Merge branch 'rd/tag-doc-lightweight'
[git] / builtin / pack-objects.c
1 #include "builtin.h"
2 #include "cache.h"
3 #include "repository.h"
4 #include "config.h"
5 #include "attr.h"
6 #include "object.h"
7 #include "blob.h"
8 #include "commit.h"
9 #include "tag.h"
10 #include "tree.h"
11 #include "delta.h"
12 #include "pack.h"
13 #include "pack-revindex.h"
14 #include "csum-file.h"
15 #include "tree-walk.h"
16 #include "diff.h"
17 #include "revision.h"
18 #include "list-objects.h"
19 #include "list-objects-filter.h"
20 #include "list-objects-filter-options.h"
21 #include "pack-objects.h"
22 #include "progress.h"
23 #include "refs.h"
24 #include "streaming.h"
25 #include "thread-utils.h"
26 #include "pack-bitmap.h"
27 #include "reachable.h"
28 #include "sha1-array.h"
29 #include "argv-array.h"
30 #include "list.h"
31 #include "packfile.h"
32 #include "object-store.h"
33 #include "dir.h"
34
35 #define IN_PACK(obj) oe_in_pack(&to_pack, obj)
36 #define SIZE(obj) oe_size(&to_pack, obj)
37 #define SET_SIZE(obj,size) oe_set_size(&to_pack, obj, size)
38 #define DELTA_SIZE(obj) oe_delta_size(&to_pack, obj)
39 #define DELTA(obj) oe_delta(&to_pack, obj)
40 #define DELTA_CHILD(obj) oe_delta_child(&to_pack, obj)
41 #define DELTA_SIBLING(obj) oe_delta_sibling(&to_pack, obj)
42 #define SET_DELTA(obj, val) oe_set_delta(&to_pack, obj, val)
43 #define SET_DELTA_SIZE(obj, val) oe_set_delta_size(&to_pack, obj, val)
44 #define SET_DELTA_CHILD(obj, val) oe_set_delta_child(&to_pack, obj, val)
45 #define SET_DELTA_SIBLING(obj, val) oe_set_delta_sibling(&to_pack, obj, val)
46
47 static const char *pack_usage[] = {
48         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
49         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
50         NULL
51 };
52
53 /*
54  * Objects we are going to pack are collected in the `to_pack` structure.
55  * It contains an array (dynamically expanded) of the object data, and a map
56  * that can resolve SHA1s to their position in the array.
57  */
58 static struct packing_data to_pack;
59
60 static struct pack_idx_entry **written_list;
61 static uint32_t nr_result, nr_written, nr_seen;
62
63 static int non_empty;
64 static int reuse_delta = 1, reuse_object = 1;
65 static int keep_unreachable, unpack_unreachable, include_tag;
66 static timestamp_t unpack_unreachable_expiration;
67 static int pack_loose_unreachable;
68 static int local;
69 static int have_non_local_packs;
70 static int incremental;
71 static int ignore_packed_keep_on_disk;
72 static int ignore_packed_keep_in_core;
73 static int allow_ofs_delta;
74 static struct pack_idx_option pack_idx_opts;
75 static const char *base_name;
76 static int progress = 1;
77 static int window = 10;
78 static unsigned long pack_size_limit;
79 static int depth = 50;
80 static int delta_search_threads;
81 static int pack_to_stdout;
82 static int num_preferred_base;
83 static struct progress *progress_state;
84
85 static struct packed_git *reuse_packfile;
86 static uint32_t reuse_packfile_objects;
87 static off_t reuse_packfile_offset;
88
89 static int use_bitmap_index_default = 1;
90 static int use_bitmap_index = -1;
91 static int write_bitmap_index;
92 static uint16_t write_bitmap_options;
93
94 static int exclude_promisor_objects;
95
96 static unsigned long delta_cache_size = 0;
97 static unsigned long max_delta_cache_size = DEFAULT_DELTA_CACHE_SIZE;
98 static unsigned long cache_max_small_delta_size = 1000;
99
100 static unsigned long window_memory_limit = 0;
101
102 static struct list_objects_filter_options filter_options;
103
104 enum missing_action {
105         MA_ERROR = 0,      /* fail if any missing objects are encountered */
106         MA_ALLOW_ANY,      /* silently allow ALL missing objects */
107         MA_ALLOW_PROMISOR, /* silently allow all missing PROMISOR objects */
108 };
109 static enum missing_action arg_missing_action;
110 static show_object_fn fn_show_object;
111
112 /*
113  * stats
114  */
115 static uint32_t written, written_delta;
116 static uint32_t reused, reused_delta;
117
118 /*
119  * Indexed commits
120  */
121 static struct commit **indexed_commits;
122 static unsigned int indexed_commits_nr;
123 static unsigned int indexed_commits_alloc;
124
125 static void index_commit_for_bitmap(struct commit *commit)
126 {
127         if (indexed_commits_nr >= indexed_commits_alloc) {
128                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
129                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
130         }
131
132         indexed_commits[indexed_commits_nr++] = commit;
133 }
134
135 static void *get_delta(struct object_entry *entry)
136 {
137         unsigned long size, base_size, delta_size;
138         void *buf, *base_buf, *delta_buf;
139         enum object_type type;
140
141         buf = read_object_file(&entry->idx.oid, &type, &size);
142         if (!buf)
143                 die("unable to read %s", oid_to_hex(&entry->idx.oid));
144         base_buf = read_object_file(&DELTA(entry)->idx.oid, &type,
145                                     &base_size);
146         if (!base_buf)
147                 die("unable to read %s",
148                     oid_to_hex(&DELTA(entry)->idx.oid));
149         delta_buf = diff_delta(base_buf, base_size,
150                                buf, size, &delta_size, 0);
151         if (!delta_buf || delta_size != DELTA_SIZE(entry))
152                 die("delta size changed");
153         free(buf);
154         free(base_buf);
155         return delta_buf;
156 }
157
158 static unsigned long do_compress(void **pptr, unsigned long size)
159 {
160         git_zstream stream;
161         void *in, *out;
162         unsigned long maxsize;
163
164         git_deflate_init(&stream, pack_compression_level);
165         maxsize = git_deflate_bound(&stream, size);
166
167         in = *pptr;
168         out = xmalloc(maxsize);
169         *pptr = out;
170
171         stream.next_in = in;
172         stream.avail_in = size;
173         stream.next_out = out;
174         stream.avail_out = maxsize;
175         while (git_deflate(&stream, Z_FINISH) == Z_OK)
176                 ; /* nothing */
177         git_deflate_end(&stream);
178
179         free(in);
180         return stream.total_out;
181 }
182
183 static unsigned long write_large_blob_data(struct git_istream *st, struct hashfile *f,
184                                            const struct object_id *oid)
185 {
186         git_zstream stream;
187         unsigned char ibuf[1024 * 16];
188         unsigned char obuf[1024 * 16];
189         unsigned long olen = 0;
190
191         git_deflate_init(&stream, pack_compression_level);
192
193         for (;;) {
194                 ssize_t readlen;
195                 int zret = Z_OK;
196                 readlen = read_istream(st, ibuf, sizeof(ibuf));
197                 if (readlen == -1)
198                         die(_("unable to read %s"), oid_to_hex(oid));
199
200                 stream.next_in = ibuf;
201                 stream.avail_in = readlen;
202                 while ((stream.avail_in || readlen == 0) &&
203                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
204                         stream.next_out = obuf;
205                         stream.avail_out = sizeof(obuf);
206                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
207                         hashwrite(f, obuf, stream.next_out - obuf);
208                         olen += stream.next_out - obuf;
209                 }
210                 if (stream.avail_in)
211                         die(_("deflate error (%d)"), zret);
212                 if (readlen == 0) {
213                         if (zret != Z_STREAM_END)
214                                 die(_("deflate error (%d)"), zret);
215                         break;
216                 }
217         }
218         git_deflate_end(&stream);
219         return olen;
220 }
221
222 /*
223  * we are going to reuse the existing object data as is.  make
224  * sure it is not corrupt.
225  */
226 static int check_pack_inflate(struct packed_git *p,
227                 struct pack_window **w_curs,
228                 off_t offset,
229                 off_t len,
230                 unsigned long expect)
231 {
232         git_zstream stream;
233         unsigned char fakebuf[4096], *in;
234         int st;
235
236         memset(&stream, 0, sizeof(stream));
237         git_inflate_init(&stream);
238         do {
239                 in = use_pack(p, w_curs, offset, &stream.avail_in);
240                 stream.next_in = in;
241                 stream.next_out = fakebuf;
242                 stream.avail_out = sizeof(fakebuf);
243                 st = git_inflate(&stream, Z_FINISH);
244                 offset += stream.next_in - in;
245         } while (st == Z_OK || st == Z_BUF_ERROR);
246         git_inflate_end(&stream);
247         return (st == Z_STREAM_END &&
248                 stream.total_out == expect &&
249                 stream.total_in == len) ? 0 : -1;
250 }
251
252 static void copy_pack_data(struct hashfile *f,
253                 struct packed_git *p,
254                 struct pack_window **w_curs,
255                 off_t offset,
256                 off_t len)
257 {
258         unsigned char *in;
259         unsigned long avail;
260
261         while (len) {
262                 in = use_pack(p, w_curs, offset, &avail);
263                 if (avail > len)
264                         avail = (unsigned long)len;
265                 hashwrite(f, in, avail);
266                 offset += avail;
267                 len -= avail;
268         }
269 }
270
271 /* Return 0 if we will bust the pack-size limit */
272 static unsigned long write_no_reuse_object(struct hashfile *f, struct object_entry *entry,
273                                            unsigned long limit, int usable_delta)
274 {
275         unsigned long size, datalen;
276         unsigned char header[MAX_PACK_OBJECT_HEADER],
277                       dheader[MAX_PACK_OBJECT_HEADER];
278         unsigned hdrlen;
279         enum object_type type;
280         void *buf;
281         struct git_istream *st = NULL;
282         const unsigned hashsz = the_hash_algo->rawsz;
283
284         if (!usable_delta) {
285                 if (oe_type(entry) == OBJ_BLOB &&
286                     oe_size_greater_than(&to_pack, entry, big_file_threshold) &&
287                     (st = open_istream(&entry->idx.oid, &type, &size, NULL)) != NULL)
288                         buf = NULL;
289                 else {
290                         buf = read_object_file(&entry->idx.oid, &type, &size);
291                         if (!buf)
292                                 die(_("unable to read %s"),
293                                     oid_to_hex(&entry->idx.oid));
294                 }
295                 /*
296                  * make sure no cached delta data remains from a
297                  * previous attempt before a pack split occurred.
298                  */
299                 FREE_AND_NULL(entry->delta_data);
300                 entry->z_delta_size = 0;
301         } else if (entry->delta_data) {
302                 size = DELTA_SIZE(entry);
303                 buf = entry->delta_data;
304                 entry->delta_data = NULL;
305                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
306                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
307         } else {
308                 buf = get_delta(entry);
309                 size = DELTA_SIZE(entry);
310                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
311                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
312         }
313
314         if (st) /* large blob case, just assume we don't compress well */
315                 datalen = size;
316         else if (entry->z_delta_size)
317                 datalen = entry->z_delta_size;
318         else
319                 datalen = do_compress(&buf, size);
320
321         /*
322          * The object header is a byte of 'type' followed by zero or
323          * more bytes of length.
324          */
325         hdrlen = encode_in_pack_object_header(header, sizeof(header),
326                                               type, size);
327
328         if (type == OBJ_OFS_DELTA) {
329                 /*
330                  * Deltas with relative base contain an additional
331                  * encoding of the relative offset for the delta
332                  * base from this object's position in the pack.
333                  */
334                 off_t ofs = entry->idx.offset - DELTA(entry)->idx.offset;
335                 unsigned pos = sizeof(dheader) - 1;
336                 dheader[pos] = ofs & 127;
337                 while (ofs >>= 7)
338                         dheader[--pos] = 128 | (--ofs & 127);
339                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + hashsz >= limit) {
340                         if (st)
341                                 close_istream(st);
342                         free(buf);
343                         return 0;
344                 }
345                 hashwrite(f, header, hdrlen);
346                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
347                 hdrlen += sizeof(dheader) - pos;
348         } else if (type == OBJ_REF_DELTA) {
349                 /*
350                  * Deltas with a base reference contain
351                  * additional bytes for the base object ID.
