sha1_file: squelch "packfile cannot be accessed" warnings
[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, total;
712         int fd;
713
714         if (!is_pack_valid(reuse_packfile))
715                 die("packfile is invalid: %s", reuse_packfile->pack_name);
716
717         fd = git_open_noatime(reuse_packfile->pack_name);
718         if (fd < 0)
719                 die_errno("unable to open packfile for reuse: %s",
720                           reuse_packfile->pack_name);
721
722         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
723                 die_errno("unable to seek in reused packfile");
724
725         if (reuse_packfile_offset < 0)
726                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
727
728         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
729
730         while (to_write) {
731                 int read_pack = xread(fd, buffer, sizeof(buffer));
732
733                 if (read_pack <= 0)
734                         die_errno("unable to read from reused packfile");
735
736                 if (read_pack > to_write)
737                         read_pack = to_write;
738
739                 sha1write(f, buffer, read_pack);
740                 to_write -= read_pack;
741
742                 /*
743                  * We don't know the actual number of objects written,
744                  * only how many bytes written, how many bytes total, and
745                  * how many objects total. So we can fake it by pretending all
746                  * objects we are writing are the same size. This gives us a
747                  * smooth progress meter, and at the end it matches the true
748                  * answer.
749                  */
750                 written = reuse_packfile_objects *
751                                 (((double)(total - to_write)) / total);
752                 display_progress(progress_state, written);
753         }
754
755         close(fd);
756         written = reuse_packfile_objects;
757         display_progress(progress_state, written);
758         return reuse_packfile_offset - sizeof(struct pack_header);
759 }
760
761 static void write_pack_file(void)
762 {
763         uint32_t i = 0, j;
764         struct sha1file *f;
765         off_t offset;
766         uint32_t nr_remaining = nr_result;
767         time_t last_mtime = 0;
768         struct object_entry **write_order;
769
770         if (progress > pack_to_stdout)
771                 progress_state = start_progress(_("Writing objects"), nr_result);
772         written_list = xmalloc(to_pack.nr_objects * sizeof(*written_list));
773         write_order = compute_write_order();
774
775         do {
776                 unsigned char sha1[20];
777                 char *pack_tmp_name = NULL;
778
779                 if (pack_to_stdout)
780                         f = sha1fd_throughput(1, "<stdout>", progress_state);
781                 else
782                         f = create_tmp_packfile(&pack_tmp_name);
783
784                 offset = write_pack_header(f, nr_remaining);
785
786                 if (reuse_packfile) {
787                         off_t packfile_size;
788                         assert(pack_to_stdout);
789
790                         packfile_size = write_reused_pack(f);
791                         offset += packfile_size;
792                 }
793
794                 nr_written = 0;
795                 for (; i < to_pack.nr_objects; i++) {
796                         struct object_entry *e = write_order[i];
797                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
798                                 break;
799                         display_progress(progress_state, written);
800                 }
801
802                 /*
803                  * Did we write the wrong # entries in the header?
804                  * If so, rewrite it like in fast-import
805                  */
806                 if (pack_to_stdout) {
807                         sha1close(f, sha1, CSUM_CLOSE);
808                 } else if (nr_written == nr_remaining) {
809                         sha1close(f, sha1, CSUM_FSYNC);
810                 } else {
811                         int fd = sha1close(f, sha1, 0);
812                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
813                                                  nr_written, sha1, offset);
814                         close(fd);
815                 }
816
817                 if (!pack_to_stdout) {
818                         struct stat st;
819                         struct strbuf tmpname = STRBUF_INIT;
820
821                         /*
822                          * Packs are runtime accessed in their mtime
823                          * order since newer packs are more likely to contain
824                          * younger objects.  So if we are creating multiple
825                          * packs then we should modify the mtime of later ones
826                          * to preserve this property.
827                          */
828                         if (stat(pack_tmp_name, &st) < 0) {
829                                 warning("failed to stat %s: %s",
830                                         pack_tmp_name, strerror(errno));
831                         } else if (!last_mtime) {
832                                 last_mtime = st.st_mtime;
833                         } else {
834                                 struct utimbuf utb;
835                                 utb.actime = st.st_atime;
836                                 utb.modtime = --last_mtime;
837                                 if (utime(pack_tmp_name, &utb) < 0)
838                                         warning("failed utime() on %s: %s",
839                                                 pack_tmp_name, strerror(errno));
840                         }
841
842                         strbuf_addf(&tmpname, "%s-", base_name);
843
844                         if (write_bitmap_index) {
845                                 bitmap_writer_set_checksum(sha1);
846                                 bitmap_writer_build_type_index(written_list, nr_written);
847                         }
848
849                         finish_tmp_packfile(&tmpname, pack_tmp_name,
850                                             written_list, nr_written,
851                                             &pack_idx_opts, sha1);
852
853                         if (write_bitmap_index) {
854                                 strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
855
856                                 stop_progress(&progress_state);
857
858                                 bitmap_writer_show_progress(progress);
859                                 bitmap_writer_reuse_bitmaps(&to_pack);
860                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
861                                 bitmap_writer_build(&to_pack);
862                                 bitmap_writer_finish(written_list, nr_written,
863                                                      tmpname.buf, write_bitmap_options);
864                                 write_bitmap_index = 0;
865                         }
866
867                         strbuf_release(&tmpname);
868                         free(pack_tmp_name);
869                         puts(sha1_to_hex(sha1));
870                 }
871
872                 /* mark written objects as written to previous pack */
873                 for (j = 0; j < nr_written; j++) {
874                         written_list[j]->offset = (off_t)-1;
875                 }
876                 nr_remaining -= nr_written;
877         } while (nr_remaining && i < to_pack.nr_objects);
878
879         free(written_list);
880         free(write_order);
881         stop_progress(&progress_state);
882         if (written != nr_result)
883                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
884                         written, nr_result);
885 }
886
887 static void setup_delta_attr_check(struct git_attr_check *check)
888 {
889         static struct git_attr *attr_delta;
890
891         if (!attr_delta)
892                 attr_delta = git_attr("delta");
893
894         check[0].attr = attr_delta;
895 }
896
897 static int no_try_delta(const char *path)
898 {
899         struct git_attr_check check[1];
900
901         setup_delta_attr_check(check);
902         if (git_check_attr(path, ARRAY_SIZE(check), check))
903                 return 0;
904         if (ATTR_FALSE(check->value))
905                 return 1;
906         return 0;
907 }
908
909 /*
910  * When adding an object, check whether we have already added it
911  * to our packing list. If so, we can skip. However, if we are
912  * being asked to excludei t, but the previous mention was to include
913  * it, make sure to adjust its flags and tweak our numbers accordingly.
914  *
915  * As an optimization, we pass out the index position where we would have
916  * found the item, since that saves us from having to look it up again a
917  * few lines later when we want to add the new entry.
918  */
919 static int have_duplicate_entry(const unsigned char *sha1,
920                                 int exclude,
921                                 uint32_t *index_pos)
922 {
923         struct object_entry *entry;
924
925         entry = packlist_find(&to_pack, sha1, index_pos);
926         if (!entry)
927                 return 0;
928
929         if (exclude) {
930                 if (!entry->preferred_base)
931                         nr_result--;
932                 entry->preferred_base = 1;
933         }
934
935         return 1;
936 }
937
938 /*
939  * Check whether we want the object in the pack (e.g., we do not want
940  * objects found in non-local stores if the "--local" option was used).
941  *
942  * As a side effect of this check, we will find the packed version of this
943  * object, if any. We therefore pass out the pack information to avoid having
944  * to look it up again later.
