Merge branch 'dl/doc-stash-remove-mention-of-reflog'
[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 "delta-islands.h"
28 #include "reachable.h"
29 #include "oid-array.h"
30 #include "argv-array.h"
31 #include "list.h"
32 #include "packfile.h"
33 #include "object-store.h"
34 #include "dir.h"
35 #include "midx.h"
36 #include "trace2.h"
37
38 #define IN_PACK(obj) oe_in_pack(&to_pack, obj)
39 #define SIZE(obj) oe_size(&to_pack, obj)
40 #define SET_SIZE(obj,size) oe_set_size(&to_pack, obj, size)
41 #define DELTA_SIZE(obj) oe_delta_size(&to_pack, obj)
42 #define DELTA(obj) oe_delta(&to_pack, obj)
43 #define DELTA_CHILD(obj) oe_delta_child(&to_pack, obj)
44 #define DELTA_SIBLING(obj) oe_delta_sibling(&to_pack, obj)
45 #define SET_DELTA(obj, val) oe_set_delta(&to_pack, obj, val)
46 #define SET_DELTA_EXT(obj, oid) oe_set_delta_ext(&to_pack, obj, oid)
47 #define SET_DELTA_SIZE(obj, val) oe_set_delta_size(&to_pack, obj, val)
48 #define SET_DELTA_CHILD(obj, val) oe_set_delta_child(&to_pack, obj, val)
49 #define SET_DELTA_SIBLING(obj, val) oe_set_delta_sibling(&to_pack, obj, val)
50
51 static const char *pack_usage[] = {
52         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
53         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
54         NULL
55 };
56
57 /*
58  * Objects we are going to pack are collected in the `to_pack` structure.
59  * It contains an array (dynamically expanded) of the object data, and a map
60  * that can resolve SHA1s to their position in the array.
61  */
62 static struct packing_data to_pack;
63
64 static struct pack_idx_entry **written_list;
65 static uint32_t nr_result, nr_written, nr_seen;
66 static struct bitmap_index *bitmap_git;
67 static uint32_t write_layer;
68
69 static int non_empty;
70 static int reuse_delta = 1, reuse_object = 1;
71 static int keep_unreachable, unpack_unreachable, include_tag;
72 static timestamp_t unpack_unreachable_expiration;
73 static int pack_loose_unreachable;
74 static int local;
75 static int have_non_local_packs;
76 static int incremental;
77 static int ignore_packed_keep_on_disk;
78 static int ignore_packed_keep_in_core;
79 static int allow_ofs_delta;
80 static struct pack_idx_option pack_idx_opts;
81 static const char *base_name;
82 static int progress = 1;
83 static int window = 10;
84 static unsigned long pack_size_limit;
85 static int depth = 50;
86 static int delta_search_threads;
87 static int pack_to_stdout;
88 static int sparse;
89 static int thin;
90 static int num_preferred_base;
91 static struct progress *progress_state;
92
93 static struct packed_git *reuse_packfile;
94 static uint32_t reuse_packfile_objects;
95 static struct bitmap *reuse_packfile_bitmap;
96
97 static int use_bitmap_index_default = 1;
98 static int use_bitmap_index = -1;
99 static int allow_pack_reuse = 1;
100 static enum {
101         WRITE_BITMAP_FALSE = 0,
102         WRITE_BITMAP_QUIET,
103         WRITE_BITMAP_TRUE,
104 } write_bitmap_index;
105 static uint16_t write_bitmap_options = BITMAP_OPT_HASH_CACHE;
106
107 static int exclude_promisor_objects;
108
109 static int use_delta_islands;
110
111 static unsigned long delta_cache_size = 0;
112 static unsigned long max_delta_cache_size = DEFAULT_DELTA_CACHE_SIZE;
113 static unsigned long cache_max_small_delta_size = 1000;
114
115 static unsigned long window_memory_limit = 0;
116
117 static struct list_objects_filter_options filter_options;
118
119 enum missing_action {
120         MA_ERROR = 0,      /* fail if any missing objects are encountered */
121         MA_ALLOW_ANY,      /* silently allow ALL missing objects */
122         MA_ALLOW_PROMISOR, /* silently allow all missing PROMISOR objects */
123 };
124 static enum missing_action arg_missing_action;
125 static show_object_fn fn_show_object;
126
127 /*
128  * stats
129  */
130 static uint32_t written, written_delta;
131 static uint32_t reused, reused_delta;
132
133 /*
134  * Indexed commits
135  */
136 static struct commit **indexed_commits;
137 static unsigned int indexed_commits_nr;
138 static unsigned int indexed_commits_alloc;
139
140 static void index_commit_for_bitmap(struct commit *commit)
141 {
142         if (indexed_commits_nr >= indexed_commits_alloc) {
143                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
144                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
145         }
146
147         indexed_commits[indexed_commits_nr++] = commit;
148 }
149
150 static void *get_delta(struct object_entry *entry)
151 {
152         unsigned long size, base_size, delta_size;
153         void *buf, *base_buf, *delta_buf;
154         enum object_type type;
155
156         buf = read_object_file(&entry->idx.oid, &type, &size);
157         if (!buf)
158                 die(_("unable to read %s"), oid_to_hex(&entry->idx.oid));
159         base_buf = read_object_file(&DELTA(entry)->idx.oid, &type,
160                                     &base_size);
161         if (!base_buf)
162                 die("unable to read %s",
163                     oid_to_hex(&DELTA(entry)->idx.oid));
164         delta_buf = diff_delta(base_buf, base_size,
165                                buf, size, &delta_size, 0);
166         /*
167          * We successfully computed this delta once but dropped it for
168          * memory reasons. Something is very wrong if this time we
169          * recompute and create a different delta.
170          */
171         if (!delta_buf || delta_size != DELTA_SIZE(entry))
172                 BUG("delta size changed");
173         free(buf);
174         free(base_buf);
175         return delta_buf;
176 }
177
178 static unsigned long do_compress(void **pptr, unsigned long size)
179 {
180         git_zstream stream;
181         void *in, *out;
182         unsigned long maxsize;
183
184         git_deflate_init(&stream, pack_compression_level);
185         maxsize = git_deflate_bound(&stream, size);
186
187         in = *pptr;
188         out = xmalloc(maxsize);
189         *pptr = out;
190
191         stream.next_in = in;
192         stream.avail_in = size;
193         stream.next_out = out;
194         stream.avail_out = maxsize;
195         while (git_deflate(&stream, Z_FINISH) == Z_OK)
196                 ; /* nothing */
197         git_deflate_end(&stream);
198
199         free(in);
200         return stream.total_out;
201 }
202
203 static unsigned long write_large_blob_data(struct git_istream *st, struct hashfile *f,
204                                            const struct object_id *oid)
205 {
206         git_zstream stream;
207         unsigned char ibuf[1024 * 16];
208         unsigned char obuf[1024 * 16];
209         unsigned long olen = 0;
210
211         git_deflate_init(&stream, pack_compression_level);
212
213         for (;;) {
214                 ssize_t readlen;
215                 int zret = Z_OK;
216                 readlen = read_istream(st, ibuf, sizeof(ibuf));
217                 if (readlen == -1)
218                         die(_("unable to read %s"), oid_to_hex(oid));
219
220                 stream.next_in = ibuf;
221                 stream.avail_in = readlen;
222                 while ((stream.avail_in || readlen == 0) &&
223                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
224                         stream.next_out = obuf;
225                         stream.avail_out = sizeof(obuf);
226                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
227                         hashwrite(f, obuf, stream.next_out - obuf);
228                         olen += stream.next_out - obuf;
229                 }
230                 if (stream.avail_in)
231                         die(_("deflate error (%d)"), zret);
232                 if (readlen == 0) {
233                         if (zret != Z_STREAM_END)
234                                 die(_("deflate error (%d)"), zret);
235                         break;
236                 }
237         }
238         git_deflate_end(&stream);
239         return olen;
240 }
241
242 /*
243  * we are going to reuse the existing object data as is.  make
244  * sure it is not corrupt.
245  */
246 static int check_pack_inflate(struct packed_git *p,
247                 struct pack_window **w_curs,
248                 off_t offset,
249                 off_t len,
250                 unsigned long expect)
251 {
252         git_zstream stream;
253         unsigned char fakebuf[4096], *in;
254         int st;
255
256         memset(&stream, 0, sizeof(stream));
257         git_inflate_init(&stream);
258         do {
259                 in = use_pack(p, w_curs, offset, &stream.avail_in);
260                 stream.next_in = in;
261                 stream.next_out = fakebuf;
262                 stream.avail_out = sizeof(fakebuf);
263                 st = git_inflate(&stream, Z_FINISH);
264                 offset += stream.next_in - in;
265         } while (st == Z_OK || st == Z_BUF_ERROR);
266         git_inflate_end(&stream);
267         return (st == Z_STREAM_END &&
268                 stream.total_out == expect &&
269                 stream.total_in == len) ? 0 : -1;
270 }
271
272 static void copy_pack_data(struct hashfile *f,
273                 struct packed_git *p,
274                 struct pack_window **w_curs,
275                 off_t offset,
276                 off_t len)
277 {
278         unsigned char *in;
279         unsigned long avail;
280
281         while (len) {
282                 in = use_pack(p, w_curs, offset, &avail);
283                 if (avail > len)
284                         avail = (unsigned long)len;
285                 hashwrite(f, in, avail);
286                 offset += avail;
287                 len -= avail;
288         }
289 }
290
291 /* Return 0 if we will bust the pack-size limit */
292 static unsigned long write_no_reuse_object(struct hashfile *f, struct object_entry *entry,
293                                            unsigned long limit, int usable_delta)
294 {
295         unsigned long size, datalen;
296         unsigned char header[MAX_PACK_OBJECT_HEADER],
297                       dheader[MAX_PACK_OBJECT_HEADER];
298         unsigned hdrlen;
299         enum object_type type;
300         void *buf;
301         struct git_istream *st = NULL;
302         const unsigned hashsz = the_hash_algo->rawsz;
303
304         if (!usable_delta) {
305                 if (oe_type(entry) == OBJ_BLOB &&
306                     oe_size_greater_than(&to_pack, entry, big_file_threshold) &&
307                     (st = open_istream(the_repository, &entry->idx.oid, &type,
308                                        &size, NULL)) != NULL)
309                         buf = NULL;
310                 else {
311                         buf = read_object_file(&entry->idx.oid, &type, &size);
312                         if (!buf)
313                                 die(_("unable to read %s"),
314                                     oid_to_hex(&entry->idx.oid));
315                 }
316                 /*
317                  * make sure no cached delta data remains from a
318                  * previous attempt before a pack split occurred.
319                  */
320                 FREE_AND_NULL(entry->delta_data);
321                 entry->z_delta_size = 0;
322         } else if (entry->delta_data) {
323                 size = DELTA_SIZE(entry);
324                 buf = entry->delta_data;
325                 entry->delta_data = NULL;
326                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
327                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
328         } else {
329                 buf = get_delta(entry);
330                 size = DELTA_SIZE(entry);
331                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
332                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
333         }
334
335         if (st) /* large blob case, just assume we don't compress well */
336                 datalen = size;
337         else if (entry->z_delta_size)
338                 datalen = entry->z_delta_size;
339         else
340                 datalen = do_compress(&buf, size);
341
342         /*
343          * The object header is a byte of 'type' followed by zero or
344          * more bytes of length.
345          */
346         hdrlen = encode_in_pack_object_header(header, sizeof(header),
347                                               type, size);
348
349         if (type == OBJ_OFS_DELTA) {
350                 /*
351                  * Deltas with relative base contain an additional
352                  * encoding of the relative offset for the delta
353                  * base from this object's position in the pack.
354                  */
355                 off_t ofs = entry->idx.offset - DELTA(entry)->idx.offset;
356                 unsigned pos = sizeof(dheader) - 1;
357                 dheader[pos] = ofs & 127;
358                 while (ofs >>= 7)
359                         dheader[--pos] = 128 | (--ofs & 127);
360                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + hashsz >= limit) {
361                         if (st)
362                                 close_istream(st);
363                         free(buf);
364                         return 0;
365                 }
366                 hashwrite(f, header, hdrlen);
367                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
368                 hdrlen += sizeof(dheader) - pos;
369         } else if (type == OBJ_REF_DELTA) {
370                 /*
371                  * Deltas with a base reference contain
372                  * additional bytes for the base object ID.
373                  */
374                 if (limit && hdrlen + hashsz + datalen + hashsz >= limit) {
375                         if (st)
376                                 close_istream(st);
377                         free(buf);
378                         return 0;
379                 }
380                 hashwrite(f, header, hdrlen);
381                 hashwrite(f, DELTA(entry)->idx.oid.hash, hashsz);
382                 hdrlen += hashsz;
383         } else {
384                 if (limit && hdrlen + datalen + hashsz >= limit) {
385                         if (st)
386                                 close_istream(st);
387                         free(buf);
388                         return 0;
389                 }
390                 hashwrite(f, header, hdrlen);
391         }
392         if (st) {
393                 datalen = write_large_blob_data(st, f, &entry->idx.oid);
394                 close_istream(st);
395         } else {
396                 hashwrite(f, buf, datalen);
397                 free(buf);
398         }
399
400         return hdrlen + datalen;
401 }
402
403 /* Return 0 if we will bust the pack-size limit */
404 static off_t write_reuse_object(struct hashfile *f, struct object_entry *entry,
405                                 unsigned long limit, int usable_delta)
406 {
407         struct packed_git *p = IN_PACK(entry);
408         struct pack_window *w_curs = NULL;
409         struct revindex_entry *revidx;
410         off_t offset;
411         enum object_type type = oe_type(entry);
412         off_t datalen;
413         unsigned char header[MAX_PACK_OBJECT_HEADER],
414                       dheader[MAX_PACK_OBJECT_HEADER];
415         unsigned hdrlen;
416         const unsigned hashsz = the_hash_algo->rawsz;
417         unsigned long entry_size = SIZE(entry);
418
419         if (DELTA(entry))
420                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
421                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
422         hdrlen = encode_in_pack_object_header(header, sizeof(header),
423                                               type, entry_size);
424
425         offset = entry->in_pack_offset;
426         revidx = find_pack_revindex(p, offset);
427         datalen = revidx[1].offset - offset;
428         if (!pack_to_stdout && p->index_version > 1 &&
429             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
430                 error(_("bad packed object CRC for %s"),
431                       oid_to_hex(&entry->idx.oid));
432                 unuse_pack(&w_curs);
433                 return write_no_reuse_object(f, entry, limit, usable_delta);
434         }
435
436         offset += entry->in_pack_header_size;
437         datalen -= entry->in_pack_header_size;
438
439         if (!pack_to_stdout && p->index_version == 1 &&
440             check_pack_inflate(p, &w_curs, offset, datalen, entry_size)) {
441                 error(_("corrupt packed object for %s"),
442                       oid_to_hex(&entry->idx.oid));
443                 unuse_pack(&w_curs);
444                 return write_no_reuse_object(f, entry, limit, usable_delta);
445         }
446
447         if (type == OBJ_OFS_DELTA) {
448                 off_t ofs = entry->idx.offset - DELTA(entry)->idx.offset;
449                 unsigned pos = sizeof(dheader) - 1;
450                 dheader[pos] = ofs & 127;
451                 while (ofs >>= 7)
452                         dheader[--pos] = 128 | (--ofs & 127);
453                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + hashsz >= limit) {
454                         unuse_pack(&w_curs);
455                         return 0;
456                 }
457                 hashwrite(f, header, hdrlen);
458                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
459                 hdrlen += sizeof(dheader) - pos;
460                 reused_delta++;
461         } else if (type == OBJ_REF_DELTA) {
462                 if (limit && hdrlen + hashsz + datalen + hashsz >= limit) {
463                         unuse_pack(&w_curs);
464                         return 0;
465                 }
466                 hashwrite(f, header, hdrlen);
467                 hashwrite(f, DELTA(entry)->idx.oid.hash, hashsz);
468                 hdrlen += hashsz;
469                 reused_delta++;
470         } else {
471                 if (limit && hdrlen + datalen + hashsz >= limit) {
472                         unuse_pack(&w_curs);
473                         return 0;
474                 }
475                 hashwrite(f, header, hdrlen);
476         }
477         copy_pack_data(f, p, &w_curs, offset, datalen);
478         unuse_pack(&w_curs);
479         reused++;
480         return hdrlen + datalen;
481 }
482
483 /* Return 0 if we will bust the pack-size limit */
484 static off_t write_object(struct hashfile *f,
485                           struct object_entry *entry,
486                           off_t write_offset)
487 {
488         unsigned long limit;
489         off_t len;
490         int usable_delta, to_reuse;
491
492         if (!pack_to_stdout)
493                 crc32_begin(f);
494
495         /* apply size limit if limited packsize and not first object */
496         if (!pack_size_limit || !nr_written)
497                 limit = 0;
498         else if (pack_size_limit <= write_offset)
499                 /*
500                  * the earlier object did not fit the limit; avoid
501                  * mistaking this with unlimited (i.e. limit = 0).