352                  */
353                 if (limit && hdrlen + hashsz + datalen + hashsz >= limit) {
354                         if (st)
355                                 close_istream(st);
356                         free(buf);
357                         return 0;
358                 }
359                 hashwrite(f, header, hdrlen);
360                 hashwrite(f, DELTA(entry)->idx.oid.hash, hashsz);
361                 hdrlen += hashsz;
362         } else {
363                 if (limit && hdrlen + datalen + hashsz >= limit) {
364                         if (st)
365                                 close_istream(st);
366                         free(buf);
367                         return 0;
368                 }
369                 hashwrite(f, header, hdrlen);
370         }
371         if (st) {
372                 datalen = write_large_blob_data(st, f, &entry->idx.oid);
373                 close_istream(st);
374         } else {
375                 hashwrite(f, buf, datalen);
376                 free(buf);
377         }
378
379         return hdrlen + datalen;
380 }
381
382 /* Return 0 if we will bust the pack-size limit */
383 static off_t write_reuse_object(struct hashfile *f, struct object_entry *entry,
384                                 unsigned long limit, int usable_delta)
385 {
386         struct packed_git *p = IN_PACK(entry);
387         struct pack_window *w_curs = NULL;
388         struct revindex_entry *revidx;
389         off_t offset;
390         enum object_type type = oe_type(entry);
391         off_t datalen;
392         unsigned char header[MAX_PACK_OBJECT_HEADER],
393                       dheader[MAX_PACK_OBJECT_HEADER];
394         unsigned hdrlen;
395         const unsigned hashsz = the_hash_algo->rawsz;
396         unsigned long entry_size = SIZE(entry);
397
398         if (DELTA(entry))
399                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
400                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
401         hdrlen = encode_in_pack_object_header(header, sizeof(header),
402                                               type, entry_size);
403
404         offset = entry->in_pack_offset;
405         revidx = find_pack_revindex(p, offset);
406         datalen = revidx[1].offset - offset;
407         if (!pack_to_stdout && p->index_version > 1 &&
408             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
409                 error("bad packed object CRC for %s",
410                       oid_to_hex(&entry->idx.oid));
411                 unuse_pack(&w_curs);
412                 return write_no_reuse_object(f, entry, limit, usable_delta);
413         }
414
415         offset += entry->in_pack_header_size;
416         datalen -= entry->in_pack_header_size;
417
418         if (!pack_to_stdout && p->index_version == 1 &&
419             check_pack_inflate(p, &w_curs, offset, datalen, entry_size)) {
420                 error("corrupt packed object for %s",
421                       oid_to_hex(&entry->idx.oid));
422                 unuse_pack(&w_curs);
423                 return write_no_reuse_object(f, entry, limit, usable_delta);
424         }
425
426         if (type == OBJ_OFS_DELTA) {
427                 off_t ofs = entry->idx.offset - DELTA(entry)->idx.offset;
428                 unsigned pos = sizeof(dheader) - 1;
429                 dheader[pos] = ofs & 127;
430                 while (ofs >>= 7)
431                         dheader[--pos] = 128 | (--ofs & 127);
432                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + hashsz >= limit) {
433                         unuse_pack(&w_curs);
434                         return 0;
435                 }
436                 hashwrite(f, header, hdrlen);
437                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
438                 hdrlen += sizeof(dheader) - pos;
439                 reused_delta++;
440         } else if (type == OBJ_REF_DELTA) {
441                 if (limit && hdrlen + hashsz + datalen + hashsz >= limit) {
442                         unuse_pack(&w_curs);
443                         return 0;
444                 }
445                 hashwrite(f, header, hdrlen);
446                 hashwrite(f, DELTA(entry)->idx.oid.hash, hashsz);
447                 hdrlen += hashsz;
448                 reused_delta++;
449         } else {
450                 if (limit && hdrlen + datalen + hashsz >= limit) {
451                         unuse_pack(&w_curs);
452                         return 0;
453                 }
454                 hashwrite(f, header, hdrlen);
455         }
456         copy_pack_data(f, p, &w_curs, offset, datalen);
457         unuse_pack(&w_curs);
458         reused++;
459         return hdrlen + datalen;
460 }
461
462 /* Return 0 if we will bust the pack-size limit */
463 static off_t write_object(struct hashfile *f,
464                           struct object_entry *entry,
465                           off_t write_offset)
466 {
467         unsigned long limit;
468         off_t len;
469         int usable_delta, to_reuse;
470
471         if (!pack_to_stdout)
472                 crc32_begin(f);
473
474         /* apply size limit if limited packsize and not first object */
475         if (!pack_size_limit || !nr_written)
476                 limit = 0;
477         else if (pack_size_limit <= write_offset)
478                 /*
479                  * the earlier object did not fit the limit; avoid
480                  * mistaking this with unlimited (i.e. limit = 0).
481                  */
482                 limit = 1;
483         else
484                 limit = pack_size_limit - write_offset;
485
486         if (!DELTA(entry))
487                 usable_delta = 0;       /* no delta */
488         else if (!pack_size_limit)
489                usable_delta = 1;        /* unlimited packfile */
490         else if (DELTA(entry)->idx.offset == (off_t)-1)
491                 usable_delta = 0;       /* base was written to another pack */
492         else if (DELTA(entry)->idx.offset)
493                 usable_delta = 1;       /* base already exists in this pack */
494         else
495                 usable_delta = 0;       /* base could end up in another pack */
496
497         if (!reuse_object)
498                 to_reuse = 0;   /* explicit */
499         else if (!IN_PACK(entry))
500                 to_reuse = 0;   /* can't reuse what we don't have */
501         else if (oe_type(entry) == OBJ_REF_DELTA ||
502                  oe_type(entry) == OBJ_OFS_DELTA)
503                                 /* check_object() decided it for us ... */
504                 to_reuse = usable_delta;
505                                 /* ... but pack split may override that */
506         else if (oe_type(entry) != entry->in_pack_type)
507                 to_reuse = 0;   /* pack has delta which is unusable */
508         else if (DELTA(entry))
509                 to_reuse = 0;   /* we want to pack afresh */
510         else
511                 to_reuse = 1;   /* we have it in-pack undeltified,
512                                  * and we do not need to deltify it.
513                                  */
514
515         if (!to_reuse)
516                 len = write_no_reuse_object(f, entry, limit, usable_delta);
517         else
518                 len = write_reuse_object(f, entry, limit, usable_delta);
519         if (!len)
520                 return 0;
521
522         if (usable_delta)
523                 written_delta++;
524         written++;
525         if (!pack_to_stdout)
526                 entry->idx.crc32 = crc32_end(f);
527         return len;
528 }
529
530 enum write_one_status {
531         WRITE_ONE_SKIP = -1, /* already written */
532         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
533         WRITE_ONE_WRITTEN = 1, /* normal */
534         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
535 };
536
537 static enum write_one_status write_one(struct hashfile *f,
538                                        struct object_entry *e,
539                                        off_t *offset)
540 {
541         off_t size;
542         int recursing;
543
544         /*
545          * we set offset to 1 (which is an impossible value) to mark
546          * the fact that this object is involved in "write its base
547          * first before writing a deltified object" recursion.
548          */
549         recursing = (e->idx.offset == 1);
550         if (recursing) {
551                 warning("recursive delta detected for object %s",
552                         oid_to_hex(&e->idx.oid));
553                 return WRITE_ONE_RECURSIVE;
554         } else if (e->idx.offset || e->preferred_base) {
555                 /* offset is non zero if object is written already. */
556                 return WRITE_ONE_SKIP;
557         }
558
559         /* if we are deltified, write out base object first. */
560         if (DELTA(e)) {
561                 e->idx.offset = 1; /* now recurse */
562                 switch (write_one(f, DELTA(e), offset)) {
563                 case WRITE_ONE_RECURSIVE:
564                         /* we cannot depend on this one */
565                         SET_DELTA(e, NULL);
566                         break;
567                 default:
568                         break;
569                 case WRITE_ONE_BREAK:
570                         e->idx.offset = recursing;
571                         return WRITE_ONE_BREAK;
572                 }
573         }
574
575         e->idx.offset = *offset;
576         size = write_object(f, e, *offset);
577         if (!size) {
578                 e->idx.offset = recursing;
579                 return WRITE_ONE_BREAK;
580         }
581         written_list[nr_written++] = &e->idx;
582
583         /* make sure off_t is sufficiently large not to wrap */
584         if (signed_add_overflows(*offset, size))
585                 die("pack too large for current definition of off_t");
586         *offset += size;
587         return WRITE_ONE_WRITTEN;
588 }
589
590 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
591                        void *cb_data)
592 {
593         struct object_id peeled;
594         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
595
596         if (entry)
597                 entry->tagged = 1;
598         if (!peel_ref(path, &peeled)) {
599                 entry = packlist_find(&to_pack, peeled.hash, NULL);
600                 if (entry)
601                         entry->tagged = 1;
602         }
603         return 0;
604 }
605
606 static inline void add_to_write_order(struct object_entry **wo,
607                                unsigned int *endp,
608                                struct object_entry *e)
609 {
610         if (e->filled)
611                 return;
612         wo[(*endp)++] = e;
613         e->filled = 1;
614 }
615
616 static void add_descendants_to_write_order(struct object_entry **wo,
617                                            unsigned int *endp,
618                                            struct object_entry *e)
619 {
620         int add_to_order = 1;
621         while (e) {
622                 if (add_to_order) {
623                         struct object_entry *s;
624                         /* add this node... */
625                         add_to_write_order(wo, endp, e);
626                         /* all its siblings... */
627                         for (s = DELTA_SIBLING(e); s; s = DELTA_SIBLING(s)) {
628                                 add_to_write_order(wo, endp, s);
629                         }
630                 }
631                 /* drop down a level to add left subtree nodes if possible */
632                 if (DELTA_CHILD(e)) {
633                         add_to_order = 1;
634                         e = DELTA_CHILD(e);
635                 } else {
636                         add_to_order = 0;
637                         /* our sibling might have some children, it is next */
638                         if (DELTA_SIBLING(e)) {
639                                 e = DELTA_SIBLING(e);
640                                 continue;
641                         }
642                         /* go back to our parent node */
643                         e = DELTA(e);
644                         while (e && !DELTA_SIBLING(e)) {
645                                 /* we're on the right side of a subtree, keep
646                                  * going up until we can go right again */
647                                 e = DELTA(e);
648                         }
649                         if (!e) {
650                                 /* done- we hit our original root node */
651                                 return;
652                         }
653                         /* pass it off to sibling at this level */
654                         e = DELTA_SIBLING(e);
655                 }
656         };
657 }
658
659 static void add_family_to_write_order(struct object_entry **wo,
660                                       unsigned int *endp,
661                                       struct object_entry *e)
662 {
663         struct object_entry *root;
664
665         for (root = e; DELTA(root); root = DELTA(root))
666                 ; /* nothing */
667         add_descendants_to_write_order(wo, endp, root);
668 }
669
670 static struct object_entry **compute_write_order(void)
671 {
672         unsigned int i, wo_end, last_untagged;
673
674         struct object_entry **wo;
675         struct object_entry *objects = to_pack.objects;
676
677         for (i = 0; i < to_pack.nr_objects; i++) {
678                 objects[i].tagged = 0;
679                 objects[i].filled = 0;
680                 SET_DELTA_CHILD(&objects[i], NULL);
681                 SET_DELTA_SIBLING(&objects[i], NULL);
682         }
683
684         /*
685          * Fully connect delta_child/delta_sibling network.
686          * Make sure delta_sibling is sorted in the original
687          * recency order.
688          */
689         for (i = to_pack.nr_objects; i > 0;) {
690                 struct object_entry *e = &objects[--i];
691                 if (!DELTA(e))
692                         continue;
693                 /* Mark me as the first child */
694                 e->delta_sibling_idx = DELTA(e)->delta_child_idx;
695                 SET_DELTA_CHILD(DELTA(e), e);
696         }
697
698         /*
699          * Mark objects that are at the tip of tags.
700          */
701         for_each_tag_ref(mark_tagged, NULL);
702
703         /*
704          * Give the objects in the original recency order until
705          * we see a tagged tip.
706          */
707         ALLOC_ARRAY(wo, to_pack.nr_objects);
708         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
709                 if (objects[i].tagged)
710                         break;
711                 add_to_write_order(wo, &wo_end, &objects[i]);
712         }
713         last_untagged = i;
714
715         /*
716          * Then fill all the tagged tips.
717          */
718         for (; i < to_pack.nr_objects; i++) {
719                 if (objects[i].tagged)
720                         add_to_write_order(wo, &wo_end, &objects[i]);
721         }
722
723         /*
724          * And then all remaining commits and tags.
725          */
726         for (i = last_untagged; i < to_pack.nr_objects; i++) {
727                 if (oe_type(&objects[i]) != OBJ_COMMIT &&
728                     oe_type(&objects[i]) != OBJ_TAG)
729                         continue;
730                 add_to_write_order(wo, &wo_end, &objects[i]);
731         }
732
733         /*
734          * And then all the trees.
735          */
736         for (i = last_untagged; i < to_pack.nr_objects; i++) {
737                 if (oe_type(&objects[i]) != OBJ_TREE)
738                         continue;
739                 add_to_write_order(wo, &wo_end, &objects[i]);
740         }
741
742         /*
743          * Finally all the rest in really tight order
744          */
745         for (i = last_untagged; i < to_pack.nr_objects; i++) {
746                 if (!objects[i].filled)
747                         add_family_to_write_order(wo, &wo_end, &objects[i]);
748         }
749
750         if (wo_end != to_pack.nr_objects)
751                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
752
753         return wo;
754 }
755
756 static off_t write_reused_pack(struct hashfile *f)
757 {
758         unsigned char buffer[8192];
759         off_t to_write, total;
760         int fd;
761
762         if (!is_pack_valid(reuse_packfile))
763                 die("packfile is invalid: %s", reuse_packfile->pack_name);
764
765         fd = git_open(reuse_packfile->pack_name);
766         if (fd < 0)
767                 die_errno("unable to open packfile for reuse: %s",
768                           reuse_packfile->pack_name);
769
770         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
771                 die_errno("unable to seek in reused packfile");
772
773         if (reuse_packfile_offset < 0)
774                 reuse_packfile_offset = reuse_packfile->pack_size - the_hash_algo->rawsz;
775
776         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
777
778         while (to_write) {
779                 int read_pack = xread(fd, buffer, sizeof(buffer));
780
781                 if (read_pack <= 0)
782                         die_errno("unable to read from reused packfile");
783
784                 if (read_pack > to_write)
785                         read_pack = to_write;
786
787                 hashwrite(f, buffer, read_pack);
788                 to_write -= read_pack;
789
790                 /*
791                  * We don't know the actual number of objects written,
792                  * only how many bytes written, how many bytes total, and
793                  * how many objects total. So we can fake it by pretending all
794                  * objects we are writing are the same size. This gives us a
795                  * smooth progress meter, and at the end it matches the true
796                  * answer.
797                  */
798                 written = reuse_packfile_objects *
799                                 (((double)(total - to_write)) / total);
800                 display_progress(progress_state, written);
801         }
802
803         close(fd);
804         written = reuse_packfile_objects;
805         display_progress(progress_state, written);
806         return reuse_packfile_offset - sizeof(struct pack_header);
807 }
808
809 static const char no_split_warning[] = N_(
810 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
811 );
812
813 static void write_pack_file(void)
814 {
815         uint32_t i = 0, j;
816         struct hashfile *f;
817         off_t offset;
818         uint32_t nr_remaining = nr_result;
819         time_t last_mtime = 0;
820         struct object_entry **write_order;
821
822         if (progress > pack_to_stdout)
823                 progress_state = start_progress(_("Writing objects"), nr_result);
824         ALLOC_ARRAY(written_list, to_pack.nr_objects);
825         write_order = compute_write_order();
826
827         do {
828                 struct object_id oid;
829                 char *pack_tmp_name = NULL;
830
831                 if (pack_to_stdout)
832                         f = hashfd_throughput(1, "<stdout>", progress_state);
833                 else
834                         f = create_tmp_packfile(&pack_tmp_name);
835
836                 offset = write_pack_header(f, nr_remaining);
837
838                 if (reuse_packfile) {
839                         off_t packfile_size;
840                         assert(pack_to_stdout);
841
842                         packfile_size = write_reused_pack(f);
843                         offset += packfile_size;
844                 }
845
846                 nr_written = 0;
847                 for (; i < to_pack.nr_objects; i++) {
848                         struct object_entry *e = write_order[i];
849                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
850                                 break;
851                         display_progress(progress_state, written);
852                 }
853
854                 /*
855                  * Did we write the wrong # entries in the header?