945  */
946 static int want_object_in_pack(const unsigned char *sha1,
947                                int exclude,
948                                struct packed_git **found_pack,
949                                off_t *found_offset)
950 {
951         struct packed_git *p;
952
953         if (!exclude && local && has_loose_object_nonlocal(sha1))
954                 return 0;
955
956         *found_pack = NULL;
957         *found_offset = 0;
958
959         for (p = packed_git; p; p = p->next) {
960                 off_t offset = find_pack_entry_one(sha1, p);
961                 if (offset) {
962                         if (!*found_pack) {
963                                 if (!is_pack_valid(p))
964                                         continue;
965                                 *found_offset = offset;
966                                 *found_pack = p;
967                         }
968                         if (exclude)
969                                 return 1;
970                         if (incremental)
971                                 return 0;
972                         if (local && !p->pack_local)
973                                 return 0;
974                         if (ignore_packed_keep && p->pack_local && p->pack_keep)
975                                 return 0;
976                 }
977         }
978
979         return 1;
980 }
981
982 static void create_object_entry(const unsigned char *sha1,
983                                 enum object_type type,
984                                 uint32_t hash,
985                                 int exclude,
986                                 int no_try_delta,
987                                 uint32_t index_pos,
988                                 struct packed_git *found_pack,
989                                 off_t found_offset)
990 {
991         struct object_entry *entry;
992
993         entry = packlist_alloc(&to_pack, sha1, index_pos);
994         entry->hash = hash;
995         if (type)
996                 entry->type = type;
997         if (exclude)
998                 entry->preferred_base = 1;
999         else
1000                 nr_result++;
1001         if (found_pack) {
1002                 entry->in_pack = found_pack;
1003                 entry->in_pack_offset = found_offset;
1004         }
1005
1006         entry->no_try_delta = no_try_delta;
1007 }
1008
1009 static const char no_closure_warning[] = N_(
1010 "disabling bitmap writing, as some objects are not being packed"
1011 );
1012
1013 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1014                             const char *name, int exclude)
1015 {
1016         struct packed_git *found_pack;
1017         off_t found_offset;
1018         uint32_t index_pos;
1019
1020         if (have_duplicate_entry(sha1, exclude, &index_pos))
1021                 return 0;
1022
1023         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1024                 /* The pack is missing an object, so it will not have closure */
1025                 if (write_bitmap_index) {
1026                         warning(_(no_closure_warning));
1027                         write_bitmap_index = 0;
1028                 }
1029                 return 0;
1030         }
1031
1032         create_object_entry(sha1, type, pack_name_hash(name),
1033                             exclude, name && no_try_delta(name),
1034                             index_pos, found_pack, found_offset);
1035
1036         display_progress(progress_state, nr_result);
1037         return 1;
1038 }
1039
1040 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1041                                         enum object_type type,
1042                                         int flags, uint32_t name_hash,
1043                                         struct packed_git *pack, off_t offset)
1044 {
1045         uint32_t index_pos;
1046
1047         if (have_duplicate_entry(sha1, 0, &index_pos))
1048                 return 0;
1049
1050         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1051
1052         display_progress(progress_state, nr_result);
1053         return 1;
1054 }
1055
1056 struct pbase_tree_cache {
1057         unsigned char sha1[20];
1058         int ref;
1059         int temporary;
1060         void *tree_data;
1061         unsigned long tree_size;
1062 };
1063
1064 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1065 static int pbase_tree_cache_ix(const unsigned char *sha1)
1066 {
1067         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1068 }
1069 static int pbase_tree_cache_ix_incr(int ix)
1070 {
1071         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1072 }
1073
1074 static struct pbase_tree {
1075         struct pbase_tree *next;
1076         /* This is a phony "cache" entry; we are not
1077          * going to evict it or find it through _get()
1078          * mechanism -- this is for the toplevel node that
1079          * would almost always change with any commit.
1080          */
1081         struct pbase_tree_cache pcache;
1082 } *pbase_tree;
1083
1084 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1085 {
1086         struct pbase_tree_cache *ent, *nent;
1087         void *data;
1088         unsigned long size;
1089         enum object_type type;
1090         int neigh;
1091         int my_ix = pbase_tree_cache_ix(sha1);
1092         int available_ix = -1;
1093
1094         /* pbase-tree-cache acts as a limited hashtable.
1095          * your object will be found at your index or within a few
1096          * slots after that slot if it is cached.
1097          */
1098         for (neigh = 0; neigh < 8; neigh++) {
1099                 ent = pbase_tree_cache[my_ix];
1100                 if (ent && !hashcmp(ent->sha1, sha1)) {
1101                         ent->ref++;
1102                         return ent;
1103                 }
1104                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1105                          ((0 <= available_ix) &&
1106                           (!ent && pbase_tree_cache[available_ix])))
1107                         available_ix = my_ix;
1108                 if (!ent)
1109                         break;
1110                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1111         }
1112
1113         /* Did not find one.  Either we got a bogus request or
1114          * we need to read and perhaps cache.
1115          */
1116         data = read_sha1_file(sha1, &type, &size);
1117         if (!data)
1118                 return NULL;
1119         if (type != OBJ_TREE) {
1120                 free(data);
1121                 return NULL;
1122         }
1123
1124         /* We need to either cache or return a throwaway copy */
1125
1126         if (available_ix < 0)
1127                 ent = NULL;
1128         else {
1129                 ent = pbase_tree_cache[available_ix];
1130                 my_ix = available_ix;
1131         }
1132
1133         if (!ent) {
1134                 nent = xmalloc(sizeof(*nent));
1135                 nent->temporary = (available_ix < 0);
1136         }
1137         else {
1138                 /* evict and reuse */
1139                 free(ent->tree_data);
1140                 nent = ent;
1141         }
1142         hashcpy(nent->sha1, sha1);
1143         nent->tree_data = data;
1144         nent->tree_size = size;
1145         nent->ref = 1;
1146         if (!nent->temporary)
1147                 pbase_tree_cache[my_ix] = nent;
1148         return nent;
1149 }
1150
1151 static void pbase_tree_put(struct pbase_tree_cache *cache)
1152 {
1153         if (!cache->temporary) {
1154                 cache->ref--;
1155                 return;
1156         }
1157         free(cache->tree_data);
1158         free(cache);
1159 }
1160
1161 static int name_cmp_len(const char *name)
1162 {
1163         int i;
1164         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1165                 ;
1166         return i;
1167 }
1168
1169 static void add_pbase_object(struct tree_desc *tree,
1170                              const char *name,
1171                              int cmplen,
1172                              const char *fullname)
1173 {
1174         struct name_entry entry;
1175         int cmp;
1176
1177         while (tree_entry(tree,&entry)) {
1178                 if (S_ISGITLINK(entry.mode))
1179                         continue;
1180                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1181                       memcmp(name, entry.path, cmplen);
1182                 if (cmp > 0)
1183                         continue;
1184                 if (cmp < 0)
1185                         return;
1186                 if (name[cmplen] != '/') {
1187                         add_object_entry(entry.sha1,
1188                                          object_type(entry.mode),
1189                                          fullname, 1);
1190                         return;
1191                 }
1192                 if (S_ISDIR(entry.mode)) {
1193                         struct tree_desc sub;
1194                         struct pbase_tree_cache *tree;
1195                         const char *down = name+cmplen+1;
1196                         int downlen = name_cmp_len(down);
1197
1198                         tree = pbase_tree_get(entry.sha1);
1199                         if (!tree)
1200                                 return;
1201                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1202
1203                         add_pbase_object(&sub, down, downlen, fullname);
1204                         pbase_tree_put(tree);
1205                 }
1206         }
1207 }
1208
1209 static unsigned *done_pbase_paths;
1210 static int done_pbase_paths_num;
1211 static int done_pbase_paths_alloc;
1212 static int done_pbase_path_pos(unsigned hash)
1213 {
1214         int lo = 0;
1215         int hi = done_pbase_paths_num;
1216         while (lo < hi) {
1217                 int mi = (hi + lo) / 2;
1218                 if (done_pbase_paths[mi] == hash)
1219                         return mi;
1220                 if (done_pbase_paths[mi] < hash)
1221                         hi = mi;
1222                 else
1223                         lo = mi + 1;
1224         }
1225         return -lo-1;
1226 }
1227
1228 static int check_pbase_path(unsigned hash)
1229 {
1230         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1231         if (0 <= pos)
1232                 return 1;
1233         pos = -pos - 1;
1234         ALLOC_GROW(done_pbase_paths,
1235                    done_pbase_paths_num + 1,
1236                    done_pbase_paths_alloc);
1237         done_pbase_paths_num++;
1238         if (pos < done_pbase_paths_num)
1239                 memmove(done_pbase_paths + pos + 1,
1240                         done_pbase_paths + pos,
1241                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1242         done_pbase_paths[pos] = hash;
1243         return 0;
1244 }
1245
1246 static void add_preferred_base_object(const char *name)
1247 {
1248         struct pbase_tree *it;
1249         int cmplen;
1250         unsigned hash = pack_name_hash(name);
1251
1252         if (!num_preferred_base || check_pbase_path(hash))
1253                 return;
1254
1255         cmplen = name_cmp_len(name);
1256         for (it = pbase_tree; it; it = it->next) {
1257                 if (cmplen == 0) {
1258                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1259                 }
1260                 else {
1261                         struct tree_desc tree;
1262                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1263                         add_pbase_object(&tree, name, cmplen, name);
1264                 }
1265         }
1266 }
1267
1268 static void add_preferred_base(unsigned char *sha1)
1269 {
1270         struct pbase_tree *it;
1271         void *data;
1272         unsigned long size;
1273         unsigned char tree_sha1[20];
1274
1275         if (window <= num_preferred_base++)
1276                 return;
1277
1278         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1279         if (!data)
1280                 return;
1281
1282         for (it = pbase_tree; it; it = it->next) {
1283                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1284                         free(data);
1285                         return;
1286                 }
1287         }
1288
1289         it = xcalloc(1, sizeof(*it));
1290         it->next = pbase_tree;
1291         pbase_tree = it;
1292
1293         hashcpy(it->pcache.sha1, tree_sha1);
1294         it->pcache.tree_data = data;
1295         it->pcache.tree_size = size;
1296 }
1297
1298 static void cleanup_preferred_base(void)
1299 {
1300         struct pbase_tree *it;
1301         unsigned i;
1302
1303         it = pbase_tree;
1304         pbase_tree = NULL;
1305         while (it) {
1306                 struct pbase_tree *this = it;
1307                 it = this->next;
1308                 free(this->pcache.tree_data);
1309                 free(this);
1310         }
1311
1312         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1313                 if (!pbase_tree_cache[i])
1314                         continue;
1315                 free(pbase_tree_cache[i]->tree_data);
1316                 free(pbase_tree_cache[i]);
1317                 pbase_tree_cache[i] = NULL;
1318         }
1319
1320         free(done_pbase_paths);
1321         done_pbase_paths = NULL;
1322         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1323 }
1324
1325 static void check_object(struct object_entry *entry)
1326 {
1327         if (entry->in_pack) {
1328                 struct packed_git *p = entry->in_pack;
1329                 struct pack_window *w_curs = NULL;
1330                 const unsigned char *base_ref = NULL;
1331                 struct object_entry *base_entry;
1332                 unsigned long used, used_0;
1333                 unsigned long avail;
1334                 off_t ofs;
1335                 unsigned char *buf, c;
1336
1337                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1338
1339                 /*
1340                  * We want in_pack_type even if we do not reuse delta
1341                  * since non-delta representations could still be reused.