502                  */
503                 limit = 1;
504         else
505                 limit = pack_size_limit - write_offset;
506
507         if (!DELTA(entry))
508                 usable_delta = 0;       /* no delta */
509         else if (!pack_size_limit)
510                usable_delta = 1;        /* unlimited packfile */
511         else if (DELTA(entry)->idx.offset == (off_t)-1)
512                 usable_delta = 0;       /* base was written to another pack */
513         else if (DELTA(entry)->idx.offset)
514                 usable_delta = 1;       /* base already exists in this pack */
515         else
516                 usable_delta = 0;       /* base could end up in another pack */
517
518         if (!reuse_object)
519                 to_reuse = 0;   /* explicit */
520         else if (!IN_PACK(entry))
521                 to_reuse = 0;   /* can't reuse what we don't have */
522         else if (oe_type(entry) == OBJ_REF_DELTA ||
523                  oe_type(entry) == OBJ_OFS_DELTA)
524                                 /* check_object() decided it for us ... */
525                 to_reuse = usable_delta;
526                                 /* ... but pack split may override that */
527         else if (oe_type(entry) != entry->in_pack_type)
528                 to_reuse = 0;   /* pack has delta which is unusable */
529         else if (DELTA(entry))
530                 to_reuse = 0;   /* we want to pack afresh */
531         else
532                 to_reuse = 1;   /* we have it in-pack undeltified,
533                                  * and we do not need to deltify it.
534                                  */
535
536         if (!to_reuse)
537                 len = write_no_reuse_object(f, entry, limit, usable_delta);
538         else
539                 len = write_reuse_object(f, entry, limit, usable_delta);
540         if (!len)
541                 return 0;
542
543         if (usable_delta)
544                 written_delta++;
545         written++;
546         if (!pack_to_stdout)
547                 entry->idx.crc32 = crc32_end(f);
548         return len;
549 }
550
551 enum write_one_status {
552         WRITE_ONE_SKIP = -1, /* already written */
553         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
554         WRITE_ONE_WRITTEN = 1, /* normal */
555         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
556 };
557
558 static enum write_one_status write_one(struct hashfile *f,
559                                        struct object_entry *e,
560                                        off_t *offset)
561 {
562         off_t size;
563         int recursing;
564
565         /*
566          * we set offset to 1 (which is an impossible value) to mark
567          * the fact that this object is involved in "write its base
568          * first before writing a deltified object" recursion.
569          */
570         recursing = (e->idx.offset == 1);
571         if (recursing) {
572                 warning(_("recursive delta detected for object %s"),
573                         oid_to_hex(&e->idx.oid));
574                 return WRITE_ONE_RECURSIVE;
575         } else if (e->idx.offset || e->preferred_base) {
576                 /* offset is non zero if object is written already. */
577                 return WRITE_ONE_SKIP;
578         }
579
580         /* if we are deltified, write out base object first. */
581         if (DELTA(e)) {
582                 e->idx.offset = 1; /* now recurse */
583                 switch (write_one(f, DELTA(e), offset)) {
584                 case WRITE_ONE_RECURSIVE:
585                         /* we cannot depend on this one */
586                         SET_DELTA(e, NULL);
587                         break;
588                 default:
589                         break;
590                 case WRITE_ONE_BREAK:
591                         e->idx.offset = recursing;
592                         return WRITE_ONE_BREAK;
593                 }
594         }
595
596         e->idx.offset = *offset;
597         size = write_object(f, e, *offset);
598         if (!size) {
599                 e->idx.offset = recursing;
600                 return WRITE_ONE_BREAK;
601         }
602         written_list[nr_written++] = &e->idx;
603
604         /* make sure off_t is sufficiently large not to wrap */
605         if (signed_add_overflows(*offset, size))
606                 die(_("pack too large for current definition of off_t"));
607         *offset += size;
608         return WRITE_ONE_WRITTEN;
609 }
610
611 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
612                        void *cb_data)
613 {
614         struct object_id peeled;
615         struct object_entry *entry = packlist_find(&to_pack, oid);
616
617         if (entry)
618                 entry->tagged = 1;
619         if (!peel_ref(path, &peeled)) {
620                 entry = packlist_find(&to_pack, &peeled);
621                 if (entry)
622                         entry->tagged = 1;
623         }
624         return 0;
625 }
626
627 static inline void add_to_write_order(struct object_entry **wo,
628                                unsigned int *endp,
629                                struct object_entry *e)
630 {
631         if (e->filled || oe_layer(&to_pack, e) != write_layer)
632                 return;
633         wo[(*endp)++] = e;
634         e->filled = 1;
635 }
636
637 static void add_descendants_to_write_order(struct object_entry **wo,
638                                            unsigned int *endp,
639                                            struct object_entry *e)
640 {
641         int add_to_order = 1;
642         while (e) {
643                 if (add_to_order) {
644                         struct object_entry *s;
645                         /* add this node... */
646                         add_to_write_order(wo, endp, e);
647                         /* all its siblings... */
648                         for (s = DELTA_SIBLING(e); s; s = DELTA_SIBLING(s)) {
649                                 add_to_write_order(wo, endp, s);
650                         }
651                 }
652                 /* drop down a level to add left subtree nodes if possible */
653                 if (DELTA_CHILD(e)) {
654                         add_to_order = 1;
655                         e = DELTA_CHILD(e);
656                 } else {
657                         add_to_order = 0;
658                         /* our sibling might have some children, it is next */
659                         if (DELTA_SIBLING(e)) {
660                                 e = DELTA_SIBLING(e);
661                                 continue;
662                         }
663                         /* go back to our parent node */
664                         e = DELTA(e);
665                         while (e && !DELTA_SIBLING(e)) {
666                                 /* we're on the right side of a subtree, keep
667                                  * going up until we can go right again */
668                                 e = DELTA(e);
669                         }
670                         if (!e) {
671                                 /* done- we hit our original root node */
672                                 return;
673                         }
674                         /* pass it off to sibling at this level */
675                         e = DELTA_SIBLING(e);
676                 }
677         };
678 }
679
680 static void add_family_to_write_order(struct object_entry **wo,
681                                       unsigned int *endp,
682                                       struct object_entry *e)
683 {
684         struct object_entry *root;
685
686         for (root = e; DELTA(root); root = DELTA(root))
687                 ; /* nothing */
688         add_descendants_to_write_order(wo, endp, root);
689 }
690
691 static void compute_layer_order(struct object_entry **wo, unsigned int *wo_end)
692 {
693         unsigned int i, last_untagged;
694         struct object_entry *objects = to_pack.objects;
695
696         for (i = 0; i < to_pack.nr_objects; i++) {
697                 if (objects[i].tagged)
698                         break;
699                 add_to_write_order(wo, wo_end, &objects[i]);
700         }
701         last_untagged = i;
702
703         /*
704          * Then fill all the tagged tips.
705          */
706         for (; i < to_pack.nr_objects; i++) {
707                 if (objects[i].tagged)
708                         add_to_write_order(wo, wo_end, &objects[i]);
709         }
710
711         /*
712          * And then all remaining commits and tags.
713          */
714         for (i = last_untagged; i < to_pack.nr_objects; i++) {
715                 if (oe_type(&objects[i]) != OBJ_COMMIT &&
716                     oe_type(&objects[i]) != OBJ_TAG)
717                         continue;
718                 add_to_write_order(wo, wo_end, &objects[i]);
719         }
720
721         /*
722          * And then all the trees.
723          */
724         for (i = last_untagged; i < to_pack.nr_objects; i++) {
725                 if (oe_type(&objects[i]) != OBJ_TREE)
726                         continue;
727                 add_to_write_order(wo, wo_end, &objects[i]);
728         }
729
730         /*
731          * Finally all the rest in really tight order
732          */
733         for (i = last_untagged; i < to_pack.nr_objects; i++) {
734                 if (!objects[i].filled && oe_layer(&to_pack, &objects[i]) == write_layer)
735                         add_family_to_write_order(wo, wo_end, &objects[i]);
736         }
737 }
738
739 static struct object_entry **compute_write_order(void)
740 {
741         uint32_t max_layers = 1;
742         unsigned int i, wo_end;
743
744         struct object_entry **wo;
745         struct object_entry *objects = to_pack.objects;
746
747         for (i = 0; i < to_pack.nr_objects; i++) {
748                 objects[i].tagged = 0;
749                 objects[i].filled = 0;
750                 SET_DELTA_CHILD(&objects[i], NULL);
751                 SET_DELTA_SIBLING(&objects[i], NULL);
752         }
753
754         /*
755          * Fully connect delta_child/delta_sibling network.
756          * Make sure delta_sibling is sorted in the original
757          * recency order.
758          */
759         for (i = to_pack.nr_objects; i > 0;) {
760                 struct object_entry *e = &objects[--i];
761                 if (!DELTA(e))
762                         continue;
763                 /* Mark me as the first child */
764                 e->delta_sibling_idx = DELTA(e)->delta_child_idx;
765                 SET_DELTA_CHILD(DELTA(e), e);
766         }
767
768         /*
769          * Mark objects that are at the tip of tags.
770          */
771         for_each_tag_ref(mark_tagged, NULL);
772
773         if (use_delta_islands)
774                 max_layers = compute_pack_layers(&to_pack);
775
776         ALLOC_ARRAY(wo, to_pack.nr_objects);
777         wo_end = 0;
778
779         for (; write_layer < max_layers; ++write_layer)
780                 compute_layer_order(wo, &wo_end);
781
782         if (wo_end != to_pack.nr_objects)
783                 die(_("ordered %u objects, expected %"PRIu32),
784                     wo_end, to_pack.nr_objects);
785
786         return wo;
787 }
788
789
790 /*
791  * A reused set of objects. All objects in a chunk have the same
792  * relative position in the original packfile and the generated
793  * packfile.
794  */
795
796 static struct reused_chunk {
797         /* The offset of the first object of this chunk in the original
798          * packfile. */
799         off_t original;
800         /* The offset of the first object of this chunk in the generated
801          * packfile minus "original". */
802         off_t difference;
803 } *reused_chunks;
804 static int reused_chunks_nr;
805 static int reused_chunks_alloc;
806
807 static void record_reused_object(off_t where, off_t offset)
808 {
809         if (reused_chunks_nr && reused_chunks[reused_chunks_nr-1].difference == offset)
810                 return;
811
812         ALLOC_GROW(reused_chunks, reused_chunks_nr + 1,
813                    reused_chunks_alloc);
814         reused_chunks[reused_chunks_nr].original = where;
815         reused_chunks[reused_chunks_nr].difference = offset;
816         reused_chunks_nr++;
817 }
818
819 /*
820  * Binary search to find the chunk that "where" is in. Note
821  * that we're not looking for an exact match, just the first
822  * chunk that contains it (which implicitly ends at the start
823  * of the next chunk.
824  */
825 static off_t find_reused_offset(off_t where)
826 {
827         int lo = 0, hi = reused_chunks_nr;
828         while (lo < hi) {
829                 int mi = lo + ((hi - lo) / 2);
830                 if (where == reused_chunks[mi].original)
831                         return reused_chunks[mi].difference;
832                 if (where < reused_chunks[mi].original)
833                         hi = mi;
834                 else
835                         lo = mi + 1;
836         }
837
838         /*
839          * The first chunk starts at zero, so we can't have gone below
840          * there.
841          */
842         assert(lo);
843         return reused_chunks[lo-1].difference;
844 }
845
846 static void write_reused_pack_one(size_t pos, struct hashfile *out,
847                                   struct pack_window **w_curs)
848 {
849         off_t offset, next, cur;
850         enum object_type type;
851         unsigned long size;
852
853         offset = reuse_packfile->revindex[pos].offset;
854         next = reuse_packfile->revindex[pos + 1].offset;
855
856         record_reused_object(offset, offset - hashfile_total(out));
857
858         cur = offset;
859         type = unpack_object_header(reuse_packfile, w_curs, &cur, &size);
860         assert(type >= 0);
861
862         if (type == OBJ_OFS_DELTA) {
863                 off_t base_offset;
864                 off_t fixup;
865
866                 unsigned char header[MAX_PACK_OBJECT_HEADER];
867                 unsigned len;
868
869                 base_offset = get_delta_base(reuse_packfile, w_curs, &cur, type, offset);
870                 assert(base_offset != 0);
871
872                 /* Convert to REF_DELTA if we must... */
873                 if (!allow_ofs_delta) {
874                         int base_pos = find_revindex_position(reuse_packfile, base_offset);
875                         struct object_id base_oid;
876
877                         nth_packed_object_id(&base_oid, reuse_packfile,
878                                              reuse_packfile->revindex[base_pos].nr);
879
880                         len = encode_in_pack_object_header(header, sizeof(header),
881                                                            OBJ_REF_DELTA, size);
882                         hashwrite(out, header, len);
883                         hashwrite(out, base_oid.hash, the_hash_algo->rawsz);
884                         copy_pack_data(out, reuse_packfile, w_curs, cur, next - cur);
885                         return;
886                 }
887
888                 /* Otherwise see if we need to rewrite the offset... */
889                 fixup = find_reused_offset(offset) -
890                         find_reused_offset(base_offset);
891                 if (fixup) {
892                         unsigned char ofs_header[10];
893                         unsigned i, ofs_len;
894                         off_t ofs = offset - base_offset - fixup;
895
896                         len = encode_in_pack_object_header(header, sizeof(header),
897                                                            OBJ_OFS_DELTA, size);
898
899                         i = sizeof(ofs_header) - 1;
900                         ofs_header[i] = ofs & 127;
901                         while (ofs >>= 7)
902                                 ofs_header[--i] = 128 | (--ofs & 127);
903
904                         ofs_len = sizeof(ofs_header) - i;
905
906                         hashwrite(out, header, len);
907                         hashwrite(out, ofs_header + sizeof(ofs_header) - ofs_len, ofs_len);
908                         copy_pack_data(out, reuse_packfile, w_curs, cur, next - cur);
909                         return;
910                 }
911
912                 /* ...otherwise we have no fixup, and can write it verbatim */
913         }
914
915         copy_pack_data(out, reuse_packfile, w_curs, offset, next - offset);
916 }
917
918 static size_t write_reused_pack_verbatim(struct hashfile *out,
919                                          struct pack_window **w_curs)
920 {
921         size_t pos = 0;
922
923         while (pos < reuse_packfile_bitmap->word_alloc &&
924                         reuse_packfile_bitmap->words[pos] == (eword_t)~0)
925                 pos++;
926
927         if (pos) {
928                 off_t to_write;
929
930                 written = (pos * BITS_IN_EWORD);
931                 to_write = reuse_packfile->revindex[written].offset
932                         - sizeof(struct pack_header);
933
934                 /* We're recording one chunk, not one object. */
935                 record_reused_object(sizeof(struct pack_header), 0);
936                 hashflush(out);
937                 copy_pack_data(out, reuse_packfile, w_curs,
938                         sizeof(struct pack_header), to_write);
939
940                 display_progress(progress_state, written);
941         }
942         return pos;
943 }
944
945 static void write_reused_pack(struct hashfile *f)
946 {
947         size_t i = 0;
948         uint32_t offset;
949         struct pack_window *w_curs = NULL;
950
951         if (allow_ofs_delta)
952                 i = write_reused_pack_verbatim(f, &w_curs);
953
954         for (; i < reuse_packfile_bitmap->word_alloc; ++i) {
955                 eword_t word = reuse_packfile_bitmap->words[i];
956                 size_t pos = (i * BITS_IN_EWORD);
957
958                 for (offset = 0; offset < BITS_IN_EWORD; ++offset) {
959                         if ((word >> offset) == 0)
960                                 break;
961
962                         offset += ewah_bit_ctz64(word >> offset);
963                         write_reused_pack_one(pos + offset, f, &w_curs);
964                         display_progress(progress_state, ++written);
965                 }
966         }
967
968         unuse_pack(&w_curs);
969 }
970
971 static const char no_split_warning[] = N_(
972 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
973 );
974
975 static void write_pack_file(void)
976 {
977         uint32_t i = 0, j;
978         struct hashfile *f;
979         off_t offset;
980         uint32_t nr_remaining = nr_result;
981         time_t last_mtime = 0;
982         struct object_entry **write_order;
983
984         if (progress > pack_to_stdout)
985                 progress_state = start_progress(_("Writing objects"), nr_result);
986         ALLOC_ARRAY(written_list, to_pack.nr_objects);
987         write_order = compute_write_order();
988
989         do {
990                 struct object_id oid;
991                 char *pack_tmp_name = NULL;
992
993                 if (pack_to_stdout)
994                         f = hashfd_throughput(1, "<stdout>", progress_state);
995                 else
996                         f = create_tmp_packfile(&pack_tmp_name);
997
998                 offset = write_pack_header(f, nr_remaining);
999
1000                 if (reuse_packfile) {
1001                         assert(pack_to_stdout);
1002                         write_reused_pack(f);
1003                         offset = hashfile_total(f);
1004                 }
1005
1006                 nr_written = 0;
1007                 for (; i < to_pack.nr_objects; i++) {
1008                         struct object_entry *e = write_order[i];
1009                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
1010                                 break;
1011                         display_progress(progress_state, written);
1012                 }
1013
1014                 /*
1015                  * Did we write the wrong # entries in the header?
