git.oblomov.eu Git - git/blob - builtin/pack-objects.c

   1 #include "builtin.h"
   2 #include "cache.h"
   3 #include "attr.h"
   4 #include "object.h"
   5 #include "blob.h"
   6 #include "commit.h"
   7 #include "tag.h"
   8 #include "tree.h"
   9 #include "delta.h"
  10 #include "pack.h"
  11 #include "pack-revindex.h"
  12 #include "csum-file.h"
  13 #include "tree-walk.h"
  14 #include "diff.h"
  15 #include "revision.h"
  16 #include "list-objects.h"
  17 #include "pack-objects.h"
  18 #include "progress.h"
  19 #include "refs.h"
  20 #include "streaming.h"
  21 #include "thread-utils.h"
  22 #include "pack-bitmap.h"
  23 #include "reachable.h"
  24 #include "sha1-array.h"
  25 #include "argv-array.h"
  26
  27 static const char *pack_usage[] = {
  28         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
  29         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
  30         NULL
  31 };
  32
  33 /*
  34  * Objects we are going to pack are collected in the `to_pack` structure.
  35  * It contains an array (dynamically expanded) of the object data, and a map
  36  * that can resolve SHA1s to their position in the array.
  37  */
  38 static struct packing_data to_pack;
  39
  40 static struct pack_idx_entry **written_list;
  41 static uint32_t nr_result, nr_written;
  42
  43 static int non_empty;
  44 static int reuse_delta = 1, reuse_object = 1;
  45 static int keep_unreachable, unpack_unreachable, include_tag;
  46 static unsigned long unpack_unreachable_expiration;
  47 static int pack_loose_unreachable;
  48 static int local;
  49 static int have_non_local_packs;
  50 static int incremental;
  51 static int ignore_packed_keep;
  52 static int allow_ofs_delta;
  53 static struct pack_idx_option pack_idx_opts;
  54 static const char *base_name;
  55 static int progress = 1;
  56 static int window = 10;
  57 static unsigned long pack_size_limit;
  58 static int depth = 50;
  59 static int delta_search_threads;
  60 static int pack_to_stdout;
  61 static int num_preferred_base;
  62 static struct progress *progress_state;
  63 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
  64 static int pack_compression_seen;
  65
  66 static struct packed_git *reuse_packfile;
  67 static uint32_t reuse_packfile_objects;
  68 static off_t reuse_packfile_offset;
  69
  70 static int use_bitmap_index_default = 1;
  71 static int use_bitmap_index = -1;
  72 static int write_bitmap_index;
  73 static uint16_t write_bitmap_options;
  74
  75 static unsigned long delta_cache_size = 0;
  76 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
  77 static unsigned long cache_max_small_delta_size = 1000;
  78
  79 static unsigned long window_memory_limit = 0;
  80
  81 /*
  82  * stats
  83  */
  84 static uint32_t written, written_delta;
  85 static uint32_t reused, reused_delta;
  86
  87 /*
  88  * Indexed commits
  89  */
  90 static struct commit **indexed_commits;
  91 static unsigned int indexed_commits_nr;
  92 static unsigned int indexed_commits_alloc;
  93
  94 static void index_commit_for_bitmap(struct commit *commit)
  95 {
  96         if (indexed_commits_nr >= indexed_commits_alloc) {
  97                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
  98                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
  99         }
 100
 101         indexed_commits[indexed_commits_nr++] = commit;
 102 }
 103
 104 static void *get_delta(struct object_entry *entry)
 105 {
 106         unsigned long size, base_size, delta_size;
 107         void *buf, *base_buf, *delta_buf;
 108         enum object_type type;
 109
 110         buf = read_sha1_file(entry->idx.sha1, &type, &size);
 111         if (!buf)
 112                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
 113         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
 114         if (!base_buf)
 115                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
 116         delta_buf = diff_delta(base_buf, base_size,
 117                                buf, size, &delta_size, 0);
 118         if (!delta_buf || delta_size != entry->delta_size)
 119                 die("delta size changed");
 120         free(buf);
 121         free(base_buf);
 122         return delta_buf;
 123 }
 124
 125 static unsigned long do_compress(void **pptr, unsigned long size)
 126 {
 127         git_zstream stream;
 128         void *in, *out;
 129         unsigned long maxsize;
 130
 131         git_deflate_init(&stream, pack_compression_level);
 132         maxsize = git_deflate_bound(&stream, size);
 133
 134         in = *pptr;
 135         out = xmalloc(maxsize);
 136         *pptr = out;
 137
 138         stream.next_in = in;
 139         stream.avail_in = size;
 140         stream.next_out = out;
 141         stream.avail_out = maxsize;
 142         while (git_deflate(&stream, Z_FINISH) == Z_OK)
 143                 ; /* nothing */
 144         git_deflate_end(&stream);
 145
 146         free(in);
 147         return stream.total_out;
 148 }
 149
 150 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
 151                                            const unsigned char *sha1)
 152 {
 153         git_zstream stream;
 154         unsigned char ibuf[1024 * 16];
 155         unsigned char obuf[1024 * 16];
 156         unsigned long olen = 0;
 157
 158         git_deflate_init(&stream, pack_compression_level);
 159
 160         for (;;) {
 161                 ssize_t readlen;
 162                 int zret = Z_OK;
 163                 readlen = read_istream(st, ibuf, sizeof(ibuf));
 164                 if (readlen == -1)
 165                         die(_("unable to read %s"), sha1_to_hex(sha1));
 166
 167                 stream.next_in = ibuf;
 168                 stream.avail_in = readlen;
 169                 while ((stream.avail_in || readlen == 0) &&
 170                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
 171                         stream.next_out = obuf;
 172                         stream.avail_out = sizeof(obuf);
 173                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 174                         sha1write(f, obuf, stream.next_out - obuf);
 175                         olen += stream.next_out - obuf;
 176                 }
 177                 if (stream.avail_in)
 178                         die(_("deflate error (%d)"), zret);
 179                 if (readlen == 0) {
 180                         if (zret != Z_STREAM_END)
 181                                 die(_("deflate error (%d)"), zret);
 182                         break;
 183                 }
 184         }
 185         git_deflate_end(&stream);
 186         return olen;
 187 }
 188
 189 /*
 190  * we are going to reuse the existing object data as is.  make
 191  * sure it is not corrupt.
 192  */
 193 static int check_pack_inflate(struct packed_git *p,
 194                 struct pack_window **w_curs,
 195                 off_t offset,
 196                 off_t len,
 197                 unsigned long expect)
 198 {
 199         git_zstream stream;
 200         unsigned char fakebuf[4096], *in;
 201         int st;
 202
 203         memset(&stream, 0, sizeof(stream));
 204         git_inflate_init(&stream);
 205         do {
 206                 in = use_pack(p, w_curs, offset, &stream.avail_in);
 207                 stream.next_in = in;
 208                 stream.next_out = fakebuf;
 209                 stream.avail_out = sizeof(fakebuf);
 210                 st = git_inflate(&stream, Z_FINISH);
 211                 offset += stream.next_in - in;
 212         } while (st == Z_OK || st == Z_BUF_ERROR);
 213         git_inflate_end(&stream);
 214         return (st == Z_STREAM_END &&
 215                 stream.total_out == expect &&
 216                 stream.total_in == len) ? 0 : -1;
 217 }
 218
 219 static void copy_pack_data(struct sha1file *f,
 220                 struct packed_git *p,
 221                 struct pack_window **w_curs,
 222                 off_t offset,
 223                 off_t len)
 224 {
 225         unsigned char *in;
 226         unsigned long avail;
 227
 228         while (len) {
 229                 in = use_pack(p, w_curs, offset, &avail);
 230                 if (avail > len)
 231                         avail = (unsigned long)len;
 232                 sha1write(f, in, avail);
 233                 offset += avail;
 234                 len -= avail;
 235         }
 236 }
 237
 238 /* Return 0 if we will bust the pack-size limit */
 239 static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
 240                                            unsigned long limit, int usable_delta)
 241 {
 242         unsigned long size, datalen;
 243         unsigned char header[10], dheader[10];
 244         unsigned hdrlen;
 245         enum object_type type;
 246         void *buf;
 247         struct git_istream *st = NULL;
 248
 249         if (!usable_delta) {
 250                 if (entry->type == OBJ_BLOB &&
 251                     entry->size > big_file_threshold &&
 252                     (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
 253                         buf = NULL;
 254                 else {
 255                         buf = read_sha1_file(entry->idx.sha1, &type, &size);
 256                         if (!buf)
 257                                 die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
 258                 }
 259                 /*
 260                  * make sure no cached delta data remains from a
 261                  * previous attempt before a pack split occurred.
 262                  */
 263                 free(entry->delta_data);
 264                 entry->delta_data = NULL;
 265                 entry->z_delta_size = 0;
 266         } else if (entry->delta_data) {
 267                 size = entry->delta_size;
 268                 buf = entry->delta_data;
 269                 entry->delta_data = NULL;
 270                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 271                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 272         } else {
 273                 buf = get_delta(entry);
 274                 size = entry->delta_size;
 275                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 276                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 277         }
 278
 279         if (st) /* large blob case, just assume we don't compress well */
 280                 datalen = size;
 281         else if (entry->z_delta_size)
 282                 datalen = entry->z_delta_size;
 283         else
 284                 datalen = do_compress(&buf, size);
 285
 286         /*
 287          * The object header is a byte of 'type' followed by zero or
 288          * more bytes of length.
 289          */
 290         hdrlen = encode_in_pack_object_header(type, size, header);
 291
 292         if (type == OBJ_OFS_DELTA) {
 293                 /*
 294                  * Deltas with relative base contain an additional
 295                  * encoding of the relative offset for the delta
 296                  * base from this object's position in the pack.
 297                  */
 298                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 299                 unsigned pos = sizeof(dheader) - 1;
 300                 dheader[pos] = ofs & 127;
 301                 while (ofs >>= 7)
 302                         dheader[--pos] = 128 | (--ofs & 127);
 303                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 304                         if (st)
 305                                 close_istream(st);
 306                         free(buf);
 307                         return 0;
 308                 }
 309                 sha1write(f, header, hdrlen);
 310                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 311                 hdrlen += sizeof(dheader) - pos;
 312         } else if (type == OBJ_REF_DELTA) {
 313                 /*
 314                  * Deltas with a base reference contain
 315                  * an additional 20 bytes for the base sha1.