1342                  */
1343                 used = unpack_object_header_buffer(buf, avail,
1344                                                    &entry->in_pack_type,
1345                                                    &entry->size);
1346                 if (used == 0)
1347                         goto give_up;
1348
1349                 /*
1350                  * Determine if this is a delta and if so whether we can
1351                  * reuse it or not.  Otherwise let's find out as cheaply as
1352                  * possible what the actual type and size for this object is.
1353                  */
1354                 switch (entry->in_pack_type) {
1355                 default:
1356                         /* Not a delta hence we've already got all we need. */
1357                         entry->type = entry->in_pack_type;
1358                         entry->in_pack_header_size = used;
1359                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1360                                 goto give_up;
1361                         unuse_pack(&w_curs);
1362                         return;
1363                 case OBJ_REF_DELTA:
1364                         if (reuse_delta && !entry->preferred_base)
1365                                 base_ref = use_pack(p, &w_curs,
1366                                                 entry->in_pack_offset + used, NULL);
1367                         entry->in_pack_header_size = used + 20;
1368                         break;
1369                 case OBJ_OFS_DELTA:
1370                         buf = use_pack(p, &w_curs,
1371                                        entry->in_pack_offset + used, NULL);
1372                         used_0 = 0;
1373                         c = buf[used_0++];
1374                         ofs = c & 127;
1375                         while (c & 128) {
1376                                 ofs += 1;
1377                                 if (!ofs || MSB(ofs, 7)) {
1378                                         error("delta base offset overflow in pack for %s",
1379                                               sha1_to_hex(entry->idx.sha1));
1380                                         goto give_up;
1381                                 }
1382                                 c = buf[used_0++];
1383                                 ofs = (ofs << 7) + (c & 127);
1384                         }
1385                         ofs = entry->in_pack_offset - ofs;
1386                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1387                                 error("delta base offset out of bound for %s",
1388                                       sha1_to_hex(entry->idx.sha1));
1389                                 goto give_up;
1390                         }
1391                         if (reuse_delta && !entry->preferred_base) {
1392                                 struct revindex_entry *revidx;
1393                                 revidx = find_pack_revindex(p, ofs);
1394                                 if (!revidx)
1395                                         goto give_up;
1396                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1397                         }
1398                         entry->in_pack_header_size = used + used_0;
1399                         break;
1400                 }
1401
1402                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1403                         /*
1404                          * If base_ref was set above that means we wish to
1405                          * reuse delta data, and we even found that base
1406                          * in the list of objects we want to pack. Goodie!
1407                          *
1408                          * Depth value does not matter - find_deltas() will
1409                          * never consider reused delta as the base object to
1410                          * deltify other objects against, in order to avoid
1411                          * circular deltas.
1412                          */
1413                         entry->type = entry->in_pack_type;
1414                         entry->delta = base_entry;
1415                         entry->delta_size = entry->size;
1416                         entry->delta_sibling = base_entry->delta_child;
1417                         base_entry->delta_child = entry;
1418                         unuse_pack(&w_curs);
1419                         return;
1420                 }
1421
1422                 if (entry->type) {
1423                         /*
1424                          * This must be a delta and we already know what the
1425                          * final object type is.  Let's extract the actual
1426                          * object size from the delta header.
1427                          */
1428                         entry->size = get_size_from_delta(p, &w_curs,
1429                                         entry->in_pack_offset + entry->in_pack_header_size);
1430                         if (entry->size == 0)
1431                                 goto give_up;
1432                         unuse_pack(&w_curs);
1433                         return;
1434                 }
1435
1436                 /*
1437                  * No choice but to fall back to the recursive delta walk
1438                  * with sha1_object_info() to find about the object type
1439                  * at this point...
1440                  */
1441                 give_up:
1442                 unuse_pack(&w_curs);
1443         }
1444
1445         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1446         /*
1447          * The error condition is checked in prepare_pack().  This is
1448          * to permit a missing preferred base object to be ignored
1449          * as a preferred base.  Doing so can result in a larger
1450          * pack file, but the transfer will still take place.
1451          */
1452 }
1453
1454 static int pack_offset_sort(const void *_a, const void *_b)
1455 {
1456         const struct object_entry *a = *(struct object_entry **)_a;
1457         const struct object_entry *b = *(struct object_entry **)_b;
1458
1459         /* avoid filesystem trashing with loose objects */
1460         if (!a->in_pack && !b->in_pack)
1461                 return hashcmp(a->idx.sha1, b->idx.sha1);
1462
1463         if (a->in_pack < b->in_pack)
1464                 return -1;
1465         if (a->in_pack > b->in_pack)
1466                 return 1;
1467         return a->in_pack_offset < b->in_pack_offset ? -1 :
1468                         (a->in_pack_offset > b->in_pack_offset);
1469 }
1470
1471 static void get_object_details(void)
1472 {
1473         uint32_t i;
1474         struct object_entry **sorted_by_offset;
1475
1476         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1477         for (i = 0; i < to_pack.nr_objects; i++)
1478                 sorted_by_offset[i] = to_pack.objects + i;
1479         qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1480
1481         for (i = 0; i < to_pack.nr_objects; i++) {
1482                 struct object_entry *entry = sorted_by_offset[i];
1483                 check_object(entry);
1484                 if (big_file_threshold < entry->size)
1485                         entry->no_try_delta = 1;
1486         }
1487
1488         free(sorted_by_offset);
1489 }
1490
1491 /*
1492  * We search for deltas in a list sorted by type, by filename hash, and then
1493  * by size, so that we see progressively smaller and smaller files.
1494  * That's because we prefer deltas to be from the bigger file
1495  * to the smaller -- deletes are potentially cheaper, but perhaps
1496  * more importantly, the bigger file is likely the more recent
1497  * one.  The deepest deltas are therefore the oldest objects which are
1498  * less susceptible to be accessed often.