1016                  * If so, rewrite it like in fast-import
1017                  */
1018                 if (pack_to_stdout) {
1019                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_CLOSE);
1020                 } else if (nr_written == nr_remaining) {
1021                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE);
1022                 } else {
1023                         int fd = finalize_hashfile(f, oid.hash, 0);
1024                         fixup_pack_header_footer(fd, oid.hash, pack_tmp_name,
1025                                                  nr_written, oid.hash, offset);
1026                         close(fd);
1027                         if (write_bitmap_index) {
1028                                 if (write_bitmap_index != WRITE_BITMAP_QUIET)
1029                                         warning(_(no_split_warning));
1030                                 write_bitmap_index = 0;
1031                         }
1032                 }
1033
1034                 if (!pack_to_stdout) {
1035                         struct stat st;
1036                         struct strbuf tmpname = STRBUF_INIT;
1037
1038                         /*
1039                          * Packs are runtime accessed in their mtime
1040                          * order since newer packs are more likely to contain
1041                          * younger objects.  So if we are creating multiple
1042                          * packs then we should modify the mtime of later ones
1043                          * to preserve this property.
1044                          */
1045                         if (stat(pack_tmp_name, &st) < 0) {
1046                                 warning_errno(_("failed to stat %s"), pack_tmp_name);
1047                         } else if (!last_mtime) {
1048                                 last_mtime = st.st_mtime;
1049                         } else {
1050                                 struct utimbuf utb;
1051                                 utb.actime = st.st_atime;
1052                                 utb.modtime = --last_mtime;
1053                                 if (utime(pack_tmp_name, &utb) < 0)
1054                                         warning_errno(_("failed utime() on %s"), pack_tmp_name);
1055                         }
1056
1057                         strbuf_addf(&tmpname, "%s-", base_name);
1058
1059                         if (write_bitmap_index) {
1060                                 bitmap_writer_set_checksum(oid.hash);
1061                                 bitmap_writer_build_type_index(
1062                                         &to_pack, written_list, nr_written);
1063                         }
1064
1065                         finish_tmp_packfile(&tmpname, pack_tmp_name,
1066                                             written_list, nr_written,
1067                                             &pack_idx_opts, oid.hash);
1068
1069                         if (write_bitmap_index) {
1070                                 strbuf_addf(&tmpname, "%s.bitmap", oid_to_hex(&oid));
1071
1072                                 stop_progress(&progress_state);
1073
1074                                 bitmap_writer_show_progress(progress);
1075                                 bitmap_writer_reuse_bitmaps(&to_pack);
1076                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
1077                                 bitmap_writer_build(&to_pack);
1078                                 bitmap_writer_finish(written_list, nr_written,
1079                                                      tmpname.buf, write_bitmap_options);
1080                                 write_bitmap_index = 0;
1081                         }
1082
1083                         strbuf_release(&tmpname);
1084                         free(pack_tmp_name);
1085                         puts(oid_to_hex(&oid));
1086                 }
1087
1088                 /* mark written objects as written to previous pack */
1089                 for (j = 0; j < nr_written; j++) {
1090                         written_list[j]->offset = (off_t)-1;
1091                 }
1092                 nr_remaining -= nr_written;
1093         } while (nr_remaining && i < to_pack.nr_objects);
1094
1095         free(written_list);
1096         free(write_order);
1097         stop_progress(&progress_state);
1098         if (written != nr_result)
1099                 die(_("wrote %"PRIu32" objects while expecting %"PRIu32),
1100                     written, nr_result);
1101         trace2_data_intmax("pack-objects", the_repository,
1102                            "write_pack_file/wrote", nr_result);
1103 }
1104
1105 static int no_try_delta(const char *path)
1106 {
1107         static struct attr_check *check;
1108
1109         if (!check)
1110                 check = attr_check_initl("delta", NULL);
1111         git_check_attr(the_repository->index, path, check);
1112         if (ATTR_FALSE(check->items[0].value))
1113                 return 1;
1114         return 0;
1115 }
1116
1117 /*
1118  * When adding an object, check whether we have already added it
1119  * to our packing list. If so, we can skip. However, if we are
1120  * being asked to excludei t, but the previous mention was to include
1121  * it, make sure to adjust its flags and tweak our numbers accordingly.
1122  *
1123  * As an optimization, we pass out the index position where we would have
1124  * found the item, since that saves us from having to look it up again a
1125  * few lines later when we want to add the new entry.
1126  */
1127 static int have_duplicate_entry(const struct object_id *oid,
1128                                 int exclude)
1129 {
1130         struct object_entry *entry;
1131
1132         if (reuse_packfile_bitmap &&
1133             bitmap_walk_contains(bitmap_git, reuse_packfile_bitmap, oid))
1134                 return 1;
1135
1136         entry = packlist_find(&to_pack, oid);
1137         if (!entry)
1138                 return 0;
1139
1140         if (exclude) {
1141                 if (!entry->preferred_base)
1142                         nr_result--;
1143                 entry->preferred_base = 1;
1144         }
1145
1146         return 1;
1147 }
1148
1149 static int want_found_object(int exclude, struct packed_git *p)
1150 {
1151         if (exclude)
1152                 return 1;
1153         if (incremental)
1154                 return 0;
1155
1156         /*
1157          * When asked to do --local (do not include an object that appears in a
1158          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
1159          * an object that appears in a pack marked with .keep), finding a pack
1160          * that matches the criteria is sufficient for us to decide to omit it.
1161          * However, even if this pack does not satisfy the criteria, we need to
1162          * make sure no copy of this object appears in _any_ pack that makes us
1163          * to omit the object, so we need to check all the packs.
1164          *
1165          * We can however first check whether these options can possible matter;
1166          * if they do not matter we know we want the object in generated pack.
1167          * Otherwise, we signal "-1" at the end to tell the caller that we do
1168          * not know either way, and it needs to check more packs.
1169          */
1170         if (!ignore_packed_keep_on_disk &&
1171             !ignore_packed_keep_in_core &&
1172             (!local || !have_non_local_packs))
1173                 return 1;
1174
1175         if (local && !p->pack_local)
1176                 return 0;
1177         if (p->pack_local &&
1178             ((ignore_packed_keep_on_disk && p->pack_keep) ||
1179              (ignore_packed_keep_in_core && p->pack_keep_in_core)))
1180                 return 0;
1181
1182         /* we don't know yet; keep looking for more packs */
1183         return -1;
1184 }
1185
1186 /*
1187  * Check whether we want the object in the pack (e.g., we do not want
1188  * objects found in non-local stores if the "--local" option was used).
1189  *
1190  * If the caller already knows an existing pack it wants to take the object
1191  * from, that is passed in *found_pack and *found_offset; otherwise this
1192  * function finds if there is any pack that has the object and returns the pack
1193  * and its offset in these variables.
1194  */
1195 static int want_object_in_pack(const struct object_id *oid,
1196                                int exclude,
1197                                struct packed_git **found_pack,
1198                                off_t *found_offset)
1199 {
1200         int want;
1201         struct list_head *pos;
1202         struct multi_pack_index *m;
1203
1204         if (!exclude && local && has_loose_object_nonlocal(oid))
1205                 return 0;
1206
1207         /*
1208          * If we already know the pack object lives in, start checks from that
1209          * pack - in the usual case when neither --local was given nor .keep files
1210          * are present we will determine the answer right now.
1211          */
1212         if (*found_pack) {
1213                 want = want_found_object(exclude, *found_pack);
1214                 if (want != -1)
1215                         return want;
1216         }
1217
1218         for (m = get_multi_pack_index(the_repository); m; m = m->next) {
1219                 struct pack_entry e;
1220                 if (fill_midx_entry(the_repository, oid, &e, m)) {
1221                         struct packed_git *p = e.p;
1222                         off_t offset;
1223
1224                         if (p == *found_pack)
1225                                 offset = *found_offset;
1226                         else
1227                                 offset = find_pack_entry_one(oid->hash, p);
1228
1229                         if (offset) {
1230                                 if (!*found_pack) {
1231                                         if (!is_pack_valid(p))
1232                                                 continue;
1233                                         *found_offset = offset;
1234                                         *found_pack = p;
1235                                 }
1236                                 want = want_found_object(exclude, p);
1237                                 if (want != -1)
1238                                         return want;
1239                         }
1240                 }
1241         }
1242
1243         list_for_each(pos, get_packed_git_mru(the_repository)) {
1244                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
1245                 off_t offset;
1246
1247                 if (p == *found_pack)
1248                         offset = *found_offset;
1249                 else
1250                         offset = find_pack_entry_one(oid->hash, p);
1251
1252                 if (offset) {
1253                         if (!*found_pack) {
1254                                 if (!is_pack_valid(p))
1255                                         continue;
1256                                 *found_offset = offset;
1257                                 *found_pack = p;
1258                         }
1259                         want = want_found_object(exclude, p);
1260                         if (!exclude && want > 0)
1261                                 list_move(&p->mru,
1262                                           get_packed_git_mru(the_repository));
1263                         if (want != -1)
1264                                 return want;
1265                 }
1266         }
1267
1268         return 1;
1269 }
1270
1271 static void create_object_entry(const struct object_id *oid,
1272                                 enum object_type type,
1273                                 uint32_t hash,
1274                                 int exclude,
1275                                 int no_try_delta,
1276                                 struct packed_git *found_pack,
1277                                 off_t found_offset)
1278 {
1279         struct object_entry *entry;
1280
1281         entry = packlist_alloc(&to_pack, oid);
1282         entry->hash = hash;
1283         oe_set_type(entry, type);
1284         if (exclude)
1285                 entry->preferred_base = 1;
1286         else
1287                 nr_result++;
1288         if (found_pack) {
1289                 oe_set_in_pack(&to_pack, entry, found_pack);
1290                 entry->in_pack_offset = found_offset;
1291         }
1292
1293         entry->no_try_delta = no_try_delta;
1294 }
1295
1296 static const char no_closure_warning[] = N_(
1297 "disabling bitmap writing, as some objects are not being packed"
1298 );
1299
1300 static int add_object_entry(const struct object_id *oid, enum object_type type,
1301                             const char *name, int exclude)
1302 {
1303         struct packed_git *found_pack = NULL;
1304         off_t found_offset = 0;
1305
1306         display_progress(progress_state, ++nr_seen);
1307
1308         if (have_duplicate_entry(oid, exclude))
1309                 return 0;
1310
1311         if (!want_object_in_pack(oid, exclude, &found_pack, &found_offset)) {
1312                 /* The pack is missing an object, so it will not have closure */
1313                 if (write_bitmap_index) {
1314                         if (write_bitmap_index != WRITE_BITMAP_QUIET)
1315                                 warning(_(no_closure_warning));
1316                         write_bitmap_index = 0;
1317                 }
1318                 return 0;
1319         }
1320
1321         create_object_entry(oid, type, pack_name_hash(name),
1322                             exclude, name && no_try_delta(name),
1323                             found_pack, found_offset);
1324         return 1;
1325 }
1326
1327 static int add_object_entry_from_bitmap(const struct object_id *oid,
1328                                         enum object_type type,
1329                                         int flags, uint32_t name_hash,
1330                                         struct packed_git *pack, off_t offset)
1331 {
1332         display_progress(progress_state, ++nr_seen);
1333
1334         if (have_duplicate_entry(oid, 0))
1335                 return 0;
1336
1337         if (!want_object_in_pack(oid, 0, &pack, &offset))
1338                 return 0;
1339
1340         create_object_entry(oid, type, name_hash, 0, 0, pack, offset);
1341         return 1;
1342 }
1343
1344 struct pbase_tree_cache {
1345         struct object_id oid;
1346         int ref;
1347         int temporary;
1348         void *tree_data;
1349         unsigned long tree_size;
1350 };
1351
1352 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1353 static int pbase_tree_cache_ix(const struct object_id *oid)
1354 {
1355         return oid->hash[0] % ARRAY_SIZE(pbase_tree_cache);
1356 }
1357 static int pbase_tree_cache_ix_incr(int ix)
1358 {
1359         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1360 }
1361
1362 static struct pbase_tree {
1363         struct pbase_tree *next;
1364         /* This is a phony "cache" entry; we are not
1365          * going to evict it or find it through _get()
1366          * mechanism -- this is for the toplevel node that
1367          * would almost always change with any commit.
1368          */
1369         struct pbase_tree_cache pcache;
1370 } *pbase_tree;
1371
1372 static struct pbase_tree_cache *pbase_tree_get(const struct object_id *oid)
1373 {
1374         struct pbase_tree_cache *ent, *nent;
1375         void *data;
1376         unsigned long size;
1377         enum object_type type;
1378         int neigh;
1379         int my_ix = pbase_tree_cache_ix(oid);
1380         int available_ix = -1;
1381
1382         /* pbase-tree-cache acts as a limited hashtable.
1383          * your object will be found at your index or within a few
1384          * slots after that slot if it is cached.
1385          */
1386         for (neigh = 0; neigh < 8; neigh++) {
1387                 ent = pbase_tree_cache[my_ix];
1388                 if (ent && oideq(&ent->oid, oid)) {
1389                         ent->ref++;
1390                         return ent;
1391                 }
1392                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1393                          ((0 <= available_ix) &&
1394                           (!ent && pbase_tree_cache[available_ix])))
1395                         available_ix = my_ix;
1396                 if (!ent)
1397                         break;
1398                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1399         }
1400
1401         /* Did not find one.  Either we got a bogus request or
1402          * we need to read and perhaps cache.