856                  * If so, rewrite it like in fast-import
857                  */
858                 if (pack_to_stdout) {
859                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_CLOSE);
860                 } else if (nr_written == nr_remaining) {
861                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE);
862                 } else {
863                         int fd = finalize_hashfile(f, oid.hash, 0);
864                         fixup_pack_header_footer(fd, oid.hash, pack_tmp_name,
865                                                  nr_written, oid.hash, offset);
866                         close(fd);
867                         if (write_bitmap_index) {
868                                 warning(_(no_split_warning));
869                                 write_bitmap_index = 0;
870                         }
871                 }
872
873                 if (!pack_to_stdout) {
874                         struct stat st;
875                         struct strbuf tmpname = STRBUF_INIT;
876
877                         /*
878                          * Packs are runtime accessed in their mtime
879                          * order since newer packs are more likely to contain
880                          * younger objects.  So if we are creating multiple
881                          * packs then we should modify the mtime of later ones
882                          * to preserve this property.
883                          */
884                         if (stat(pack_tmp_name, &st) < 0) {
885                                 warning_errno("failed to stat %s", pack_tmp_name);
886                         } else if (!last_mtime) {
887                                 last_mtime = st.st_mtime;
888                         } else {
889                                 struct utimbuf utb;
890                                 utb.actime = st.st_atime;
891                                 utb.modtime = --last_mtime;
892                                 if (utime(pack_tmp_name, &utb) < 0)
893                                         warning_errno("failed utime() on %s", pack_tmp_name);
894                         }
895
896                         strbuf_addf(&tmpname, "%s-", base_name);
897
898                         if (write_bitmap_index) {
899                                 bitmap_writer_set_checksum(oid.hash);
900                                 bitmap_writer_build_type_index(
901                                         &to_pack, written_list, nr_written);
902                         }
903
904                         finish_tmp_packfile(&tmpname, pack_tmp_name,
905                                             written_list, nr_written,
906                                             &pack_idx_opts, oid.hash);
907
908                         if (write_bitmap_index) {
909                                 strbuf_addf(&tmpname, "%s.bitmap", oid_to_hex(&oid));
910
911                                 stop_progress(&progress_state);
912
913                                 bitmap_writer_show_progress(progress);
914                                 bitmap_writer_reuse_bitmaps(&to_pack);
915                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
916                                 bitmap_writer_build(&to_pack);
917                                 bitmap_writer_finish(written_list, nr_written,
918                                                      tmpname.buf, write_bitmap_options);
919                                 write_bitmap_index = 0;
920                         }
921
922                         strbuf_release(&tmpname);
923                         free(pack_tmp_name);
924                         puts(oid_to_hex(&oid));
925                 }
926
927                 /* mark written objects as written to previous pack */
928                 for (j = 0; j < nr_written; j++) {
929                         written_list[j]->offset = (off_t)-1;
930                 }
931                 nr_remaining -= nr_written;
932         } while (nr_remaining && i < to_pack.nr_objects);
933
934         free(written_list);
935         free(write_order);
936         stop_progress(&progress_state);
937         if (written != nr_result)
938                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
939                         written, nr_result);
940 }
941
942 static int no_try_delta(const char *path)
943 {
944         static struct attr_check *check;
945
946         if (!check)
947                 check = attr_check_initl("delta", NULL);
948         if (git_check_attr(path, check))
949                 return 0;
950         if (ATTR_FALSE(check->items[0].value))
951                 return 1;
952         return 0;
953 }
954
955 /*
956  * When adding an object, check whether we have already added it
957  * to our packing list. If so, we can skip. However, if we are
958  * being asked to excludei t, but the previous mention was to include
959  * it, make sure to adjust its flags and tweak our numbers accordingly.
960  *
961  * As an optimization, we pass out the index position where we would have
962  * found the item, since that saves us from having to look it up again a
963  * few lines later when we want to add the new entry.
964  */
965 static int have_duplicate_entry(const struct object_id *oid,
966                                 int exclude,
967                                 uint32_t *index_pos)
968 {
969         struct object_entry *entry;
970
971         entry = packlist_find(&to_pack, oid->hash, index_pos);
972         if (!entry)
973                 return 0;
974
975         if (exclude) {
976                 if (!entry->preferred_base)
977                         nr_result--;
978                 entry->preferred_base = 1;
979         }
980
981         return 1;
982 }
983
984 static int want_found_object(int exclude, struct packed_git *p)
985 {
986         if (exclude)
987                 return 1;
988         if (incremental)
989                 return 0;
990
991         /*
992          * When asked to do --local (do not include an object that appears in a
993          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
994          * an object that appears in a pack marked with .keep), finding a pack
995          * that matches the criteria is sufficient for us to decide to omit it.
996          * However, even if this pack does not satisfy the criteria, we need to
997          * make sure no copy of this object appears in _any_ pack that makes us
998          * to omit the object, so we need to check all the packs.
999          *
1000          * We can however first check whether these options can possible matter;
1001          * if they do not matter we know we want the object in generated pack.
1002          * Otherwise, we signal "-1" at the end to tell the caller that we do
1003          * not know either way, and it needs to check more packs.
1004          */
1005         if (!ignore_packed_keep_on_disk &&
1006             !ignore_packed_keep_in_core &&
1007             (!local || !have_non_local_packs))
1008                 return 1;
1009
1010         if (local && !p->pack_local)
1011                 return 0;
1012         if (p->pack_local &&
1013             ((ignore_packed_keep_on_disk && p->pack_keep) ||
1014              (ignore_packed_keep_in_core && p->pack_keep_in_core)))
1015                 return 0;
1016
1017         /* we don't know yet; keep looking for more packs */
1018         return -1;
1019 }
1020
1021 /*
1022  * Check whether we want the object in the pack (e.g., we do not want
1023  * objects found in non-local stores if the "--local" option was used).
1024  *
1025  * If the caller already knows an existing pack it wants to take the object
1026  * from, that is passed in *found_pack and *found_offset; otherwise this
1027  * function finds if there is any pack that has the object and returns the pack
1028  * and its offset in these variables.
1029  */
1030 static int want_object_in_pack(const struct object_id *oid,
1031                                int exclude,
1032                                struct packed_git **found_pack,
1033                                off_t *found_offset)
1034 {
1035         int want;
1036         struct list_head *pos;
1037
1038         if (!exclude && local && has_loose_object_nonlocal(oid))
1039                 return 0;
1040
1041         /*
1042          * If we already know the pack object lives in, start checks from that
1043          * pack - in the usual case when neither --local was given nor .keep files
1044          * are present we will determine the answer right now.
1045          */
1046         if (*found_pack) {
1047                 want = want_found_object(exclude, *found_pack);
1048                 if (want != -1)
1049                         return want;
1050         }
1051         list_for_each(pos, get_packed_git_mru(the_repository)) {
1052                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
1053                 off_t offset;
1054
1055                 if (p == *found_pack)
1056                         offset = *found_offset;
1057                 else
1058                         offset = find_pack_entry_one(oid->hash, p);
1059
1060                 if (offset) {
1061                         if (!*found_pack) {
1062                                 if (!is_pack_valid(p))
1063                                         continue;
1064                                 *found_offset = offset;
1065                                 *found_pack = p;
1066                         }
1067                         want = want_found_object(exclude, p);
1068                         if (!exclude && want > 0)
1069                                 list_move(&p->mru,
1070                                           get_packed_git_mru(the_repository));
1071                         if (want != -1)
1072                                 return want;
1073                 }
1074         }
1075
1076         return 1;
1077 }
1078
1079 static void create_object_entry(const struct object_id *oid,
1080                                 enum object_type type,
1081                                 uint32_t hash,
1082                                 int exclude,
1083                                 int no_try_delta,
1084                                 uint32_t index_pos,
1085                                 struct packed_git *found_pack,
1086                                 off_t found_offset)
1087 {
1088         struct object_entry *entry;
1089
1090         entry = packlist_alloc(&to_pack, oid->hash, index_pos);
1091         entry->hash = hash;
1092         oe_set_type(entry, type);
1093         if (exclude)
1094                 entry->preferred_base = 1;
1095         else
1096                 nr_result++;
1097         if (found_pack) {
1098                 oe_set_in_pack(&to_pack, entry, found_pack);
1099                 entry->in_pack_offset = found_offset;
1100         }
1101
1102         entry->no_try_delta = no_try_delta;
1103 }
1104
1105 static const char no_closure_warning[] = N_(
1106 "disabling bitmap writing, as some objects are not being packed"
1107 );
1108
1109 static int add_object_entry(const struct object_id *oid, enum object_type type,
1110                             const char *name, int exclude)
1111 {
1112         struct packed_git *found_pack = NULL;
1113         off_t found_offset = 0;
1114         uint32_t index_pos;
1115
1116         display_progress(progress_state, ++nr_seen);
1117
1118         if (have_duplicate_entry(oid, exclude, &index_pos))
1119                 return 0;
1120
1121         if (!want_object_in_pack(oid, exclude, &found_pack, &found_offset)) {
1122                 /* The pack is missing an object, so it will not have closure */
1123                 if (write_bitmap_index) {
1124                         warning(_(no_closure_warning));
1125                         write_bitmap_index = 0;
1126                 }
1127                 return 0;
1128         }
1129
1130         create_object_entry(oid, type, pack_name_hash(name),
1131                             exclude, name && no_try_delta(name),
1132                             index_pos, found_pack, found_offset);
1133         return 1;
1134 }
1135
1136 static int add_object_entry_from_bitmap(const struct object_id *oid,
1137                                         enum object_type type,
1138                                         int flags, uint32_t name_hash,
1139                                         struct packed_git *pack, off_t offset)
1140 {
1141         uint32_t index_pos;
1142
1143         display_progress(progress_state, ++nr_seen);
1144
1145         if (have_duplicate_entry(oid, 0, &index_pos))
1146                 return 0;
1147
1148         if (!want_object_in_pack(oid, 0, &pack, &offset))
1149                 return 0;
1150
1151         create_object_entry(oid, type, name_hash, 0, 0, index_pos, pack, offset);
1152         return 1;
1153 }
1154
1155 struct pbase_tree_cache {
1156         struct object_id oid;
1157         int ref;
1158         int temporary;
1159         void *tree_data;
1160         unsigned long tree_size;
1161 };
1162
1163 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1164 static int pbase_tree_cache_ix(const struct object_id *oid)
1165 {
1166         return oid->hash[0] % ARRAY_SIZE(pbase_tree_cache);
1167 }
1168 static int pbase_tree_cache_ix_incr(int ix)
1169 {
1170         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1171 }
1172
1173 static struct pbase_tree {
1174         struct pbase_tree *next;
1175         /* This is a phony "cache" entry; we are not
1176          * going to evict it or find it through _get()
1177          * mechanism -- this is for the toplevel node that
1178          * would almost always change with any commit.
1179          */
1180         struct pbase_tree_cache pcache;
1181 } *pbase_tree;
1182
1183 static struct pbase_tree_cache *pbase_tree_get(const struct object_id *oid)
1184 {
1185         struct pbase_tree_cache *ent, *nent;
1186         void *data;
1187         unsigned long size;
1188         enum object_type type;
1189         int neigh;
1190         int my_ix = pbase_tree_cache_ix(oid);
1191         int available_ix = -1;
1192
1193         /* pbase-tree-cache acts as a limited hashtable.
1194          * your object will be found at your index or within a few
1195          * slots after that slot if it is cached.
1196          */
1197         for (neigh = 0; neigh < 8; neigh++) {
1198                 ent = pbase_tree_cache[my_ix];
1199                 if (ent && !oidcmp(&ent->oid, oid)) {
1200                         ent->ref++;
1201                         return ent;
1202                 }
1203                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1204                          ((0 <= available_ix) &&
1205                           (!ent && pbase_tree_cache[available_ix])))
1206                         available_ix = my_ix;
1207                 if (!ent)
1208                         break;
1209                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1210         }
1211
1212         /* Did not find one.  Either we got a bogus request or
1213          * we need to read and perhaps cache.