1499  */
1500 static int type_size_sort(const void *_a, const void *_b)
1501 {
1502         const struct object_entry *a = *(struct object_entry **)_a;
1503         const struct object_entry *b = *(struct object_entry **)_b;
1504
1505         if (a->type > b->type)
1506                 return -1;
1507         if (a->type < b->type)
1508                 return 1;
1509         if (a->hash > b->hash)
1510                 return -1;
1511         if (a->hash < b->hash)
1512                 return 1;
1513         if (a->preferred_base > b->preferred_base)
1514                 return -1;
1515         if (a->preferred_base < b->preferred_base)
1516                 return 1;
1517         if (a->size > b->size)
1518                 return -1;
1519         if (a->size < b->size)
1520                 return 1;
1521         return a < b ? -1 : (a > b);  /* newest first */
1522 }
1523
1524 struct unpacked {
1525         struct object_entry *entry;
1526         void *data;
1527         struct delta_index *index;
1528         unsigned depth;
1529 };
1530
1531 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1532                            unsigned long delta_size)
1533 {
1534         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1535                 return 0;
1536
1537         if (delta_size < cache_max_small_delta_size)
1538                 return 1;
1539
1540         /* cache delta, if objects are large enough compared to delta size */
1541         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1542                 return 1;
1543
1544         return 0;
1545 }
1546
1547 #ifndef NO_PTHREADS
1548
1549 static pthread_mutex_t read_mutex;
1550 #define read_lock()             pthread_mutex_lock(&read_mutex)
1551 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1552
1553 static pthread_mutex_t cache_mutex;
1554 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1555 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1556
1557 static pthread_mutex_t progress_mutex;
1558 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1559 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1560
1561 #else
1562
1563 #define read_lock()             (void)0
1564 #define read_unlock()           (void)0
1565 #define cache_lock()            (void)0
1566 #define cache_unlock()          (void)0
1567 #define progress_lock()         (void)0
1568 #define progress_unlock()       (void)0
1569
1570 #endif
1571
1572 static int try_delta(struct unpacked *trg, struct unpacked *src,
1573                      unsigned max_depth, unsigned long *mem_usage)
1574 {
1575         struct object_entry *trg_entry = trg->entry;
1576         struct object_entry *src_entry = src->entry;
1577         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1578         unsigned ref_depth;
1579         enum object_type type;
1580         void *delta_buf;
1581
1582         /* Don't bother doing diffs between different types */
1583         if (trg_entry->type != src_entry->type)
1584                 return -1;
1585
1586         /*
1587          * We do not bother to try a delta that we discarded on an
1588          * earlier try, but only when reusing delta data.  Note that
1589          * src_entry that is marked as the preferred_base should always
1590          * be considered, as even if we produce a suboptimal delta against
1591          * it, we will still save the transfer cost, as we already know
1592          * the other side has it and we won't send src_entry at all.
1593          */
1594         if (reuse_delta && trg_entry->in_pack &&
1595             trg_entry->in_pack == src_entry->in_pack &&
1596             !src_entry->preferred_base &&
1597             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1598             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1599                 return 0;
1600
1601         /* Let's not bust the allowed depth. */
1602         if (src->depth >= max_depth)
1603                 return 0;
1604
1605         /* Now some size filtering heuristics. */
1606         trg_size = trg_entry->size;
1607         if (!trg_entry->delta) {
1608                 max_size = trg_size/2 - 20;
1609                 ref_depth = 1;
1610         } else {
1611                 max_size = trg_entry->delta_size;
1612                 ref_depth = trg->depth;
1613         }
1614         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1615                                                 (max_depth - ref_depth + 1);
1616         if (max_size == 0)
1617                 return 0;
1618         src_size = src_entry->size;
1619         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1620         if (sizediff >= max_size)
1621                 return 0;
1622         if (trg_size < src_size / 32)
1623                 return 0;
1624
1625         /* Load data if not already done */
1626         if (!trg->data) {
1627                 read_lock();
1628                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1629                 read_unlock();
1630                 if (!trg->data)
1631                         die("object %s cannot be read",
1632                             sha1_to_hex(trg_entry->idx.sha1));
1633                 if (sz != trg_size)
1634                         die("object %s inconsistent object length (%lu vs %lu)",
1635                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1636                 *mem_usage += sz;
1637         }
1638         if (!src->data) {
1639                 read_lock();
1640                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1641                 read_unlock();
1642                 if (!src->data) {
1643                         if (src_entry->preferred_base) {
1644                                 static int warned = 0;
1645                                 if (!warned++)
1646                                         warning("object %s cannot be read",
1647                                                 sha1_to_hex(src_entry->idx.sha1));
1648                                 /*
1649                                  * Those objects are not included in the
1650                                  * resulting pack.  Be resilient and ignore
1651                                  * them if they can't be read, in case the
1652                                  * pack could be created nevertheless.
1653                                  */
1654                                 return 0;
1655                         }
1656                         die("object %s cannot be read",
1657                             sha1_to_hex(src_entry->idx.sha1));
1658                 }
1659                 if (sz != src_size)
1660                         die("object %s inconsistent object length (%lu vs %lu)",
1661                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1662                 *mem_usage += sz;
1663         }
1664         if (!src->index) {
1665                 src->index = create_delta_index(src->data, src_size);
1666                 if (!src->index) {
1667                         static int warned = 0;
1668                         if (!warned++)
1669                                 warning("suboptimal pack - out of memory");
1670                         return 0;
1671                 }
1672                 *mem_usage += sizeof_delta_index(src->index);
1673         }
1674
1675         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1676         if (!delta_buf)
1677                 return 0;
1678
1679         if (trg_entry->delta) {
1680                 /* Prefer only shallower same-sized deltas. */
1681                 if (delta_size == trg_entry->delta_size &&
1682                     src->depth + 1 >= trg->depth) {
1683                         free(delta_buf);
1684                         return 0;
1685                 }
1686         }
1687
1688         /*
1689          * Handle memory allocation outside of the cache
1690          * accounting lock.  Compiler will optimize the strangeness
1691          * away when NO_PTHREADS is defined.
1692          */
1693         free(trg_entry->delta_data);
1694         cache_lock();
1695         if (trg_entry->delta_data) {
1696                 delta_cache_size -= trg_entry->delta_size;
1697                 trg_entry->delta_data = NULL;
1698         }
1699         if (delta_cacheable(src_size, trg_size, delta_size)) {
1700                 delta_cache_size += delta_size;
1701                 cache_unlock();
1702                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1703         } else {
1704                 cache_unlock();
1705                 free(delta_buf);
1706         }
1707
1708         trg_entry->delta = src_entry;
1709         trg_entry->delta_size = delta_size;
1710         trg->depth = src->depth + 1;
1711
1712         return 1;
1713 }
1714
1715 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1716 {
1717         struct object_entry *child = me->delta_child;
1718         unsigned int m = n;
1719         while (child) {
1720                 unsigned int c = check_delta_limit(child, n + 1);
1721                 if (m < c)
1722                         m = c;
1723                 child = child->delta_sibling;
1724         }
1725         return m;
1726 }
1727
1728 static unsigned long free_unpacked(struct unpacked *n)
1729 {
1730         unsigned long freed_mem = sizeof_delta_index(n->index);
1731         free_delta_index(n->index);
1732         n->index = NULL;
1733         if (n->data) {
1734                 freed_mem += n->entry->size;
1735                 free(n->data);
1736                 n->data = NULL;
1737         }
1738         n->entry = NULL;
1739         n->depth = 0;
1740         return freed_mem;
1741 }
1742
1743 static void find_deltas(struct object_entry **list, unsigned *list_size,
1744                         int window, int depth, unsigned *processed)
1745 {
1746         uint32_t i, idx = 0, count = 0;
1747         struct unpacked *array;
1748         unsigned long mem_usage = 0;
1749
1750         array = xcalloc(window, sizeof(struct unpacked));
1751
1752         for (;;) {
1753                 struct object_entry *entry;
1754                 struct unpacked *n = array + idx;
1755                 int j, max_depth, best_base = -1;
1756
1757                 progress_lock();
1758                 if (!*list_size) {
1759                         progress_unlock();
1760                         break;
1761                 }
1762                 entry = *list++;
1763                 (*list_size)--;
1764                 if (!entry->preferred_base) {
1765                         (*processed)++;
1766                         display_progress(progress_state, *processed);
1767                 }
1768                 progress_unlock();
1769
1770                 mem_usage -= free_unpacked(n);
1771                 n->entry = entry;
1772
1773                 while (window_memory_limit &&
1774                        mem_usage > window_memory_limit &&
1775                        count > 1) {
1776                         uint32_t tail = (idx + window - count) % window;
1777                         mem_usage -= free_unpacked(array + tail);
1778                         count--;
1779                 }
1780
1781                 /* We do not compute delta to *create* objects we are not
1782                  * going to pack.