 316                  */
 317                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 318                         if (st)
 319                                 close_istream(st);
 320                         free(buf);
 321                         return 0;
 322                 }
 323                 sha1write(f, header, hdrlen);
 324                 sha1write(f, entry->delta->idx.sha1, 20);
 325                 hdrlen += 20;
 326         } else {
 327                 if (limit && hdrlen + datalen + 20 >= limit) {
 328                         if (st)
 329                                 close_istream(st);
 330                         free(buf);
 331                         return 0;
 332                 }
 333                 sha1write(f, header, hdrlen);
 334         }
 335         if (st) {
 336                 datalen = write_large_blob_data(st, f, entry->idx.sha1);
 337                 close_istream(st);
 338         } else {
 339                 sha1write(f, buf, datalen);
 340                 free(buf);
 341         }
 342
 343         return hdrlen + datalen;
 344 }
 345
 346 /* Return 0 if we will bust the pack-size limit */
 347 static unsigned long write_reuse_object(struct sha1file *f, struct object_entry *entry,
 348                                         unsigned long limit, int usable_delta)
 349 {
 350         struct packed_git *p = entry->in_pack;
 351         struct pack_window *w_curs = NULL;
 352         struct revindex_entry *revidx;
 353         off_t offset;
 354         enum object_type type = entry->type;
 355         unsigned long datalen;
 356         unsigned char header[10], dheader[10];
 357         unsigned hdrlen;
 358
 359         if (entry->delta)
 360                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 361                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 362         hdrlen = encode_in_pack_object_header(type, entry->size, header);
 363
 364         offset = entry->in_pack_offset;
 365         revidx = find_pack_revindex(p, offset);
 366         datalen = revidx[1].offset - offset;
 367         if (!pack_to_stdout && p->index_version > 1 &&
 368             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 369                 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
 370                 unuse_pack(&w_curs);
 371                 return write_no_reuse_object(f, entry, limit, usable_delta);
 372         }
 373
 374         offset += entry->in_pack_header_size;
 375         datalen -= entry->in_pack_header_size;
 376
 377         if (!pack_to_stdout && p->index_version == 1 &&
 378             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 379                 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
 380                 unuse_pack(&w_curs);
 381                 return write_no_reuse_object(f, entry, limit, usable_delta);
 382         }
 383
 384         if (type == OBJ_OFS_DELTA) {
 385                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 386                 unsigned pos = sizeof(dheader) - 1;
 387                 dheader[pos] = ofs & 127;
 388                 while (ofs >>= 7)
 389                         dheader[--pos] = 128 | (--ofs & 127);
 390                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 391                         unuse_pack(&w_curs);
 392                         return 0;
 393                 }
 394                 sha1write(f, header, hdrlen);
 395                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 396                 hdrlen += sizeof(dheader) - pos;
 397                 reused_delta++;
 398         } else if (type == OBJ_REF_DELTA) {
 399                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 400                         unuse_pack(&w_curs);
 401                         return 0;
 402                 }
 403                 sha1write(f, header, hdrlen);
 404                 sha1write(f, entry->delta->idx.sha1, 20);
 405                 hdrlen += 20;
 406                 reused_delta++;
 407         } else {
 408                 if (limit && hdrlen + datalen + 20 >= limit) {
 409                         unuse_pack(&w_curs);
 410                         return 0;
 411                 }
 412                 sha1write(f, header, hdrlen);
 413         }
 414         copy_pack_data(f, p, &w_curs, offset, datalen);
 415         unuse_pack(&w_curs);
 416         reused++;
 417         return hdrlen + datalen;
 418 }
 419
 420 /* Return 0 if we will bust the pack-size limit */
 421 static unsigned long write_object(struct sha1file *f,
 422                                   struct object_entry *entry,
 423                                   off_t write_offset)
 424 {
 425         unsigned long limit, len;
 426         int usable_delta, to_reuse;
 427
 428         if (!pack_to_stdout)
 429                 crc32_begin(f);
 430
 431         /* apply size limit if limited packsize and not first object */
 432         if (!pack_size_limit || !nr_written)
 433                 limit = 0;
 434         else if (pack_size_limit <= write_offset)
 435                 /*
 436                  * the earlier object did not fit the limit; avoid
 437                  * mistaking this with unlimited (i.e. limit = 0).
 438                  */
 439                 limit = 1;
 440         else
 441                 limit = pack_size_limit - write_offset;
 442
 443         if (!entry->delta)
 444                 usable_delta = 0;       /* no delta */
 445         else if (!pack_size_limit)
 446                usable_delta = 1;        /* unlimited packfile */
 447         else if (entry->delta->idx.offset == (off_t)-1)
 448                 usable_delta = 0;       /* base was written to another pack */
 449         else if (entry->delta->idx.offset)
 450                 usable_delta = 1;       /* base already exists in this pack */
 451         else
 452                 usable_delta = 0;       /* base could end up in another pack */
 453
 454         if (!reuse_object)
 455                 to_reuse = 0;   /* explicit */
 456         else if (!entry->in_pack)
 457                 to_reuse = 0;   /* can't reuse what we don't have */
 458         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
 459                                 /* check_object() decided it for us ... */
 460                 to_reuse = usable_delta;
 461                                 /* ... but pack split may override that */
 462         else if (entry->type != entry->in_pack_type)
 463                 to_reuse = 0;   /* pack has delta which is unusable */
 464         else if (entry->delta)
 465                 to_reuse = 0;   /* we want to pack afresh */
 466         else
 467                 to_reuse = 1;   /* we have it in-pack undeltified,
 468                                  * and we do not need to deltify it.
 469                                  */
 470
 471         if (!to_reuse)
 472                 len = write_no_reuse_object(f, entry, limit, usable_delta);
 473         else
 474                 len = write_reuse_object(f, entry, limit, usable_delta);
 475         if (!len)
 476                 return 0;
 477
 478         if (usable_delta)
 479                 written_delta++;
 480         written++;
 481         if (!pack_to_stdout)
 482                 entry->idx.crc32 = crc32_end(f);
 483         return len;
 484 }
 485
 486 enum write_one_status {
 487         WRITE_ONE_SKIP = -1, /* already written */
 488         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 489         WRITE_ONE_WRITTEN = 1, /* normal */
 490         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 491 };
 492
 493 static enum write_one_status write_one(struct sha1file *f,
 494                                        struct object_entry *e,
 495                                        off_t *offset)
 496 {
 497         unsigned long size;
 498         int recursing;
 499
 500         /*
 501          * we set offset to 1 (which is an impossible value) to mark
 502          * the fact that this object is involved in "write its base
 503          * first before writing a deltified object" recursion.
 504          */
 505         recursing = (e->idx.offset == 1);
 506         if (recursing) {
 507                 warning("recursive delta detected for object %s",
 508                         sha1_to_hex(e->idx.sha1));
 509                 return WRITE_ONE_RECURSIVE;
 510         } else if (e->idx.offset || e->preferred_base) {
 511                 /* offset is non zero if object is written already. */
 512                 return WRITE_ONE_SKIP;
 513         }
 514
 515         /* if we are deltified, write out base object first. */
 516         if (e->delta) {
 517                 e->idx.offset = 1; /* now recurse */
 518                 switch (write_one(f, e->delta, offset)) {
 519                 case WRITE_ONE_RECURSIVE:
 520                         /* we cannot depend on this one */
 521                         e->delta = NULL;
 522                         break;
 523                 default:
 524                         break;
 525                 case WRITE_ONE_BREAK:
 526                         e->idx.offset = recursing;
 527                         return WRITE_ONE_BREAK;
 528                 }
 529         }
 530
 531         e->idx.offset = *offset;
 532         size = write_object(f, e, *offset);
 533         if (!size) {
 534                 e->idx.offset = recursing;
 535                 return WRITE_ONE_BREAK;
 536         }
 537         written_list[nr_written++] = &e->idx;
 538
 539         /* make sure off_t is sufficiently large not to wrap */
 540         if (signed_add_overflows(*offset, size))
 541                 die("pack too large for current definition of off_t");
 542         *offset += size;
 543         return WRITE_ONE_WRITTEN;
 544 }
 545
 546 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
 547                        void *cb_data)
 548 {
 549         unsigned char peeled[20];
 550         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
 551
 552         if (entry)
 553                 entry->tagged = 1;
 554         if (!peel_ref(path, peeled)) {
 555                 entry = packlist_find(&to_pack, peeled, NULL);
 556                 if (entry)
 557                         entry->tagged = 1;
 558         }
 559         return 0;
 560 }
 561
 562 static inline void add_to_write_order(struct object_entry **wo,
 563                                unsigned int *endp,
 564                                struct object_entry *e)
 565 {
 566         if (e->filled)
 567                 return;
 568         wo[(*endp)++] = e;
 569         e->filled = 1;
 570 }
 571
 572 static void add_descendants_to_write_order(struct object_entry **wo,
 573                                            unsigned int *endp,
 574                                            struct object_entry *e)
 575 {
 576         int add_to_order = 1;
 577         while (e) {
 578                 if (add_to_order) {
 579                         struct object_entry *s;
 580                         /* add this node... */
 581                         add_to_write_order(wo, endp, e);
 582                         /* all its siblings... */
 583                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
 584                                 add_to_write_order(wo, endp, s);
 585                         }
 586                 }
 587                 /* drop down a level to add left subtree nodes if possible */
 588                 if (e->delta_child) {
 589                         add_to_order = 1;
 590                         e = e->delta_child;
 591                 } else {
 592                         add_to_order = 0;
 593                         /* our sibling might have some children, it is next */
 594                         if (e->delta_sibling) {
 595                                 e = e->delta_sibling;
 596                                 continue;
 597                         }
 598                         /* go back to our parent node */
 599                         e = e->delta;
 600                         while (e && !e->delta_sibling) {
 601                                 /* we're on the right side of a subtree, keep
 602                                  * going up until we can go right again */
 603                                 e = e->delta;
 604                         }
 605                         if (!e) {
 606                                 /* done- we hit our original root node */
 607                                 return;
 608                         }
 609                         /* pass it off to sibling at this level */
 610                         e = e->delta_sibling;
 611                 }
 612         };
 613 }
 614
 615 static void add_family_to_write_order(struct object_entry **wo,
 616                                       unsigned int *endp,
 617                                       struct object_entry *e)
 618 {
 619         struct object_entry *root;
 620
 621         for (root = e; root->delta; root = root->delta)
 622                 ; /* nothing */
 623         add_descendants_to_write_order(wo, endp, root);
 624 }
 625
 626 static struct object_entry **compute_write_order(void)
 627 {
 628         unsigned int i, wo_end, last_untagged;
 629
 630         struct object_entry **wo;
 631         struct object_entry *objects = to_pack.objects;
 632
 633         for (i = 0; i < to_pack.nr_objects; i++) {
 634                 objects[i].tagged = 0;
 635                 objects[i].filled = 0;
 636                 objects[i].delta_child = NULL;
 637                 objects[i].delta_sibling = NULL;
 638         }
 639
 640         /*
 641          * Fully connect delta_child/delta_sibling network.
 642          * Make sure delta_sibling is sorted in the original
 643          * recency order.
 644          */
 645         for (i = to_pack.nr_objects; i > 0;) {
 646                 struct object_entry *e = &objects[--i];
 647                 if (!e->delta)
 648                         continue;
 649                 /* Mark me as the first child */
 650                 e->delta_sibling = e->delta->delta_child;
 651                 e->delta->delta_child = e;
 652         }
 653
 654         /*
 655          * Mark objects that are at the tip of tags.
 656          */
 657         for_each_tag_ref(mark_tagged, NULL);
 658
 659         /*
 660          * Give the objects in the original recency order until
 661          * we see a tagged tip.