1214          */
1215         data = read_object_file(oid, &type, &size);
1216         if (!data)
1217                 return NULL;
1218         if (type != OBJ_TREE) {
1219                 free(data);
1220                 return NULL;
1221         }
1222
1223         /* We need to either cache or return a throwaway copy */
1224
1225         if (available_ix < 0)
1226                 ent = NULL;
1227         else {
1228                 ent = pbase_tree_cache[available_ix];
1229                 my_ix = available_ix;
1230         }
1231
1232         if (!ent) {
1233                 nent = xmalloc(sizeof(*nent));
1234                 nent->temporary = (available_ix < 0);
1235         }
1236         else {
1237                 /* evict and reuse */
1238                 free(ent->tree_data);
1239                 nent = ent;
1240         }
1241         oidcpy(&nent->oid, oid);
1242         nent->tree_data = data;
1243         nent->tree_size = size;
1244         nent->ref = 1;
1245         if (!nent->temporary)
1246                 pbase_tree_cache[my_ix] = nent;
1247         return nent;
1248 }
1249
1250 static void pbase_tree_put(struct pbase_tree_cache *cache)
1251 {
1252         if (!cache->temporary) {
1253                 cache->ref--;
1254                 return;
1255         }
1256         free(cache->tree_data);
1257         free(cache);
1258 }
1259
1260 static int name_cmp_len(const char *name)
1261 {
1262         int i;
1263         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1264                 ;
1265         return i;
1266 }
1267
1268 static void add_pbase_object(struct tree_desc *tree,
1269                              const char *name,
1270                              int cmplen,
1271                              const char *fullname)
1272 {
1273         struct name_entry entry;
1274         int cmp;
1275
1276         while (tree_entry(tree,&entry)) {
1277                 if (S_ISGITLINK(entry.mode))
1278                         continue;
1279                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1280                       memcmp(name, entry.path, cmplen);
1281                 if (cmp > 0)
1282                         continue;
1283                 if (cmp < 0)
1284                         return;
1285                 if (name[cmplen] != '/') {
1286                         add_object_entry(entry.oid,
1287                                          object_type(entry.mode),
1288                                          fullname, 1);
1289                         return;
1290                 }
1291                 if (S_ISDIR(entry.mode)) {
1292                         struct tree_desc sub;
1293                         struct pbase_tree_cache *tree;
1294                         const char *down = name+cmplen+1;
1295                         int downlen = name_cmp_len(down);
1296
1297                         tree = pbase_tree_get(entry.oid);
1298                         if (!tree)
1299                                 return;
1300                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1301
1302                         add_pbase_object(&sub, down, downlen, fullname);
1303                         pbase_tree_put(tree);
1304                 }
1305         }
1306 }
1307
1308 static unsigned *done_pbase_paths;
1309 static int done_pbase_paths_num;
1310 static int done_pbase_paths_alloc;
1311 static int done_pbase_path_pos(unsigned hash)
1312 {
1313         int lo = 0;
1314         int hi = done_pbase_paths_num;
1315         while (lo < hi) {
1316                 int mi = lo + (hi - lo) / 2;
1317                 if (done_pbase_paths[mi] == hash)
1318                         return mi;
1319                 if (done_pbase_paths[mi] < hash)
1320                         hi = mi;
1321                 else
1322                         lo = mi + 1;
1323         }
1324         return -lo-1;
1325 }
1326
1327 static int check_pbase_path(unsigned hash)
1328 {
1329         int pos = done_pbase_path_pos(hash);
1330         if (0 <= pos)
1331                 return 1;
1332         pos = -pos - 1;
1333         ALLOC_GROW(done_pbase_paths,
1334                    done_pbase_paths_num + 1,
1335                    done_pbase_paths_alloc);
1336         done_pbase_paths_num++;
1337         if (pos < done_pbase_paths_num)
1338                 MOVE_ARRAY(done_pbase_paths + pos + 1, done_pbase_paths + pos,
1339                            done_pbase_paths_num - pos - 1);
1340         done_pbase_paths[pos] = hash;
1341         return 0;
1342 }
1343
1344 static void add_preferred_base_object(const char *name)
1345 {
1346         struct pbase_tree *it;
1347         int cmplen;
1348         unsigned hash = pack_name_hash(name);
1349
1350         if (!num_preferred_base || check_pbase_path(hash))
1351                 return;
1352
1353         cmplen = name_cmp_len(name);
1354         for (it = pbase_tree; it; it = it->next) {
1355                 if (cmplen == 0) {
1356                         add_object_entry(&it->pcache.oid, OBJ_TREE, NULL, 1);
1357                 }
1358                 else {
1359                         struct tree_desc tree;
1360                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1361                         add_pbase_object(&tree, name, cmplen, name);
1362                 }
1363         }
1364 }
1365
1366 static void add_preferred_base(struct object_id *oid)
1367 {
1368         struct pbase_tree *it;
1369         void *data;
1370         unsigned long size;
1371         struct object_id tree_oid;
1372
1373         if (window <= num_preferred_base++)
1374                 return;
1375
1376         data = read_object_with_reference(oid, tree_type, &size, &tree_oid);
1377         if (!data)
1378                 return;
1379
1380         for (it = pbase_tree; it; it = it->next) {
1381                 if (!oidcmp(&it->pcache.oid, &tree_oid)) {
1382                         free(data);
1383                         return;
1384                 }
1385         }
1386
1387         it = xcalloc(1, sizeof(*it));
1388         it->next = pbase_tree;
1389         pbase_tree = it;
1390
1391         oidcpy(&it->pcache.oid, &tree_oid);
1392         it->pcache.tree_data = data;
1393         it->pcache.tree_size = size;
1394 }
1395
1396 static void cleanup_preferred_base(void)
1397 {
1398         struct pbase_tree *it;
1399         unsigned i;
1400
1401         it = pbase_tree;
1402         pbase_tree = NULL;
1403         while (it) {
1404                 struct pbase_tree *tmp = it;
1405                 it = tmp->next;
1406                 free(tmp->pcache.tree_data);
1407                 free(tmp);
1408         }
1409
1410         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1411                 if (!pbase_tree_cache[i])
1412                         continue;
1413                 free(pbase_tree_cache[i]->tree_data);
1414                 FREE_AND_NULL(pbase_tree_cache[i]);
1415         }
1416
1417         FREE_AND_NULL(done_pbase_paths);
1418         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1419 }
1420
1421 static void check_object(struct object_entry *entry)
1422 {
1423         unsigned long canonical_size;
1424
1425         if (IN_PACK(entry)) {
1426                 struct packed_git *p = IN_PACK(entry);
1427                 struct pack_window *w_curs = NULL;
1428                 const unsigned char *base_ref = NULL;
1429                 struct object_entry *base_entry;
1430                 unsigned long used, used_0;
1431                 unsigned long avail;
1432                 off_t ofs;
1433                 unsigned char *buf, c;
1434                 enum object_type type;
1435                 unsigned long in_pack_size;
1436
1437                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1438
1439                 /*
1440                  * We want in_pack_type even if we do not reuse delta
1441                  * since non-delta representations could still be reused.
1442                  */
1443                 used = unpack_object_header_buffer(buf, avail,
1444                                                    &type,
1445                                                    &in_pack_size);
1446                 if (used == 0)
1447                         goto give_up;
1448
1449                 if (type < 0)
1450                         BUG("invalid type %d", type);
1451                 entry->in_pack_type = type;
1452
1453                 /*
1454                  * Determine if this is a delta and if so whether we can
1455                  * reuse it or not.  Otherwise let's find out as cheaply as
1456                  * possible what the actual type and size for this object is.
1457                  */
1458                 switch (entry->in_pack_type) {
1459                 default:
1460                         /* Not a delta hence we've already got all we need. */
1461                         oe_set_type(entry, entry->in_pack_type);
1462                         SET_SIZE(entry, in_pack_size);
1463                         entry->in_pack_header_size = used;
1464                         if (oe_type(entry) < OBJ_COMMIT || oe_type(entry) > OBJ_BLOB)
1465                                 goto give_up;
1466                         unuse_pack(&w_curs);
1467                         return;
1468                 case OBJ_REF_DELTA:
1469                         if (reuse_delta && !entry->preferred_base)
1470                                 base_ref = use_pack(p, &w_curs,
1471                                                 entry->in_pack_offset + used, NULL);
1472                         entry->in_pack_header_size = used + the_hash_algo->rawsz;
1473                         break;
1474                 case OBJ_OFS_DELTA:
1475                         buf = use_pack(p, &w_curs,
1476                                        entry->in_pack_offset + used, NULL);
1477                         used_0 = 0;
1478                         c = buf[used_0++];
1479                         ofs = c & 127;
1480                         while (c & 128) {
1481                                 ofs += 1;
1482                                 if (!ofs || MSB(ofs, 7)) {
1483                                         error("delta base offset overflow in pack for %s",
1484                                               oid_to_hex(&entry->idx.oid));
1485                                         goto give_up;
1486                                 }
1487                                 c = buf[used_0++];
1488                                 ofs = (ofs << 7) + (c & 127);
1489                         }
1490                         ofs = entry->in_pack_offset - ofs;
1491                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1492                                 error("delta base offset out of bound for %s",
1493                                       oid_to_hex(&entry->idx.oid));
1494                                 goto give_up;
1495                         }
1496                         if (reuse_delta && !entry->preferred_base) {
1497                                 struct revindex_entry *revidx;
1498                                 revidx = find_pack_revindex(p, ofs);
1499                                 if (!revidx)
1500                                         goto give_up;
1501                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1502                         }
1503                         entry->in_pack_header_size = used + used_0;
1504                         break;
1505                 }
1506
1507                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1508                         /*
1509                          * If base_ref was set above that means we wish to
1510                          * reuse delta data, and we even found that base
1511                          * in the list of objects we want to pack. Goodie!
1512                          *
1513                          * Depth value does not matter - find_deltas() will
1514                          * never consider reused delta as the base object to
1515                          * deltify other objects against, in order to avoid
1516                          * circular deltas.
1517                          */
1518                         oe_set_type(entry, entry->in_pack_type);
1519                         SET_SIZE(entry, in_pack_size); /* delta size */
1520                         SET_DELTA(entry, base_entry);
1521                         SET_DELTA_SIZE(entry, in_pack_size);
1522                         entry->delta_sibling_idx = base_entry->delta_child_idx;
1523                         SET_DELTA_CHILD(base_entry, entry);
1524                         unuse_pack(&w_curs);
1525                         return;
1526                 }
1527
1528                 if (oe_type(entry)) {
1529                         off_t delta_pos;
1530
1531                         /*
1532                          * This must be a delta and we already know what the
1533                          * final object type is.  Let's extract the actual
1534                          * object size from the delta header.
1535                          */
1536                         delta_pos = entry->in_pack_offset + entry->in_pack_header_size;
1537                         canonical_size = get_size_from_delta(p, &w_curs, delta_pos);
1538                         if (canonical_size == 0)
1539                                 goto give_up;
1540                         SET_SIZE(entry, canonical_size);
1541                         unuse_pack(&w_curs);
1542                         return;
1543                 }
1544
1545                 /*
1546                  * No choice but to fall back to the recursive delta walk
1547                  * with sha1_object_info() to find about the object type
1548                  * at this point...
1549                  */
1550                 give_up:
1551                 unuse_pack(&w_curs);
1552         }
1553
1554         oe_set_type(entry,
1555                     oid_object_info(the_repository, &entry->idx.oid, &canonical_size));
1556         if (entry->type_valid) {
1557                 SET_SIZE(entry, canonical_size);
1558         } else {
1559                 /*
1560                  * Bad object type is checked in prepare_pack().  This is
1561                  * to permit a missing preferred base object to be ignored
1562                  * as a preferred base.  Doing so can result in a larger
1563                  * pack file, but the transfer will still take place.
1564                  */
1565         }
1566 }
1567
1568 static int pack_offset_sort(const void *_a, const void *_b)
1569 {
1570         const struct object_entry *a = *(struct object_entry **)_a;
1571         const struct object_entry *b = *(struct object_entry **)_b;
1572         const struct packed_git *a_in_pack = IN_PACK(a);
1573         const struct packed_git *b_in_pack = IN_PACK(b);
1574
1575         /* avoid filesystem trashing with loose objects */
1576         if (!a_in_pack && !b_in_pack)
1577                 return oidcmp(&a->idx.oid, &b->idx.oid);
1578
1579         if (a_in_pack < b_in_pack)
1580                 return -1;
1581         if (a_in_pack > b_in_pack)
1582                 return 1;
1583         return a->in_pack_offset < b->in_pack_offset ? -1 :
1584                         (a->in_pack_offset > b->in_pack_offset);
1585 }
1586
1587 /*
1588  * Drop an on-disk delta we were planning to reuse. Naively, this would
1589  * just involve blanking out the "delta" field, but we have to deal
1590  * with some extra book-keeping:
1591  *
1592  *   1. Removing ourselves from the delta_sibling linked list.
1593  *
1594  *   2. Updating our size/type to the non-delta representation. These were
1595  *      either not recorded initially (size) or overwritten with the delta type
1596  *      (type) when check_object() decided to reuse the delta.
1597  *
1598  *   3. Resetting our delta depth, as we are now a base object.
1599  */
1600 static void drop_reused_delta(struct object_entry *entry)
1601 {
1602         unsigned *idx = &to_pack.objects[entry->delta_idx - 1].delta_child_idx;
1603         struct object_info oi = OBJECT_INFO_INIT;
1604         enum object_type type;
1605         unsigned long size;
1606
1607         while (*idx) {
1608                 struct object_entry *oe = &to_pack.objects[*idx - 1];
1609
1610                 if (oe == entry)
1611                         *idx = oe->delta_sibling_idx;
1612                 else
1613                         idx = &oe->delta_sibling_idx;
1614         }
1615         SET_DELTA(entry, NULL);
1616         entry->depth = 0;
1617
1618         oi.sizep = &size;
1619         oi.typep = &type;
1620         if (packed_object_info(the_repository, IN_PACK(entry), entry->in_pack_offset, &oi) < 0) {
1621                 /*
1622                  * We failed to get the info from this pack for some reason;
1623                  * fall back to sha1_object_info, which may find another copy.
1624                  * And if that fails, the error will be recorded in oe_type(entry)
1625                  * and dealt with in prepare_pack().
1626                  */
1627                 oe_set_type(entry,
1628                             oid_object_info(the_repository, &entry->idx.oid, &size));
1629         } else {
1630                 oe_set_type(entry, type);
1631         }
1632         SET_SIZE(entry, size);
1633 }
1634
1635 /*
1636  * Follow the chain of deltas from this entry onward, throwing away any links
1637  * that cause us to hit a cycle (as determined by the DFS state flags in
1638  * the entries).
1639  *
1640  * We also detect too-long reused chains that would violate our --depth
1641  * limit.
1642  */
1643 static void break_delta_chains(struct object_entry *entry)
1644 {
1645         /*
1646          * The actual depth of each object we will write is stored as an int,
1647          * as it cannot exceed our int "depth" limit. But before we break
1648          * changes based no that limit, we may potentially go as deep as the
1649          * number of objects, which is elsewhere bounded to a uint32_t.
1650          */
1651         uint32_t total_depth;
1652         struct object_entry *cur, *next;
1653
1654         for (cur = entry, total_depth = 0;
1655              cur;
1656              cur = DELTA(cur), total_depth++) {
1657                 if (cur->dfs_state == DFS_DONE) {
1658                         /*
1659                          * We've already seen this object and know it isn't
1660                          * part of a cycle. We do need to append its depth
1661                          * to our count.
1662                          */
1663                         total_depth += cur->depth;
1664                         break;
1665                 }
1666
1667                 /*
1668                  * We break cycles before looping, so an ACTIVE state (or any
1669                  * other cruft which made its way into the state variable)
1670                  * is a bug.
1671                  */
1672                 if (cur->dfs_state != DFS_NONE)
1673                         BUG("confusing delta dfs state in first pass: %d",
1674                             cur->dfs_state);
1675
1676                 /*
1677                  * Now we know this is the first time we've seen the object. If
1678                  * it's not a delta, we're done traversing, but we'll mark it
1679                  * done to save time on future traversals.
1680                  */
1681                 if (!DELTA(cur)) {
1682                         cur->dfs_state = DFS_DONE;
1683                         break;
1684                 }
1685
1686                 /*
1687                  * Mark ourselves as active and see if the next step causes
1688                  * us to cycle to another active object. It's important to do
1689                  * this _before_ we loop, because it impacts where we make the
1690                  * cut, and thus how our total_depth counter works.
1691                  * E.g., We may see a partial loop like:
1692                  *
1693                  *   A -> B -> C -> D -> B
1694                  *
1695                  * Cutting B->C breaks the cycle. But now the depth of A is
1696                  * only 1, and our total_depth counter is at 3. The size of the
1697                  * error is always one less than the size of the cycle we
1698                  * broke. Commits C and D were "lost" from A's chain.
1699                  *
1700                  * If we instead cut D->B, then the depth of A is correct at 3.
1701                  * We keep all commits in the chain that we examined.