1783                  */
1784                 if (entry->preferred_base)
1785                         goto next;
1786
1787                 /*
1788                  * If the current object is at pack edge, take the depth the
1789                  * objects that depend on the current object into account
1790                  * otherwise they would become too deep.
1791                  */
1792                 max_depth = depth;
1793                 if (entry->delta_child) {
1794                         max_depth -= check_delta_limit(entry, 0);
1795                         if (max_depth <= 0)
1796                                 goto next;
1797                 }
1798
1799                 j = window;
1800                 while (--j > 0) {
1801                         int ret;
1802                         uint32_t other_idx = idx + j;
1803                         struct unpacked *m;
1804                         if (other_idx >= window)
1805                                 other_idx -= window;
1806                         m = array + other_idx;
1807                         if (!m->entry)
1808                                 break;
1809                         ret = try_delta(n, m, max_depth, &mem_usage);
1810                         if (ret < 0)
1811                                 break;
1812                         else if (ret > 0)
1813                                 best_base = other_idx;
1814                 }
1815
1816                 /*
1817                  * If we decided to cache the delta data, then it is best
1818                  * to compress it right away.  First because we have to do
1819                  * it anyway, and doing it here while we're threaded will
1820                  * save a lot of time in the non threaded write phase,
1821                  * as well as allow for caching more deltas within
1822                  * the same cache size limit.
1823                  * ...
1824                  * But only if not writing to stdout, since in that case
1825                  * the network is most likely throttling writes anyway,
1826                  * and therefore it is best to go to the write phase ASAP
1827                  * instead, as we can afford spending more time compressing
1828                  * between writes at that moment.
1829                  */
1830                 if (entry->delta_data && !pack_to_stdout) {
1831                         entry->z_delta_size = do_compress(&entry->delta_data,
1832                                                           entry->delta_size);
1833                         cache_lock();
1834                         delta_cache_size -= entry->delta_size;
1835                         delta_cache_size += entry->z_delta_size;
1836                         cache_unlock();
1837                 }
1838
1839                 /* if we made n a delta, and if n is already at max
1840                  * depth, leaving it in the window is pointless.  we
1841                  * should evict it first.
1842                  */
1843                 if (entry->delta && max_depth <= n->depth)
1844                         continue;
1845
1846                 /*
1847                  * Move the best delta base up in the window, after the
1848                  * currently deltified object, to keep it longer.  It will
1849                  * be the first base object to be attempted next.
1850                  */
1851                 if (entry->delta) {
1852                         struct unpacked swap = array[best_base];
1853                         int dist = (window + idx - best_base) % window;
1854                         int dst = best_base;
1855                         while (dist--) {
1856                                 int src = (dst + 1) % window;
1857                                 array[dst] = array[src];
1858                                 dst = src;
1859                         }
1860                         array[dst] = swap;
1861                 }
1862
1863                 next:
1864                 idx++;
1865                 if (count + 1 < window)
1866                         count++;
1867                 if (idx >= window)
1868                         idx = 0;
1869         }
1870
1871         for (i = 0; i < window; ++i) {
1872                 free_delta_index(array[i].index);
1873                 free(array[i].data);
1874         }
1875         free(array);
1876 }
1877
1878 #ifndef NO_PTHREADS
1879
1880 static void try_to_free_from_threads(size_t size)
1881 {
1882         read_lock();
1883         release_pack_memory(size);
1884         read_unlock();
1885 }
1886
1887 static try_to_free_t old_try_to_free_routine;
1888
1889 /*
1890  * The main thread waits on the condition that (at least) one of the workers
1891  * has stopped working (which is indicated in the .working member of
1892  * struct thread_params).
1893  * When a work thread has completed its work, it sets .working to 0 and
1894  * signals the main thread and waits on the condition that .data_ready
1895  * becomes 1.
1896  */
1897
1898 struct thread_params {
1899         pthread_t thread;
1900         struct object_entry **list;
1901         unsigned list_size;
1902         unsigned remaining;
1903         int window;
1904         int depth;
1905         int working;
1906         int data_ready;
1907         pthread_mutex_t mutex;
1908         pthread_cond_t cond;
1909         unsigned *processed;
1910 };
1911
1912 static pthread_cond_t progress_cond;
1913
1914 /*
1915  * Mutex and conditional variable can't be statically-initialized on Windows.
1916  */
1917 static void init_threaded_search(void)
1918 {
1919         init_recursive_mutex(&read_mutex);
1920         pthread_mutex_init(&cache_mutex, NULL);
1921         pthread_mutex_init(&progress_mutex, NULL);
1922         pthread_cond_init(&progress_cond, NULL);
1923         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1924 }
1925
1926 static void cleanup_threaded_search(void)
1927 {
1928         set_try_to_free_routine(old_try_to_free_routine);
1929         pthread_cond_destroy(&progress_cond);
1930         pthread_mutex_destroy(&read_mutex);
1931         pthread_mutex_destroy(&cache_mutex);
1932         pthread_mutex_destroy(&progress_mutex);
1933 }
1934
1935 static void *threaded_find_deltas(void *arg)
1936 {
1937         struct thread_params *me = arg;
1938
1939         while (me->remaining) {
1940                 find_deltas(me->list, &me->remaining,
1941                             me->window, me->depth, me->processed);
1942
1943                 progress_lock();
1944                 me->working = 0;
1945                 pthread_cond_signal(&progress_cond);
1946                 progress_unlock();
1947
1948                 /*
1949                  * We must not set ->data_ready before we wait on the
1950                  * condition because the main thread may have set it to 1
1951                  * before we get here. In order to be sure that new
1952                  * work is available if we see 1 in ->data_ready, it
1953                  * was initialized to 0 before this thread was spawned
1954                  * and we reset it to 0 right away.
1955                  */
1956                 pthread_mutex_lock(&me->mutex);
1957                 while (!me->data_ready)
1958                         pthread_cond_wait(&me->cond, &me->mutex);
1959                 me->data_ready = 0;
1960                 pthread_mutex_unlock(&me->mutex);
1961         }
1962         /* leave ->working 1 so that this doesn't get more work assigned */
1963         return NULL;
1964 }
1965
1966 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1967                            int window, int depth, unsigned *processed)
1968 {
1969         struct thread_params *p;
1970         int i, ret, active_threads = 0;
1971
1972         init_threaded_search();
1973
1974         if (!delta_search_threads)      /* --threads=0 means autodetect */
1975                 delta_search_threads = online_cpus();
1976         if (delta_search_threads <= 1) {
1977                 find_deltas(list, &list_size, window, depth, processed);
1978                 cleanup_threaded_search();
1979                 return;
1980         }
1981         if (progress > pack_to_stdout)
1982                 fprintf(stderr, "Delta compression using up to %d threads.\n",
1983                                 delta_search_threads);
1984         p = xcalloc(delta_search_threads, sizeof(*p));
1985
1986         /* Partition the work amongst work threads. */
1987         for (i = 0; i < delta_search_threads; i++) {
1988                 unsigned sub_size = list_size / (delta_search_threads - i);
1989
1990                 /* don't use too small segments or no deltas will be found */
1991                 if (sub_size < 2*window && i+1 < delta_search_threads)
1992                         sub_size = 0;
1993
1994                 p[i].window = window;
1995                 p[i].depth = depth;
1996                 p[i].processed = processed;
1997                 p[i].working = 1;
1998                 p[i].data_ready = 0;
1999
2000                 /* try to split chunks on "path" boundaries */
2001                 while (sub_size && sub_size < list_size &&
2002                        list[sub_size]->hash &&
2003                        list[sub_size]->hash == list[sub_size-1]->hash)
2004                         sub_size++;
2005
2006                 p[i].list = list;
2007                 p[i].list_size = sub_size;
2008                 p[i].remaining = sub_size;
2009
2010                 list += sub_size;
2011                 list_size -= sub_size;
2012         }
2013
2014         /* Start work threads. */
2015         for (i = 0; i < delta_search_threads; i++) {
2016                 if (!p[i].list_size)
2017                         continue;
2018                 pthread_mutex_init(&p[i].mutex, NULL);
2019                 pthread_cond_init(&p[i].cond, NULL);
2020                 ret = pthread_create(&p[i].thread, NULL,
2021                                      threaded_find_deltas, &p[i]);
2022                 if (ret)
2023                         die("unable to create thread: %s", strerror(ret));
2024                 active_threads++;
2025         }
2026
2027         /*
2028          * Now let's wait for work completion.  Each time a thread is done
2029          * with its work, we steal half of the remaining work from the
2030          * thread with the largest number of unprocessed objects and give
2031          * it to that newly idle thread.  This ensure good load balancing
2032          * until the remaining object list segments are simply too short
2033          * to be worth splitting anymore.