1403          */
1404         data = read_object_file(oid, &type, &size);
1405         if (!data)
1406                 return NULL;
1407         if (type != OBJ_TREE) {
1408                 free(data);
1409                 return NULL;
1410         }
1411
1412         /* We need to either cache or return a throwaway copy */
1413
1414         if (available_ix < 0)
1415                 ent = NULL;
1416         else {
1417                 ent = pbase_tree_cache[available_ix];
1418                 my_ix = available_ix;
1419         }
1420
1421         if (!ent) {
1422                 nent = xmalloc(sizeof(*nent));
1423                 nent->temporary = (available_ix < 0);
1424         }
1425         else {
1426                 /* evict and reuse */
1427                 free(ent->tree_data);
1428                 nent = ent;
1429         }
1430         oidcpy(&nent->oid, oid);
1431         nent->tree_data = data;
1432         nent->tree_size = size;
1433         nent->ref = 1;
1434         if (!nent->temporary)
1435                 pbase_tree_cache[my_ix] = nent;
1436         return nent;
1437 }
1438
1439 static void pbase_tree_put(struct pbase_tree_cache *cache)
1440 {
1441         if (!cache->temporary) {
1442                 cache->ref--;
1443                 return;
1444         }
1445         free(cache->tree_data);
1446         free(cache);
1447 }
1448
1449 static int name_cmp_len(const char *name)
1450 {
1451         int i;
1452         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1453                 ;
1454         return i;
1455 }
1456
1457 static void add_pbase_object(struct tree_desc *tree,
1458                              const char *name,
1459                              int cmplen,
1460                              const char *fullname)
1461 {
1462         struct name_entry entry;
1463         int cmp;
1464
1465         while (tree_entry(tree,&entry)) {
1466                 if (S_ISGITLINK(entry.mode))
1467                         continue;
1468                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1469                       memcmp(name, entry.path, cmplen);
1470                 if (cmp > 0)
1471                         continue;
1472                 if (cmp < 0)
1473                         return;
1474                 if (name[cmplen] != '/') {
1475                         add_object_entry(&entry.oid,
1476                                          object_type(entry.mode),
1477                                          fullname, 1);
1478                         return;
1479                 }
1480                 if (S_ISDIR(entry.mode)) {
1481                         struct tree_desc sub;
1482                         struct pbase_tree_cache *tree;
1483                         const char *down = name+cmplen+1;
1484                         int downlen = name_cmp_len(down);
1485
1486                         tree = pbase_tree_get(&entry.oid);
1487                         if (!tree)
1488                                 return;
1489                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1490
1491                         add_pbase_object(&sub, down, downlen, fullname);
1492                         pbase_tree_put(tree);
1493                 }
1494         }
1495 }
1496
1497 static unsigned *done_pbase_paths;
1498 static int done_pbase_paths_num;
1499 static int done_pbase_paths_alloc;
1500 static int done_pbase_path_pos(unsigned hash)
1501 {
1502         int lo = 0;
1503         int hi = done_pbase_paths_num;
1504         while (lo < hi) {
1505                 int mi = lo + (hi - lo) / 2;
1506                 if (done_pbase_paths[mi] == hash)
1507                         return mi;
1508                 if (done_pbase_paths[mi] < hash)
1509                         hi = mi;
1510                 else
1511                         lo = mi + 1;
1512         }
1513         return -lo-1;
1514 }
1515
1516 static int check_pbase_path(unsigned hash)
1517 {
1518         int pos = done_pbase_path_pos(hash);
1519         if (0 <= pos)
1520                 return 1;
1521         pos = -pos - 1;
1522         ALLOC_GROW(done_pbase_paths,
1523                    done_pbase_paths_num + 1,
1524                    done_pbase_paths_alloc);
1525         done_pbase_paths_num++;
1526         if (pos < done_pbase_paths_num)
1527                 MOVE_ARRAY(done_pbase_paths + pos + 1, done_pbase_paths + pos,
1528                            done_pbase_paths_num - pos - 1);
1529         done_pbase_paths[pos] = hash;
1530         return 0;
1531 }
1532
1533 static void add_preferred_base_object(const char *name)
1534 {
1535         struct pbase_tree *it;
1536         int cmplen;
1537         unsigned hash = pack_name_hash(name);
1538
1539         if (!num_preferred_base || check_pbase_path(hash))
1540                 return;
1541
1542         cmplen = name_cmp_len(name);
1543         for (it = pbase_tree; it; it = it->next) {
1544                 if (cmplen == 0) {
1545                         add_object_entry(&it->pcache.oid, OBJ_TREE, NULL, 1);
1546                 }
1547                 else {
1548                         struct tree_desc tree;
1549                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1550                         add_pbase_object(&tree, name, cmplen, name);
1551                 }
1552         }
1553 }
1554
1555 static void add_preferred_base(struct object_id *oid)
1556 {
1557         struct pbase_tree *it;
1558         void *data;
1559         unsigned long size;
1560         struct object_id tree_oid;
1561
1562         if (window <= num_preferred_base++)
1563                 return;
1564
1565         data = read_object_with_reference(the_repository, oid,
1566                                           tree_type, &size, &tree_oid);
1567         if (!data)
1568                 return;
1569
1570         for (it = pbase_tree; it; it = it->next) {
1571                 if (oideq(&it->pcache.oid, &tree_oid)) {
1572                         free(data);
1573                         return;
1574                 }
1575         }
1576
1577         it = xcalloc(1, sizeof(*it));
1578         it->next = pbase_tree;
1579         pbase_tree = it;
1580
1581         oidcpy(&it->pcache.oid, &tree_oid);
1582         it->pcache.tree_data = data;
1583         it->pcache.tree_size = size;
1584 }
1585
1586 static void cleanup_preferred_base(void)
1587 {
1588         struct pbase_tree *it;
1589         unsigned i;
1590
1591         it = pbase_tree;
1592         pbase_tree = NULL;
1593         while (it) {
1594                 struct pbase_tree *tmp = it;
1595                 it = tmp->next;
1596                 free(tmp->pcache.tree_data);
1597                 free(tmp);
1598         }
1599
1600         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1601                 if (!pbase_tree_cache[i])
1602                         continue;
1603                 free(pbase_tree_cache[i]->tree_data);
1604                 FREE_AND_NULL(pbase_tree_cache[i]);
1605         }
1606
1607         FREE_AND_NULL(done_pbase_paths);
1608         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1609 }
1610
1611 /*
1612  * Return 1 iff the object specified by "delta" can be sent
1613  * literally as a delta against the base in "base_sha1". If
1614  * so, then *base_out will point to the entry in our packing
1615  * list, or NULL if we must use the external-base list.
1616  *
1617  * Depth value does not matter - find_deltas() will
1618  * never consider reused delta as the base object to
1619  * deltify other objects against, in order to avoid
1620  * circular deltas.
1621  */
1622 static int can_reuse_delta(const struct object_id *base_oid,
1623                            struct object_entry *delta,
1624                            struct object_entry **base_out)
1625 {
1626         struct object_entry *base;
1627
1628         /*
1629          * First see if we're already sending the base (or it's explicitly in
1630          * our "excluded" list).
1631          */
1632         base = packlist_find(&to_pack, base_oid);
1633         if (base) {
1634                 if (!in_same_island(&delta->idx.oid, &base->idx.oid))
1635                         return 0;
1636                 *base_out = base;
1637                 return 1;
1638         }
1639
1640         /*
1641          * Otherwise, reachability bitmaps may tell us if the receiver has it,
1642          * even if it was buried too deep in history to make it into the
1643          * packing list.
1644          */
1645         if (thin && bitmap_has_oid_in_uninteresting(bitmap_git, base_oid)) {
1646                 if (use_delta_islands) {
1647                         if (!in_same_island(&delta->idx.oid, base_oid))
1648                                 return 0;
1649                 }
1650                 *base_out = NULL;
1651                 return 1;
1652         }
1653
1654         return 0;
1655 }
1656
1657 static void check_object(struct object_entry *entry)
1658 {
1659         unsigned long canonical_size;
1660
1661         if (IN_PACK(entry)) {
1662                 struct packed_git *p = IN_PACK(entry);
1663                 struct pack_window *w_curs = NULL;
1664                 int have_base = 0;
1665                 struct object_id base_ref;
1666                 struct object_entry *base_entry;
1667                 unsigned long used, used_0;
1668                 unsigned long avail;
1669                 off_t ofs;
1670                 unsigned char *buf, c;
1671                 enum object_type type;
1672                 unsigned long in_pack_size;
1673
1674                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1675
1676                 /*
1677                  * We want in_pack_type even if we do not reuse delta
1678                  * since non-delta representations could still be reused.
1679                  */
1680                 used = unpack_object_header_buffer(buf, avail,
1681                                                    &type,
1682                                                    &in_pack_size);
1683                 if (used == 0)
1684                         goto give_up;
1685
1686                 if (type < 0)
1687                         BUG("invalid type %d", type);
1688                 entry->in_pack_type = type;
1689
1690                 /*
1691                  * Determine if this is a delta and if so whether we can
1692                  * reuse it or not.  Otherwise let's find out as cheaply as
1693                  * possible what the actual type and size for this object is.
1694                  */
1695                 switch (entry->in_pack_type) {
1696                 default:
1697                         /* Not a delta hence we've already got all we need. */
1698                         oe_set_type(entry, entry->in_pack_type);
1699                         SET_SIZE(entry, in_pack_size);
1700                         entry->in_pack_header_size = used;
1701                         if (oe_type(entry) < OBJ_COMMIT || oe_type(entry) > OBJ_BLOB)
1702                                 goto give_up;
1703                         unuse_pack(&w_curs);
1704                         return;
1705                 case OBJ_REF_DELTA:
1706                         if (reuse_delta && !entry->preferred_base) {
1707                                 oidread(&base_ref,
1708                                         use_pack(p, &w_curs,
1709                                                  entry->in_pack_offset + used,
1710                                                  NULL));
1711                                 have_base = 1;
1712                         }
1713                         entry->in_pack_header_size = used + the_hash_algo->rawsz;
1714                         break;
1715                 case OBJ_OFS_DELTA:
1716                         buf = use_pack(p, &w_curs,
1717                                        entry->in_pack_offset + used, NULL);
1718                         used_0 = 0;
1719                         c = buf[used_0++];
1720                         ofs = c & 127;
1721                         while (c & 128) {
1722                                 ofs += 1;
1723                                 if (!ofs || MSB(ofs, 7)) {
1724                                         error(_("delta base offset overflow in pack for %s"),
1725                                               oid_to_hex(&entry->idx.oid));
1726                                         goto give_up;
1727                                 }
1728                                 c = buf[used_0++];
1729                                 ofs = (ofs << 7) + (c & 127);
1730                         }
1731                         ofs = entry->in_pack_offset - ofs;
1732                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1733                                 error(_("delta base offset out of bound for %s"),
1734                                       oid_to_hex(&entry->idx.oid));
1735                                 goto give_up;
1736                         }
1737                         if (reuse_delta && !entry->preferred_base) {
1738                                 struct revindex_entry *revidx;
1739                                 revidx = find_pack_revindex(p, ofs);
1740                                 if (!revidx)
1741                                         goto give_up;
1742                                 if (!nth_packed_object_id(&base_ref, p, revidx->nr))
1743                                         have_base = 1;
1744                         }
1745                         entry->in_pack_header_size = used + used_0;
1746                         break;
1747                 }
1748
1749                 if (have_base &&
1750                     can_reuse_delta(&base_ref, entry, &base_entry)) {
1751                         oe_set_type(entry, entry->in_pack_type);
1752                         SET_SIZE(entry, in_pack_size); /* delta size */
1753                         SET_DELTA_SIZE(entry, in_pack_size);
1754
1755                         if (base_entry) {
1756                                 SET_DELTA(entry, base_entry);
1757                                 entry->delta_sibling_idx = base_entry->delta_child_idx;
1758                                 SET_DELTA_CHILD(base_entry, entry);
1759                         } else {
1760                                 SET_DELTA_EXT(entry, &base_ref);
1761                         }
1762
1763                         unuse_pack(&w_curs);
1764                         return;
1765                 }
1766
1767                 if (oe_type(entry)) {
1768                         off_t delta_pos;
1769
1770                         /*
1771                          * This must be a delta and we already know what the
1772                          * final object type is.  Let's extract the actual
1773                          * object size from the delta header.
1774                          */
1775                         delta_pos = entry->in_pack_offset + entry->in_pack_header_size;
1776                         canonical_size = get_size_from_delta(p, &w_curs, delta_pos);
1777                         if (canonical_size == 0)
1778                                 goto give_up;
1779                         SET_SIZE(entry, canonical_size);
1780                         unuse_pack(&w_curs);
1781                         return;
1782                 }
1783
1784                 /*
1785                  * No choice but to fall back to the recursive delta walk
1786                  * with oid_object_info() to find about the object type
1787                  * at this point...
1788                  */
1789                 give_up:
1790                 unuse_pack(&w_curs);
1791         }
1792
1793         oe_set_type(entry,
1794                     oid_object_info(the_repository, &entry->idx.oid, &canonical_size));
1795         if (entry->type_valid) {
1796                 SET_SIZE(entry, canonical_size);
1797         } else {
1798                 /*
1799                  * Bad object type is checked in prepare_pack().  This is
1800                  * to permit a missing preferred base object to be ignored
1801                  * as a preferred base.  Doing so can result in a larger
1802                  * pack file, but the transfer will still take place.
1803                  */
1804         }
1805 }
1806
1807 static int pack_offset_sort(const void *_a, const void *_b)
1808 {
1809         const struct object_entry *a = *(struct object_entry **)_a;
1810         const struct object_entry *b = *(struct object_entry **)_b;
1811         const struct packed_git *a_in_pack = IN_PACK(a);
1812         const struct packed_git *b_in_pack = IN_PACK(b);
1813
1814         /* avoid filesystem trashing with loose objects */
1815         if (!a_in_pack && !b_in_pack)
1816                 return oidcmp(&a->idx.oid, &b->idx.oid);
1817
1818         if (a_in_pack < b_in_pack)
1819                 return -1;
1820         if (a_in_pack > b_in_pack)
1821                 return 1;
1822         return a->in_pack_offset < b->in_pack_offset ? -1 :
1823                         (a->in_pack_offset > b->in_pack_offset);
1824 }
1825
1826 /*
1827  * Drop an on-disk delta we were planning to reuse. Naively, this would
1828  * just involve blanking out the "delta" field, but we have to deal
1829  * with some extra book-keeping:
1830  *
1831  *   1. Removing ourselves from the delta_sibling linked list.
1832  *
1833  *   2. Updating our size/type to the non-delta representation. These were
1834  *      either not recorded initially (size) or overwritten with the delta type
1835  *      (type) when check_object() decided to reuse the delta.
1836  *
1837  *   3. Resetting our delta depth, as we are now a base object.
1838  */
1839 static void drop_reused_delta(struct object_entry *entry)
1840 {
1841         unsigned *idx = &to_pack.objects[entry->delta_idx - 1].delta_child_idx;
1842         struct object_info oi = OBJECT_INFO_INIT;
1843         enum object_type type;
1844         unsigned long size;
1845
1846         while (*idx) {
1847                 struct object_entry *oe = &to_pack.objects[*idx - 1];
1848
1849                 if (oe == entry)
1850                         *idx = oe->delta_sibling_idx;
1851                 else
1852                         idx = &oe->delta_sibling_idx;
1853         }
1854         SET_DELTA(entry, NULL);
1855         entry->depth = 0;
1856
1857         oi.sizep = &size;
1858         oi.typep = &type;
1859         if (packed_object_info(the_repository, IN_PACK(entry), entry->in_pack_offset, &oi) < 0) {
1860                 /*
1861                  * We failed to get the info from this pack for some reason;
1862                  * fall back to oid_object_info, which may find another copy.
1863                  * And if that fails, the error will be recorded in oe_type(entry)
1864                  * and dealt with in prepare_pack().
1865                  */
1866                 oe_set_type(entry,
1867                             oid_object_info(the_repository, &entry->idx.oid, &size));
1868         } else {
1869                 oe_set_type(entry, type);
1870         }
1871         SET_SIZE(entry, size);
1872 }
1873
1874 /*
1875  * Follow the chain of deltas from this entry onward, throwing away any links
1876  * that cause us to hit a cycle (as determined by the DFS state flags in
1877  * the entries).
1878  *
1879  * We also detect too-long reused chains that would violate our --depth
1880  * limit.
1881  */
1882 static void break_delta_chains(struct object_entry *entry)
1883 {
1884         /*
1885          * The actual depth of each object we will write is stored as an int,
1886          * as it cannot exceed our int "depth" limit. But before we break
1887          * changes based no that limit, we may potentially go as deep as the
1888          * number of objects, which is elsewhere bounded to a uint32_t.