 662          */
 663         ALLOC_ARRAY(wo, to_pack.nr_objects);
 664         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
 665                 if (objects[i].tagged)
 666                         break;
 667                 add_to_write_order(wo, &wo_end, &objects[i]);
 668         }
 669         last_untagged = i;
 670
 671         /*
 672          * Then fill all the tagged tips.
 673          */
 674         for (; i < to_pack.nr_objects; i++) {
 675                 if (objects[i].tagged)
 676                         add_to_write_order(wo, &wo_end, &objects[i]);
 677         }
 678
 679         /*
 680          * And then all remaining commits and tags.
 681          */
 682         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 683                 if (objects[i].type != OBJ_COMMIT &&
 684                     objects[i].type != OBJ_TAG)
 685                         continue;
 686                 add_to_write_order(wo, &wo_end, &objects[i]);
 687         }
 688
 689         /*
 690          * And then all the trees.
 691          */
 692         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 693                 if (objects[i].type != OBJ_TREE)
 694                         continue;
 695                 add_to_write_order(wo, &wo_end, &objects[i]);
 696         }
 697
 698         /*
 699          * Finally all the rest in really tight order
 700          */
 701         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 702                 if (!objects[i].filled)
 703                         add_family_to_write_order(wo, &wo_end, &objects[i]);
 704         }
 705
 706         if (wo_end != to_pack.nr_objects)
 707                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
 708
 709         return wo;
 710 }
 711
 712 static off_t write_reused_pack(struct sha1file *f)
 713 {
 714         unsigned char buffer[8192];
 715         off_t to_write, total;
 716         int fd;
 717
 718         if (!is_pack_valid(reuse_packfile))
 719                 die("packfile is invalid: %s", reuse_packfile->pack_name);
 720
 721         fd = git_open_noatime(reuse_packfile->pack_name);
 722         if (fd < 0)
 723                 die_errno("unable to open packfile for reuse: %s",
 724                           reuse_packfile->pack_name);
 725
 726         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
 727                 die_errno("unable to seek in reused packfile");
 728
 729         if (reuse_packfile_offset < 0)
 730                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
 731
 732         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
 733
 734         while (to_write) {
 735                 int read_pack = xread(fd, buffer, sizeof(buffer));
 736
 737                 if (read_pack <= 0)
 738                         die_errno("unable to read from reused packfile");
 739
 740                 if (read_pack > to_write)
 741                         read_pack = to_write;
 742
 743                 sha1write(f, buffer, read_pack);
 744                 to_write -= read_pack;
 745
 746                 /*
 747                  * We don't know the actual number of objects written,
 748                  * only how many bytes written, how many bytes total, and
 749                  * how many objects total. So we can fake it by pretending all
 750                  * objects we are writing are the same size. This gives us a
 751                  * smooth progress meter, and at the end it matches the true
 752                  * answer.
 753                  */
 754                 written = reuse_packfile_objects *
 755                                 (((double)(total - to_write)) / total);
 756                 display_progress(progress_state, written);
 757         }
 758
 759         close(fd);
 760         written = reuse_packfile_objects;
 761         display_progress(progress_state, written);
 762         return reuse_packfile_offset - sizeof(struct pack_header);
 763 }
 764
 765 static const char no_split_warning[] = N_(
 766 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
 767 );
 768
 769 static void write_pack_file(void)
 770 {
 771         uint32_t i = 0, j;
 772         struct sha1file *f;
 773         off_t offset;
 774         uint32_t nr_remaining = nr_result;
 775         time_t last_mtime = 0;
 776         struct object_entry **write_order;
 777
 778         if (progress > pack_to_stdout)
 779                 progress_state = start_progress(_("Writing objects"), nr_result);
 780         ALLOC_ARRAY(written_list, to_pack.nr_objects);
 781         write_order = compute_write_order();
 782
 783         do {
 784                 unsigned char sha1[20];
 785                 char *pack_tmp_name = NULL;
 786
 787                 if (pack_to_stdout)
 788                         f = sha1fd_throughput(1, "<stdout>", progress_state);
 789                 else
 790                         f = create_tmp_packfile(&pack_tmp_name);
 791
 792                 offset = write_pack_header(f, nr_remaining);
 793
 794                 if (reuse_packfile) {
 795                         off_t packfile_size;
 796                         assert(pack_to_stdout);
 797
 798                         packfile_size = write_reused_pack(f);
 799                         offset += packfile_size;
 800                 }
 801
 802                 nr_written = 0;
 803                 for (; i < to_pack.nr_objects; i++) {
 804                         struct object_entry *e = write_order[i];
 805                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 806                                 break;
 807                         display_progress(progress_state, written);
 808                 }
 809
 810                 /*
 811                  * Did we write the wrong # entries in the header?
 812                  * If so, rewrite it like in fast-import
 813                  */
 814                 if (pack_to_stdout) {
 815                         sha1close(f, sha1, CSUM_CLOSE);
 816                 } else if (nr_written == nr_remaining) {
 817                         sha1close(f, sha1, CSUM_FSYNC);
 818                 } else {
 819                         int fd = sha1close(f, sha1, 0);
 820                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
 821                                                  nr_written, sha1, offset);
 822                         close(fd);
 823                         if (write_bitmap_index) {
 824                                 warning(_(no_split_warning));
 825                                 write_bitmap_index = 0;
 826                         }
 827                 }
 828
 829                 if (!pack_to_stdout) {
 830                         struct stat st;
 831                         struct strbuf tmpname = STRBUF_INIT;
 832
 833                         /*
 834                          * Packs are runtime accessed in their mtime
 835                          * order since newer packs are more likely to contain
 836                          * younger objects.  So if we are creating multiple
 837                          * packs then we should modify the mtime of later ones
 838                          * to preserve this property.
 839                          */
 840                         if (stat(pack_tmp_name, &st) < 0) {
 841                                 warning_errno("failed to stat %s", pack_tmp_name);
 842                         } else if (!last_mtime) {
 843                                 last_mtime = st.st_mtime;
 844                         } else {
 845                                 struct utimbuf utb;
 846                                 utb.actime = st.st_atime;
 847                                 utb.modtime = --last_mtime;
 848                                 if (utime(pack_tmp_name, &utb) < 0)
 849                                         warning_errno("failed utime() on %s", pack_tmp_name);
 850                         }
 851
 852                         strbuf_addf(&tmpname, "%s-", base_name);
 853
 854                         if (write_bitmap_index) {
 855                                 bitmap_writer_set_checksum(sha1);
 856                                 bitmap_writer_build_type_index(written_list, nr_written);
 857                         }
 858
 859                         finish_tmp_packfile(&tmpname, pack_tmp_name,
 860                                             written_list, nr_written,
 861                                             &pack_idx_opts, sha1);
 862
 863                         if (write_bitmap_index) {
 864                                 strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
 865
 866                                 stop_progress(&progress_state);
 867
 868                                 bitmap_writer_show_progress(progress);
 869                                 bitmap_writer_reuse_bitmaps(&to_pack);
 870                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
 871                                 bitmap_writer_build(&to_pack);
 872                                 bitmap_writer_finish(written_list, nr_written,
 873                                                      tmpname.buf, write_bitmap_options);
 874                                 write_bitmap_index = 0;
 875                         }
 876
 877                         strbuf_release(&tmpname);
 878                         free(pack_tmp_name);
 879                         puts(sha1_to_hex(sha1));
 880                 }
 881
 882                 /* mark written objects as written to previous pack */
 883                 for (j = 0; j < nr_written; j++) {
 884                         written_list[j]->offset = (off_t)-1;
 885                 }
 886                 nr_remaining -= nr_written;
 887         } while (nr_remaining && i < to_pack.nr_objects);
 888
 889         free(written_list);
 890         free(write_order);
 891         stop_progress(&progress_state);
 892         if (written != nr_result)
 893                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
 894                         written, nr_result);
 895 }
 896
 897 static void setup_delta_attr_check(struct git_attr_check *check)
 898 {
 899         static struct git_attr *attr_delta;
 900
 901         if (!attr_delta)
 902                 attr_delta = git_attr("delta");
 903
 904         check[0].attr = attr_delta;
 905 }
 906
 907 static int no_try_delta(const char *path)
 908 {
 909         struct git_attr_check check[1];
 910
 911         setup_delta_attr_check(check);
 912         if (git_check_attr(path, ARRAY_SIZE(check), check))
 913                 return 0;
 914         if (ATTR_FALSE(check->value))
 915                 return 1;
 916         return 0;
 917 }
 918
 919 /*
 920  * When adding an object, check whether we have already added it
 921  * to our packing list. If so, we can skip. However, if we are
 922  * being asked to excludei t, but the previous mention was to include
 923  * it, make sure to adjust its flags and tweak our numbers accordingly.
 924  *
 925  * As an optimization, we pass out the index position where we would have
 926  * found the item, since that saves us from having to look it up again a
 927  * few lines later when we want to add the new entry.
 928  */
 929 static int have_duplicate_entry(const unsigned char *sha1,
 930                                 int exclude,
 931                                 uint32_t *index_pos)
 932 {
 933         struct object_entry *entry;
 934
 935         entry = packlist_find(&to_pack, sha1, index_pos);
 936         if (!entry)
 937                 return 0;
 938
 939         if (exclude) {
 940                 if (!entry->preferred_base)
 941                         nr_result--;
 942                 entry->preferred_base = 1;
 943         }
 944
 945         return 1;
 946 }
 947
 948 static int want_found_object(int exclude, struct packed_git *p)
 949 {
 950         if (exclude)
 951                 return 1;
 952         if (incremental)
 953                 return 0;
 954
 955         /*
 956          * When asked to do --local (do not include an object that appears in a
 957          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
 958          * an object that appears in a pack marked with .keep), finding a pack
 959          * that matches the criteria is sufficient for us to decide to omit it.
 960          * However, even if this pack does not satisfy the criteria, we need to
 961          * make sure no copy of this object appears in _any_ pack that makes us
 962          * to omit the object, so we need to check all the packs.
 963          *
 964          * We can however first check whether these options can possible matter;
 965          * if they do not matter we know we want the object in generated pack.
 966          * Otherwise, we signal "-1" at the end to tell the caller that we do
 967          * not know either way, and it needs to check more packs.
 968          */
 969         if (!ignore_packed_keep &&
 970             (!local || !have_non_local_packs))
 971                 return 1;
 972
 973         if (local && !p->pack_local)
 974                 return 0;
 975         if (ignore_packed_keep && p->pack_local && p->pack_keep)
 976                 return 0;
 977
 978         /* we don't know yet; keep looking for more packs */
 979         return -1;
 980 }
 981
 982 /*
 983  * Check whether we want the object in the pack (e.g., we do not want
 984  * objects found in non-local stores if the "--local" option was used).
 985  *
 986  * If the caller already knows an existing pack it wants to take the object
 987  * from, that is passed in *found_pack and *found_offset; otherwise this
 988  * function finds if there is any pack that has the object and returns the pack
 989  * and its offset in these variables.
 990  */
 991 static int want_object_in_pack(const unsigned char *sha1,
 992                                int exclude,
 993                                struct packed_git **found_pack,
 994                                off_t *found_offset)
 995 {
 996         struct packed_git *p;
 997         int want;
 998
 999         if (!exclude && local && has_loose_object_nonlocal(sha1))
1000                 return 0;
1001
1002         /*
1003          * If we already know the pack object lives in, start checks from that
1004          * pack - in the usual case when neither --local was given nor .keep files
1005          * are present we will determine the answer right now.