2034          */
2035         while (active_threads) {
2036                 struct thread_params *target = NULL;
2037                 struct thread_params *victim = NULL;
2038                 unsigned sub_size = 0;
2039
2040                 progress_lock();
2041                 for (;;) {
2042                         for (i = 0; !target && i < delta_search_threads; i++)
2043                                 if (!p[i].working)
2044                                         target = &p[i];
2045                         if (target)
2046                                 break;
2047                         pthread_cond_wait(&progress_cond, &progress_mutex);
2048                 }
2049
2050                 for (i = 0; i < delta_search_threads; i++)
2051                         if (p[i].remaining > 2*window &&
2052                             (!victim || victim->remaining < p[i].remaining))
2053                                 victim = &p[i];
2054                 if (victim) {
2055                         sub_size = victim->remaining / 2;
2056                         list = victim->list + victim->list_size - sub_size;
2057                         while (sub_size && list[0]->hash &&
2058                                list[0]->hash == list[-1]->hash) {
2059                                 list++;
2060                                 sub_size--;
2061                         }
2062                         if (!sub_size) {
2063                                 /*
2064                                  * It is possible for some "paths" to have
2065                                  * so many objects that no hash boundary
2066                                  * might be found.  Let's just steal the
2067                                  * exact half in that case.
2068                                  */
2069                                 sub_size = victim->remaining / 2;
2070                                 list -= sub_size;
2071                         }
2072                         target->list = list;
2073                         victim->list_size -= sub_size;
2074                         victim->remaining -= sub_size;
2075                 }
2076                 target->list_size = sub_size;
2077                 target->remaining = sub_size;
2078                 target->working = 1;
2079                 progress_unlock();
2080
2081                 pthread_mutex_lock(&target->mutex);
2082                 target->data_ready = 1;
2083                 pthread_cond_signal(&target->cond);
2084                 pthread_mutex_unlock(&target->mutex);
2085
2086                 if (!sub_size) {
2087                         pthread_join(target->thread, NULL);
2088                         pthread_cond_destroy(&target->cond);
2089                         pthread_mutex_destroy(&target->mutex);
2090                         active_threads--;
2091                 }
2092         }
2093         cleanup_threaded_search();
2094         free(p);
2095 }
2096
2097 #else
2098 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2099 #endif
2100
2101 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
2102 {
2103         unsigned char peeled[20];
2104
2105         if (starts_with(path, "refs/tags/") && /* is a tag? */
2106             !peel_ref(path, peeled)        && /* peelable? */
2107             packlist_find(&to_pack, peeled, NULL))      /* object packed? */
2108                 add_object_entry(sha1, OBJ_TAG, NULL, 0);
2109         return 0;
2110 }
2111
2112 static void prepare_pack(int window, int depth)
2113 {
2114         struct object_entry **delta_list;
2115         uint32_t i, nr_deltas;
2116         unsigned n;
2117
2118         get_object_details();
2119
2120         /*
2121          * If we're locally repacking then we need to be doubly careful
2122          * from now on in order to make sure no stealth corruption gets
2123          * propagated to the new pack.  Clients receiving streamed packs
2124          * should validate everything they get anyway so no need to incur
2125          * the additional cost here in that case.
2126          */
2127         if (!pack_to_stdout)
2128                 do_check_packed_object_crc = 1;
2129
2130         if (!to_pack.nr_objects || !window || !depth)
2131                 return;
2132
2133         delta_list = xmalloc(to_pack.nr_objects * sizeof(*delta_list));
2134         nr_deltas = n = 0;
2135
2136         for (i = 0; i < to_pack.nr_objects; i++) {
2137                 struct object_entry *entry = to_pack.objects + i;
2138
2139                 if (entry->delta)
2140                         /* This happens if we decided to reuse existing
2141                          * delta from a pack.  "reuse_delta &&" is implied.
2142                          */
2143                         continue;
2144
2145                 if (entry->size < 50)
2146                         continue;
2147
2148                 if (entry->no_try_delta)
2149                         continue;
2150
2151                 if (!entry->preferred_base) {
2152                         nr_deltas++;
2153                         if (entry->type < 0)
2154                                 die("unable to get type of object %s",
2155                                     sha1_to_hex(entry->idx.sha1));
2156                 } else {
2157                         if (entry->type < 0) {
2158                                 /*
2159                                  * This object is not found, but we
2160                                  * don't have to include it anyway.
2161                                  */
2162                                 continue;
2163                         }
2164                 }
2165
2166                 delta_list[n++] = entry;
2167         }
2168
2169         if (nr_deltas && n > 1) {
2170                 unsigned nr_done = 0;
2171                 if (progress)
2172                         progress_state = start_progress(_("Compressing objects"),
2173                                                         nr_deltas);
2174                 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2175                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2176                 stop_progress(&progress_state);
2177                 if (nr_done != nr_deltas)
2178                         die("inconsistency with delta count");
2179         }
2180         free(delta_list);
2181 }
2182
2183 static int git_pack_config(const char *k, const char *v, void *cb)
2184 {
2185         if (!strcmp(k, "pack.window")) {
2186                 window = git_config_int(k, v);
2187                 return 0;
2188         }
2189         if (!strcmp(k, "pack.windowmemory")) {
2190                 window_memory_limit = git_config_ulong(k, v);
2191                 return 0;
2192         }
2193         if (!strcmp(k, "pack.depth")) {
2194                 depth = git_config_int(k, v);
2195                 return 0;
2196         }
2197         if (!strcmp(k, "pack.compression")) {
2198                 int level = git_config_int(k, v);
2199                 if (level == -1)
2200                         level = Z_DEFAULT_COMPRESSION;
2201                 else if (level < 0 || level > Z_BEST_COMPRESSION)
2202                         die("bad pack compression level %d", level);
2203                 pack_compression_level = level;
2204                 pack_compression_seen = 1;
2205                 return 0;
2206         }
2207         if (!strcmp(k, "pack.deltacachesize")) {
2208                 max_delta_cache_size = git_config_int(k, v);
2209                 return 0;
2210         }
2211         if (!strcmp(k, "pack.deltacachelimit")) {
2212                 cache_max_small_delta_size = git_config_int(k, v);
2213                 return 0;
2214         }
2215         if (!strcmp(k, "pack.writebitmaps")) {
2216                 write_bitmap_index = git_config_bool(k, v);
2217                 return 0;
2218         }
2219         if (!strcmp(k, "pack.writebitmaphashcache")) {
2220                 if (git_config_bool(k, v))
2221                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2222                 else
2223                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2224         }
2225         if (!strcmp(k, "pack.usebitmaps")) {
2226                 use_bitmap_index = git_config_bool(k, v);
2227                 return 0;
2228         }
2229         if (!strcmp(k, "pack.threads")) {
2230                 delta_search_threads = git_config_int(k, v);
2231                 if (delta_search_threads < 0)
2232                         die("invalid number of threads specified (%d)",
2233                             delta_search_threads);
2234 #ifdef NO_PTHREADS
2235                 if (delta_search_threads != 1)
2236                         warning("no threads support, ignoring %s", k);
2237 #endif
2238                 return 0;
2239         }
2240         if (!strcmp(k, "pack.indexversion")) {
2241                 pack_idx_opts.version = git_config_int(k, v);
2242                 if (pack_idx_opts.version > 2)
2243                         die("bad pack.indexversion=%"PRIu32,
2244                             pack_idx_opts.version);
2245                 return 0;
2246         }
2247         return git_default_config(k, v, cb);
2248 }
2249
2250 static void read_object_list_from_stdin(void)
2251 {
2252         char line[40 + 1 + PATH_MAX + 2];
2253         unsigned char sha1[20];
2254
2255         for (;;) {
2256                 if (!fgets(line, sizeof(line), stdin)) {
2257                         if (feof(stdin))
2258                                 break;
2259                         if (!ferror(stdin))
2260                                 die("fgets returned NULL, not EOF, not error!");
2261                         if (errno != EINTR)
2262                                 die_errno("fgets");
2263                         clearerr(stdin);
2264                         continue;
2265                 }
2266                 if (line[0] == '-') {
2267                         if (get_sha1_hex(line+1, sha1))
2268                                 die("expected edge sha1, got garbage:\n %s",
2269                                     line);
2270                         add_preferred_base(sha1);
2271                         continue;
2272                 }
2273                 if (get_sha1_hex(line, sha1))
2274                         die("expected sha1, got garbage:\n %s", line);
2275
2276                 add_preferred_base_object(line+41);
2277                 add_object_entry(sha1, 0, line+41, 0);
2278         }
2279 }
2280
2281 #define OBJECT_ADDED (1u<<20)
2282
2283 static void show_commit(struct commit *commit, void *data)
2284 {
2285         add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
2286         commit->object.flags |= OBJECT_ADDED;
2287
2288         if (write_bitmap_index)
2289                 index_commit_for_bitmap(commit);
2290 }
2291
2292 static void show_object(struct object *obj,
2293                         const struct name_path *path, const char *last,
2294                         void *data)
2295 {
2296         char *name = path_name(path, last);
2297
2298         add_preferred_base_object(name);
2299         add_object_entry(obj->sha1, obj->type, name, 0);
2300         obj->flags |= OBJECT_ADDED;
2301
2302         /*
2303          * We will have generated the hash from the name,
2304          * but not saved a pointer to it - we can free it
2305          */
2306         free((char *)name);
2307 }
2308
2309 static void show_edge(struct commit *commit)
2310 {
2311         add_preferred_base(commit->object.sha1);
2312 }
2313
2314 struct in_pack_object {
2315         off_t offset;
2316         struct object *object;
2317 };
2318
2319 struct in_pack {
2320         int alloc;
2321         int nr;
2322         struct in_pack_object *array;
2323 };
2324
2325 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2326 {
2327         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
2328         in_pack->array[in_pack->nr].object = object;
2329         in_pack->nr++;
2330 }
2331
2332 /*
2333  * Compare the objects in the offset order, in order to emulate the
2334  * "git rev-list --objects" output that produced the pack originally.