1889          */
1890         uint32_t total_depth;
1891         struct object_entry *cur, *next;
1892
1893         for (cur = entry, total_depth = 0;
1894              cur;
1895              cur = DELTA(cur), total_depth++) {
1896                 if (cur->dfs_state == DFS_DONE) {
1897                         /*
1898                          * We've already seen this object and know it isn't
1899                          * part of a cycle. We do need to append its depth
1900                          * to our count.
1901                          */
1902                         total_depth += cur->depth;
1903                         break;
1904                 }
1905
1906                 /*
1907                  * We break cycles before looping, so an ACTIVE state (or any
1908                  * other cruft which made its way into the state variable)
1909                  * is a bug.
1910                  */
1911                 if (cur->dfs_state != DFS_NONE)
1912                         BUG("confusing delta dfs state in first pass: %d",
1913                             cur->dfs_state);
1914
1915                 /*
1916                  * Now we know this is the first time we've seen the object. If
1917                  * it's not a delta, we're done traversing, but we'll mark it
1918                  * done to save time on future traversals.
1919                  */
1920                 if (!DELTA(cur)) {
1921                         cur->dfs_state = DFS_DONE;
1922                         break;
1923                 }
1924
1925                 /*
1926                  * Mark ourselves as active and see if the next step causes
1927                  * us to cycle to another active object. It's important to do
1928                  * this _before_ we loop, because it impacts where we make the
1929                  * cut, and thus how our total_depth counter works.
1930                  * E.g., We may see a partial loop like:
1931                  *
1932                  *   A -> B -> C -> D -> B
1933                  *
1934                  * Cutting B->C breaks the cycle. But now the depth of A is
1935                  * only 1, and our total_depth counter is at 3. The size of the
1936                  * error is always one less than the size of the cycle we
1937                  * broke. Commits C and D were "lost" from A's chain.
1938                  *
1939                  * If we instead cut D->B, then the depth of A is correct at 3.
1940                  * We keep all commits in the chain that we examined.
1941                  */
1942                 cur->dfs_state = DFS_ACTIVE;
1943                 if (DELTA(cur)->dfs_state == DFS_ACTIVE) {
1944                         drop_reused_delta(cur);
1945                         cur->dfs_state = DFS_DONE;
1946                         break;
1947                 }
1948         }
1949
1950         /*
1951          * And now that we've gone all the way to the bottom of the chain, we
1952          * need to clear the active flags and set the depth fields as
1953          * appropriate. Unlike the loop above, which can quit when it drops a
1954          * delta, we need to keep going to look for more depth cuts. So we need
1955          * an extra "next" pointer to keep going after we reset cur->delta.
1956          */
1957         for (cur = entry; cur; cur = next) {
1958                 next = DELTA(cur);
1959
1960                 /*
1961                  * We should have a chain of zero or more ACTIVE states down to
1962                  * a final DONE. We can quit after the DONE, because either it
1963                  * has no bases, or we've already handled them in a previous
1964                  * call.
1965                  */
1966                 if (cur->dfs_state == DFS_DONE)
1967                         break;
1968                 else if (cur->dfs_state != DFS_ACTIVE)
1969                         BUG("confusing delta dfs state in second pass: %d",
1970                             cur->dfs_state);
1971
1972                 /*
1973                  * If the total_depth is more than depth, then we need to snip
1974                  * the chain into two or more smaller chains that don't exceed
1975                  * the maximum depth. Most of the resulting chains will contain
1976                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1977                  * the last chain (i.e., the one containing entry) will contain
1978                  * whatever entries are left over, namely
1979                  * (total_depth % (depth + 1)) of them.
1980                  *
1981                  * Since we are iterating towards decreasing depth, we need to
1982                  * decrement total_depth as we go, and we need to write to the
1983                  * entry what its final depth will be after all of the
1984                  * snipping. Since we're snipping into chains of length (depth
1985                  * + 1) entries, the final depth of an entry will be its
1986                  * original depth modulo (depth + 1). Any time we encounter an
1987                  * entry whose final depth is supposed to be zero, we snip it
1988                  * from its delta base, thereby making it so.
1989                  */
1990                 cur->depth = (total_depth--) % (depth + 1);
1991                 if (!cur->depth)
1992                         drop_reused_delta(cur);
1993
1994                 cur->dfs_state = DFS_DONE;
1995         }
1996 }
1997
1998 static void get_object_details(void)
1999 {
2000         uint32_t i;
2001         struct object_entry **sorted_by_offset;
2002
2003         if (progress)
2004                 progress_state = start_progress(_("Counting objects"),
2005                                                 to_pack.nr_objects);
2006
2007         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
2008         for (i = 0; i < to_pack.nr_objects; i++)
2009                 sorted_by_offset[i] = to_pack.objects + i;
2010         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
2011
2012         for (i = 0; i < to_pack.nr_objects; i++) {
2013                 struct object_entry *entry = sorted_by_offset[i];
2014                 check_object(entry);
2015                 if (entry->type_valid &&
2016                     oe_size_greater_than(&to_pack, entry, big_file_threshold))
2017                         entry->no_try_delta = 1;
2018                 display_progress(progress_state, i + 1);
2019         }
2020         stop_progress(&progress_state);
2021
2022         /*
2023          * This must happen in a second pass, since we rely on the delta
2024          * information for the whole list being completed.
2025          */
2026         for (i = 0; i < to_pack.nr_objects; i++)
2027                 break_delta_chains(&to_pack.objects[i]);
2028
2029         free(sorted_by_offset);
2030 }
2031
2032 /*
2033  * We search for deltas in a list sorted by type, by filename hash, and then
2034  * by size, so that we see progressively smaller and smaller files.
2035  * That's because we prefer deltas to be from the bigger file
2036  * to the smaller -- deletes are potentially cheaper, but perhaps
2037  * more importantly, the bigger file is likely the more recent
2038  * one.  The deepest deltas are therefore the oldest objects which are
2039  * less susceptible to be accessed often.
2040  */
2041 static int type_size_sort(const void *_a, const void *_b)
2042 {
2043         const struct object_entry *a = *(struct object_entry **)_a;
2044         const struct object_entry *b = *(struct object_entry **)_b;
2045         const enum object_type a_type = oe_type(a);
2046         const enum object_type b_type = oe_type(b);
2047         const unsigned long a_size = SIZE(a);
2048         const unsigned long b_size = SIZE(b);
2049
2050         if (a_type > b_type)
2051                 return -1;
2052         if (a_type < b_type)
2053                 return 1;
2054         if (a->hash > b->hash)
2055                 return -1;
2056         if (a->hash < b->hash)
2057                 return 1;
2058         if (a->preferred_base > b->preferred_base)
2059                 return -1;
2060         if (a->preferred_base < b->preferred_base)
2061                 return 1;
2062         if (use_delta_islands) {
2063                 const int island_cmp = island_delta_cmp(&a->idx.oid, &b->idx.oid);
2064                 if (island_cmp)
2065                         return island_cmp;
2066         }
2067         if (a_size > b_size)
2068                 return -1;
2069         if (a_size < b_size)
2070                 return 1;
2071         return a < b ? -1 : (a > b);  /* newest first */
2072 }
2073
2074 struct unpacked {
2075         struct object_entry *entry;
2076         void *data;
2077         struct delta_index *index;
2078         unsigned depth;
2079 };
2080
2081 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
2082                            unsigned long delta_size)
2083 {
2084         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
2085                 return 0;
2086
2087         if (delta_size < cache_max_small_delta_size)
2088                 return 1;
2089
2090         /* cache delta, if objects are large enough compared to delta size */
2091         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
2092                 return 1;
2093
2094         return 0;
2095 }
2096
2097 /* Protect delta_cache_size */
2098 static pthread_mutex_t cache_mutex;
2099 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
2100 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
2101
2102 /*
2103  * Protect object list partitioning (e.g. struct thread_param) and
2104  * progress_state
2105  */
2106 static pthread_mutex_t progress_mutex;
2107 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
2108 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
2109
2110 /*
2111  * Access to struct object_entry is unprotected since each thread owns
2112  * a portion of the main object list. Just don't access object entries
2113  * ahead in the list because they can be stolen and would need
2114  * progress_mutex for protection.
2115  */
2116
2117 /*
2118  * Return the size of the object without doing any delta
2119  * reconstruction (so non-deltas are true object sizes, but deltas
2120  * return the size of the delta data).
2121  */
2122 unsigned long oe_get_size_slow(struct packing_data *pack,
2123                                const struct object_entry *e)
2124 {
2125         struct packed_git *p;
2126         struct pack_window *w_curs;
2127         unsigned char *buf;
2128         enum object_type type;
2129         unsigned long used, avail, size;
2130
2131         if (e->type_ != OBJ_OFS_DELTA && e->type_ != OBJ_REF_DELTA) {
2132                 packing_data_lock(&to_pack);
2133                 if (oid_object_info(the_repository, &e->idx.oid, &size) < 0)
2134                         die(_("unable to get size of %s"),
2135                             oid_to_hex(&e->idx.oid));
2136                 packing_data_unlock(&to_pack);
2137                 return size;
2138         }
2139
2140         p = oe_in_pack(pack, e);
2141         if (!p)
2142                 BUG("when e->type is a delta, it must belong to a pack");
2143
2144         packing_data_lock(&to_pack);
2145         w_curs = NULL;
2146         buf = use_pack(p, &w_curs, e->in_pack_offset, &avail);
2147         used = unpack_object_header_buffer(buf, avail, &type, &size);
2148         if (used == 0)
2149                 die(_("unable to parse object header of %s"),
2150                     oid_to_hex(&e->idx.oid));
2151
2152         unuse_pack(&w_curs);
2153         packing_data_unlock(&to_pack);
2154         return size;
2155 }
2156
2157 static int try_delta(struct unpacked *trg, struct unpacked *src,
2158                      unsigned max_depth, unsigned long *mem_usage)
2159 {
2160         struct object_entry *trg_entry = trg->entry;
2161         struct object_entry *src_entry = src->entry;
2162         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
2163         unsigned ref_depth;
2164         enum object_type type;
2165         void *delta_buf;
2166
2167         /* Don't bother doing diffs between different types */
2168         if (oe_type(trg_entry) != oe_type(src_entry))
2169                 return -1;
2170
2171         /*
2172          * We do not bother to try a delta that we discarded on an
2173          * earlier try, but only when reusing delta data.  Note that
2174          * src_entry that is marked as the preferred_base should always
2175          * be considered, as even if we produce a suboptimal delta against
2176          * it, we will still save the transfer cost, as we already know
2177          * the other side has it and we won't send src_entry at all.
2178          */
2179         if (reuse_delta && IN_PACK(trg_entry) &&
2180             IN_PACK(trg_entry) == IN_PACK(src_entry) &&
2181             !src_entry->preferred_base &&
2182             trg_entry->in_pack_type != OBJ_REF_DELTA &&
2183             trg_entry->in_pack_type != OBJ_OFS_DELTA)
2184                 return 0;
2185
2186         /* Let's not bust the allowed depth. */
2187         if (src->depth >= max_depth)
2188                 return 0;
2189
2190         /* Now some size filtering heuristics. */
2191         trg_size = SIZE(trg_entry);
2192         if (!DELTA(trg_entry)) {
2193                 max_size = trg_size/2 - the_hash_algo->rawsz;
2194                 ref_depth = 1;
2195         } else {
2196                 max_size = DELTA_SIZE(trg_entry);
2197                 ref_depth = trg->depth;
2198         }
2199         max_size = (uint64_t)max_size * (max_depth - src->depth) /
2200                                                 (max_depth - ref_depth + 1);
2201         if (max_size == 0)
2202                 return 0;
2203         src_size = SIZE(src_entry);
2204         sizediff = src_size < trg_size ? trg_size - src_size : 0;
2205         if (sizediff >= max_size)
2206                 return 0;
2207         if (trg_size < src_size / 32)
2208                 return 0;
2209
2210         if (!in_same_island(&trg->entry->idx.oid, &src->entry->idx.oid))
2211                 return 0;
2212
2213         /* Load data if not already done */
2214         if (!trg->data) {
2215                 packing_data_lock(&to_pack);
2216                 trg->data = read_object_file(&trg_entry->idx.oid, &type, &sz);
2217                 packing_data_unlock(&to_pack);
2218                 if (!trg->data)
2219                         die(_("object %s cannot be read"),
2220                             oid_to_hex(&trg_entry->idx.oid));
2221                 if (sz != trg_size)
2222                         die(_("object %s inconsistent object length (%"PRIuMAX" vs %"PRIuMAX")"),
2223                             oid_to_hex(&trg_entry->idx.oid), (uintmax_t)sz,
2224                             (uintmax_t)trg_size);
2225                 *mem_usage += sz;
2226         }
2227         if (!src->data) {
2228                 packing_data_lock(&to_pack);
2229                 src->data = read_object_file(&src_entry->idx.oid, &type, &sz);
2230                 packing_data_unlock(&to_pack);
2231                 if (!src->data) {
2232                         if (src_entry->preferred_base) {
2233                                 static int warned = 0;
2234                                 if (!warned++)
2235                                         warning(_("object %s cannot be read"),
2236                                                 oid_to_hex(&src_entry->idx.oid));
2237                                 /*
2238                                  * Those objects are not included in the
2239                                  * resulting pack.  Be resilient and ignore
2240                                  * them if they can't be read, in case the
2241                                  * pack could be created nevertheless.
2242                                  */
2243                                 return 0;
2244                         }
2245                         die(_("object %s cannot be read"),
2246                             oid_to_hex(&src_entry->idx.oid));
2247                 }
2248                 if (sz != src_size)
2249                         die(_("object %s inconsistent object length (%"PRIuMAX" vs %"PRIuMAX")"),
2250                             oid_to_hex(&src_entry->idx.oid), (uintmax_t)sz,
2251                             (uintmax_t)src_size);
2252                 *mem_usage += sz;
2253         }
2254         if (!src->index) {
2255                 src->index = create_delta_index(src->data, src_size);
2256                 if (!src->index) {
2257                         static int warned = 0;
2258                         if (!warned++)
2259                                 warning(_("suboptimal pack - out of memory"));
2260                         return 0;
2261                 }
2262                 *mem_usage += sizeof_delta_index(src->index);
2263         }
2264
2265         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
2266         if (!delta_buf)
2267                 return 0;
2268
2269         if (DELTA(trg_entry)) {
2270                 /* Prefer only shallower same-sized deltas. */
2271                 if (delta_size == DELTA_SIZE(trg_entry) &&
2272                     src->depth + 1 >= trg->depth) {
2273                         free(delta_buf);
2274                         return 0;
2275                 }
2276         }
2277
2278         /*
2279          * Handle memory allocation outside of the cache
2280          * accounting lock.  Compiler will optimize the strangeness
2281          * away when NO_PTHREADS is defined.