1006          */
1007         if (*found_pack) {
1008                 want = want_found_object(exclude, *found_pack);
1009                 if (want != -1)
1010                         return want;
1011         }
1012
1013         for (p = packed_git; p; p = p->next) {
1014                 off_t offset;
1015
1016                 if (p == *found_pack)
1017                         offset = *found_offset;
1018                 else
1019                         offset = find_pack_entry_one(sha1, p);
1020
1021                 if (offset) {
1022                         if (!*found_pack) {
1023                                 if (!is_pack_valid(p))
1024                                         continue;
1025                                 *found_offset = offset;
1026                                 *found_pack = p;
1027                         }
1028                         want = want_found_object(exclude, p);
1029                         if (want != -1)
1030                                 return want;
1031                 }
1032         }
1033
1034         return 1;
1035 }
1036
1037 static void create_object_entry(const unsigned char *sha1,
1038                                 enum object_type type,
1039                                 uint32_t hash,
1040                                 int exclude,
1041                                 int no_try_delta,
1042                                 uint32_t index_pos,
1043                                 struct packed_git *found_pack,
1044                                 off_t found_offset)
1045 {
1046         struct object_entry *entry;
1047
1048         entry = packlist_alloc(&to_pack, sha1, index_pos);
1049         entry->hash = hash;
1050         if (type)
1051                 entry->type = type;
1052         if (exclude)
1053                 entry->preferred_base = 1;
1054         else
1055                 nr_result++;
1056         if (found_pack) {
1057                 entry->in_pack = found_pack;
1058                 entry->in_pack_offset = found_offset;
1059         }
1060
1061         entry->no_try_delta = no_try_delta;
1062 }
1063
1064 static const char no_closure_warning[] = N_(
1065 "disabling bitmap writing, as some objects are not being packed"
1066 );
1067
1068 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1069                             const char *name, int exclude)
1070 {
1071         struct packed_git *found_pack = NULL;
1072         off_t found_offset = 0;
1073         uint32_t index_pos;
1074
1075         if (have_duplicate_entry(sha1, exclude, &index_pos))
1076                 return 0;
1077
1078         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1079                 /* The pack is missing an object, so it will not have closure */
1080                 if (write_bitmap_index) {
1081                         warning(_(no_closure_warning));
1082                         write_bitmap_index = 0;
1083                 }
1084                 return 0;
1085         }
1086
1087         create_object_entry(sha1, type, pack_name_hash(name),
1088                             exclude, name && no_try_delta(name),
1089                             index_pos, found_pack, found_offset);
1090
1091         display_progress(progress_state, nr_result);
1092         return 1;
1093 }
1094
1095 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1096                                         enum object_type type,
1097                                         int flags, uint32_t name_hash,
1098                                         struct packed_git *pack, off_t offset)
1099 {
1100         uint32_t index_pos;
1101
1102         if (have_duplicate_entry(sha1, 0, &index_pos))
1103                 return 0;
1104
1105         if (!want_object_in_pack(sha1, 0, &pack, &offset))
1106                 return 0;
1107
1108         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1109
1110         display_progress(progress_state, nr_result);
1111         return 1;
1112 }
1113
1114 struct pbase_tree_cache {
1115         unsigned char sha1[20];
1116         int ref;
1117         int temporary;
1118         void *tree_data;
1119         unsigned long tree_size;
1120 };
1121
1122 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1123 static int pbase_tree_cache_ix(const unsigned char *sha1)
1124 {
1125         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1126 }
1127 static int pbase_tree_cache_ix_incr(int ix)
1128 {
1129         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1130 }
1131
1132 static struct pbase_tree {
1133         struct pbase_tree *next;
1134         /* This is a phony "cache" entry; we are not
1135          * going to evict it or find it through _get()
1136          * mechanism -- this is for the toplevel node that
1137          * would almost always change with any commit.
1138          */
1139         struct pbase_tree_cache pcache;
1140 } *pbase_tree;
1141
1142 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1143 {
1144         struct pbase_tree_cache *ent, *nent;
1145         void *data;
1146         unsigned long size;
1147         enum object_type type;
1148         int neigh;
1149         int my_ix = pbase_tree_cache_ix(sha1);
1150         int available_ix = -1;
1151
1152         /* pbase-tree-cache acts as a limited hashtable.
1153          * your object will be found at your index or within a few
1154          * slots after that slot if it is cached.
1155          */
1156         for (neigh = 0; neigh < 8; neigh++) {
1157                 ent = pbase_tree_cache[my_ix];
1158                 if (ent && !hashcmp(ent->sha1, sha1)) {
1159                         ent->ref++;
1160                         return ent;
1161                 }
1162                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1163                          ((0 <= available_ix) &&
1164                           (!ent && pbase_tree_cache[available_ix])))
1165                         available_ix = my_ix;
1166                 if (!ent)
1167                         break;
1168                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1169         }
1170
1171         /* Did not find one.  Either we got a bogus request or
1172          * we need to read and perhaps cache.
1173          */
1174         data = read_sha1_file(sha1, &type, &size);
1175         if (!data)
1176                 return NULL;
1177         if (type != OBJ_TREE) {
1178                 free(data);
1179                 return NULL;
1180         }
1181
1182         /* We need to either cache or return a throwaway copy */
1183
1184         if (available_ix < 0)
1185                 ent = NULL;
1186         else {
1187                 ent = pbase_tree_cache[available_ix];
1188                 my_ix = available_ix;
1189         }
1190
1191         if (!ent) {
1192                 nent = xmalloc(sizeof(*nent));
1193                 nent->temporary = (available_ix < 0);
1194         }
1195         else {
1196                 /* evict and reuse */
1197                 free(ent->tree_data);
1198                 nent = ent;
1199         }
1200         hashcpy(nent->sha1, sha1);
1201         nent->tree_data = data;
1202         nent->tree_size = size;
1203         nent->ref = 1;
1204         if (!nent->temporary)
1205                 pbase_tree_cache[my_ix] = nent;
1206         return nent;
1207 }
1208
1209 static void pbase_tree_put(struct pbase_tree_cache *cache)
1210 {
1211         if (!cache->temporary) {
1212                 cache->ref--;
1213                 return;
1214         }
1215         free(cache->tree_data);
1216         free(cache);
1217 }
1218
1219 static int name_cmp_len(const char *name)
1220 {
1221         int i;
1222         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1223                 ;
1224         return i;
1225 }
1226
1227 static void add_pbase_object(struct tree_desc *tree,
1228                              const char *name,
1229                              int cmplen,
1230                              const char *fullname)
1231 {
1232         struct name_entry entry;
1233         int cmp;
1234
1235         while (tree_entry(tree,&entry)) {
1236                 if (S_ISGITLINK(entry.mode))
1237                         continue;
1238                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1239                       memcmp(name, entry.path, cmplen);
1240                 if (cmp > 0)
1241                         continue;
1242                 if (cmp < 0)
1243                         return;
1244                 if (name[cmplen] != '/') {
1245                         add_object_entry(entry.oid->hash,
1246                                          object_type(entry.mode),
1247                                          fullname, 1);
1248                         return;
1249                 }
1250                 if (S_ISDIR(entry.mode)) {
1251                         struct tree_desc sub;
1252                         struct pbase_tree_cache *tree;
1253                         const char *down = name+cmplen+1;
1254                         int downlen = name_cmp_len(down);
1255
1256                         tree = pbase_tree_get(entry.oid->hash);
1257                         if (!tree)
1258                                 return;
1259                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1260
1261                         add_pbase_object(&sub, down, downlen, fullname);
1262                         pbase_tree_put(tree);
1263                 }
1264         }
1265 }
1266
1267 static unsigned *done_pbase_paths;
1268 static int done_pbase_paths_num;
1269 static int done_pbase_paths_alloc;
1270 static int done_pbase_path_pos(unsigned hash)
1271 {
1272         int lo = 0;
1273         int hi = done_pbase_paths_num;
1274         while (lo < hi) {
1275                 int mi = (hi + lo) / 2;
1276                 if (done_pbase_paths[mi] == hash)
1277                         return mi;
1278                 if (done_pbase_paths[mi] < hash)
1279                         hi = mi;
1280                 else
1281                         lo = mi + 1;
1282         }
1283         return -lo-1;
1284 }
1285
1286 static int check_pbase_path(unsigned hash)
1287 {
1288         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1289         if (0 <= pos)
1290                 return 1;
1291         pos = -pos - 1;
1292         ALLOC_GROW(done_pbase_paths,
1293                    done_pbase_paths_num + 1,
1294                    done_pbase_paths_alloc);
1295         done_pbase_paths_num++;
1296         if (pos < done_pbase_paths_num)
1297                 memmove(done_pbase_paths + pos + 1,
1298                         done_pbase_paths + pos,
1299                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1300         done_pbase_paths[pos] = hash;
1301         return 0;
1302 }
1303
1304 static void add_preferred_base_object(const char *name)
1305 {
1306         struct pbase_tree *it;
1307         int cmplen;
1308         unsigned hash = pack_name_hash(name);
1309
1310         if (!num_preferred_base || check_pbase_path(hash))
1311                 return;
1312
1313         cmplen = name_cmp_len(name);
1314         for (it = pbase_tree; it; it = it->next) {
1315                 if (cmplen == 0) {
1316                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1317                 }
1318                 else {
1319                         struct tree_desc tree;
1320                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1321                         add_pbase_object(&tree, name, cmplen, name);
1322                 }
1323         }
1324 }
1325
1326 static void add_preferred_base(unsigned char *sha1)
1327 {
1328         struct pbase_tree *it;
1329         void *data;
1330         unsigned long size;
1331         unsigned char tree_sha1[20];
1332
1333         if (window <= num_preferred_base++)
1334                 return;
1335
1336         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1337         if (!data)
1338                 return;
1339
1340         for (it = pbase_tree; it; it = it->next) {
1341                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1342                         free(data);
1343                         return;
1344                 }
1345         }
1346
1347         it = xcalloc(1, sizeof(*it));
1348         it->next = pbase_tree;
1349         pbase_tree = it;
1350
1351         hashcpy(it->pcache.sha1, tree_sha1);
1352         it->pcache.tree_data = data;
1353         it->pcache.tree_size = size;
1354 }
1355
1356 static void cleanup_preferred_base(void)
1357 {
1358         struct pbase_tree *it;
1359         unsigned i;
1360
1361         it = pbase_tree;
1362         pbase_tree = NULL;
1363         while (it) {
1364                 struct pbase_tree *this = it;
1365                 it = this->next;
1366                 free(this->pcache.tree_data);
1367                 free(this);
1368         }
1369
1370         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1371                 if (!pbase_tree_cache[i])
1372                         continue;
1373                 free(pbase_tree_cache[i]->tree_data);
1374                 free(pbase_tree_cache[i]);
1375                 pbase_tree_cache[i] = NULL;
1376         }
1377
1378         free(done_pbase_paths);
1379         done_pbase_paths = NULL;
1380         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1381 }
1382
1383 static void check_object(struct object_entry *entry)
1384 {
1385         if (entry->in_pack) {
1386                 struct packed_git *p = entry->in_pack;
1387                 struct pack_window *w_curs = NULL;
1388                 const unsigned char *base_ref = NULL;
1389                 struct object_entry *base_entry;
1390                 unsigned long used, used_0;
1391                 unsigned long avail;
1392                 off_t ofs;
1393                 unsigned char *buf, c;
1394
1395                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1396
1397                 /*
1398                  * We want in_pack_type even if we do not reuse delta
1399                  * since non-delta representations could still be reused.