1702                  */
1703                 cur->dfs_state = DFS_ACTIVE;
1704                 if (DELTA(cur)->dfs_state == DFS_ACTIVE) {
1705                         drop_reused_delta(cur);
1706                         cur->dfs_state = DFS_DONE;
1707                         break;
1708                 }
1709         }
1710
1711         /*
1712          * And now that we've gone all the way to the bottom of the chain, we
1713          * need to clear the active flags and set the depth fields as
1714          * appropriate. Unlike the loop above, which can quit when it drops a
1715          * delta, we need to keep going to look for more depth cuts. So we need
1716          * an extra "next" pointer to keep going after we reset cur->delta.
1717          */
1718         for (cur = entry; cur; cur = next) {
1719                 next = DELTA(cur);
1720
1721                 /*
1722                  * We should have a chain of zero or more ACTIVE states down to
1723                  * a final DONE. We can quit after the DONE, because either it
1724                  * has no bases, or we've already handled them in a previous
1725                  * call.
1726                  */
1727                 if (cur->dfs_state == DFS_DONE)
1728                         break;
1729                 else if (cur->dfs_state != DFS_ACTIVE)
1730                         BUG("confusing delta dfs state in second pass: %d",
1731                             cur->dfs_state);
1732
1733                 /*
1734                  * If the total_depth is more than depth, then we need to snip
1735                  * the chain into two or more smaller chains that don't exceed
1736                  * the maximum depth. Most of the resulting chains will contain
1737                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1738                  * the last chain (i.e., the one containing entry) will contain
1739                  * whatever entries are left over, namely
1740                  * (total_depth % (depth + 1)) of them.
1741                  *
1742                  * Since we are iterating towards decreasing depth, we need to
1743                  * decrement total_depth as we go, and we need to write to the
1744                  * entry what its final depth will be after all of the
1745                  * snipping. Since we're snipping into chains of length (depth
1746                  * + 1) entries, the final depth of an entry will be its
1747                  * original depth modulo (depth + 1). Any time we encounter an
1748                  * entry whose final depth is supposed to be zero, we snip it
1749                  * from its delta base, thereby making it so.
1750                  */
1751                 cur->depth = (total_depth--) % (depth + 1);
1752                 if (!cur->depth)
1753                         drop_reused_delta(cur);
1754
1755                 cur->dfs_state = DFS_DONE;
1756         }
1757 }
1758
1759 static void get_object_details(void)
1760 {
1761         uint32_t i;
1762         struct object_entry **sorted_by_offset;
1763
1764         if (progress)
1765                 progress_state = start_progress(_("Counting objects"),
1766                                                 to_pack.nr_objects);
1767
1768         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1769         for (i = 0; i < to_pack.nr_objects; i++)
1770                 sorted_by_offset[i] = to_pack.objects + i;
1771         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1772
1773         for (i = 0; i < to_pack.nr_objects; i++) {
1774                 struct object_entry *entry = sorted_by_offset[i];
1775                 check_object(entry);
1776                 if (entry->type_valid &&
1777                     oe_size_greater_than(&to_pack, entry, big_file_threshold))
1778                         entry->no_try_delta = 1;
1779                 display_progress(progress_state, i + 1);
1780         }
1781         stop_progress(&progress_state);
1782
1783         /*
1784          * This must happen in a second pass, since we rely on the delta
1785          * information for the whole list being completed.
1786          */
1787         for (i = 0; i < to_pack.nr_objects; i++)
1788                 break_delta_chains(&to_pack.objects[i]);
1789
1790         free(sorted_by_offset);
1791 }
1792
1793 /*
1794  * We search for deltas in a list sorted by type, by filename hash, and then
1795  * by size, so that we see progressively smaller and smaller files.
1796  * That's because we prefer deltas to be from the bigger file
1797  * to the smaller -- deletes are potentially cheaper, but perhaps
1798  * more importantly, the bigger file is likely the more recent
1799  * one.  The deepest deltas are therefore the oldest objects which are
1800  * less susceptible to be accessed often.
1801  */
1802 static int type_size_sort(const void *_a, const void *_b)
1803 {
1804         const struct object_entry *a = *(struct object_entry **)_a;
1805         const struct object_entry *b = *(struct object_entry **)_b;
1806         enum object_type a_type = oe_type(a);
1807         enum object_type b_type = oe_type(b);
1808         unsigned long a_size = SIZE(a);
1809         unsigned long b_size = SIZE(b);
1810
1811         if (a_type > b_type)
1812                 return -1;
1813         if (a_type < b_type)
1814                 return 1;
1815         if (a->hash > b->hash)
1816                 return -1;
1817         if (a->hash < b->hash)
1818                 return 1;
1819         if (a->preferred_base > b->preferred_base)
1820                 return -1;
1821         if (a->preferred_base < b->preferred_base)
1822                 return 1;
1823         if (a_size > b_size)
1824                 return -1;
1825         if (a_size < b_size)
1826                 return 1;
1827         return a < b ? -1 : (a > b);  /* newest first */
1828 }
1829
1830 struct unpacked {
1831         struct object_entry *entry;
1832         void *data;
1833         struct delta_index *index;
1834         unsigned depth;
1835 };
1836
1837 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1838                            unsigned long delta_size)
1839 {
1840         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1841                 return 0;
1842
1843         if (delta_size < cache_max_small_delta_size)
1844                 return 1;
1845
1846         /* cache delta, if objects are large enough compared to delta size */
1847         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1848                 return 1;
1849
1850         return 0;
1851 }
1852
1853 #ifndef NO_PTHREADS
1854
1855 static pthread_mutex_t read_mutex;
1856 #define read_lock()             pthread_mutex_lock(&read_mutex)
1857 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1858
1859 static pthread_mutex_t cache_mutex;
1860 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1861 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1862
1863 static pthread_mutex_t progress_mutex;
1864 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1865 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1866
1867 #else
1868
1869 #define read_lock()             (void)0
1870 #define read_unlock()           (void)0
1871 #define cache_lock()            (void)0
1872 #define cache_unlock()          (void)0
1873 #define progress_lock()         (void)0
1874 #define progress_unlock()       (void)0
1875
1876 #endif
1877
1878 /*
1879  * Return the size of the object without doing any delta
1880  * reconstruction (so non-deltas are true object sizes, but deltas
1881  * return the size of the delta data).
1882  */
1883 unsigned long oe_get_size_slow(struct packing_data *pack,
1884                                const struct object_entry *e)
1885 {
1886         struct packed_git *p;
1887         struct pack_window *w_curs;
1888         unsigned char *buf;
1889         enum object_type type;
1890         unsigned long used, avail, size;
1891
1892         if (e->type_ != OBJ_OFS_DELTA && e->type_ != OBJ_REF_DELTA) {
1893                 read_lock();
1894                 if (oid_object_info(the_repository, &e->idx.oid, &size) < 0)
1895                         die(_("unable to get size of %s"),
1896                             oid_to_hex(&e->idx.oid));
1897                 read_unlock();
1898                 return size;
1899         }
1900
1901         p = oe_in_pack(pack, e);
1902         if (!p)
1903                 BUG("when e->type is a delta, it must belong to a pack");
1904
1905         read_lock();
1906         w_curs = NULL;
1907         buf = use_pack(p, &w_curs, e->in_pack_offset, &avail);
1908         used = unpack_object_header_buffer(buf, avail, &type, &size);
1909         if (used == 0)
1910                 die(_("unable to parse object header of %s"),
1911                     oid_to_hex(&e->idx.oid));
1912
1913         unuse_pack(&w_curs);
1914         read_unlock();
1915         return size;
1916 }
1917
1918 static int try_delta(struct unpacked *trg, struct unpacked *src,
1919                      unsigned max_depth, unsigned long *mem_usage)
1920 {
1921         struct object_entry *trg_entry = trg->entry;
1922         struct object_entry *src_entry = src->entry;
1923         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1924         unsigned ref_depth;
1925         enum object_type type;
1926         void *delta_buf;
1927
1928         /* Don't bother doing diffs between different types */
1929         if (oe_type(trg_entry) != oe_type(src_entry))
1930                 return -1;
1931
1932         /*
1933          * We do not bother to try a delta that we discarded on an
1934          * earlier try, but only when reusing delta data.  Note that
1935          * src_entry that is marked as the preferred_base should always
1936          * be considered, as even if we produce a suboptimal delta against
1937          * it, we will still save the transfer cost, as we already know
1938          * the other side has it and we won't send src_entry at all.
1939          */
1940         if (reuse_delta && IN_PACK(trg_entry) &&
1941             IN_PACK(trg_entry) == IN_PACK(src_entry) &&
1942             !src_entry->preferred_base &&
1943             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1944             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1945                 return 0;
1946
1947         /* Let's not bust the allowed depth. */
1948         if (src->depth >= max_depth)
1949                 return 0;
1950
1951         /* Now some size filtering heuristics. */
1952         trg_size = SIZE(trg_entry);
1953         if (!DELTA(trg_entry)) {
1954                 max_size = trg_size/2 - the_hash_algo->rawsz;
1955                 ref_depth = 1;
1956         } else {
1957                 max_size = DELTA_SIZE(trg_entry);
1958                 ref_depth = trg->depth;
1959         }
1960         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1961                                                 (max_depth - ref_depth + 1);
1962         if (max_size == 0)
1963                 return 0;
1964         src_size = SIZE(src_entry);
1965         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1966         if (sizediff >= max_size)
1967                 return 0;
1968         if (trg_size < src_size / 32)
1969                 return 0;
1970
1971         /* Load data if not already done */
1972         if (!trg->data) {
1973                 read_lock();
1974                 trg->data = read_object_file(&trg_entry->idx.oid, &type, &sz);
1975                 read_unlock();
1976                 if (!trg->data)
1977                         die("object %s cannot be read",
1978                             oid_to_hex(&trg_entry->idx.oid));
1979                 if (sz != trg_size)
1980                         die("object %s inconsistent object length (%lu vs %lu)",
1981                             oid_to_hex(&trg_entry->idx.oid), sz,
1982                             trg_size);
1983                 *mem_usage += sz;
1984         }
1985         if (!src->data) {
1986                 read_lock();
1987                 src->data = read_object_file(&src_entry->idx.oid, &type, &sz);
1988                 read_unlock();
1989                 if (!src->data) {
1990                         if (src_entry->preferred_base) {
1991                                 static int warned = 0;
1992                                 if (!warned++)
1993                                         warning("object %s cannot be read",
1994                                                 oid_to_hex(&src_entry->idx.oid));
1995                                 /*
1996                                  * Those objects are not included in the
1997                                  * resulting pack.  Be resilient and ignore
1998                                  * them if they can't be read, in case the
1999                                  * pack could be created nevertheless.
2000                                  */
2001                                 return 0;
2002                         }
2003                         die("object %s cannot be read",
2004                             oid_to_hex(&src_entry->idx.oid));
2005                 }
2006                 if (sz != src_size)
2007                         die("object %s inconsistent object length (%lu vs %lu)",
2008                             oid_to_hex(&src_entry->idx.oid), sz,
2009                             src_size);
2010                 *mem_usage += sz;
2011         }
2012         if (!src->index) {
2013                 src->index = create_delta_index(src->data, src_size);
2014                 if (!src->index) {
2015                         static int warned = 0;
2016                         if (!warned++)
2017                                 warning("suboptimal pack - out of memory");
2018                         return 0;
2019                 }
2020                 *mem_usage += sizeof_delta_index(src->index);
2021         }
2022
2023         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
2024         if (!delta_buf)
2025                 return 0;
2026         if (delta_size >= (1U << OE_DELTA_SIZE_BITS)) {
2027                 free(delta_buf);
2028                 return 0;
2029         }
2030
2031         if (DELTA(trg_entry)) {
2032                 /* Prefer only shallower same-sized deltas. */
2033                 if (delta_size == DELTA_SIZE(trg_entry) &&
2034                     src->depth + 1 >= trg->depth) {
2035                         free(delta_buf);
2036                         return 0;
2037                 }
2038         }
2039
2040         /*
2041          * Handle memory allocation outside of the cache
2042          * accounting lock.  Compiler will optimize the strangeness
2043          * away when NO_PTHREADS is defined.
2044          */
2045         free(trg_entry->delta_data);
2046         cache_lock();
2047         if (trg_entry->delta_data) {
2048                 delta_cache_size -= DELTA_SIZE(trg_entry);
2049                 trg_entry->delta_data = NULL;
2050         }
2051         if (delta_cacheable(src_size, trg_size, delta_size)) {
2052                 delta_cache_size += delta_size;
2053                 cache_unlock();
2054                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
2055         } else {
2056                 cache_unlock();
2057                 free(delta_buf);
2058         }
2059
2060         SET_DELTA(trg_entry, src_entry);
2061         SET_DELTA_SIZE(trg_entry, delta_size);
2062         trg->depth = src->depth + 1;
2063
2064         return 1;
2065 }
2066
2067 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
2068 {
2069         struct object_entry *child = DELTA_CHILD(me);
2070         unsigned int m = n;
2071         while (child) {
2072                 unsigned int c = check_delta_limit(child, n + 1);
2073                 if (m < c)
2074                         m = c;
2075                 child = DELTA_SIBLING(child);
2076         }
2077         return m;
2078 }
2079
2080 static unsigned long free_unpacked(struct unpacked *n)
2081 {
2082         unsigned long freed_mem = sizeof_delta_index(n->index);
2083         free_delta_index(n->index);
2084         n->index = NULL;
2085         if (n->data) {
2086                 freed_mem += SIZE(n->entry);
2087                 FREE_AND_NULL(n->data);
2088         }
2089         n->entry = NULL;
2090         n->depth = 0;
2091         return freed_mem;
2092 }
2093
2094 static void find_deltas(struct object_entry **list, unsigned *list_size,
2095                         int window, int depth, unsigned *processed)
2096 {
2097         uint32_t i, idx = 0, count = 0;
2098         struct unpacked *array;
2099         unsigned long mem_usage = 0;
2100
2101         array = xcalloc(window, sizeof(struct unpacked));
2102
2103         for (;;) {
2104                 struct object_entry *entry;
2105                 struct unpacked *n = array + idx;
2106                 int j, max_depth, best_base = -1;
2107
2108                 progress_lock();
2109                 if (!*list_size) {
2110                         progress_unlock();
2111                         break;
2112                 }
2113                 entry = *list++;
2114                 (*list_size)--;
2115                 if (!entry->preferred_base) {
2116                         (*processed)++;
2117                         display_progress(progress_state, *processed);
2118                 }
2119                 progress_unlock();
2120
2121                 mem_usage -= free_unpacked(n);
2122                 n->entry = entry;
2123
2124                 while (window_memory_limit &&
2125                        mem_usage > window_memory_limit &&
2126                        count > 1) {
2127                         uint32_t tail = (idx + window - count) % window;
2128                         mem_usage -= free_unpacked(array + tail);
2129                         count--;
2130                 }
2131
2132                 /* We do not compute delta to *create* objects we are not
2133                  * going to pack.