2335  */
2336 static int ofscmp(const void *a_, const void *b_)
2337 {
2338         struct in_pack_object *a = (struct in_pack_object *)a_;
2339         struct in_pack_object *b = (struct in_pack_object *)b_;
2340
2341         if (a->offset < b->offset)
2342                 return -1;
2343         else if (a->offset > b->offset)
2344                 return 1;
2345         else
2346                 return hashcmp(a->object->sha1, b->object->sha1);
2347 }
2348
2349 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2350 {
2351         struct packed_git *p;
2352         struct in_pack in_pack;
2353         uint32_t i;
2354
2355         memset(&in_pack, 0, sizeof(in_pack));
2356
2357         for (p = packed_git; p; p = p->next) {
2358                 const unsigned char *sha1;
2359                 struct object *o;
2360
2361                 if (!p->pack_local || p->pack_keep)
2362                         continue;
2363                 if (open_pack_index(p))
2364                         die("cannot open pack index");
2365
2366                 ALLOC_GROW(in_pack.array,
2367                            in_pack.nr + p->num_objects,
2368                            in_pack.alloc);
2369
2370                 for (i = 0; i < p->num_objects; i++) {
2371                         sha1 = nth_packed_object_sha1(p, i);
2372                         o = lookup_unknown_object(sha1);
2373                         if (!(o->flags & OBJECT_ADDED))
2374                                 mark_in_pack_object(o, p, &in_pack);
2375                         o->flags |= OBJECT_ADDED;
2376                 }
2377         }
2378
2379         if (in_pack.nr) {
2380                 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2381                       ofscmp);
2382                 for (i = 0; i < in_pack.nr; i++) {
2383                         struct object *o = in_pack.array[i].object;
2384                         add_object_entry(o->sha1, o->type, "", 0);
2385                 }
2386         }
2387         free(in_pack.array);
2388 }
2389
2390 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2391 {
2392         static struct packed_git *last_found = (void *)1;
2393         struct packed_git *p;
2394
2395         p = (last_found != (void *)1) ? last_found : packed_git;
2396
2397         while (p) {
2398                 if ((!p->pack_local || p->pack_keep) &&
2399                         find_pack_entry_one(sha1, p)) {
2400                         last_found = p;
2401                         return 1;
2402                 }
2403                 if (p == last_found)
2404                         p = packed_git;
2405                 else
2406                         p = p->next;
2407                 if (p == last_found)
2408                         p = p->next;
2409         }
2410         return 0;
2411 }
2412
2413 static void loosen_unused_packed_objects(struct rev_info *revs)
2414 {
2415         struct packed_git *p;
2416         uint32_t i;
2417         const unsigned char *sha1;
2418
2419         for (p = packed_git; p; p = p->next) {
2420                 if (!p->pack_local || p->pack_keep)
2421                         continue;
2422
2423                 if (unpack_unreachable_expiration &&
2424                     p->mtime < unpack_unreachable_expiration)
2425                         continue;
2426
2427                 if (open_pack_index(p))
2428                         die("cannot open pack index");
2429
2430                 for (i = 0; i < p->num_objects; i++) {
2431                         sha1 = nth_packed_object_sha1(p, i);
2432                         if (!packlist_find(&to_pack, sha1, NULL) &&
2433                                 !has_sha1_pack_kept_or_nonlocal(sha1))
2434                                 if (force_object_loose(sha1, p->mtime))
2435                                         die("unable to force loose object");
2436                 }
2437         }
2438 }
2439
2440 /*
2441  * This tracks any options which a reader of the pack might
2442  * not understand, and which would therefore prevent blind reuse
2443  * of what we have on disk.