1400                  */
1401                 used = unpack_object_header_buffer(buf, avail,
1402                                                    &entry->in_pack_type,
1403                                                    &entry->size);
1404                 if (used == 0)
1405                         goto give_up;
1406
1407                 /*
1408                  * Determine if this is a delta and if so whether we can
1409                  * reuse it or not.  Otherwise let's find out as cheaply as
1410                  * possible what the actual type and size for this object is.
1411                  */
1412                 switch (entry->in_pack_type) {
1413                 default:
1414                         /* Not a delta hence we've already got all we need. */
1415                         entry->type = entry->in_pack_type;
1416                         entry->in_pack_header_size = used;
1417                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1418                                 goto give_up;
1419                         unuse_pack(&w_curs);
1420                         return;
1421                 case OBJ_REF_DELTA:
1422                         if (reuse_delta && !entry->preferred_base)
1423                                 base_ref = use_pack(p, &w_curs,
1424                                                 entry->in_pack_offset + used, NULL);
1425                         entry->in_pack_header_size = used + 20;
1426                         break;
1427                 case OBJ_OFS_DELTA:
1428                         buf = use_pack(p, &w_curs,
1429                                        entry->in_pack_offset + used, NULL);
1430                         used_0 = 0;
1431                         c = buf[used_0++];
1432                         ofs = c & 127;
1433                         while (c & 128) {
1434                                 ofs += 1;
1435                                 if (!ofs || MSB(ofs, 7)) {
1436                                         error("delta base offset overflow in pack for %s",
1437                                               sha1_to_hex(entry->idx.sha1));
1438                                         goto give_up;
1439                                 }
1440                                 c = buf[used_0++];
1441                                 ofs = (ofs << 7) + (c & 127);
1442                         }
1443                         ofs = entry->in_pack_offset - ofs;
1444                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1445                                 error("delta base offset out of bound for %s",
1446                                       sha1_to_hex(entry->idx.sha1));
1447                                 goto give_up;
1448                         }
1449                         if (reuse_delta && !entry->preferred_base) {
1450                                 struct revindex_entry *revidx;
1451                                 revidx = find_pack_revindex(p, ofs);
1452                                 if (!revidx)
1453                                         goto give_up;
1454                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1455                         }
1456                         entry->in_pack_header_size = used + used_0;
1457                         break;
1458                 }
1459
1460                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1461                         /*
1462                          * If base_ref was set above that means we wish to
1463                          * reuse delta data, and we even found that base
1464                          * in the list of objects we want to pack. Goodie!
1465                          *
1466                          * Depth value does not matter - find_deltas() will
1467                          * never consider reused delta as the base object to
1468                          * deltify other objects against, in order to avoid
1469                          * circular deltas.
1470                          */
1471                         entry->type = entry->in_pack_type;
1472                         entry->delta = base_entry;
1473                         entry->delta_size = entry->size;
1474                         entry->delta_sibling = base_entry->delta_child;
1475                         base_entry->delta_child = entry;
1476                         unuse_pack(&w_curs);
1477                         return;
1478                 }
1479
1480                 if (entry->type) {
1481                         /*
1482                          * This must be a delta and we already know what the
1483                          * final object type is.  Let's extract the actual
1484                          * object size from the delta header.
1485                          */
1486                         entry->size = get_size_from_delta(p, &w_curs,
1487                                         entry->in_pack_offset + entry->in_pack_header_size);
1488                         if (entry->size == 0)
1489                                 goto give_up;
1490                         unuse_pack(&w_curs);
1491                         return;
1492                 }
1493
1494                 /*
1495                  * No choice but to fall back to the recursive delta walk
1496                  * with sha1_object_info() to find about the object type
1497                  * at this point...
1498                  */
1499                 give_up:
1500                 unuse_pack(&w_curs);
1501         }
1502
1503         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1504         /*
1505          * The error condition is checked in prepare_pack().  This is
1506          * to permit a missing preferred base object to be ignored
1507          * as a preferred base.  Doing so can result in a larger
1508          * pack file, but the transfer will still take place.
1509          */
1510 }
1511
1512 static int pack_offset_sort(const void *_a, const void *_b)
1513 {
1514         const struct object_entry *a = *(struct object_entry **)_a;
1515         const struct object_entry *b = *(struct object_entry **)_b;
1516
1517         /* avoid filesystem trashing with loose objects */
1518         if (!a->in_pack && !b->in_pack)
1519                 return hashcmp(a->idx.sha1, b->idx.sha1);
1520
1521         if (a->in_pack < b->in_pack)
1522                 return -1;
1523         if (a->in_pack > b->in_pack)
1524                 return 1;
1525         return a->in_pack_offset < b->in_pack_offset ? -1 :
1526                         (a->in_pack_offset > b->in_pack_offset);
1527 }
1528
1529 static void get_object_details(void)
1530 {
1531         uint32_t i;
1532         struct object_entry **sorted_by_offset;
1533
1534         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1535         for (i = 0; i < to_pack.nr_objects; i++)
1536                 sorted_by_offset[i] = to_pack.objects + i;
1537         qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1538
1539         for (i = 0; i < to_pack.nr_objects; i++) {
1540                 struct object_entry *entry = sorted_by_offset[i];
1541                 check_object(entry);
1542                 if (big_file_threshold < entry->size)
1543                         entry->no_try_delta = 1;
1544         }
1545
1546         free(sorted_by_offset);
1547 }
1548
1549 /*
1550  * We search for deltas in a list sorted by type, by filename hash, and then
1551  * by size, so that we see progressively smaller and smaller files.
1552  * That's because we prefer deltas to be from the bigger file
1553  * to the smaller -- deletes are potentially cheaper, but perhaps
1554  * more importantly, the bigger file is likely the more recent
1555  * one.  The deepest deltas are therefore the oldest objects which are
1556  * less susceptible to be accessed often.
1557  */
1558 static int type_size_sort(const void *_a, const void *_b)
1559 {
1560         const struct object_entry *a = *(struct object_entry **)_a;
1561         const struct object_entry *b = *(struct object_entry **)_b;
1562
1563         if (a->type > b->type)
1564                 return -1;
1565         if (a->type < b->type)
1566                 return 1;
1567         if (a->hash > b->hash)
1568                 return -1;
1569         if (a->hash < b->hash)
1570                 return 1;
1571         if (a->preferred_base > b->preferred_base)
1572                 return -1;
1573         if (a->preferred_base < b->preferred_base)
1574                 return 1;
1575         if (a->size > b->size)
1576                 return -1;
1577         if (a->size < b->size)
1578                 return 1;
1579         return a < b ? -1 : (a > b);  /* newest first */
1580 }
1581
1582 struct unpacked {
1583         struct object_entry *entry;
1584         void *data;
1585         struct delta_index *index;
1586         unsigned depth;
1587 };
1588
1589 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1590                            unsigned long delta_size)
1591 {
1592         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1593                 return 0;
1594
1595         if (delta_size < cache_max_small_delta_size)
1596                 return 1;
1597
1598         /* cache delta, if objects are large enough compared to delta size */
1599         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1600                 return 1;
1601
1602         return 0;
1603 }
1604
1605 #ifndef NO_PTHREADS
1606
1607 static pthread_mutex_t read_mutex;
1608 #define read_lock()             pthread_mutex_lock(&read_mutex)
1609 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1610
1611 static pthread_mutex_t cache_mutex;
1612 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1613 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1614
1615 static pthread_mutex_t progress_mutex;
1616 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1617 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1618
1619 #else
1620
1621 #define read_lock()             (void)0
1622 #define read_unlock()           (void)0
1623 #define cache_lock()            (void)0
1624 #define cache_unlock()          (void)0
1625 #define progress_lock()         (void)0
1626 #define progress_unlock()       (void)0
1627
1628 #endif
1629
1630 static int try_delta(struct unpacked *trg, struct unpacked *src,
1631                      unsigned max_depth, unsigned long *mem_usage)
1632 {
1633         struct object_entry *trg_entry = trg->entry;
1634         struct object_entry *src_entry = src->entry;
1635         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1636         unsigned ref_depth;
1637         enum object_type type;
1638         void *delta_buf;
1639
1640         /* Don't bother doing diffs between different types */
1641         if (trg_entry->type != src_entry->type)
1642                 return -1;
1643
1644         /*
1645          * We do not bother to try a delta that we discarded on an
1646          * earlier try, but only when reusing delta data.  Note that
1647          * src_entry that is marked as the preferred_base should always
1648          * be considered, as even if we produce a suboptimal delta against
1649          * it, we will still save the transfer cost, as we already know
1650          * the other side has it and we won't send src_entry at all.
1651          */
1652         if (reuse_delta && trg_entry->in_pack &&
1653             trg_entry->in_pack == src_entry->in_pack &&
1654             !src_entry->preferred_base &&
1655             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1656             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1657                 return 0;
1658
1659         /* Let's not bust the allowed depth. */
1660         if (src->depth >= max_depth)
1661                 return 0;
1662
1663         /* Now some size filtering heuristics. */
1664         trg_size = trg_entry->size;
1665         if (!trg_entry->delta) {
1666                 max_size = trg_size/2 - 20;
1667                 ref_depth = 1;
1668         } else {
1669                 max_size = trg_entry->delta_size;
1670                 ref_depth = trg->depth;
1671         }
1672         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1673                                                 (max_depth - ref_depth + 1);
1674         if (max_size == 0)
1675                 return 0;
1676         src_size = src_entry->size;
1677         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1678         if (sizediff >= max_size)
1679                 return 0;
1680         if (trg_size < src_size / 32)
1681                 return 0;
1682
1683         /* Load data if not already done */
1684         if (!trg->data) {
1685                 read_lock();
1686                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1687                 read_unlock();
1688                 if (!trg->data)
1689                         die("object %s cannot be read",
1690                             sha1_to_hex(trg_entry->idx.sha1));
1691                 if (sz != trg_size)
1692                         die("object %s inconsistent object length (%lu vs %lu)",
1693                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1694                 *mem_usage += sz;
1695         }
1696         if (!src->data) {
1697                 read_lock();
1698                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1699                 read_unlock();
1700                 if (!src->data) {
1701                         if (src_entry->preferred_base) {
1702                                 static int warned = 0;
1703                                 if (!warned++)
1704                                         warning("object %s cannot be read",
1705                                                 sha1_to_hex(src_entry->idx.sha1));
1706                                 /*
1707                                  * Those objects are not included in the
1708                                  * resulting pack.  Be resilient and ignore
1709                                  * them if they can't be read, in case the
1710                                  * pack could be created nevertheless.