2134                  */
2135                 if (entry->preferred_base)
2136                         goto next;
2137
2138                 /*
2139                  * If the current object is at pack edge, take the depth the
2140                  * objects that depend on the current object into account
2141                  * otherwise they would become too deep.
2142                  */
2143                 max_depth = depth;
2144                 if (DELTA_CHILD(entry)) {
2145                         max_depth -= check_delta_limit(entry, 0);
2146                         if (max_depth <= 0)
2147                                 goto next;
2148                 }
2149
2150                 j = window;
2151                 while (--j > 0) {
2152                         int ret;
2153                         uint32_t other_idx = idx + j;
2154                         struct unpacked *m;
2155                         if (other_idx >= window)
2156                                 other_idx -= window;
2157                         m = array + other_idx;
2158                         if (!m->entry)
2159                                 break;
2160                         ret = try_delta(n, m, max_depth, &mem_usage);
2161                         if (ret < 0)
2162                                 break;
2163                         else if (ret > 0)
2164                                 best_base = other_idx;
2165                 }
2166
2167                 /*
2168                  * If we decided to cache the delta data, then it is best
2169                  * to compress it right away.  First because we have to do
2170                  * it anyway, and doing it here while we're threaded will
2171                  * save a lot of time in the non threaded write phase,
2172                  * as well as allow for caching more deltas within
2173                  * the same cache size limit.
2174                  * ...
2175                  * But only if not writing to stdout, since in that case
2176                  * the network is most likely throttling writes anyway,
2177                  * and therefore it is best to go to the write phase ASAP
2178                  * instead, as we can afford spending more time compressing
2179                  * between writes at that moment.
2180                  */
2181                 if (entry->delta_data && !pack_to_stdout) {
2182                         unsigned long size;
2183
2184                         size = do_compress(&entry->delta_data, DELTA_SIZE(entry));
2185                         if (size < (1U << OE_Z_DELTA_BITS)) {
2186                                 entry->z_delta_size = size;
2187                                 cache_lock();
2188                                 delta_cache_size -= DELTA_SIZE(entry);
2189                                 delta_cache_size += entry->z_delta_size;
2190                                 cache_unlock();
2191                         } else {
2192                                 FREE_AND_NULL(entry->delta_data);
2193                                 entry->z_delta_size = 0;
2194                         }
2195                 }
2196
2197                 /* if we made n a delta, and if n is already at max
2198                  * depth, leaving it in the window is pointless.  we
2199                  * should evict it first.
2200                  */
2201                 if (DELTA(entry) && max_depth <= n->depth)
2202                         continue;
2203
2204                 /*
2205                  * Move the best delta base up in the window, after the
2206                  * currently deltified object, to keep it longer.  It will
2207                  * be the first base object to be attempted next.
2208                  */
2209                 if (DELTA(entry)) {
2210                         struct unpacked swap = array[best_base];
2211                         int dist = (window + idx - best_base) % window;
2212                         int dst = best_base;
2213                         while (dist--) {
2214                                 int src = (dst + 1) % window;
2215                                 array[dst] = array[src];
2216                                 dst = src;
2217                         }
2218                         array[dst] = swap;
2219                 }
2220
2221                 next:
2222                 idx++;
2223                 if (count + 1 < window)
2224                         count++;
2225                 if (idx >= window)
2226                         idx = 0;
2227         }
2228
2229         for (i = 0; i < window; ++i) {
2230                 free_delta_index(array[i].index);
2231                 free(array[i].data);
2232         }
2233         free(array);
2234 }
2235
2236 #ifndef NO_PTHREADS
2237
2238 static void try_to_free_from_threads(size_t size)
2239 {
2240         read_lock();
2241         release_pack_memory(size);
2242         read_unlock();
2243 }
2244
2245 static try_to_free_t old_try_to_free_routine;
2246
2247 /*
2248  * The main thread waits on the condition that (at least) one of the workers
2249  * has stopped working (which is indicated in the .working member of
2250  * struct thread_params).
2251  * When a work thread has completed its work, it sets .working to 0 and
2252  * signals the main thread and waits on the condition that .data_ready
2253  * becomes 1.
2254  */
2255
2256 struct thread_params {
2257         pthread_t thread;
2258         struct object_entry **list;
2259         unsigned list_size;
2260         unsigned remaining;
2261         int window;
2262         int depth;
2263         int working;
2264         int data_ready;
2265         pthread_mutex_t mutex;
2266         pthread_cond_t cond;
2267         unsigned *processed;
2268 };
2269
2270 static pthread_cond_t progress_cond;
2271
2272 /*
2273  * Mutex and conditional variable can't be statically-initialized on Windows.
2274  */
2275 static void init_threaded_search(void)
2276 {
2277         init_recursive_mutex(&read_mutex);
2278         pthread_mutex_init(&cache_mutex, NULL);
2279         pthread_mutex_init(&progress_mutex, NULL);
2280         pthread_cond_init(&progress_cond, NULL);
2281         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2282 }
2283
2284 static void cleanup_threaded_search(void)
2285 {
2286         set_try_to_free_routine(old_try_to_free_routine);
2287         pthread_cond_destroy(&progress_cond);
2288         pthread_mutex_destroy(&read_mutex);
2289         pthread_mutex_destroy(&cache_mutex);
2290         pthread_mutex_destroy(&progress_mutex);
2291 }
2292
2293 static void *threaded_find_deltas(void *arg)
2294 {
2295         struct thread_params *me = arg;
2296
2297         progress_lock();
2298         while (me->remaining) {
2299                 progress_unlock();
2300
2301                 find_deltas(me->list, &me->remaining,
2302                             me->window, me->depth, me->processed);
2303
2304                 progress_lock();
2305                 me->working = 0;
2306                 pthread_cond_signal(&progress_cond);
2307                 progress_unlock();
2308
2309                 /*
2310                  * We must not set ->data_ready before we wait on the
2311                  * condition because the main thread may have set it to 1
2312                  * before we get here. In order to be sure that new
2313                  * work is available if we see 1 in ->data_ready, it
2314                  * was initialized to 0 before this thread was spawned
2315                  * and we reset it to 0 right away.
2316                  */
2317                 pthread_mutex_lock(&me->mutex);
2318                 while (!me->data_ready)
2319                         pthread_cond_wait(&me->cond, &me->mutex);
2320                 me->data_ready = 0;
2321                 pthread_mutex_unlock(&me->mutex);
2322
2323                 progress_lock();
2324         }
2325         progress_unlock();
2326         /* leave ->working 1 so that this doesn't get more work assigned */
2327         return NULL;
2328 }
2329
2330 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2331                            int window, int depth, unsigned *processed)
2332 {
2333         struct thread_params *p;
2334         int i, ret, active_threads = 0;
2335
2336         init_threaded_search();
2337
2338         if (delta_search_threads <= 1) {
2339                 find_deltas(list, &list_size, window, depth, processed);
2340                 cleanup_threaded_search();
2341                 return;
2342         }
2343         if (progress > pack_to_stdout)
2344                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2345                                 delta_search_threads);
2346         p = xcalloc(delta_search_threads, sizeof(*p));
2347
2348         /* Partition the work amongst work threads. */
2349         for (i = 0; i < delta_search_threads; i++) {
2350                 unsigned sub_size = list_size / (delta_search_threads - i);
2351
2352                 /* don't use too small segments or no deltas will be found */
2353                 if (sub_size < 2*window && i+1 < delta_search_threads)
2354                         sub_size = 0;
2355
2356                 p[i].window = window;
2357                 p[i].depth = depth;
2358                 p[i].processed = processed;
2359                 p[i].working = 1;
2360                 p[i].data_ready = 0;
2361
2362                 /* try to split chunks on "path" boundaries */
2363                 while (sub_size && sub_size < list_size &&
2364                        list[sub_size]->hash &&
2365                        list[sub_size]->hash == list[sub_size-1]->hash)
2366                         sub_size++;
2367
2368                 p[i].list = list;
2369                 p[i].list_size = sub_size;
2370                 p[i].remaining = sub_size;
2371
2372                 list += sub_size;
2373                 list_size -= sub_size;
2374         }
2375
2376         /* Start work threads. */
2377         for (i = 0; i < delta_search_threads; i++) {
2378                 if (!p[i].list_size)
2379                         continue;
2380                 pthread_mutex_init(&p[i].mutex, NULL);
2381                 pthread_cond_init(&p[i].cond, NULL);
2382                 ret = pthread_create(&p[i].thread, NULL,
2383                                      threaded_find_deltas, &p[i]);
2384                 if (ret)
2385                         die("unable to create thread: %s", strerror(ret));
2386                 active_threads++;
2387         }
2388
2389         /*
2390          * Now let's wait for work completion.  Each time a thread is done
2391          * with its work, we steal half of the remaining work from the
2392          * thread with the largest number of unprocessed objects and give
2393          * it to that newly idle thread.  This ensure good load balancing
2394          * until the remaining object list segments are simply too short
2395          * to be worth splitting anymore.
2396          */
2397         while (active_threads) {
2398                 struct thread_params *target = NULL;
2399                 struct thread_params *victim = NULL;
2400                 unsigned sub_size = 0;
2401
2402                 progress_lock();
2403                 for (;;) {
2404                         for (i = 0; !target && i < delta_search_threads; i++)
2405                                 if (!p[i].working)
2406                                         target = &p[i];
2407                         if (target)
2408                                 break;
2409                         pthread_cond_wait(&progress_cond, &progress_mutex);
2410                 }
2411
2412                 for (i = 0; i < delta_search_threads; i++)
2413                         if (p[i].remaining > 2*window &&
2414                             (!victim || victim->remaining < p[i].remaining))
2415                                 victim = &p[i];
2416                 if (victim) {
2417                         sub_size = victim->remaining / 2;
2418                         list = victim->list + victim->list_size - sub_size;
2419                         while (sub_size && list[0]->hash &&
2420                                list[0]->hash == list[-1]->hash) {
2421                                 list++;
2422                                 sub_size--;
2423                         }
2424                         if (!sub_size) {
2425                                 /*
2426                                  * It is possible for some "paths" to have
2427                                  * so many objects that no hash boundary
2428                                  * might be found.  Let's just steal the
2429                                  * exact half in that case.
2430                                  */
2431                                 sub_size = victim->remaining / 2;
2432                                 list -= sub_size;
2433                         }
2434                         target->list = list;
2435                         victim->list_size -= sub_size;
2436                         victim->remaining -= sub_size;
2437                 }
2438                 target->list_size = sub_size;
2439                 target->remaining = sub_size;
2440                 target->working = 1;
2441                 progress_unlock();
2442
2443                 pthread_mutex_lock(&target->mutex);
2444                 target->data_ready = 1;
2445                 pthread_cond_signal(&target->cond);
2446                 pthread_mutex_unlock(&target->mutex);
2447
2448                 if (!sub_size) {
2449                         pthread_join(target->thread, NULL);
2450                         pthread_cond_destroy(&target->cond);
2451                         pthread_mutex_destroy(&target->mutex);
2452                         active_threads--;
2453                 }
2454         }
2455         cleanup_threaded_search();
2456         free(p);
2457 }
2458
2459 #else
2460 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2461 #endif
2462
2463 static void add_tag_chain(const struct object_id *oid)
2464 {
2465         struct tag *tag;
2466
2467         /*
2468          * We catch duplicates already in add_object_entry(), but we'd
2469          * prefer to do this extra check to avoid having to parse the
2470          * tag at all if we already know that it's being packed (e.g., if
2471          * it was included via bitmaps, we would not have parsed it
2472          * previously).
2473          */
2474         if (packlist_find(&to_pack, oid->hash, NULL))
2475                 return;
2476
2477         tag = lookup_tag(oid);
2478         while (1) {
2479                 if (!tag || parse_tag(tag) || !tag->tagged)
2480                         die("unable to pack objects reachable from tag %s",
2481                             oid_to_hex(oid));
2482
2483                 add_object_entry(&tag->object.oid, OBJ_TAG, NULL, 0);
2484
2485                 if (tag->tagged->type != OBJ_TAG)
2486                         return;
2487
2488                 tag = (struct tag *)tag->tagged;
2489         }
2490 }
2491
2492 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2493 {
2494         struct object_id peeled;
2495
2496         if (starts_with(path, "refs/tags/") && /* is a tag? */
2497             !peel_ref(path, &peeled)    && /* peelable? */
2498             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2499                 add_tag_chain(oid);
2500         return 0;
2501 }
2502
2503 static void prepare_pack(int window, int depth)
2504 {
2505         struct object_entry **delta_list;
2506         uint32_t i, nr_deltas;
2507         unsigned n;
2508
2509         get_object_details();
2510
2511         /*
2512          * If we're locally repacking then we need to be doubly careful
2513          * from now on in order to make sure no stealth corruption gets
2514          * propagated to the new pack.  Clients receiving streamed packs
2515          * should validate everything they get anyway so no need to incur
2516          * the additional cost here in that case.
2517          */
2518         if (!pack_to_stdout)
2519                 do_check_packed_object_crc = 1;
2520
2521         if (!to_pack.nr_objects || !window || !depth)
2522                 return;
2523
2524         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2525         nr_deltas = n = 0;
2526
2527         for (i = 0; i < to_pack.nr_objects; i++) {
2528                 struct object_entry *entry = to_pack.objects + i;
2529
2530                 if (DELTA(entry))
2531                         /* This happens if we decided to reuse existing
2532                          * delta from a pack.  "reuse_delta &&" is implied.
2533                          */
2534                         continue;
2535
2536                 if (!entry->type_valid ||
2537                     oe_size_less_than(&to_pack, entry, 50))
2538                         continue;
2539
2540                 if (entry->no_try_delta)
2541                         continue;
2542
2543                 if (!entry->preferred_base) {
2544                         nr_deltas++;
2545                         if (oe_type(entry) < 0)
2546                                 die("unable to get type of object %s",
2547                                     oid_to_hex(&entry->idx.oid));
2548                 } else {
2549                         if (oe_type(entry) < 0) {
2550                                 /*
2551                                  * This object is not found, but we
2552                                  * don't have to include it anyway.
