),
2561                 OPT_INTEGER(0, "window", &window,
2562                             N_("limit pack window by objects")),
2563                 OPT_ULONG(0, "window-memory", &window_memory_limit,
2564                           N_("limit pack window by memory in addition to object limit")),
2565                 OPT_INTEGER(0, "depth", &depth,
2566                             N_("maximum length of delta chain allowed in the resulting pack")),
2567                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2568                          N_("reuse existing deltas")),
2569                 OPT_BOOL(0, "reuse-object", &reuse_object,
2570                          N_("reuse existing objects")),
2571                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2572                          N_("use OFS_DELTA objects")),
2573                 OPT_INTEGER(0, "threads", &delta_search_threads,
2574                             N_("use threads when searching for best delta matches")),
2575                 OPT_BOOL(0, "non-empty", &non_empty,
2576                          N_("do not create an empty pack output")),
2577                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2578                          N_("read revision arguments from standard input")),
2579                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2580                   N_("limit the objects to those that are not yet packed"),
2581                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2582                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2583                   N_("include objects reachable from any reference"),
2584                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2585                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2586                   N_("include objects referred by reflog entries"),
2587                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2588                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2589                          N_("output pack to stdout")),
2590                 OPT_BOOL(0, "include-tag", &include_tag,
2591                          N_("include tag objects that refer to objects to be packed")),
2592                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2593                          N_("keep unreachable objects")),
2594                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2595                   N_("unpack unreachable objects newer than <time>"),
2596                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2597                 OPT_BOOL(0, "thin", &thin,
2598                          N_("create thin packs")),
2599                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2600                          N_("ignore packs that have companion .keep file")),
2601                 OPT_INTEGER(0, "compression", &pack_compression_level,
2602                             N_("pack compression level")),
2603                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2604                             N_("do not hide commits by grafts"), 0),
2605                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2606                          N_("use a bitmap index if available to speed up counting objects")),
2607                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2608                          N_("write a bitmap index together with the pack index")),
2609                 OPT_END(),
2610         };
2611
2612         check_replace_refs = 0;
2613
2614         reset_pack_idx_option(&pack_idx_opts);
2615         git_config(git_pack_config, NULL);
2616         if (!pack_compression_seen && core_compression_seen)
2617                 pack_compression_level = core_compression_level;
2618
2619         progress = isatty(2);
2620         argc = parse_options(argc, argv, prefix, pack_objects_options,
2621                              pack_usage, 0);
2622
2623         if (argc) {
2624                 base_name = argv[0];
2625                 argc--;
2626         }
2627         if (pack_to_stdout != !base_name || argc)
2628                 usage_with_options(pack_usage, pack_objects_options);
2629
2630         rp_av[rp_ac++] = "pack-objects";
2631         if (thin) {
2632                 use_internal_rev_list = 1;
2633                 rp_av[rp_ac++] = "--objects-edge";
2634         } else
2635                 rp_av[rp_ac++] = "--objects";
2636
2637         if (rev_list_all) {
2638                 use_internal_rev_list = 1;
2639                 rp_av[rp_ac++] = "--all";
2640         }
2641         if (rev_list_reflog) {
2642                 use_internal_rev_list = 1;
2643                 rp_av[rp_ac++] = "--reflog";
2644         }
2645         if (rev_list_unpacked) {
2646                 use_internal_rev_list = 1;
2647                 rp_av[rp_ac++] = "--unpacked";
2648         }
2649
2650         if (!reuse_object)
2651                 reuse_delta = 0;
2652         if (pack_compression_level == -1)
2653                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2654         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2655                 die("bad pack compression level %d", pack_compression_level);
2656 #ifdef NO_PTHREADS
2657         if (delta_search_threads != 1)
2658                 warning("no threads support, ignoring --threads");
2659 #endif
2660         if (!pack_to_stdout && !pack_size_limit)
2661                 pack_size_limit = pack_size_limit_cfg;
2662         if (pack_to_stdout && pack_size_limit)
2663                 die("--max-pack-size cannot be used to build a pack for transfer.");
2664         if (pack_size_limit && pack_size_limit < 1024*1024) {
2665                 warning("minimum pack size limit is 1 MiB");
2666                 pack_size_limit = 1024*1024;
2667         }
2668
2669         if (!pack_to_stdout && thin)
2670                 die("--thin cannot be used to build an indexable pack.");
2671
2672         if (keep_unreachable && unpack_unreachable)
2673                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2674
2675         if (!use_internal_rev_list || !pack_to_stdout || is_repository_shallow())
2676                 use_bitmap_index = 0;
2677
2678         if (pack_to_stdout || !rev_list_all)
2679                 write_bitmap_index = 0;
2680
2681         if (progress && all_progress_implied)
2682                 progress = 2;
2683
2684         prepare_packed_git();
2685
2686         if (progress)
2687                 progress_state = start_progress(_("Counting objects"), 0);
2688         if (!use_internal_rev_list)
2689                 read_object_list_from_stdin();
2690         else {
2691                 rp_av[rp_ac] = NULL;
2692                 get_object_list(rp_ac, rp_av);
2693         }
2694         cleanup_preferred_base();
2695         if (include_tag && nr_result)
2696                 for_each_ref(add_ref_tag, NULL);
2697         stop_progress(&progress_state);
2698
2699         if (non_empty && !nr_result)
2700                 return 0;
2701         if (nr_result)
2702                 prepare_pack(window, depth);
2703         write_pack_file();
2704         if (progress)
2705                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2706                         " reused %"PRIu32" (delta %"PRIu32")\n",
2707                         written, written_delta, reused, reused_delta);
2708         return 0;
2709 }