1711                                  */
1712                                 return 0;
1713                         }
1714                         die("object %s cannot be read",
1715                             sha1_to_hex(src_entry->idx.sha1));
1716                 }
1717                 if (sz != src_size)
1718                         die("object %s inconsistent object length (%lu vs %lu)",
1719                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1720                 *mem_usage += sz;
1721         }
1722         if (!src->index) {
1723                 src->index = create_delta_index(src->data, src_size);
1724                 if (!src->index) {
1725                         static int warned = 0;
1726                         if (!warned++)
1727                                 warning("suboptimal pack - out of memory");
1728                         return 0;
1729                 }
1730                 *mem_usage += sizeof_delta_index(src->index);
1731         }
1732
1733         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1734         if (!delta_buf)
1735                 return 0;
1736
1737         if (trg_entry->delta) {
1738                 /* Prefer only shallower same-sized deltas. */
1739                 if (delta_size == trg_entry->delta_size &&
1740                     src->depth + 1 >= trg->depth) {
1741                         free(delta_buf);
1742                         return 0;
1743                 }
1744         }
1745
1746         /*
1747          * Handle memory allocation outside of the cache
1748          * accounting lock.  Compiler will optimize the strangeness
1749          * away when NO_PTHREADS is defined.
1750          */
1751         free(trg_entry->delta_data);
1752         cache_lock();
1753         if (trg_entry->delta_data) {
1754                 delta_cache_size -= trg_entry->delta_size;
1755                 trg_entry->delta_data = NULL;
1756         }
1757         if (delta_cacheable(src_size, trg_size, delta_size)) {
1758                 delta_cache_size += delta_size;
1759                 cache_unlock();
1760                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1761         } else {
1762                 cache_unlock();
1763                 free(delta_buf);
1764         }
1765
1766         trg_entry->delta = src_entry;
1767         trg_entry->delta_size = delta_size;
1768         trg->depth = src->depth + 1;
1769
1770         return 1;
1771 }
1772
1773 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1774 {
1775         struct object_entry *child = me->delta_child;
1776         unsigned int m = n;
1777         while (child) {
1778                 unsigned int c = check_delta_limit(child, n + 1);
1779                 if (m < c)
1780                         m = c;
1781                 child = child->delta_sibling;
1782         }
1783         return m;
1784 }
1785
1786 static unsigned long free_unpacked(struct unpacked *n)
1787 {
1788         unsigned long freed_mem = sizeof_delta_index(n->index);
1789         free_delta_index(n->index);
1790         n->index = NULL;
1791         if (n->data) {
1792                 freed_mem += n->entry->size;
1793                 free(n->data);
1794                 n->data = NULL;
1795         }
1796         n->entry = NULL;
1797         n->depth = 0;
1798         return freed_mem;
1799 }
1800
1801 static void find_deltas(struct object_entry **list, unsigned *list_size,
1802                         int window, int depth, unsigned *processed)
1803 {
1804         uint32_t i, idx = 0, count = 0;
1805         struct unpacked *array;
1806         unsigned long mem_usage = 0;
1807
1808         array = xcalloc(window, sizeof(struct unpacked));
1809
1810         for (;;) {
1811                 struct object_entry *entry;
1812                 struct unpacked *n = array + idx;
1813                 int j, max_depth, best_base = -1;
1814
1815                 progress_lock();
1816                 if (!*list_size) {
1817                         progress_unlock();
1818                         break;
1819                 }
1820                 entry = *list++;
1821                 (*list_size)--;
1822                 if (!entry->preferred_base) {
1823                         (*processed)++;
1824                         display_progress(progress_state, *processed);
1825                 }
1826                 progress_unlock();
1827
1828                 mem_usage -= free_unpacked(n);
1829                 n->entry = entry;
1830
1831                 while (window_memory_limit &&
1832                        mem_usage > window_memory_limit &&
1833                        count > 1) {
1834                         uint32_t tail = (idx + window - count) % window;
1835                         mem_usage -= free_unpacked(array + tail);
1836                         count--;
1837                 }
1838
1839                 /* We do not compute delta to *create* objects we are not
1840                  * going to pack.
1841                  */
1842                 if (entry->preferred_base)
1843                         goto next;
1844
1845                 /*
1846                  * If the current object is at pack edge, take the depth the
1847                  * objects that depend on the current object into account
1848                  * otherwise they would become too deep.
1849                  */
1850                 max_depth = depth;
1851                 if (entry->delta_child) {
1852                         max_depth -= check_delta_limit(entry, 0);
1853                         if (max_depth <= 0)
1854                                 goto next;
1855                 }
1856
1857                 j = window;
1858                 while (--j > 0) {
1859                         int ret;
1860                         uint32_t other_idx = idx + j;
1861                         struct unpacked *m;
1862                         if (other_idx >= window)
1863                                 other_idx -= window;
1864                         m = array + other_idx;
1865                         if (!m->entry)
1866                                 break;
1867                         ret = try_delta(n, m, max_depth, &mem_usage);
1868                         if (ret < 0)
1869                                 break;
1870                         else if (ret > 0)
1871                                 best_base = other_idx;
1872                 }
1873
1874                 /*
1875                  * If we decided to cache the delta data, then it is best
1876                  * to compress it right away.  First because we have to do
1877                  * it anyway, and doing it here while we're threaded will
1878                  * save a lot of time in the non threaded write phase,
1879                  * as well as allow for caching more deltas within
1880                  * the same cache size limit.
1881                  * ...
1882                  * But only if not writing to stdout, since in that case
1883                  * the network is most likely throttling writes anyway,
1884                  * and therefore it is best to go to the write phase ASAP
1885                  * instead, as we can afford spending more time compressing
1886                  * between writes at that moment.
1887                  */
1888                 if (entry->delta_data && !pack_to_stdout) {
1889                         entry->z_delta_size = do_compress(&entry->delta_data,
1890                                                           entry->delta_size);
1891                         cache_lock();
1892                         delta_cache_size -= entry->delta_size;
1893                         delta_cache_size += entry->z_delta_size;
1894                         cache_unlock();
1895                 }
1896
1897                 /* if we made n a delta, and if n is already at max
1898                  * depth, leaving it in the window is pointless.  we
1899                  * should evict it first.
1900                  */
1901                 if (entry->delta && max_depth <= n->depth)
1902                         continue;
1903
1904                 /*
1905                  * Move the best delta base up in the window, after the
1906                  * currently deltified object, to keep it longer.  It will
1907                  * be the first base object to be attempted next.
1908                  */
1909                 if (entry->delta) {
1910                         struct unpacked swap = array[best_base];
1911                         int dist = (window + idx - best_base) % window;
1912                         int dst = best_base;
1913                         while (dist--) {
1914                                 int src = (dst + 1) % window;
1915                                 array[dst] = array[src];
1916                                 dst = src;
1917                         }
1918                         array[dst] = swap;
1919                 }
1920
1921                 next:
1922                 idx++;
1923                 if (count + 1 < window)
1924                         count++;
1925                 if (idx >= window)
1926                         idx = 0;
1927         }
1928
1929         for (i = 0; i < window; ++i) {
1930                 free_delta_index(array[i].index);
1931                 free(array[i].data);
1932         }
1933         free(array);
1934 }
1935
1936 #ifndef NO_PTHREADS
1937
1938 static void try_to_free_from_threads(size_t size)
1939 {
1940         read_lock();
1941         release_pack_memory(size);
1942         read_unlock();
1943 }
1944
1945 static try_to_free_t old_try_to_free_routine;
1946
1947 /*
1948  * The main thread waits on the condition that (at least) one of the workers
1949  * has stopped working (which is indicated in the .working member of
1950  * struct thread_params).
1951  * When a work thread has completed its work, it sets .working to 0 and
1952  * signals the main thread and waits on the condition that .data_ready
1953  * becomes 1.
1954  */
1955
1956 struct thread_params {
1957         pthread_t thread;
1958         struct object_entry **list;
1959         unsigned list_size;
1960         unsigned remaining;
1961         int window;
1962         int depth;
1963         int working;
1964         int data_ready;
1965         pthread_mutex_t mutex;
1966         pthread_cond_t cond;
1967         unsigned *processed;
1968 };
1969
1970 static pthread_cond_t progress_cond;
1971
1972 /*
1973  * Mutex and conditional variable can't be statically-initialized on Windows.
1974  */
1975 static void init_threaded_search(void)
1976 {
1977         init_recursive_mutex(&read_mutex);
1978         pthread_mutex_init(&cache_mutex, NULL);
1979         pthread_mutex_init(&progress_mutex, NULL);
1980         pthread_cond_init(&progress_cond, NULL);
1981         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1982 }
1983
1984 static void cleanup_threaded_search(void)
1985 {
1986         set_try_to_free_routine(old_try_to_free_routine);
1987         pthread_cond_destroy(&progress_cond);
1988         pthread_mutex_destroy(&read_mutex);
1989         pthread_mutex_destroy(&cache_mutex);
1990         pthread_mutex_destroy(&progress_mutex);
1991 }
1992
1993 static void *threaded_find_deltas(void *arg)
1994 {
1995         struct thread_params *me = arg;
1996
1997         while (me->remaining) {
1998                 find_deltas(me->list, &me->remaining,
1999                             me->window, me->depth, me->processed);
2000
2001                 progress_lock();
2002                 me->working = 0;
2003                 pthread_cond_signal(&progress_cond);
2004                 progress_unlock();
2005
2006                 /*
2007                  * We must not set ->data_ready before we wait on the
2008                  * condition because the main thread may have set it to 1
2009                  * before we get here. In order to be sure that new
2010                  * work is available if we see 1 in ->data_ready, it
2011                  * was initialized to 0 before this thread was spawned
2012                  * and we reset it to 0 right away.