2553                                  */
2554                                 continue;
2555                         }
2556                 }
2557
2558                 delta_list[n++] = entry;
2559         }
2560
2561         if (nr_deltas && n > 1) {
2562                 unsigned nr_done = 0;
2563                 if (progress)
2564                         progress_state = start_progress(_("Compressing objects"),
2565                                                         nr_deltas);
2566                 QSORT(delta_list, n, type_size_sort);
2567                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2568                 stop_progress(&progress_state);
2569                 if (nr_done != nr_deltas)
2570                         die("inconsistency with delta count");
2571         }
2572         free(delta_list);
2573 }
2574
2575 static int git_pack_config(const char *k, const char *v, void *cb)
2576 {
2577         if (!strcmp(k, "pack.window")) {
2578                 window = git_config_int(k, v);
2579                 return 0;
2580         }
2581         if (!strcmp(k, "pack.windowmemory")) {
2582                 window_memory_limit = git_config_ulong(k, v);
2583                 return 0;
2584         }
2585         if (!strcmp(k, "pack.depth")) {
2586                 depth = git_config_int(k, v);
2587                 return 0;
2588         }
2589         if (!strcmp(k, "pack.deltacachesize")) {
2590                 max_delta_cache_size = git_config_int(k, v);
2591                 return 0;
2592         }
2593         if (!strcmp(k, "pack.deltacachelimit")) {
2594                 cache_max_small_delta_size = git_config_int(k, v);
2595                 return 0;
2596         }
2597         if (!strcmp(k, "pack.writebitmaphashcache")) {
2598                 if (git_config_bool(k, v))
2599                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2600                 else
2601                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2602         }
2603         if (!strcmp(k, "pack.usebitmaps")) {
2604                 use_bitmap_index_default = git_config_bool(k, v);
2605                 return 0;
2606         }
2607         if (!strcmp(k, "pack.threads")) {
2608                 delta_search_threads = git_config_int(k, v);
2609                 if (delta_search_threads < 0)
2610                         die("invalid number of threads specified (%d)",
2611                             delta_search_threads);
2612 #ifdef NO_PTHREADS
2613                 if (delta_search_threads != 1) {
2614                         warning("no threads support, ignoring %s", k);
2615                         delta_search_threads = 0;
2616                 }
2617 #endif
2618                 return 0;
2619         }
2620         if (!strcmp(k, "pack.indexversion")) {
2621                 pack_idx_opts.version = git_config_int(k, v);
2622                 if (pack_idx_opts.version > 2)
2623                         die("bad pack.indexversion=%"PRIu32,
2624                             pack_idx_opts.version);
2625                 return 0;
2626         }
2627         return git_default_config(k, v, cb);
2628 }
2629
2630 static void read_object_list_from_stdin(void)
2631 {
2632         char line[GIT_MAX_HEXSZ + 1 + PATH_MAX + 2];
2633         struct object_id oid;
2634         const char *p;
2635
2636         for (;;) {
2637                 if (!fgets(line, sizeof(line), stdin)) {
2638                         if (feof(stdin))
2639                                 break;
2640                         if (!ferror(stdin))
2641                                 die("fgets returned NULL, not EOF, not error!");
2642                         if (errno != EINTR)
2643                                 die_errno("fgets");
2644                         clearerr(stdin);
2645                         continue;
2646                 }
2647                 if (line[0] == '-') {
2648                         if (get_oid_hex(line+1, &oid))
2649                                 die("expected edge object ID, got garbage:\n %s",
2650                                     line);
2651                         add_preferred_base(&oid);
2652                         continue;
2653                 }
2654                 if (parse_oid_hex(line, &oid, &p))
2655                         die("expected object ID, got garbage:\n %s", line);
2656
2657                 add_preferred_base_object(p + 1);
2658                 add_object_entry(&oid, OBJ_NONE, p + 1, 0);
2659         }
2660 }
2661
2662 /* Remember to update object flag allocation in object.h */
2663 #define OBJECT_ADDED (1u<<20)
2664
2665 static void show_commit(struct commit *commit, void *data)
2666 {
2667         add_object_entry(&commit->object.oid, OBJ_COMMIT, NULL, 0);
2668         commit->object.flags |= OBJECT_ADDED;
2669
2670         if (write_bitmap_index)
2671                 index_commit_for_bitmap(commit);
2672 }
2673
2674 static void show_object(struct object *obj, const char *name, void *data)
2675 {
2676         add_preferred_base_object(name);
2677         add_object_entry(&obj->oid, obj->type, name, 0);
2678         obj->flags |= OBJECT_ADDED;
2679 }
2680
2681 static void show_object__ma_allow_any(struct object *obj, const char *name, void *data)
2682 {
2683         assert(arg_missing_action == MA_ALLOW_ANY);
2684
2685         /*
2686          * Quietly ignore ALL missing objects.  This avoids problems with
2687          * staging them now and getting an odd error later.
2688          */
2689         if (!has_object_file(&obj->oid))
2690                 return;
2691
2692         show_object(obj, name, data);
2693 }
2694
2695 static void show_object__ma_allow_promisor(struct object *obj, const char *name, void *data)
2696 {
2697         assert(arg_missing_action == MA_ALLOW_PROMISOR);
2698
2699         /*
2700          * Quietly ignore EXPECTED missing objects.  This avoids problems with
2701          * staging them now and getting an odd error later.
2702          */
2703         if (!has_object_file(&obj->oid) && is_promisor_object(&obj->oid))
2704                 return;
2705
2706         show_object(obj, name, data);
2707 }
2708
2709 static int option_parse_missing_action(const struct option *opt,
2710                                        const char *arg, int unset)
2711 {
2712         assert(arg);
2713         assert(!unset);
2714
2715         if (!strcmp(arg, "error")) {
2716                 arg_missing_action = MA_ERROR;
2717                 fn_show_object = show_object;
2718                 return 0;
2719         }
2720
2721         if (!strcmp(arg, "allow-any")) {
2722                 arg_missing_action = MA_ALLOW_ANY;
2723                 fetch_if_missing = 0;
2724                 fn_show_object = show_object__ma_allow_any;
2725                 return 0;
2726         }
2727
2728         if (!strcmp(arg, "allow-promisor")) {
2729                 arg_missing_action = MA_ALLOW_PROMISOR;
2730                 fetch_if_missing = 0;
2731                 fn_show_object = show_object__ma_allow_promisor;
2732                 return 0;
2733         }
2734
2735         die(_("invalid value for --missing"));
2736         return 0;
2737 }
2738
2739 static void show_edge(struct commit *commit)
2740 {
2741         add_preferred_base(&commit->object.oid);
2742 }
2743
2744 struct in_pack_object {
2745         off_t offset;
2746         struct object *object;
2747 };
2748
2749 struct in_pack {
2750         unsigned int alloc;
2751         unsigned int nr;
2752         struct in_pack_object *array;
2753 };
2754
2755 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2756 {
2757         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2758         in_pack->array[in_pack->nr].object = object;
2759         in_pack->nr++;
2760 }
2761
2762 /*
2763  * Compare the objects in the offset order, in order to emulate the
2764  * "git rev-list --objects" output that produced the pack originally.
2765  */
2766 static int ofscmp(const void *a_, const void *b_)
2767 {
2768         struct in_pack_object *a = (struct in_pack_object *)a_;
2769         struct in_pack_object *b = (struct in_pack_object *)b_;
2770
2771         if (a->offset < b->offset)
2772                 return -1;
2773         else if (a->offset > b->offset)
2774                 return 1;
2775         else
2776                 return oidcmp(&a->object->oid, &b->object->oid);
2777 }
2778
2779 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2780 {
2781         struct packed_git *p;
2782         struct in_pack in_pack;
2783         uint32_t i;
2784
2785         memset(&in_pack, 0, sizeof(in_pack));
2786
2787         for (p = get_packed_git(the_repository); p; p = p->next) {
2788                 struct object_id oid;
2789                 struct object *o;
2790
2791                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
2792                         continue;
2793                 if (open_pack_index(p))
2794                         die("cannot open pack index");
2795
2796                 ALLOC_GROW(in_pack.array,
2797                            in_pack.nr + p->num_objects,
2798                            in_pack.alloc);
2799
2800                 for (i = 0; i < p->num_objects; i++) {
2801                         nth_packed_object_oid(&oid, p, i);
2802                         o = lookup_unknown_object(oid.hash);
2803                         if (!(o->flags & OBJECT_ADDED))
2804                                 mark_in_pack_object(o, p, &in_pack);
2805                         o->flags |= OBJECT_ADDED;
2806                 }
2807         }
2808
2809         if (in_pack.nr) {
2810                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2811                 for (i = 0; i < in_pack.nr; i++) {
2812                         struct object *o = in_pack.array[i].object;
2813                         add_object_entry(&o->oid, o->type, "", 0);
2814                 }
2815         }
2816         free(in_pack.array);
2817 }
2818
2819 static int add_loose_object(const struct object_id *oid, const char *path,
2820                             void *data)
2821 {
2822         enum object_type type = oid_object_info(the_repository, oid, NULL);
2823
2824         if (type < 0) {
2825                 warning("loose object at %s could not be examined", path);
2826                 return 0;
2827         }
2828
2829         add_object_entry(oid, type, "", 0);
2830         return 0;
2831 }
2832
2833 /*
2834  * We actually don't even have to worry about reachability here.
2835  * add_object_entry will weed out duplicates, so we just add every
2836  * loose object we find.
2837  */
2838 static void add_unreachable_loose_objects(void)
2839 {
2840         for_each_loose_file_in_objdir(get_object_directory(),
2841                                       add_loose_object,
2842                                       NULL, NULL, NULL);
2843 }
2844
2845 static int has_sha1_pack_kept_or_nonlocal(const struct object_id *oid)
2846 {
2847         static struct packed_git *last_found = (void *)1;
2848         struct packed_git *p;
2849
2850         p = (last_found != (void *)1) ? last_found :
2851                                         get_packed_git(the_repository);
2852
2853         while (p) {
2854                 if ((!p->pack_local || p->pack_keep ||
2855                                 p->pack_keep_in_core) &&
2856                         find_pack_entry_one(oid->hash, p)) {
2857                         last_found = p;
2858                         return 1;
2859                 }
2860                 if (p == last_found)
2861                         p = get_packed_git(the_repository);
2862                 else
2863                         p = p->next;
2864                 if (p == last_found)
2865                         p = p->next;
2866         }
2867         return 0;
2868 }
2869
2870 /*
2871  * Store a list of sha1s that are should not be discarded
2872  * because they are either written too recently, or are
2873  * reachable from another object that was.
2874  *
2875  * This is filled by get_object_list.
2876  */
2877 static struct oid_array recent_objects;
2878
2879 static int loosened_object_can_be_discarded(const struct object_id *oid,
2880                                             timestamp_t mtime)
2881 {
2882         if (!unpack_unreachable_expiration)
2883                 return 0;
2884         if (mtime > unpack_unreachable_expiration)
2885                 return 0;
2886         if (oid_array_lookup(&recent_objects, oid) >= 0)
2887                 return 0;
2888         return 1;
2889 }
2890
2891 static void loosen_unused_packed_objects(struct rev_info *revs)
2892 {
2893         struct packed_git *p;
2894         uint32_t i;
2895         struct object_id oid;
2896
2897         for (p = get_packed_git(the_repository); p; p = p->next) {
2898                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
2899                         continue;
2900
2901                 if (open_pack_index(p))
2902                         die("cannot open pack index");
2903
2904                 for (i = 0; i < p->num_objects; i++) {
2905                         nth_packed_object_oid(&oid, p, i);
2906                         if (!packlist_find(&to_pack, oid.hash, NULL) &&
2907                             !has_sha1_pack_kept_or_nonlocal(&oid) &&
2908                             !loosened_object_can_be_discarded(&oid, p->mtime))
2909                                 if (force_object_loose(&oid, p->mtime))
2910                                         die("unable to force loose object");
2911                 }
2912         }
2913 }
2914
2915 /*
2916  * This tracks any options which pack-reuse code expects to be on, or which a
2917  * reader of the pack might not understand, and which would therefore prevent
2918  * blind reuse of what we have on disk.