2282          */
2283         free(trg_entry->delta_data);
2284         cache_lock();
2285         if (trg_entry->delta_data) {
2286                 delta_cache_size -= DELTA_SIZE(trg_entry);
2287                 trg_entry->delta_data = NULL;
2288         }
2289         if (delta_cacheable(src_size, trg_size, delta_size)) {
2290                 delta_cache_size += delta_size;
2291                 cache_unlock();
2292                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
2293         } else {
2294                 cache_unlock();
2295                 free(delta_buf);
2296         }
2297
2298         SET_DELTA(trg_entry, src_entry);
2299         SET_DELTA_SIZE(trg_entry, delta_size);
2300         trg->depth = src->depth + 1;
2301
2302         return 1;
2303 }
2304
2305 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
2306 {
2307         struct object_entry *child = DELTA_CHILD(me);
2308         unsigned int m = n;
2309         while (child) {
2310                 const unsigned int c = check_delta_limit(child, n + 1);
2311                 if (m < c)
2312                         m = c;
2313                 child = DELTA_SIBLING(child);
2314         }
2315         return m;
2316 }
2317
2318 static unsigned long free_unpacked(struct unpacked *n)
2319 {
2320         unsigned long freed_mem = sizeof_delta_index(n->index);
2321         free_delta_index(n->index);
2322         n->index = NULL;
2323         if (n->data) {
2324                 freed_mem += SIZE(n->entry);
2325                 FREE_AND_NULL(n->data);
2326         }
2327         n->entry = NULL;
2328         n->depth = 0;
2329         return freed_mem;
2330 }
2331
2332 static void find_deltas(struct object_entry **list, unsigned *list_size,
2333                         int window, int depth, unsigned *processed)
2334 {
2335         uint32_t i, idx = 0, count = 0;
2336         struct unpacked *array;
2337         unsigned long mem_usage = 0;
2338
2339         array = xcalloc(window, sizeof(struct unpacked));
2340
2341         for (;;) {
2342                 struct object_entry *entry;
2343                 struct unpacked *n = array + idx;
2344                 int j, max_depth, best_base = -1;
2345
2346                 progress_lock();
2347                 if (!*list_size) {
2348                         progress_unlock();
2349                         break;
2350                 }
2351                 entry = *list++;
2352                 (*list_size)--;
2353                 if (!entry->preferred_base) {
2354                         (*processed)++;
2355                         display_progress(progress_state, *processed);
2356                 }
2357                 progress_unlock();
2358
2359                 mem_usage -= free_unpacked(n);
2360                 n->entry = entry;
2361
2362                 while (window_memory_limit &&
2363                        mem_usage > window_memory_limit &&
2364                        count > 1) {
2365                         const uint32_t tail = (idx + window - count) % window;
2366                         mem_usage -= free_unpacked(array + tail);
2367                         count--;
2368                 }
2369
2370                 /* We do not compute delta to *create* objects we are not
2371                  * going to pack.
2372                  */
2373                 if (entry->preferred_base)
2374                         goto next;
2375
2376                 /*
2377                  * If the current object is at pack edge, take the depth the
2378                  * objects that depend on the current object into account
2379                  * otherwise they would become too deep.
2380                  */
2381                 max_depth = depth;
2382                 if (DELTA_CHILD(entry)) {
2383                         max_depth -= check_delta_limit(entry, 0);
2384                         if (max_depth <= 0)
2385                                 goto next;
2386                 }
2387
2388                 j = window;
2389                 while (--j > 0) {
2390                         int ret;
2391                         uint32_t other_idx = idx + j;
2392                         struct unpacked *m;
2393                         if (other_idx >= window)
2394                                 other_idx -= window;
2395                         m = array + other_idx;
2396                         if (!m->entry)
2397                                 break;
2398                         ret = try_delta(n, m, max_depth, &mem_usage);
2399                         if (ret < 0)
2400                                 break;
2401                         else if (ret > 0)
2402                                 best_base = other_idx;
2403                 }
2404
2405                 /*
2406                  * If we decided to cache the delta data, then it is best
2407                  * to compress it right away.  First because we have to do
2408                  * it anyway, and doing it here while we're threaded will
2409                  * save a lot of time in the non threaded write phase,
2410                  * as well as allow for caching more deltas within
2411                  * the same cache size limit.
2412                  * ...
2413                  * But only if not writing to stdout, since in that case
2414                  * the network is most likely throttling writes anyway,
2415                  * and therefore it is best to go to the write phase ASAP
2416                  * instead, as we can afford spending more time compressing
2417                  * between writes at that moment.
2418                  */
2419                 if (entry->delta_data && !pack_to_stdout) {
2420                         unsigned long size;
2421
2422                         size = do_compress(&entry->delta_data, DELTA_SIZE(entry));
2423                         if (size < (1U << OE_Z_DELTA_BITS)) {
2424                                 entry->z_delta_size = size;
2425                                 cache_lock();
2426                                 delta_cache_size -= DELTA_SIZE(entry);
2427                                 delta_cache_size += entry->z_delta_size;
2428                                 cache_unlock();
2429                         } else {
2430                                 FREE_AND_NULL(entry->delta_data);
2431                                 entry->z_delta_size = 0;
2432                         }
2433                 }
2434
2435                 /* if we made n a delta, and if n is already at max
2436                  * depth, leaving it in the window is pointless.  we
2437                  * should evict it first.
2438                  */
2439                 if (DELTA(entry) && max_depth <= n->depth)
2440                         continue;
2441
2442                 /*
2443                  * Move the best delta base up in the window, after the
2444                  * currently deltified object, to keep it longer.  It will
2445                  * be the first base object to be attempted next.
2446                  */
2447                 if (DELTA(entry)) {
2448                         struct unpacked swap = array[best_base];
2449                         int dist = (window + idx - best_base) % window;
2450                         int dst = best_base;
2451                         while (dist--) {
2452                                 int src = (dst + 1) % window;
2453                                 array[dst] = array[src];
2454                                 dst = src;
2455                         }
2456                         array[dst] = swap;
2457                 }
2458
2459                 next:
2460                 idx++;
2461                 if (count + 1 < window)
2462                         count++;
2463                 if (idx >= window)
2464                         idx = 0;
2465         }
2466
2467         for (i = 0; i < window; ++i) {
2468                 free_delta_index(array[i].index);
2469                 free(array[i].data);
2470         }
2471         free(array);
2472 }
2473
2474 /*
2475  * The main object list is split into smaller lists, each is handed to
2476  * one worker.
2477  *
2478  * The main thread waits on the condition that (at least) one of the workers
2479  * has stopped working (which is indicated in the .working member of
2480  * struct thread_params).
2481  *
2482  * When a work thread has completed its work, it sets .working to 0 and
2483  * signals the main thread and waits on the condition that .data_ready
2484  * becomes 1.
2485  *
2486  * The main thread steals half of the work from the worker that has
2487  * most work left to hand it to the idle worker.
2488  */
2489
2490 struct thread_params {
2491         pthread_t thread;
2492         struct object_entry **list;
2493         unsigned list_size;
2494         unsigned remaining;
2495         int window;
2496         int depth;
2497         int working;
2498         int data_ready;
2499         pthread_mutex_t mutex;
2500         pthread_cond_t cond;
2501         unsigned *processed;
2502 };
2503
2504 static pthread_cond_t progress_cond;
2505
2506 /*
2507  * Mutex and conditional variable can't be statically-initialized on Windows.
2508  */
2509 static void init_threaded_search(void)
2510 {
2511         pthread_mutex_init(&cache_mutex, NULL);
2512         pthread_mutex_init(&progress_mutex, NULL);
2513         pthread_cond_init(&progress_cond, NULL);
2514 }
2515
2516 static void cleanup_threaded_search(void)
2517 {
2518         pthread_cond_destroy(&progress_cond);
2519         pthread_mutex_destroy(&cache_mutex);
2520         pthread_mutex_destroy(&progress_mutex);
2521 }
2522
2523 static void *threaded_find_deltas(void *arg)
2524 {
2525         struct thread_params *me = arg;
2526
2527         progress_lock();
2528         while (me->remaining) {
2529                 progress_unlock();
2530
2531                 find_deltas(me->list, &me->remaining,
2532                             me->window, me->depth, me->processed);
2533
2534                 progress_lock();
2535                 me->working = 0;
2536                 pthread_cond_signal(&progress_cond);
2537                 progress_unlock();
2538
2539                 /*
2540                  * We must not set ->data_ready before we wait on the
2541                  * condition because the main thread may have set it to 1
2542                  * before we get here. In order to be sure that new
2543                  * work is available if we see 1 in ->data_ready, it
2544                  * was initialized to 0 before this thread was spawned
2545                  * and we reset it to 0 right away.
2546                  */
2547                 pthread_mutex_lock(&me->mutex);
2548                 while (!me->data_ready)
2549                         pthread_cond_wait(&me->cond, &me->mutex);
2550                 me->data_ready = 0;
2551                 pthread_mutex_unlock(&me->mutex);
2552
2553                 progress_lock();
2554         }
2555         progress_unlock();
2556         /* leave ->working 1 so that this doesn't get more work assigned */
2557         return NULL;
2558 }
2559
2560 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2561                            int window, int depth, unsigned *processed)
2562 {
2563         struct thread_params *p;
2564         int i, ret, active_threads = 0;
2565
2566         init_threaded_search();
2567
2568         if (delta_search_threads <= 1) {
2569                 find_deltas(list, &list_size, window, depth, processed);
2570                 cleanup_threaded_search();
2571                 return;
2572         }
2573         if (progress > pack_to_stdout)
2574                 fprintf_ln(stderr, _("Delta compression using up to %d threads"),
2575                            delta_search_threads);
2576         p = xcalloc(delta_search_threads, sizeof(*p));
2577
2578         /* Partition the work amongst work threads. */
2579         for (i = 0; i < delta_search_threads; i++) {
2580                 unsigned sub_size = list_size / (delta_search_threads - i);
2581
2582                 /* don't use too small segments or no deltas will be found */
2583                 if (sub_size < 2*window && i+1 < delta_search_threads)
2584                         sub_size = 0;
2585
2586                 p[i].window = window;
2587                 p[i].depth = depth;
2588                 p[i].processed = processed;
2589                 p[i].working = 1;
2590                 p[i].data_ready = 0;
2591
2592                 /* try to split chunks on "path" boundaries */
2593                 while (sub_size && sub_size < list_size &&
2594                        list[sub_size]->hash &&
2595                        list[sub_size]->hash == list[sub_size-1]->hash)
2596                         sub_size++;
2597
2598                 p[i].list = list;
2599                 p[i].list_size = sub_size;
2600                 p[i].remaining = sub_size;
2601
2602                 list += sub_size;
2603                 list_size -= sub_size;
2604         }
2605
2606         /* Start work threads. */
2607         for (i = 0; i < delta_search_threads; i++) {
2608                 if (!p[i].list_size)
2609                         continue;
2610                 pthread_mutex_init(&p[i].mutex, NULL);
2611                 pthread_cond_init(&p[i].cond, NULL);
2612                 ret = pthread_create(&p[i].thread, NULL,
2613                                      threaded_find_deltas, &p[i]);
2614                 if (ret)
2615                         die(_("unable to create thread: %s"), strerror(ret));
2616                 active_threads++;
2617         }
2618
2619         /*
2620          * Now let's wait for work completion.  Each time a thread is done
2621          * with its work, we steal half of the remaining work from the
2622          * thread with the largest number of unprocessed objects and give
2623          * it to that newly idle thread.  This ensure good load balancing
2624          * until the remaining object list segments are simply too short
2625          * to be worth splitting anymore.
2626          */
2627         while (active_threads) {
2628                 struct thread_params *target = NULL;
2629                 struct thread_params *victim = NULL;
2630                 unsigned sub_size = 0;
2631
2632                 progress_lock();
2633                 for (;;) {
2634                         for (i = 0; !target && i < delta_search_threads; i++)
2635                                 if (!p[i].working)
2636                                         target = &p[i];
2637                         if (target)
2638                                 break;
2639                         pthread_cond_wait(&progress_cond, &progress_mutex);
2640                 }
2641
2642                 for (i = 0; i < delta_search_threads; i++)
2643                         if (p[i].remaining > 2*window &&
2644                             (!victim || victim->remaining < p[i].remaining))
2645                                 victim = &p[i];
2646                 if (victim) {
2647                         sub_size = victim->remaining / 2;
2648                         list = victim->list + victim->list_size - sub_size;
2649                         while (sub_size && list[0]->hash &&
2650                                list[0]->hash == list[-1]->hash) {
2651                                 list++;
2652                                 sub_size--;
2653                         }
2654                         if (!sub_size) {
2655                                 /*
2656                                  * It is possible for some "paths" to have
2657                                  * so many objects that no hash boundary
2658                                  * might be found.  Let's just steal the
2659                                  * exact half in that case.
2660                                  */
2661                                 sub_size = victim->remaining / 2;
2662                                 list -= sub_size;
2663                         }
2664                         target->list = list;
2665                         victim->list_size -= sub_size;
2666                         victim->remaining -= sub_size;
2667                 }
2668                 target->list_size = sub_size;
2669                 target->remaining = sub_size;
2670                 target->working = 1;
2671                 progress_unlock();
2672
2673                 pthread_mutex_lock(&target->mutex);
2674                 target->data_ready = 1;
2675                 pthread_cond_signal(&target->cond);
2676                 pthread_mutex_unlock(&target->mutex);
2677
2678                 if (!sub_size) {
2679                         pthread_join(target->thread, NULL);
2680                         pthread_cond_destroy(&target->cond);
2681                         pthread_mutex_destroy(&target->mutex);
2682                         active_threads--;
2683                 }
2684         }
2685         cleanup_threaded_search();
2686         free(p);
2687 }
2688
2689 static int obj_is_packed(const struct object_id *oid)
2690 {
2691         return packlist_find(&to_pack, oid) ||
2692                 (reuse_packfile_bitmap &&
2693                  bitmap_walk_contains(bitmap_git, reuse_packfile_bitmap, oid));
2694 }
2695
2696 static void add_tag_chain(const struct object_id *oid)
2697 {
2698         struct tag *tag;
2699
2700         /*
2701          * We catch duplicates already in add_object_entry(), but we'd
2702          * prefer to do this extra check to avoid having to parse the
2703          * tag at all if we already know that it's being packed (e.g., if
2704          * it was included via bitmaps, we would not have parsed it
2705          * previously).
2706          */
2707         if (obj_is_packed(oid))
2708                 return;
2709
2710         tag = lookup_tag(the_repository, oid);
2711         while (1) {
2712                 if (!tag || parse_tag(tag) || !tag->tagged)
2713                         die(_("unable to pack objects reachable from tag %s"),
2714                             oid_to_hex(oid));
2715
2716                 add_object_entry(&tag->object.oid, OBJ_TAG, NULL, 0);
2717
2718                 if (tag->tagged->type != OBJ_TAG)
2719                         return;
2720
2721                 tag = (struct tag *)tag->tagged;
2722         }
2723 }
2724
2725 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2726 {
2727         struct object_id peeled;
2728
2729         if (starts_with(path, "refs/tags/") && /* is a tag? */
2730             !peel_ref(path, &peeled)    && /* peelable? */
2731             obj_is_packed(&peeled)) /* object packed? */
2732                 add_tag_chain(oid);
2733         return 0;
2734 }
2735
2736 static void prepare_pack(int window, int depth)
2737 {
2738         struct object_entry **delta_list;
2739         uint32_t i, nr_deltas;
2740         unsigned n;
2741
2742         if (use_delta_islands)
2743                 resolve_tree_islands(the_repository, progress, &to_pack);
2744
2745         get_object_details();
2746
2747         /*
2748          * If we're locally repacking then we need to be doubly careful
2749          * from now on in order to make sure no stealth corruption gets
2750          * propagated to the new pack.  Clients receiving streamed packs
2751          * should validate everything they get anyway so no need to incur
2752          * the additional cost here in that case.
2753          */
2754         if (!pack_to_stdout)
2755                 do_check_packed_object_crc = 1;
2756
2757         if (!to_pack.nr_objects || !window || !depth)
2758                 return;
2759
2760         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2761         nr_deltas = n = 0;
2762
2763         for (i = 0; i < to_pack.nr_objects; i++) {
2764                 struct object_entry *entry = to_pack.objects + i;
2765
2766                 if (DELTA(entry))
2767                         /* This happens if we decided to reuse existing
2768                          * delta from a pack.  "reuse_delta &&" is implied.
2769                          */
2770                         continue;
2771
2772                 if (!entry->type_valid ||
2773                     oe_size_less_than(&to_pack, entry, 50))
2774                         continue;
2775
2776                 if (entry->no_try_delta)
2777                         continue;
2778
2779                 if (!entry->preferred_base) {
2780                         nr_deltas++;
2781                         if (oe_type(entry) < 0)
2782                                 die(_("unable to get type of object %s"),
2783                                     oid_to_hex(&entry->idx.oid));
2784                 } else {
2785                         if (oe_type(entry) < 0) {
2786                                 /*
2787                                  * This object is not found, but we
2788                                  * don't have to include it anyway.