2013                  */
2014                 pthread_mutex_lock(&me->mutex);
2015                 while (!me->data_ready)
2016                         pthread_cond_wait(&me->cond, &me->mutex);
2017                 me->data_ready = 0;
2018                 pthread_mutex_unlock(&me->mutex);
2019         }
2020         /* leave ->working 1 so that this doesn't get more work assigned */
2021         return NULL;
2022 }
2023
2024 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2025                            int window, int depth, unsigned *processed)
2026 {
2027         struct thread_params *p;
2028         int i, ret, active_threads = 0;
2029
2030         init_threaded_search();
2031
2032         if (delta_search_threads <= 1) {
2033                 find_deltas(list, &list_size, window, depth, processed);
2034                 cleanup_threaded_search();
2035                 return;
2036         }
2037         if (progress > pack_to_stdout)
2038                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2039                                 delta_search_threads);
2040         p = xcalloc(delta_search_threads, sizeof(*p));
2041
2042         /* Partition the work amongst work threads. */
2043         for (i = 0; i < delta_search_threads; i++) {
2044                 unsigned sub_size = list_size / (delta_search_threads - i);
2045
2046                 /* don't use too small segments or no deltas will be found */
2047                 if (sub_size < 2*window && i+1 < delta_search_threads)
2048                         sub_size = 0;
2049
2050                 p[i].window = window;
2051                 p[i].depth = depth;
2052                 p[i].processed = processed;
2053                 p[i].working = 1;
2054                 p[i].data_ready = 0;
2055
2056                 /* try to split chunks on "path" boundaries */
2057                 while (sub_size && sub_size < list_size &&
2058                        list[sub_size]->hash &&
2059                        list[sub_size]->hash == list[sub_size-1]->hash)
2060                         sub_size++;
2061
2062                 p[i].list = list;
2063                 p[i].list_size = sub_size;
2064                 p[i].remaining = sub_size;
2065
2066                 list += sub_size;
2067                 list_size -= sub_size;
2068         }
2069
2070         /* Start work threads. */
2071         for (i = 0; i < delta_search_threads; i++) {
2072                 if (!p[i].list_size)
2073                         continue;
2074                 pthread_mutex_init(&p[i].mutex, NULL);
2075                 pthread_cond_init(&p[i].cond, NULL);
2076                 ret = pthread_create(&p[i].thread, NULL,
2077                                      threaded_find_deltas, &p[i]);
2078                 if (ret)
2079                         die("unable to create thread: %s", strerror(ret));
2080                 active_threads++;
2081         }
2082
2083         /*
2084          * Now let's wait for work completion.  Each time a thread is done
2085          * with its work, we steal half of the remaining work from the
2086          * thread with the largest number of unprocessed objects and give
2087          * it to that newly idle thread.  This ensure good load balancing
2088          * until the remaining object list segments are simply too short
2089          * to be worth splitting anymore.
2090          */
2091         while (active_threads) {
2092                 struct thread_params *target = NULL;
2093                 struct thread_params *victim = NULL;
2094                 unsigned sub_size = 0;
2095
2096                 progress_lock();
2097                 for (;;) {
2098                         for (i = 0; !target && i < delta_search_threads; i++)
2099                                 if (!p[i].working)
2100                                         target = &p[i];
2101                         if (target)
2102                                 break;
2103                         pthread_cond_wait(&progress_cond, &progress_mutex);
2104                 }
2105
2106                 for (i = 0; i < delta_search_threads; i++)
2107                         if (p[i].remaining > 2*window &&
2108                             (!victim || victim->remaining < p[i].remaining))
2109                                 victim = &p[i];
2110                 if (victim) {
2111                         sub_size = victim->remaining / 2;
2112                         list = victim->list + victim->list_size - sub_size;
2113                         while (sub_size && list[0]->hash &&
2114                                list[0]->hash == list[-1]->hash) {
2115                                 list++;
2116                                 sub_size--;
2117                         }
2118                         if (!sub_size) {
2119                                 /*
2120                                  * It is possible for some "paths" to have
2121                                  * so many objects that no hash boundary
2122                                  * might be found.  Let's just steal the
2123                                  * exact half in that case.
2124                                  */
2125                                 sub_size = victim->remaining / 2;
2126                                 list -= sub_size;
2127                         }
2128                         target->list = list;
2129                         victim->list_size -= sub_size;
2130                         victim->remaining -= sub_size;
2131                 }
2132                 target->list_size = sub_size;
2133                 target->remaining = sub_size;
2134                 target->working = 1;
2135                 progress_unlock();
2136
2137                 pthread_mutex_lock(&target->mutex);
2138                 target->data_ready = 1;
2139                 pthread_cond_signal(&target->cond);
2140                 pthread_mutex_unlock(&target->mutex);
2141
2142                 if (!sub_size) {
2143                         pthread_join(target->thread, NULL);
2144                         pthread_cond_destroy(&target->cond);
2145                         pthread_mutex_destroy(&target->mutex);
2146                         active_threads--;
2147                 }
2148         }
2149         cleanup_threaded_search();
2150         free(p);
2151 }
2152
2153 #else
2154 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2155 #endif
2156
2157 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2158 {
2159         struct object_id peeled;
2160
2161         if (starts_with(path, "refs/tags/") && /* is a tag? */
2162             !peel_ref(path, peeled.hash)    && /* peelable? */
2163             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2164                 add_object_entry(oid->hash, OBJ_TAG, NULL, 0);
2165         return 0;
2166 }
2167
2168 static void prepare_pack(int window, int depth)
2169 {
2170         struct object_entry **delta_list;
2171         uint32_t i, nr_deltas;
2172         unsigned n;
2173
2174         get_object_details();
2175
2176         /*
2177          * If we're locally repacking then we need to be doubly careful
2178          * from now on in order to make sure no stealth corruption gets
2179          * propagated to the new pack.  Clients receiving streamed packs
2180          * should validate everything they get anyway so no need to incur
2181          * the additional cost here in that case.
2182          */
2183         if (!pack_to_stdout)
2184                 do_check_packed_object_crc = 1;
2185
2186         if (!to_pack.nr_objects || !window || !depth)
2187                 return;
2188
2189         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2190         nr_deltas = n = 0;
2191
2192         for (i = 0; i < to_pack.nr_objects; i++) {
2193                 struct object_entry *entry = to_pack.objects + i;
2194
2195                 if (entry->delta)
2196                         /* This happens if we decided to reuse existing
2197                          * delta from a pack.  "reuse_delta &&" is implied.
2198                          */
2199                         continue;
2200
2201                 if (entry->size < 50)
2202                         continue;
2203
2204                 if (entry->no_try_delta)
2205                         continue;
2206
2207                 if (!entry->preferred_base) {
2208                         nr_deltas++;
2209                         if (entry->type < 0)
2210                                 die("unable to get type of object %s",
2211                                     sha1_to_hex(entry->idx.sha1));
2212                 } else {
2213                         if (entry->type < 0) {
2214                                 /*
2215                                  * This object is not found, but we
2216                                  * don't have to include it anyway.
2217                                  */
2218                                 continue;
2219                         }
2220                 }
2221
2222                 delta_list[n++] = entry;
2223         }
2224
2225         if (nr_deltas && n > 1) {
2226                 unsigned nr_done = 0;
2227                 if (progress)
2228                         progress_state = start_progress(_("Compressing objects"),
2229                                                         nr_deltas);
2230                 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2231                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2232                 stop_progress(&progress_state);
2233                 if (nr_done != nr_deltas)
2234                         die("inconsistency with delta count");
2235         }
2236         free(delta_list);
2237 }
2238
2239 static int git_pack_config(const char *k, const char *v, void *cb)
2240 {
2241         if (!strcmp(k, "pack.window")) {
2242                 window = git_config_int(k, v);
2243                 return 0;
2244         }
2245         if (!strcmp(k, "pack.windowmemory")) {
2246                 window_memory_limit = git_config_ulong(k, v);
2247                 return 0;
2248         }
2249         if (!strcmp(k, "pack.depth")) {
2250                 depth = git_config_int(k, v);
2251                 return 0;
2252         }
2253         if (!strcmp(k, "pack.compression")) {
2254                 int level = git_config_int(k, v);
2255                 if (level == -1)
2256                         level = Z_DEFAULT_COMPRESSION;
2257                 else if (level < 0 || level > Z_BEST_COMPRESSION)
2258                         die("bad pack compression level %d", level);
2259                 pack_compression_level = level;
2260                 pack_compression_seen = 1;
2261                 return 0;
2262         }
2263         if (!strcmp(k, "pack.deltacachesize")) {
2264                 max_delta_cache_size = git_config_int(k, v);
2265                 return 0;
2266         }
2267         if (!strcmp(k, "pack.deltacachelimit")) {
2268                 cache_max_small_delta_size = git_config_int(k, v);
2269                 return 0;
2270         }
2271         if (!strcmp(k, "pack.writebitmaphashcache")) {
2272                 if (git_config_bool(k, v))
2273                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2274                 else
2275                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2276         }
2277         if (!strcmp(k, "pack.usebitmaps")) {
2278                 use_bitmap_index_default = git_config_bool(k, v);
2279                 return 0;
2280         }
2281         if (!strcmp(k, "pack.threads")) {
2282                 delta_search_threads = git_config_int(k, v);
2283                 if (delta_search_threads < 0)
2284                         die("invalid number of threads specified (%d)",
2285                             delta_search_threads);
2286 #ifdef NO_PTHREADS
2287                 if (delta_search_threads != 1)
2288                         warning("no threads support, ignoring %s", k);
2289 #endif
2290                 return 0;
2291         }
2292         if (!strcmp(k, "pack.indexversion")) {
2293                 pack_idx_opts.version = git_config_int(k, v);
2294                 if (pack_idx_opts.version > 2)
2295                         die("bad pack.indexversion=%"PRIu32,
2296                             pack_idx_opts.version);
2297                 return 0;
2298         }
2299         return git_default_config(k, v, cb);
2300 }
2301
2302 static void read_object_list_from_stdin(void)
2303 {
2304         char line[40 + 1 + PATH_MAX + 2];
2305         unsigned char sha1[20];
2306
2307         for (;;) {
2308                 if (!fgets(line, sizeof(line), stdin)) {
2309                         if (feof(stdin))
2310                                 break;
2311                         if (!ferror(stdin))
2312                                 die("fgets returned NULL, not EOF, not error!");
2313                         if (errno != EINTR)
2314                                 die_errno("fgets");
2315                         clearerr(stdin);
2316                         continue;
2317                 }
2318                 if (line[0] == '-') {
2319                         if (get_sha1_hex(line+1, sha1))
2320                                 die("expected edge sha1, got garbage:\n %s",
2321                                     line);
2322                         add_preferred_base(sha1);
2323                         continue;
2324                 }
2325                 if (get_sha1_hex(line, sha1))
2326                         die("expected sha1, got garbage:\n %s", line);
2327
2328                 add_preferred_base_object(line+41);
2329                 add_object_entry(sha1, 0, line+41, 0);
2330         }
2331 }
2332
2333 #define OBJECT_ADDED (1u<<20)
2334
2335 static void show_commit(struct commit *commit, void *data)
2336 {
2337         add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2338         commit->object.flags |= OBJECT_ADDED;
2339
2340         if (write_bitmap_index)
2341                 index_commit_for_bitmap(commit);
2342 }
2343
2344 static void show_object(struct object *obj, const char *name, void *data)
2345 {
2346         add_preferred_base_object(name);
2347         add_object_entry(obj->oid.hash, obj->type, name, 0);
2348         obj->flags |= OBJECT_ADDED;
2349 }
2350
2351 static void show_edge(struct commit *commit)
2352 {
2353         add_preferred_base(commit->object.oid.hash);
2354 }
2355
2356 struct in_pack_object {
2357         off_t offset;
2358         struct object *object;
2359 };
2360
2361 struct in_pack {
2362         int alloc;
2363         int nr;
2364         struct in_pack_object *array;
2365 };
2366
2367 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2368 {
2369         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2370         in_pack->array[in_pack->nr].object = object;
2371         in_pack->nr++;
2372 }
2373
2374 /*
2375  * Compare the objects in the offset order, in order to emulate the
2376  * "git rev-list --objects" output that produced the pack originally.