2919  */
2920 static int pack_options_allow_reuse(void)
2921 {
2922         return pack_to_stdout &&
2923                allow_ofs_delta &&
2924                !ignore_packed_keep_on_disk &&
2925                !ignore_packed_keep_in_core &&
2926                (!local || !have_non_local_packs) &&
2927                !incremental;
2928 }
2929
2930 static int get_object_list_from_bitmap(struct rev_info *revs)
2931 {
2932         if (prepare_bitmap_walk(revs) < 0)
2933                 return -1;
2934
2935         if (pack_options_allow_reuse() &&
2936             !reuse_partial_packfile_from_bitmap(
2937                         &reuse_packfile,
2938                         &reuse_packfile_objects,
2939                         &reuse_packfile_offset)) {
2940                 assert(reuse_packfile_objects);
2941                 nr_result += reuse_packfile_objects;
2942                 display_progress(progress_state, nr_result);
2943         }
2944
2945         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2946         return 0;
2947 }
2948
2949 static void record_recent_object(struct object *obj,
2950                                  const char *name,
2951                                  void *data)
2952 {
2953         oid_array_append(&recent_objects, &obj->oid);
2954 }
2955
2956 static void record_recent_commit(struct commit *commit, void *data)
2957 {
2958         oid_array_append(&recent_objects, &commit->object.oid);
2959 }
2960
2961 static void get_object_list(int ac, const char **av)
2962 {
2963         struct rev_info revs;
2964         char line[1000];
2965         int flags = 0;
2966
2967         init_revisions(&revs, NULL);
2968         save_commit_buffer = 0;
2969         setup_revisions(ac, av, &revs, NULL);
2970
2971         /* make sure shallows are read */
2972         is_repository_shallow();
2973
2974         while (fgets(line, sizeof(line), stdin) != NULL) {
2975                 int len = strlen(line);
2976                 if (len && line[len - 1] == '\n')
2977                         line[--len] = 0;
2978                 if (!len)
2979                         break;
2980                 if (*line == '-') {
2981                         if (!strcmp(line, "--not")) {
2982                                 flags ^= UNINTERESTING;
2983                                 write_bitmap_index = 0;
2984                                 continue;
2985                         }
2986                         if (starts_with(line, "--shallow ")) {
2987                                 struct object_id oid;
2988                                 if (get_oid_hex(line + 10, &oid))
2989                                         die("not an SHA-1 '%s'", line + 10);
2990                                 register_shallow(&oid);
2991                                 use_bitmap_index = 0;
2992                                 continue;
2993                         }
2994                         die("not a rev '%s'", line);
2995                 }
2996                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2997                         die("bad revision '%s'", line);
2998         }
2999
3000         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
3001                 return;
3002
3003         if (prepare_revision_walk(&revs))
3004                 die("revision walk setup failed");
3005         mark_edges_uninteresting(&revs, show_edge);
3006
3007         if (!fn_show_object)
3008                 fn_show_object = show_object;
3009         traverse_commit_list_filtered(&filter_options, &revs,
3010                                       show_commit, fn_show_object, NULL,
3011                                       NULL);
3012
3013         if (unpack_unreachable_expiration) {
3014                 revs.ignore_missing_links = 1;
3015                 if (add_unseen_recent_objects_to_traversal(&revs,
3016                                 unpack_unreachable_expiration))
3017                         die("unable to add recent objects");
3018                 if (prepare_revision_walk(&revs))
3019                         die("revision walk setup failed");
3020                 traverse_commit_list(&revs, record_recent_commit,
3021                                      record_recent_object, NULL);
3022         }
3023
3024         if (keep_unreachable)
3025                 add_objects_in_unpacked_packs(&revs);
3026         if (pack_loose_unreachable)
3027                 add_unreachable_loose_objects();
3028         if (unpack_unreachable)
3029                 loosen_unused_packed_objects(&revs);
3030
3031         oid_array_clear(&recent_objects);
3032 }
3033
3034 static void add_extra_kept_packs(const struct string_list *names)
3035 {
3036         struct packed_git *p;
3037
3038         if (!names->nr)
3039                 return;
3040
3041         for (p = get_packed_git(the_repository); p; p = p->next) {
3042                 const char *name = basename(p->pack_name);
3043                 int i;
3044
3045                 if (!p->pack_local)
3046                         continue;
3047
3048                 for (i = 0; i < names->nr; i++)
3049                         if (!fspathcmp(name, names->items[i].string))
3050                                 break;
3051
3052                 if (i < names->nr) {
3053                         p->pack_keep_in_core = 1;
3054                         ignore_packed_keep_in_core = 1;
3055                         continue;
3056                 }
3057         }
3058 }
3059
3060 static int option_parse_index_version(const struct option *opt,
3061                                       const char *arg, int unset)
3062 {
3063         char *c;
3064         const char *val = arg;
3065         pack_idx_opts.version = strtoul(val, &c, 10);
3066         if (pack_idx_opts.version > 2)
3067                 die(_("unsupported index version %s"), val);
3068         if (*c == ',' && c[1])
3069                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
3070         if (*c || pack_idx_opts.off32_limit & 0x80000000)
3071                 die(_("bad index version '%s'"), val);
3072         return 0;
3073 }
3074
3075 static int option_parse_unpack_unreachable(const struct option *opt,
3076                                            const char *arg, int unset)
3077 {
3078         if (unset) {
3079                 unpack_unreachable = 0;
3080                 unpack_unreachable_expiration = 0;
3081         }
3082         else {
3083                 unpack_unreachable = 1;
3084                 if (arg)
3085                         unpack_unreachable_expiration = approxidate(arg);
3086         }
3087         return 0;
3088 }
3089
3090 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
3091 {
3092         int use_internal_rev_list = 0;
3093         int thin = 0;
3094         int shallow = 0;
3095         int all_progress_implied = 0;
3096         struct argv_array rp = ARGV_ARRAY_INIT;
3097         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
3098         int rev_list_index = 0;
3099         struct string_list keep_pack_list = STRING_LIST_INIT_NODUP;
3100         struct option pack_objects_options[] = {
3101                 OPT_SET_INT('q', "quiet", &progress,
3102                             N_("do not show progress meter"), 0),
3103                 OPT_SET_INT(0, "progress", &progress,
3104                             N_("show progress meter"), 1),
3105                 OPT_SET_INT(0, "all-progress", &progress,
3106                             N_("show progress meter during object writing phase"), 2),
3107                 OPT_BOOL(0, "all-progress-implied",
3108                          &all_progress_implied,
3109                          N_("similar to --all-progress when progress meter is shown")),
3110                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
3111                   N_("write the pack index file in the specified idx format version"),
3112                   0, option_parse_index_version },
3113                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
3114                               N_("maximum size of each output pack file")),
3115                 OPT_BOOL(0, "local", &local,
3116                          N_("ignore borrowed objects from alternate object store")),
3117                 OPT_BOOL(0, "incremental", &incremental,
3118                          N_("ignore packed objects")),
3119                 OPT_INTEGER(0, "window", &window,
3120                             N_("limit pack window by objects")),
3121                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
3122                               N_("limit pack window by memory in addition to object limit")),
3123                 OPT_INTEGER(0, "depth", &depth,
3124                             N_("maximum length of delta chain allowed in the resulting pack")),
3125                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
3126                          N_("reuse existing deltas")),
3127                 OPT_BOOL(0, "reuse-object", &reuse_object,
3128                          N_("reuse existing objects")),
3129                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
3130                          N_("use OFS_DELTA objects")),
3131                 OPT_INTEGER(0, "threads", &delta_search_threads,
3132                             N_("use threads when searching for best delta matches")),
3133                 OPT_BOOL(0, "non-empty", &non_empty,
3134                          N_("do not create an empty pack output")),
3135                 OPT_BOOL(0, "revs", &use_internal_rev_list,
3136                          N_("read revision arguments from standard input")),
3137                 OPT_SET_INT_F(0, "unpacked", &rev_list_unpacked,
3138                               N_("limit the objects to those that are not yet packed"),
3139                               1, PARSE_OPT_NONEG),
3140                 OPT_SET_INT_F(0, "all", &rev_list_all,
3141                               N_("include objects reachable from any reference"),
3142                               1, PARSE_OPT_NONEG),
3143                 OPT_SET_INT_F(0, "reflog", &rev_list_reflog,
3144                               N_("include objects referred by reflog entries"),
3145                               1, PARSE_OPT_NONEG),
3146                 OPT_SET_INT_F(0, "indexed-objects", &rev_list_index,
3147                               N_("include objects referred to by the index"),
3148                               1, PARSE_OPT_NONEG),
3149                 OPT_BOOL(0, "stdout", &pack_to_stdout,
3150                          N_("output pack to stdout")),
3151                 OPT_BOOL(0, "include-tag", &include_tag,
3152                          N_("include tag objects that refer to objects to be packed")),
3153                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
3154                          N_("keep unreachable objects")),
3155                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
3156                          N_("pack loose unreachable objects")),
3157                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
3158                   N_("unpack unreachable objects newer than <time>"),
3159                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
3160                 OPT_BOOL(0, "thin", &thin,
3161                          N_("create thin packs")),
3162                 OPT_BOOL(0, "shallow", &shallow,
3163                          N_("create packs suitable for shallow fetches")),
3164                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep_on_disk,
3165                          N_("ignore packs that have companion .keep file")),
3166                 OPT_STRING_LIST(0, "keep-pack", &keep_pack_list, N_("name"),
3167                                 N_("ignore this pack")),
3168                 OPT_INTEGER(0, "compression", &pack_compression_level,
3169                             N_("pack compression level")),
3170                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
3171                             N_("do not hide commits by grafts"), 0),
3172                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
3173                          N_("use a bitmap index if available to speed up counting objects")),
3174                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
3175                          N_("write a bitmap index together with the pack index")),
3176                 OPT_PARSE_LIST_OBJECTS_FILTER(&filter_options),
3177                 { OPTION_CALLBACK, 0, "missing", NULL, N_("action"),
3178                   N_("handling for missing objects"), PARSE_OPT_NONEG,
3179                   option_parse_missing_action },
3180                 OPT_BOOL(0, "exclude-promisor-objects", &exclude_promisor_objects,
3181                          N_("do not pack objects in promisor packfiles")),
3182                 OPT_END(),
3183         };
3184
3185         if (DFS_NUM_STATES > (1 << OE_DFS_STATE_BITS))
3186                 BUG("too many dfs states, increase OE_DFS_STATE_BITS");
3187
3188         check_replace_refs = 0;
3189
3190         reset_pack_idx_option(&pack_idx_opts);
3191         git_config(git_pack_config, NULL);
3192
3193         progress = isatty(2);
3194         argc = parse_options(argc, argv, prefix, pack_objects_options,
3195                              pack_usage, 0);
3196
3197         if (argc) {
3198                 base_name = argv[0];
3199                 argc--;
3200         }
3201         if (pack_to_stdout != !base_name || argc)
3202                 usage_with_options(pack_usage, pack_objects_options);
3203
3204         if (depth >= (1 << OE_DEPTH_BITS)) {
3205                 warning(_("delta chain depth %d is too deep, forcing %d"),
3206                         depth, (1 << OE_DEPTH_BITS) - 1);
3207                 depth = (1 << OE_DEPTH_BITS) - 1;
3208         }
3209         if (cache_max_small_delta_size >= (1U << OE_Z_DELTA_BITS)) {
3210                 warning(_("pack.deltaCacheLimit is too high, forcing %d"),
3211                         (1U << OE_Z_DELTA_BITS) - 1);
3212                 cache_max_small_delta_size = (1U << OE_Z_DELTA_BITS) - 1;
3213         }
3214
3215         argv_array_push(&rp, "pack-objects");
3216         if (thin) {
3217                 use_internal_rev_list = 1;
3218                 argv_array_push(&rp, shallow
3219                                 ? "--objects-edge-aggressive"
3220                                 : "--objects-edge");
3221         } else
3222                 argv_array_push(&rp, "--objects");
3223
3224         if (rev_list_all) {
3225                 use_internal_rev_list = 1;
3226                 argv_array_push(&rp, "--all");
3227         }
3228         if (rev_list_reflog) {
3229                 use_internal_rev_list = 1;
3230                 argv_array_push(&rp, "--reflog");
3231         }
3232         if (rev_list_index) {
3233                 use_internal_rev_list = 1;
3234                 argv_array_push(&rp, "--indexed-objects");
3235         }
3236         if (rev_list_unpacked) {
3237                 use_internal_rev_list = 1;
3238                 argv_array_push(&rp, "--unpacked");
3239         }
3240
3241         if (exclude_promisor_objects) {
3242                 use_internal_rev_list = 1;
3243                 fetch_if_missing = 0;
3244                 argv_array_push(&rp, "--exclude-promisor-objects");
3245         }
3246         if (unpack_unreachable || keep_unreachable || pack_loose_unreachable)
3247                 use_internal_rev_list = 1;
3248
3249         if (!reuse_object)
3250                 reuse_delta = 0;
3251         if (pack_compression_level == -1)
3252                 pack_compression_level = Z_DEFAULT_COMPRESSION;
3253         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
3254                 die("bad pack compression level %d", pack_compression_level);
3255
3256         if (!delta_search_threads)      /* --threads=0 means autodetect */
3257                 delta_search_threads = online_cpus();
3258
3259 #ifdef NO_PTHREADS
3260         if (delta_search_threads != 1)
3261                 warning("no threads support, ignoring --threads");
3262 #endif
3263         if (!pack_to_stdout && !pack_size_limit)
3264                 pack_size_limit = pack_size_limit_cfg;
3265         if (pack_to_stdout && pack_size_limit)
3266                 die("--max-pack-size cannot be used to build a pack for transfer.");
3267         if (pack_size_limit && pack_size_limit < 1024*1024) {
3268                 warning("minimum pack size limit is 1 MiB");
3269                 pack_size_limit = 1024*1024;
3270         }
3271
3272         if (!pack_to_stdout && thin)
3273                 die("--thin cannot be used to build an indexable pack.");
3274
3275         if (keep_unreachable && unpack_unreachable)
3276                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
3277         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3278                 unpack_unreachable_expiration = 0;
3279
3280         if (filter_options.choice) {
3281                 if (!pack_to_stdout)
3282                         die("cannot use --filter without --stdout.");
3283                 use_bitmap_index = 0;
3284         }
3285
3286         /*
3287          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3288          *
3289          * - to produce good pack (with bitmap index not-yet-packed objects are
3290          *   packed in suboptimal order).
3291          *
3292          * - to use more robust pack-generation codepath (avoiding possible
3293          *   bugs in bitmap code and possible bitmap index corruption).
3294          */
3295         if (!pack_to_stdout)
3296                 use_bitmap_index_default = 0;
3297
3298         if (use_bitmap_index < 0)
3299                 use_bitmap_index = use_bitmap_index_default;
3300
3301         /* "hard" reasons not to use bitmaps; these just won't work at all */
3302         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
3303                 use_bitmap_index = 0;
3304
3305         if (pack_to_stdout || !rev_list_all)
3306                 write_bitmap_index = 0;
3307
3308         if (progress && all_progress_implied)
3309                 progress = 2;
3310
3311         add_extra_kept_packs(&keep_pack_list);
3312         if (ignore_packed_keep_on_disk) {
3313                 struct packed_git *p;
3314                 for (p = get_packed_git(the_repository); p; p = p->next)
3315                         if (p->pack_local && p->pack_keep)
3316                                 break;
3317                 if (!p) /* no keep-able packs found */
3318                         ignore_packed_keep_on_disk = 0;
3319         }
3320         if (local) {
3321                 /*
3322                  * unlike ignore_packed_keep_on_disk above, we do not
3323                  * want to unset "local" based on looking at packs, as
3324                  * it also covers non-local objects
3325                  */
3326                 struct packed_git *p;
3327                 for (p = get_packed_git(the_repository); p; p = p->next) {
3328                         if (!p->pack_local) {
3329                                 have_non_local_packs = 1;
3330                                 break;
3331                         }
3332                 }
3333         }
3334
3335         prepare_packing_data(&to_pack);
3336
3337         if (progress)
3338                 progress_state = start_progress(_("Enumerating objects"), 0);
3339         if (!use_internal_rev_list)
3340                 read_object_list_from_stdin();
3341         else {
3342                 get_object_list(rp.argc, rp.argv);
3343                 argv_array_clear(&rp);
3344         }
3345         cleanup_preferred_base();
3346         if (include_tag && nr_result)
3347                 for_each_ref(add_ref_tag, NULL);
3348         stop_progress(&progress_state);
3349
3350         if (non_empty && !nr_result)
3351                 return 0;
3352         if (nr_result)
3353                 prepare_pack(window, depth);
3354         write_pack_file();
3355         if (progress)
3356                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
3357                         " reused %"PRIu32" (delta %"PRIu32")\n",
3358                         written, written_delta, reused, reused_delta);
3359         return 0;
3360 }