2789                                  */
2790                                 continue;
2791                         }
2792                 }
2793
2794                 delta_list[n++] = entry;
2795         }
2796
2797         if (nr_deltas && n > 1) {
2798                 unsigned nr_done = 0;
2799
2800                 if (progress)
2801                         progress_state = start_progress(_("Compressing objects"),
2802                                                         nr_deltas);
2803                 QSORT(delta_list, n, type_size_sort);
2804                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2805                 stop_progress(&progress_state);
2806                 if (nr_done != nr_deltas)
2807                         die(_("inconsistency with delta count"));
2808         }
2809         free(delta_list);
2810 }
2811
2812 static int git_pack_config(const char *k, const char *v, void *cb)
2813 {
2814         if (!strcmp(k, "pack.window")) {
2815                 window = git_config_int(k, v);
2816                 return 0;
2817         }
2818         if (!strcmp(k, "pack.windowmemory")) {
2819                 window_memory_limit = git_config_ulong(k, v);
2820                 return 0;
2821         }
2822         if (!strcmp(k, "pack.depth")) {
2823                 depth = git_config_int(k, v);
2824                 return 0;
2825         }
2826         if (!strcmp(k, "pack.deltacachesize")) {
2827                 max_delta_cache_size = git_config_int(k, v);
2828                 return 0;
2829         }
2830         if (!strcmp(k, "pack.deltacachelimit")) {
2831                 cache_max_small_delta_size = git_config_int(k, v);
2832                 return 0;
2833         }
2834         if (!strcmp(k, "pack.writebitmaphashcache")) {
2835                 if (git_config_bool(k, v))
2836                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2837                 else
2838                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2839         }
2840         if (!strcmp(k, "pack.usebitmaps")) {
2841                 use_bitmap_index_default = git_config_bool(k, v);
2842                 return 0;
2843         }
2844         if (!strcmp(k, "pack.allowpackreuse")) {
2845                 allow_pack_reuse = git_config_bool(k, v);
2846                 return 0;
2847         }
2848         if (!strcmp(k, "pack.threads")) {
2849                 delta_search_threads = git_config_int(k, v);
2850                 if (delta_search_threads < 0)
2851                         die(_("invalid number of threads specified (%d)"),
2852                             delta_search_threads);
2853                 if (!HAVE_THREADS && delta_search_threads != 1) {
2854                         warning(_("no threads support, ignoring %s"), k);
2855                         delta_search_threads = 0;
2856                 }
2857                 return 0;
2858         }
2859         if (!strcmp(k, "pack.indexversion")) {
2860                 pack_idx_opts.version = git_config_int(k, v);
2861                 if (pack_idx_opts.version > 2)
2862                         die(_("bad pack.indexversion=%"PRIu32),
2863                             pack_idx_opts.version);
2864                 return 0;
2865         }
2866         return git_default_config(k, v, cb);
2867 }
2868
2869 static void read_object_list_from_stdin(void)
2870 {
2871         char line[GIT_MAX_HEXSZ + 1 + PATH_MAX + 2];
2872         struct object_id oid;
2873         const char *p;
2874
2875         for (;;) {
2876                 if (!fgets(line, sizeof(line), stdin)) {
2877                         if (feof(stdin))
2878                                 break;
2879                         if (!ferror(stdin))
2880                                 die("BUG: fgets returned NULL, not EOF, not error!");
2881                         if (errno != EINTR)
2882                                 die_errno("fgets");
2883                         clearerr(stdin);
2884                         continue;
2885                 }
2886                 if (line[0] == '-') {
2887                         if (get_oid_hex(line+1, &oid))
2888                                 die(_("expected edge object ID, got garbage:\n %s"),
2889                                     line);
2890                         add_preferred_base(&oid);
2891                         continue;
2892                 }
2893                 if (parse_oid_hex(line, &oid, &p))
2894                         die(_("expected object ID, got garbage:\n %s"), line);
2895
2896                 add_preferred_base_object(p + 1);
2897                 add_object_entry(&oid, OBJ_NONE, p + 1, 0);
2898         }
2899 }
2900
2901 /* Remember to update object flag allocation in object.h */
2902 #define OBJECT_ADDED (1u<<20)
2903
2904 static void show_commit(struct commit *commit, void *data)
2905 {
2906         add_object_entry(&commit->object.oid, OBJ_COMMIT, NULL, 0);
2907         commit->object.flags |= OBJECT_ADDED;
2908
2909         if (write_bitmap_index)
2910                 index_commit_for_bitmap(commit);
2911
2912         if (use_delta_islands)
2913                 propagate_island_marks(commit);
2914 }
2915
2916 static void show_object(struct object *obj, const char *name, void *data)
2917 {
2918         add_preferred_base_object(name);
2919         add_object_entry(&obj->oid, obj->type, name, 0);
2920         obj->flags |= OBJECT_ADDED;
2921
2922         if (use_delta_islands) {
2923                 const char *p;
2924                 unsigned depth;
2925                 struct object_entry *ent;
2926
2927                 /* the empty string is a root tree, which is depth 0 */
2928                 depth = *name ? 1 : 0;
2929                 for (p = strchr(name, '/'); p; p = strchr(p + 1, '/'))
2930                         depth++;
2931
2932                 ent = packlist_find(&to_pack, &obj->oid);
2933                 if (ent && depth > oe_tree_depth(&to_pack, ent))
2934                         oe_set_tree_depth(&to_pack, ent, depth);
2935         }
2936 }
2937
2938 static void show_object__ma_allow_any(struct object *obj, const char *name, void *data)
2939 {
2940         assert(arg_missing_action == MA_ALLOW_ANY);
2941
2942         /*
2943          * Quietly ignore ALL missing objects.  This avoids problems with
2944          * staging them now and getting an odd error later.
2945          */
2946         if (!has_object_file(&obj->oid))
2947                 return;
2948
2949         show_object(obj, name, data);
2950 }
2951
2952 static void show_object__ma_allow_promisor(struct object *obj, const char *name, void *data)
2953 {
2954         assert(arg_missing_action == MA_ALLOW_PROMISOR);
2955
2956         /*
2957          * Quietly ignore EXPECTED missing objects.  This avoids problems with
2958          * staging them now and getting an odd error later.
2959          */
2960         if (!has_object_file(&obj->oid) && is_promisor_object(&obj->oid))
2961                 return;
2962
2963         show_object(obj, name, data);
2964 }
2965
2966 static int option_parse_missing_action(const struct option *opt,
2967                                        const char *arg, int unset)
2968 {
2969         assert(arg);
2970         assert(!unset);
2971
2972         if (!strcmp(arg, "error")) {
2973                 arg_missing_action = MA_ERROR;
2974                 fn_show_object = show_object;
2975                 return 0;
2976         }
2977
2978         if (!strcmp(arg, "allow-any")) {
2979                 arg_missing_action = MA_ALLOW_ANY;
2980                 fetch_if_missing = 0;
2981                 fn_show_object = show_object__ma_allow_any;
2982                 return 0;
2983         }
2984
2985         if (!strcmp(arg, "allow-promisor")) {
2986                 arg_missing_action = MA_ALLOW_PROMISOR;
2987                 fetch_if_missing = 0;
2988                 fn_show_object = show_object__ma_allow_promisor;
2989                 return 0;
2990         }
2991
2992         die(_("invalid value for --missing"));
2993         return 0;
2994 }
2995
2996 static void show_edge(struct commit *commit)
2997 {
2998         add_preferred_base(&commit->object.oid);
2999 }
3000
3001 struct in_pack_object {
3002         off_t offset;
3003         struct object *object;
3004 };
3005
3006 struct in_pack {
3007         unsigned int alloc;
3008         unsigned int nr;
3009         struct in_pack_object *array;
3010 };
3011
3012 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
3013 {
3014         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
3015         in_pack->array[in_pack->nr].object = object;
3016         in_pack->nr++;
3017 }
3018
3019 /*
3020  * Compare the objects in the offset order, in order to emulate the
3021  * "git rev-list --objects" output that produced the pack originally.
3022  */
3023 static int ofscmp(const void *a_, const void *b_)
3024 {
3025         struct in_pack_object *a = (struct in_pack_object *)a_;
3026         struct in_pack_object *b = (struct in_pack_object *)b_;
3027
3028         if (a->offset < b->offset)
3029                 return -1;
3030         else if (a->offset > b->offset)
3031                 return 1;
3032         else
3033                 return oidcmp(&a->object->oid, &b->object->oid);
3034 }
3035
3036 static void add_objects_in_unpacked_packs(void)
3037 {
3038         struct packed_git *p;
3039         struct in_pack in_pack;
3040         uint32_t i;
3041
3042         memset(&in_pack, 0, sizeof(in_pack));
3043
3044         for (p = get_all_packs(the_repository); p; p = p->next) {
3045                 struct object_id oid;
3046                 struct object *o;
3047
3048                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
3049                         continue;
3050                 if (open_pack_index(p))
3051                         die(_("cannot open pack index"));
3052
3053                 ALLOC_GROW(in_pack.array,
3054                            in_pack.nr + p->num_objects,
3055                            in_pack.alloc);
3056
3057                 for (i = 0; i < p->num_objects; i++) {
3058                         nth_packed_object_id(&oid, p, i);
3059                         o = lookup_unknown_object(&oid);
3060                         if (!(o->flags & OBJECT_ADDED))
3061                                 mark_in_pack_object(o, p, &in_pack);
3062                         o->flags |= OBJECT_ADDED;
3063                 }
3064         }
3065
3066         if (in_pack.nr) {
3067                 QSORT(in_pack.array, in_pack.nr, ofscmp);
3068                 for (i = 0; i < in_pack.nr; i++) {
3069                         struct object *o = in_pack.array[i].object;
3070                         add_object_entry(&o->oid, o->type, "", 0);
3071                 }
3072         }
3073         free(in_pack.array);
3074 }
3075
3076 static int add_loose_object(const struct object_id *oid, const char *path,
3077                             void *data)
3078 {
3079         enum object_type type = oid_object_info(the_repository, oid, NULL);
3080
3081         if (type < 0) {
3082                 warning(_("loose object at %s could not be examined"), path);
3083                 return 0;
3084         }
3085
3086         add_object_entry(oid, type, "", 0);
3087         return 0;
3088 }
3089
3090 /*
3091  * We actually don't even have to worry about reachability here.
3092  * add_object_entry will weed out duplicates, so we just add every
3093  * loose object we find.
3094  */
3095 static void add_unreachable_loose_objects(void)
3096 {
3097         for_each_loose_file_in_objdir(get_object_directory(),
3098                                       add_loose_object,
3099                                       NULL, NULL, NULL);
3100 }
3101
3102 static int has_sha1_pack_kept_or_nonlocal(const struct object_id *oid)
3103 {
3104         static struct packed_git *last_found = (void *)1;
3105         struct packed_git *p;
3106
3107         p = (last_found != (void *)1) ? last_found :
3108                                         get_all_packs(the_repository);
3109
3110         while (p) {
3111                 if ((!p->pack_local || p->pack_keep ||
3112                                 p->pack_keep_in_core) &&
3113                         find_pack_entry_one(oid->hash, p)) {
3114                         last_found = p;
3115                         return 1;
3116                 }
3117                 if (p == last_found)
3118                         p = get_all_packs(the_repository);
3119                 else
3120                         p = p->next;
3121                 if (p == last_found)
3122                         p = p->next;
3123         }
3124         return 0;
3125 }
3126
3127 /*
3128  * Store a list of sha1s that are should not be discarded
3129  * because they are either written too recently, or are
3130  * reachable from another object that was.
3131  *
3132  * This is filled by get_object_list.
3133  */
3134 static struct oid_array recent_objects;
3135
3136 static int loosened_object_can_be_discarded(const struct object_id *oid,
3137                                             timestamp_t mtime)
3138 {
3139         if (!unpack_unreachable_expiration)
3140                 return 0;
3141         if (mtime > unpack_unreachable_expiration)
3142                 return 0;
3143         if (oid_array_lookup(&recent_objects, oid) >= 0)
3144                 return 0;
3145         return 1;
3146 }
3147
3148 static void loosen_unused_packed_objects(void)
3149 {
3150         struct packed_git *p;
3151         uint32_t i;
3152         struct object_id oid;
3153
3154         for (p = get_all_packs(the_repository); p; p = p->next) {
3155                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
3156                         continue;
3157
3158                 if (open_pack_index(p))
3159                         die(_("cannot open pack index"));
3160
3161                 for (i = 0; i < p->num_objects; i++) {
3162                         nth_packed_object_id(&oid, p, i);
3163                         if (!packlist_find(&to_pack, &oid) &&
3164                             !has_sha1_pack_kept_or_nonlocal(&oid) &&
3165                             !loosened_object_can_be_discarded(&oid, p->mtime))
3166                                 if (force_object_loose(&oid, p->mtime))
3167                                         die(_("unable to force loose object"));
3168                 }
3169         }
3170 }
3171
3172 /*
3173  * This tracks any options which pack-reuse code expects to be on, or which a
3174  * reader of the pack might not understand, and which would therefore prevent
3175  * blind reuse of what we have on disk.