2444  */
2445 static int pack_options_allow_reuse(void)
2446 {
2447         return allow_ofs_delta;
2448 }
2449
2450 static int get_object_list_from_bitmap(struct rev_info *revs)
2451 {
2452         if (prepare_bitmap_walk(revs) < 0)
2453                 return -1;
2454
2455         if (pack_options_allow_reuse() &&
2456             !reuse_partial_packfile_from_bitmap(
2457                         &reuse_packfile,
2458                         &reuse_packfile_objects,
2459                         &reuse_packfile_offset)) {
2460                 assert(reuse_packfile_objects);
2461                 nr_result += reuse_packfile_objects;
2462                 display_progress(progress_state, nr_result);
2463         }
2464
2465         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2466         return 0;
2467 }
2468
2469 static void get_object_list(int ac, const char **av)
2470 {
2471         struct rev_info revs;
2472         char line[1000];
2473         int flags = 0;
2474
2475         init_revisions(&revs, NULL);
2476         save_commit_buffer = 0;
2477         setup_revisions(ac, av, &revs, NULL);
2478
2479         /* make sure shallows are read */
2480         is_repository_shallow();
2481
2482         while (fgets(line, sizeof(line), stdin) != NULL) {
2483                 int len = strlen(line);
2484                 if (len && line[len - 1] == '\n')
2485                         line[--len] = 0;
2486                 if (!len)
2487                         break;
2488                 if (*line == '-') {
2489                         if (!strcmp(line, "--not")) {
2490                                 flags ^= UNINTERESTING;
2491                                 write_bitmap_index = 0;
2492                                 continue;
2493                         }
2494                         if (starts_with(line, "--shallow ")) {
2495                                 unsigned char sha1[20];
2496                                 if (get_sha1_hex(line + 10, sha1))
2497                                         die("not an SHA-1 '%s'", line + 10);
2498                                 register_shallow(sha1);
2499                                 continue;
2500                         }
2501                         die("not a rev '%s'", line);
2502                 }
2503                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2504                         die("bad revision '%s'", line);
2505         }
2506
2507         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2508                 return;
2509
2510         if (prepare_revision_walk(&revs))
2511                 die("revision walk setup failed");
2512         mark_edges_uninteresting(&revs, show_edge);
2513         traverse_commit_list(&revs, show_commit, show_object, NULL);
2514
2515         if (keep_unreachable)
2516                 add_objects_in_unpacked_packs(&revs);
2517         if (unpack_unreachable)
2518                 loosen_unused_packed_objects(&revs);
2519 }
2520
2521 static int option_parse_index_version(const struct option *opt,
2522                                       const char *arg, int unset)
2523 {
2524         char *c;
2525         const char *val = arg;
2526         pack_idx_opts.version = strtoul(val, &c, 10);
2527         if (pack_idx_opts.version > 2)
2528                 die(_("unsupported index version %s"), val);
2529         if (*c == ',' && c[1])
2530                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2531         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2532                 die(_("bad index version '%s'"), val);
2533         return 0;
2534 }
2535
2536 static int option_parse_unpack_unreachable(const struct option *opt,
2537                                            const char *arg, int unset)
2538 {
2539         if (unset) {
2540                 unpack_unreachable = 0;
2541                 unpack_unreachable_expiration = 0;
2542         }
2543         else {
2544                 unpack_unreachable = 1;
2545                 if (arg)
2546                         unpack_unreachable_expiration = approxidate(arg);
2547         }
2548         return 0;
2549 }
2550
2551 static int option_parse_ulong(const struct option *opt,
2552                               const char *arg, int unset)
2553 {
2554         if (unset)
2555                 die(_("option %s does not accept negative form"),
2556                     opt->long_name);
2557
2558         if (!git_parse_ulong(arg, opt->value))
2559                 die(_("unable to parse value '%s' for option %s"),
2560                     arg, opt->long_name);
2561         return 0;
2562 }
2563
2564 #define OPT_ULONG(s, l, v, h) \
2565         { OPTION_CALLBACK, (s), (l), (v), "n", (h),     \
2566           PARSE_OPT_NONEG, option_parse_ulong }
2567
2568 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2569 {
2570         int use_internal_rev_list = 0;
2571         int thin = 0;
2572         int all_progress_implied = 0;
2573         const char *rp_av[6];
2574         int rp_ac = 0;
2575         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2576         struct option pack_objects_options[] = {
2577                 OPT_SET_INT('q', "quiet", &progress,
2578                             N_("do not show progress meter"), 0),
2579                 OPT_SET_INT(0, "progress", &progress,
2580                             N_("show progress meter"), 1),
2581                 OPT_SET_INT(0, "all-progress", &progress,
2582                             N_("show progress meter during object writing phase"), 2),
2583                 OPT_BOOL(0, "all-progress-implied",
2584                          &all_progress_implied,
2585                          N_("similar to --all-progress when progress meter is shown")),
2586                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2587                   N_("write the pack index file in the specified idx format version"),
2588                   0, option_parse_index_version },
2589                 OPT_ULONG(0, "max-pack-size", &pack_size_limit,
2590                           N_("maximum size of each output pack file")),
2591                 OPT_BOOL(0, "local", &local,
2592                          N_("ignore borrowed objects from alternate object store")),
2593                 OPT_BOOL(0, "incremental", &incremental,
2594                          N_("ignore packed objects")),
2595                 OPT_INTEGER(0, "window", &window,
2596                             N_("limit pack window by objects")),
2597                 OPT_ULONG(0, "window-memory", &window_memory_limit,
2598                           N_("limit pack window by memory in addition to object limit")),
2599                 OPT_INTEGER(0, "depth", &depth,
2600                             N_("maximum length of delta chain allowed in the resulting pack")),
2601                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2602                          N_("reuse existing deltas")),
2603                 OPT_BOOL(0, "reuse-object", &reuse_object,
2604                          N_("reuse existing objects")),
2605                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2606                          N_("use OFS_DELTA objects")),
2607                 OPT_INTEGER(0, "threads", &delta_search_threads,
2608                             N_("use threads when searching for best delta matches")),
2609                 OPT_BOOL(0, "non-empty", &non_empty,
2610                          N_("do not create an empty pack output")),
2611                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2612                          N_("read revision arguments from standard input")),
2613                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2614                   N_("limit the objects to those that are not yet packed"),
2615                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2616                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2617                   N_("include objects reachable from any reference"),
2618                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2619                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2620                   N_("include objects referred by reflog entries"),
2621                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2622                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2623                          N_("output pack to stdout")),
2624                 OPT_BOOL(0, "include-tag", &include_tag,
2625                          N_("include tag objects that refer to objects to be packed")),
2626                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2627                          N_("keep unreachable objects")),
2628                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2629                   N_("unpack unreachable objects newer than <time>"),
2630                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2631                 OPT_BOOL(0, "thin", &thin,
2632                          N_("create thin packs")),
2633                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2634                          N_("ignore packs that have companion .keep file")),
2635                 OPT_INTEGER(0, "compression", &pack_compression_level,
2636                             N_("pack compression level")),
2637                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2638                             N_("do not hide commits by grafts"), 0),
2639                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2640                          N_("use a bitmap index if available to speed up counting objects")),
2641                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2642                          N_("write a bitmap index together with the pack index")),
2643                 OPT_END(),
2644         };
2645
2646         check_replace_refs = 0;
2647
2648         reset_pack_idx_option(&pack_idx_opts);
2649         git_config(git_pack_config, NULL);
2650         if (!pack_compression_seen && core_compression_seen)
2651                 pack_compression_level = core_compression_level;
2652
2653         progress = isatty(2);
2654         argc = parse_options(argc, argv, prefix, pack_objects_options,
2655                              pack_usage, 0);
2656
2657         if (argc) {
2658                 base_name = argv[0];
2659                 argc--;
2660         }
2661         if (pack_to_stdout != !base_name || argc)
2662                 usage_with_options(pack_usage, pack_objects_options);
2663
2664         rp_av[rp_ac++] = "pack-objects";
2665         if (thin) {
2666                 use_internal_rev_list = 1;
2667                 rp_av[rp_ac++] = "--objects-edge";
2668         } else
2669                 rp_av[rp_ac++] = "--objects";
2670
2671         if (rev_list_all) {
2672                 use_internal_rev_list = 1;
2673                 rp_av[rp_ac++] = "--all";
2674         }
2675         if (rev_list_reflog) {
2676                 use_internal_rev_list = 1;
2677                 rp_av[rp_ac++] = "--reflog";
2678         }
2679         if (rev_list_unpacked) {
2680                 use_internal_rev_list = 1;
2681                 rp_av[rp_ac++] = "--unpacked";
2682         }
2683
2684         if (!reuse_object)
2685                 reuse_delta = 0;
2686         if (pack_compression_level == -1)
2687                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2688         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2689                 die("bad pack compression level %d", pack_compression_level);
2690 #ifdef NO_PTHREADS
2691         if (delta_search_threads != 1)
2692                 warning("no threads support, ignoring --threads");
2693 #endif
2694         if (!pack_to_stdout && !pack_size_limit)
2695                 pack_size_limit = pack_size_limit_cfg;
2696         if (pack_to_stdout && pack_size_limit)
2697                 die("--max-pack-size cannot be used to build a pack for transfer.");
2698         if (pack_size_limit && pack_size_limit < 1024*1024) {
2699                 warning("minimum pack size limit is 1 MiB");
2700                 pack_size_limit = 1024*1024;
2701         }
2702
2703         if (!pack_to_stdout && thin)
2704                 die("--thin cannot be used to build an indexable pack.");
2705
2706         if (keep_unreachable && unpack_unreachable)
2707                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2708
2709         if (!use_internal_rev_list || !pack_to_stdout || is_repository_shallow())
2710                 use_bitmap_index = 0;
2711
2712         if (pack_to_stdout || !rev_list_all)
2713                 write_bitmap_index = 0;
2714
2715         if (progress && all_progress_implied)
2716                 progress = 2;
2717
2718         prepare_packed_git();
2719
2720         if (progress)
2721                 progress_state = start_progress(_("Counting objects"), 0);
2722         if (!use_internal_rev_list)
2723                 read_object_list_from_stdin();
2724         else {
2725                 rp_av[rp_ac] = NULL;
2726                 get_object_list(rp_ac, rp_av);
2727         }
2728         cleanup_preferred_base();
2729         if (include_tag && nr_result)
2730                 for_each_ref(add_ref_tag, NULL);
2731         stop_progress(&progress_state);
2732
2733         if (non_empty && !nr_result)
2734                 return 0;
2735         if (nr_result)
2736                 prepare_pack(window, depth);
2737         write_pack_file();
2738         if (progress)
2739                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2740                         " reused %"PRIu32" (delta %"PRIu32")\n",
2741                         written, written_delta, reused, reused_delta);
2742         return 0;
2743 }