3176  */
3177 static int pack_options_allow_reuse(void)
3178 {
3179         return allow_pack_reuse &&
3180                pack_to_stdout &&
3181                !ignore_packed_keep_on_disk &&
3182                !ignore_packed_keep_in_core &&
3183                (!local || !have_non_local_packs) &&
3184                !incremental;
3185 }
3186
3187 static int get_object_list_from_bitmap(struct rev_info *revs)
3188 {
3189         if (!(bitmap_git = prepare_bitmap_walk(revs, &filter_options)))
3190                 return -1;
3191
3192         if (pack_options_allow_reuse() &&
3193             !reuse_partial_packfile_from_bitmap(
3194                         bitmap_git,
3195                         &reuse_packfile,
3196                         &reuse_packfile_objects,
3197                         &reuse_packfile_bitmap)) {
3198                 assert(reuse_packfile_objects);
3199                 nr_result += reuse_packfile_objects;
3200                 display_progress(progress_state, nr_result);
3201         }
3202
3203         traverse_bitmap_commit_list(bitmap_git, revs,
3204                                     &add_object_entry_from_bitmap);
3205         return 0;
3206 }
3207
3208 static void record_recent_object(struct object *obj,
3209                                  const char *name,
3210                                  void *data)
3211 {
3212         oid_array_append(&recent_objects, &obj->oid);
3213 }
3214
3215 static void record_recent_commit(struct commit *commit, void *data)
3216 {
3217         oid_array_append(&recent_objects, &commit->object.oid);
3218 }
3219
3220 static void get_object_list(int ac, const char **av)
3221 {
3222         struct rev_info revs;
3223         struct setup_revision_opt s_r_opt = {
3224                 .allow_exclude_promisor_objects = 1,
3225         };
3226         char line[1000];
3227         int flags = 0;
3228         int save_warning;
3229
3230         repo_init_revisions(the_repository, &revs, NULL);
3231         save_commit_buffer = 0;
3232         setup_revisions(ac, av, &revs, &s_r_opt);
3233
3234         /* make sure shallows are read */
3235         is_repository_shallow(the_repository);
3236
3237         save_warning = warn_on_object_refname_ambiguity;
3238         warn_on_object_refname_ambiguity = 0;
3239
3240         while (fgets(line, sizeof(line), stdin) != NULL) {
3241                 int len = strlen(line);
3242                 if (len && line[len - 1] == '\n')
3243                         line[--len] = 0;
3244                 if (!len)
3245                         break;
3246                 if (*line == '-') {
3247                         if (!strcmp(line, "--not")) {
3248                                 flags ^= UNINTERESTING;
3249                                 write_bitmap_index = 0;
3250                                 continue;
3251                         }
3252                         if (starts_with(line, "--shallow ")) {
3253                                 struct object_id oid;
3254                                 if (get_oid_hex(line + 10, &oid))
3255                                         die("not an SHA-1 '%s'", line + 10);
3256                                 register_shallow(the_repository, &oid);
3257                                 use_bitmap_index = 0;
3258                                 continue;
3259                         }
3260                         die(_("not a rev '%s'"), line);
3261                 }
3262                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
3263                         die(_("bad revision '%s'"), line);
3264         }
3265
3266         warn_on_object_refname_ambiguity = save_warning;
3267
3268         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
3269                 return;
3270
3271         if (use_delta_islands)
3272                 load_delta_islands(the_repository, progress);
3273
3274         if (prepare_revision_walk(&revs))
3275                 die(_("revision walk setup failed"));
3276         mark_edges_uninteresting(&revs, show_edge, sparse);
3277
3278         if (!fn_show_object)
3279                 fn_show_object = show_object;
3280         traverse_commit_list_filtered(&filter_options, &revs,
3281                                       show_commit, fn_show_object, NULL,
3282                                       NULL);
3283
3284         if (unpack_unreachable_expiration) {
3285                 revs.ignore_missing_links = 1;
3286                 if (add_unseen_recent_objects_to_traversal(&revs,
3287                                 unpack_unreachable_expiration))
3288                         die(_("unable to add recent objects"));
3289                 if (prepare_revision_walk(&revs))
3290                         die(_("revision walk setup failed"));
3291                 traverse_commit_list(&revs, record_recent_commit,
3292                                      record_recent_object, NULL);
3293         }
3294
3295         if (keep_unreachable)
3296                 add_objects_in_unpacked_packs();
3297         if (pack_loose_unreachable)
3298                 add_unreachable_loose_objects();
3299         if (unpack_unreachable)
3300                 loosen_unused_packed_objects();
3301
3302         oid_array_clear(&recent_objects);
3303 }
3304
3305 static void add_extra_kept_packs(const struct string_list *names)
3306 {
3307         struct packed_git *p;
3308
3309         if (!names->nr)
3310                 return;
3311
3312         for (p = get_all_packs(the_repository); p; p = p->next) {
3313                 const char *name = basename(p->pack_name);
3314                 int i;
3315
3316                 if (!p->pack_local)
3317                         continue;
3318
3319                 for (i = 0; i < names->nr; i++)
3320                         if (!fspathcmp(name, names->items[i].string))
3321                                 break;
3322
3323                 if (i < names->nr) {
3324                         p->pack_keep_in_core = 1;
3325                         ignore_packed_keep_in_core = 1;
3326                         continue;
3327                 }
3328         }
3329 }
3330
3331 static int option_parse_index_version(const struct option *opt,
3332                                       const char *arg, int unset)
3333 {
3334         char *c;
3335         const char *val = arg;
3336
3337         BUG_ON_OPT_NEG(unset);
3338
3339         pack_idx_opts.version = strtoul(val, &c, 10);
3340         if (pack_idx_opts.version > 2)
3341                 die(_("unsupported index version %s"), val);
3342         if (*c == ',' && c[1])
3343                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
3344         if (*c || pack_idx_opts.off32_limit & 0x80000000)
3345                 die(_("bad index version '%s'"), val);
3346         return 0;
3347 }
3348
3349 static int option_parse_unpack_unreachable(const struct option *opt,
3350                                            const char *arg, int unset)
3351 {
3352         if (unset) {
3353                 unpack_unreachable = 0;
3354                 unpack_unreachable_expiration = 0;
3355         }
3356         else {
3357                 unpack_unreachable = 1;
3358                 if (arg)
3359                         unpack_unreachable_expiration = approxidate(arg);
3360         }
3361         return 0;
3362 }
3363
3364 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
3365 {
3366         int use_internal_rev_list = 0;
3367         int shallow = 0;
3368         int all_progress_implied = 0;
3369         struct argv_array rp = ARGV_ARRAY_INIT;
3370         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
3371         int rev_list_index = 0;
3372         struct string_list keep_pack_list = STRING_LIST_INIT_NODUP;
3373         struct option pack_objects_options[] = {
3374                 OPT_SET_INT('q', "quiet", &progress,
3375                             N_("do not show progress meter"), 0),
3376                 OPT_SET_INT(0, "progress", &progress,
3377                             N_("show progress meter"), 1),
3378                 OPT_SET_INT(0, "all-progress", &progress,
3379                             N_("show progress meter during object writing phase"), 2),
3380                 OPT_BOOL(0, "all-progress-implied",
3381                          &all_progress_implied,
3382                          N_("similar to --all-progress when progress meter is shown")),
3383                 OPT_CALLBACK_F(0, "index-version", NULL, N_("<version>[,<offset>]"),
3384                   N_("write the pack index file in the specified idx format version"),
3385                   PARSE_OPT_NONEG, option_parse_index_version),
3386                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
3387                               N_("maximum size of each output pack file")),
3388                 OPT_BOOL(0, "local", &local,
3389                          N_("ignore borrowed objects from alternate object store")),
3390                 OPT_BOOL(0, "incremental", &incremental,
3391                          N_("ignore packed objects")),
3392                 OPT_INTEGER(0, "window", &window,
3393                             N_("limit pack window by objects")),
3394                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
3395                               N_("limit pack window by memory in addition to object limit")),
3396                 OPT_INTEGER(0, "depth", &depth,
3397                             N_("maximum length of delta chain allowed in the resulting pack")),
3398                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
3399                          N_("reuse existing deltas")),
3400                 OPT_BOOL(0, "reuse-object", &reuse_object,
3401                          N_("reuse existing objects")),
3402                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
3403                          N_("use OFS_DELTA objects")),
3404                 OPT_INTEGER(0, "threads", &delta_search_threads,
3405                             N_("use threads when searching for best delta matches")),
3406                 OPT_BOOL(0, "non-empty", &non_empty,
3407                          N_("do not create an empty pack output")),
3408                 OPT_BOOL(0, "revs", &use_internal_rev_list,
3409                          N_("read revision arguments from standard input")),
3410                 OPT_SET_INT_F(0, "unpacked", &rev_list_unpacked,
3411                               N_("limit the objects to those that are not yet packed"),
3412                               1, PARSE_OPT_NONEG),
3413                 OPT_SET_INT_F(0, "all", &rev_list_all,
3414                               N_("include objects reachable from any reference"),
3415                               1, PARSE_OPT_NONEG),
3416                 OPT_SET_INT_F(0, "reflog", &rev_list_reflog,
3417                               N_("include objects referred by reflog entries"),
3418                               1, PARSE_OPT_NONEG),
3419                 OPT_SET_INT_F(0, "indexed-objects", &rev_list_index,
3420                               N_("include objects referred to by the index"),
3421                               1, PARSE_OPT_NONEG),
3422                 OPT_BOOL(0, "stdout", &pack_to_stdout,
3423                          N_("output pack to stdout")),
3424                 OPT_BOOL(0, "include-tag", &include_tag,
3425                          N_("include tag objects that refer to objects to be packed")),
3426                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
3427                          N_("keep unreachable objects")),
3428                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
3429                          N_("pack loose unreachable objects")),
3430                 OPT_CALLBACK_F(0, "unpack-unreachable", NULL, N_("time"),
3431                   N_("unpack unreachable objects newer than <time>"),
3432                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable),
3433                 OPT_BOOL(0, "sparse", &sparse,
3434                          N_("use the sparse reachability algorithm")),
3435                 OPT_BOOL(0, "thin", &thin,
3436                          N_("create thin packs")),
3437                 OPT_BOOL(0, "shallow", &shallow,
3438                          N_("create packs suitable for shallow fetches")),
3439                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep_on_disk,
3440                          N_("ignore packs that have companion .keep file")),
3441                 OPT_STRING_LIST(0, "keep-pack", &keep_pack_list, N_("name"),
3442                                 N_("ignore this pack")),
3443                 OPT_INTEGER(0, "compression", &pack_compression_level,
3444                             N_("pack compression level")),
3445                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
3446                             N_("do not hide commits by grafts"), 0),
3447                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
3448                          N_("use a bitmap index if available to speed up counting objects")),
3449                 OPT_SET_INT(0, "write-bitmap-index", &write_bitmap_index,
3450                             N_("write a bitmap index together with the pack index"),
3451                             WRITE_BITMAP_TRUE),
3452                 OPT_SET_INT_F(0, "write-bitmap-index-quiet",
3453                               &write_bitmap_index,
3454                               N_("write a bitmap index if possible"),
3455                               WRITE_BITMAP_QUIET, PARSE_OPT_HIDDEN),
3456                 OPT_PARSE_LIST_OBJECTS_FILTER(&filter_options),
3457                 OPT_CALLBACK_F(0, "missing", NULL, N_("action"),
3458                   N_("handling for missing objects"), PARSE_OPT_NONEG,
3459                   option_parse_missing_action),
3460                 OPT_BOOL(0, "exclude-promisor-objects", &exclude_promisor_objects,
3461                          N_("do not pack objects in promisor packfiles")),
3462                 OPT_BOOL(0, "delta-islands", &use_delta_islands,
3463                          N_("respect islands during delta compression")),
3464                 OPT_END(),
3465         };
3466
3467         if (DFS_NUM_STATES > (1 << OE_DFS_STATE_BITS))
3468                 BUG("too many dfs states, increase OE_DFS_STATE_BITS");
3469
3470         read_replace_refs = 0;
3471
3472         sparse = git_env_bool("GIT_TEST_PACK_SPARSE", -1);
3473         prepare_repo_settings(the_repository);
3474         if (sparse < 0)
3475                 sparse = the_repository->settings.pack_use_sparse;
3476
3477         reset_pack_idx_option(&pack_idx_opts);
3478         git_config(git_pack_config, NULL);
3479
3480         progress = isatty(2);
3481         argc = parse_options(argc, argv, prefix, pack_objects_options,
3482                              pack_usage, 0);
3483
3484         if (argc) {
3485                 base_name = argv[0];
3486                 argc--;
3487         }
3488         if (pack_to_stdout != !base_name || argc)
3489                 usage_with_options(pack_usage, pack_objects_options);
3490
3491         if (depth >= (1 << OE_DEPTH_BITS)) {
3492                 warning(_("delta chain depth %d is too deep, forcing %d"),
3493                         depth, (1 << OE_DEPTH_BITS) - 1);
3494                 depth = (1 << OE_DEPTH_BITS) - 1;
3495         }
3496         if (cache_max_small_delta_size >= (1U << OE_Z_DELTA_BITS)) {
3497                 warning(_("pack.deltaCacheLimit is too high, forcing %d"),
3498                         (1U << OE_Z_DELTA_BITS) - 1);
3499                 cache_max_small_delta_size = (1U << OE_Z_DELTA_BITS) - 1;
3500         }
3501
3502         argv_array_push(&rp, "pack-objects");
3503         if (thin) {
3504                 use_internal_rev_list = 1;
3505                 argv_array_push(&rp, shallow
3506                                 ? "--objects-edge-aggressive"
3507                                 : "--objects-edge");
3508         } else
3509                 argv_array_push(&rp, "--objects");
3510
3511         if (rev_list_all) {
3512                 use_internal_rev_list = 1;
3513                 argv_array_push(&rp, "--all");
3514         }
3515         if (rev_list_reflog) {
3516                 use_internal_rev_list = 1;
3517                 argv_array_push(&rp, "--reflog");
3518         }
3519         if (rev_list_index) {
3520                 use_internal_rev_list = 1;
3521                 argv_array_push(&rp, "--indexed-objects");
3522         }
3523         if (rev_list_unpacked) {
3524                 use_internal_rev_list = 1;
3525                 argv_array_push(&rp, "--unpacked");
3526         }
3527
3528         if (exclude_promisor_objects) {
3529                 use_internal_rev_list = 1;
3530                 fetch_if_missing = 0;
3531                 argv_array_push(&rp, "--exclude-promisor-objects");
3532         }
3533         if (unpack_unreachable || keep_unreachable || pack_loose_unreachable)
3534                 use_internal_rev_list = 1;
3535
3536         if (!reuse_object)
3537                 reuse_delta = 0;
3538         if (pack_compression_level == -1)
3539                 pack_compression_level = Z_DEFAULT_COMPRESSION;
3540         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
3541                 die(_("bad pack compression level %d"), pack_compression_level);
3542
3543         if (!delta_search_threads)      /* --threads=0 means autodetect */
3544                 delta_search_threads = online_cpus();
3545
3546         if (!HAVE_THREADS && delta_search_threads != 1)
3547                 warning(_("no threads support, ignoring --threads"));
3548         if (!pack_to_stdout && !pack_size_limit)
3549                 pack_size_limit = pack_size_limit_cfg;
3550         if (pack_to_stdout && pack_size_limit)
3551                 die(_("--max-pack-size cannot be used to build a pack for transfer"));
3552         if (pack_size_limit && pack_size_limit < 1024*1024) {
3553                 warning(_("minimum pack size limit is 1 MiB"));
3554                 pack_size_limit = 1024*1024;
3555         }
3556
3557         if (!pack_to_stdout && thin)
3558                 die(_("--thin cannot be used to build an indexable pack"));
3559
3560         if (keep_unreachable && unpack_unreachable)
3561                 die(_("--keep-unreachable and --unpack-unreachable are incompatible"));
3562         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3563                 unpack_unreachable_expiration = 0;
3564
3565         if (filter_options.choice) {
3566                 if (!pack_to_stdout)
3567                         die(_("cannot use --filter without --stdout"));
3568         }
3569
3570         /*
3571          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3572          *
3573          * - to produce good pack (with bitmap index not-yet-packed objects are
3574          *   packed in suboptimal order).
3575          *
3576          * - to use more robust pack-generation codepath (avoiding possible
3577          *   bugs in bitmap code and possible bitmap index corruption).
3578          */
3579         if (!pack_to_stdout)
3580                 use_bitmap_index_default = 0;
3581
3582         if (use_bitmap_index < 0)
3583                 use_bitmap_index = use_bitmap_index_default;
3584
3585         /* "hard" reasons not to use bitmaps; these just won't work at all */
3586         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow(the_repository))
3587                 use_bitmap_index = 0;
3588
3589         if (pack_to_stdout || !rev_list_all)
3590                 write_bitmap_index = 0;
3591
3592         if (use_delta_islands)
3593                 argv_array_push(&rp, "--topo-order");
3594
3595         if (progress && all_progress_implied)
3596                 progress = 2;
3597
3598         add_extra_kept_packs(&keep_pack_list);
3599         if (ignore_packed_keep_on_disk) {
3600                 struct packed_git *p;
3601                 for (p = get_all_packs(the_repository); p; p = p->next)
3602                         if (p->pack_local && p->pack_keep)
3603                                 break;
3604                 if (!p) /* no keep-able packs found */
3605                         ignore_packed_keep_on_disk = 0;
3606         }
3607         if (local) {
3608                 /*
3609                  * unlike ignore_packed_keep_on_disk above, we do not
3610                  * want to unset "local" based on looking at packs, as
3611                  * it also covers non-local objects
3612                  */
3613                 struct packed_git *p;
3614                 for (p = get_all_packs(the_repository); p; p = p->next) {
3615                         if (!p->pack_local) {
3616                                 have_non_local_packs = 1;
3617                                 break;
3618                         }
3619                 }
3620         }
3621
3622         trace2_region_enter("pack-objects", "enumerate-objects",
3623                             the_repository);
3624         prepare_packing_data(the_repository, &to_pack);
3625
3626         if (progress)
3627                 progress_state = start_progress(_("Enumerating objects"), 0);
3628         if (!use_internal_rev_list)
3629                 read_object_list_from_stdin();
3630         else {
3631                 get_object_list(rp.argc, rp.argv);
3632                 argv_array_clear(&rp);
3633         }
3634         cleanup_preferred_base();
3635         if (include_tag && nr_result)
3636                 for_each_ref(add_ref_tag, NULL);
3637         stop_progress(&progress_state);
3638         trace2_region_leave("pack-objects", "enumerate-objects",
3639                             the_repository);
3640
3641         if (non_empty && !nr_result)
3642                 return 0;
3643         if (nr_result) {
3644                 trace2_region_enter("pack-objects", "prepare-pack",
3645                                     the_repository);
3646                 prepare_pack(window, depth);
3647                 trace2_region_leave("pack-objects", "prepare-pack",
3648                                     the_repository);
3649         }
3650
3651         trace2_region_enter("pack-objects", "write-pack-file", the_repository);
3652         write_pack_file();
3653         trace2_region_leave("pack-objects", "write-pack-file", the_repository);
3654
3655         if (progress)
3656                 fprintf_ln(stderr,
3657                            _("Total %"PRIu32" (delta %"PRIu32"),"
3658                              " reused %"PRIu32" (delta %"PRIu32"),"
3659                              " pack-reused %"PRIu32),
3660                            written, written_delta, reused, reused_delta,
3661                            reuse_packfile_objects);
3662         return 0;
3663 }