2377  */
2378 static int ofscmp(const void *a_, const void *b_)
2379 {
2380         struct in_pack_object *a = (struct in_pack_object *)a_;
2381         struct in_pack_object *b = (struct in_pack_object *)b_;
2382
2383         if (a->offset < b->offset)
2384                 return -1;
2385         else if (a->offset > b->offset)
2386                 return 1;
2387         else
2388                 return oidcmp(&a->object->oid, &b->object->oid);
2389 }
2390
2391 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2392 {
2393         struct packed_git *p;
2394         struct in_pack in_pack;
2395         uint32_t i;
2396
2397         memset(&in_pack, 0, sizeof(in_pack));
2398
2399         for (p = packed_git; p; p = p->next) {
2400                 const unsigned char *sha1;
2401                 struct object *o;
2402
2403                 if (!p->pack_local || p->pack_keep)
2404                         continue;
2405                 if (open_pack_index(p))
2406                         die("cannot open pack index");
2407
2408                 ALLOC_GROW(in_pack.array,
2409                            in_pack.nr + p->num_objects,
2410                            in_pack.alloc);
2411
2412                 for (i = 0; i < p->num_objects; i++) {
2413                         sha1 = nth_packed_object_sha1(p, i);
2414                         o = lookup_unknown_object(sha1);
2415                         if (!(o->flags & OBJECT_ADDED))
2416                                 mark_in_pack_object(o, p, &in_pack);
2417                         o->flags |= OBJECT_ADDED;
2418                 }
2419         }
2420
2421         if (in_pack.nr) {
2422                 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2423                       ofscmp);
2424                 for (i = 0; i < in_pack.nr; i++) {
2425                         struct object *o = in_pack.array[i].object;
2426                         add_object_entry(o->oid.hash, o->type, "", 0);
2427                 }
2428         }
2429         free(in_pack.array);
2430 }
2431
2432 static int add_loose_object(const unsigned char *sha1, const char *path,
2433                             void *data)
2434 {
2435         enum object_type type = sha1_object_info(sha1, NULL);
2436
2437         if (type < 0) {
2438                 warning("loose object at %s could not be examined", path);
2439                 return 0;
2440         }
2441
2442         add_object_entry(sha1, type, "", 0);
2443         return 0;
2444 }
2445
2446 /*
2447  * We actually don't even have to worry about reachability here.
2448  * add_object_entry will weed out duplicates, so we just add every
2449  * loose object we find.
2450  */
2451 static void add_unreachable_loose_objects(void)
2452 {
2453         for_each_loose_file_in_objdir(get_object_directory(),
2454                                       add_loose_object,
2455                                       NULL, NULL, NULL);
2456 }
2457
2458 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2459 {
2460         static struct packed_git *last_found = (void *)1;
2461         struct packed_git *p;
2462
2463         p = (last_found != (void *)1) ? last_found : packed_git;
2464
2465         while (p) {
2466                 if ((!p->pack_local || p->pack_keep) &&
2467                         find_pack_entry_one(sha1, p)) {
2468                         last_found = p;
2469                         return 1;
2470                 }
2471                 if (p == last_found)
2472                         p = packed_git;
2473                 else
2474                         p = p->next;
2475                 if (p == last_found)
2476                         p = p->next;
2477         }
2478         return 0;
2479 }
2480
2481 /*
2482  * Store a list of sha1s that are should not be discarded
2483  * because they are either written too recently, or are
2484  * reachable from another object that was.
2485  *
2486  * This is filled by get_object_list.
2487  */
2488 static struct sha1_array recent_objects;
2489
2490 static int loosened_object_can_be_discarded(const unsigned char *sha1,
2491                                             unsigned long mtime)
2492 {
2493         if (!unpack_unreachable_expiration)
2494                 return 0;
2495         if (mtime > unpack_unreachable_expiration)
2496                 return 0;
2497         if (sha1_array_lookup(&recent_objects, sha1) >= 0)
2498                 return 0;
2499         return 1;
2500 }
2501
2502 static void loosen_unused_packed_objects(struct rev_info *revs)
2503 {
2504         struct packed_git *p;
2505         uint32_t i;
2506         const unsigned char *sha1;
2507
2508         for (p = packed_git; p; p = p->next) {
2509                 if (!p->pack_local || p->pack_keep)
2510                         continue;
2511
2512                 if (open_pack_index(p))
2513                         die("cannot open pack index");
2514
2515                 for (i = 0; i < p->num_objects; i++) {
2516                         sha1 = nth_packed_object_sha1(p, i);
2517                         if (!packlist_find(&to_pack, sha1, NULL) &&
2518                             !has_sha1_pack_kept_or_nonlocal(sha1) &&
2519                             !loosened_object_can_be_discarded(sha1, p->mtime))
2520                                 if (force_object_loose(sha1, p->mtime))
2521                                         die("unable to force loose object");
2522                 }
2523         }
2524 }
2525
2526 /*
2527  * This tracks any options which pack-reuse code expects to be on, or which a
2528  * reader of the pack might not understand, and which would therefore prevent
2529  * blind reuse of what we have on disk.
2530  */
2531 static int pack_options_allow_reuse(void)
2532 {
2533         return pack_to_stdout && allow_ofs_delta;
2534 }
2535
2536 static int get_object_list_from_bitmap(struct rev_info *revs)
2537 {
2538         if (prepare_bitmap_walk(revs) < 0)
2539                 return -1;
2540
2541         if (pack_options_allow_reuse() &&
2542             !reuse_partial_packfile_from_bitmap(
2543                         &reuse_packfile,
2544                         &reuse_packfile_objects,
2545                         &reuse_packfile_offset)) {
2546                 assert(reuse_packfile_objects);
2547                 nr_result += reuse_packfile_objects;
2548                 display_progress(progress_state, nr_result);
2549         }
2550
2551         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2552         return 0;
2553 }
2554
2555 static void record_recent_object(struct object *obj,
2556                                  const char *name,
2557                                  void *data)
2558 {
2559         sha1_array_append(&recent_objects, obj->oid.hash);
2560 }
2561
2562 static void record_recent_commit(struct commit *commit, void *data)
2563 {
2564         sha1_array_append(&recent_objects, commit->object.oid.hash);
2565 }
2566
2567 static void get_object_list(int ac, const char **av)
2568 {
2569         struct rev_info revs;
2570         char line[1000];
2571         int flags = 0;
2572
2573         init_revisions(&revs, NULL);
2574         save_commit_buffer = 0;
2575         setup_revisions(ac, av, &revs, NULL);
2576
2577         /* make sure shallows are read */
2578         is_repository_shallow();
2579
2580         while (fgets(line, sizeof(line), stdin) != NULL) {
2581                 int len = strlen(line);
2582                 if (len && line[len - 1] == '\n')
2583                         line[--len] = 0;
2584                 if (!len)
2585                         break;
2586                 if (*line == '-') {
2587                         if (!strcmp(line, "--not")) {
2588                                 flags ^= UNINTERESTING;
2589                                 write_bitmap_index = 0;
2590                                 continue;
2591                         }
2592                         if (starts_with(line, "--shallow ")) {
2593                                 unsigned char sha1[20];
2594                                 if (get_sha1_hex(line + 10, sha1))
2595                                         die("not an SHA-1 '%s'", line + 10);
2596                                 register_shallow(sha1);
2597                                 use_bitmap_index = 0;
2598                                 continue;
2599                         }
2600                         die("not a rev '%s'", line);
2601                 }
2602                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2603                         die("bad revision '%s'", line);
2604         }
2605
2606         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2607                 return;
2608
2609         if (prepare_revision_walk(&revs))
2610                 die("revision walk setup failed");
2611         mark_edges_uninteresting(&revs, show_edge);
2612         traverse_commit_list(&revs, show_commit, show_object, NULL);
2613
2614         if (unpack_unreachable_expiration) {
2615                 revs.ignore_missing_links = 1;
2616                 if (add_unseen_recent_objects_to_traversal(&revs,
2617                                 unpack_unreachable_expiration))
2618                         die("unable to add recent objects");
2619                 if (prepare_revision_walk(&revs))
2620                         die("revision walk setup failed");
2621                 traverse_commit_list(&revs, record_recent_commit,
2622                                      record_recent_object, NULL);
2623         }
2624
2625         if (keep_unreachable)
2626                 add_objects_in_unpacked_packs(&revs);
2627         if (pack_loose_unreachable)
2628                 add_unreachable_loose_objects();
2629         if (unpack_unreachable)
2630                 loosen_unused_packed_objects(&revs);
2631
2632         sha1_array_clear(&recent_objects);
2633 }
2634
2635 static int option_parse_index_version(const struct option *opt,
2636                                       const char *arg, int unset)
2637 {
2638         char *c;
2639         const char *val = arg;
2640         pack_idx_opts.version = strtoul(val, &c, 10);
2641         if (pack_idx_opts.version > 2)
2642                 die(_("unsupported index version %s"), val);
2643         if (*c == ',' && c[1])
2644                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2645         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2646                 die(_("bad index version '%s'"), val);
2647         return 0;
2648 }
2649
2650 static int option_parse_unpack_unreachable(const struct option *opt,
2651                                            const char *arg, int unset)
2652 {
2653         if (unset) {
2654                 unpack_unreachable = 0;
2655                 unpack_unreachable_expiration = 0;
2656         }
2657         else {
2658                 unpack_unreachable = 1;
2659                 if (arg)
2660                         unpack_unreachable_expiration = approxidate(arg);
2661         }
2662         return 0;
2663 }
2664
2665 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2666 {
2667         int use_internal_rev_list = 0;
2668         int thin = 0;
2669         int shallow = 0;
2670         int all_progress_implied = 0;
2671         struct argv_array rp = ARGV_ARRAY_INIT;
2672         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2673         int rev_list_index = 0;
2674         struct option pack_objects_options[] = {
2675                 OPT_SET_INT('q', "quiet", &progress,
2676                             N_("do not show progress meter"), 0),
2677                 OPT_SET_INT(0, "progress", &progress,
2678                             N_("show progress meter"), 1),
2679                 OPT_SET_INT(0, "all-progress", &progress,
2680                             N_("show progress meter during object writing phase"), 2),
2681                 OPT_BOOL(0, "all-progress-implied",
2682                          &all_progress_implied,
2683                          N_("similar to --all-progress when progress meter is shown")),
2684                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2685                   N_("write the pack index file in the specified idx format version"),
2686                   0, option_parse_index_version },
2687                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2688                               N_("maximum size of each output pack file")),
2689                 OPT_BOOL(0, "local", &local,
2690                          N_("ignore borrowed objects from alternate object store")),
2691                 OPT_BOOL(0, "incremental", &incremental,
2692                          N_("ignore packed objects")),
2693                 OPT_INTEGER(0, "window", &window,
2694                             N_("limit pack window by objects")),
2695                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2696                               N_("limit pack window by memory in addition to object limit")),
2697                 OPT_INTEGER(0, "depth", &depth,
2698                             N_("maximum length of delta chain allowed in the resulting pack")),
2699                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2700                          N_("reuse existing deltas")),
2701                 OPT_BOOL(0, "reuse-object", &reuse_object,
2702                          N_("reuse existing objects")),
2703                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2704                          N_("use OFS_DELTA objects")),
2705                 OPT_INTEGER(0, "threads", &delta_search_threads,
2706                             N_("use threads when searching for best delta matches")),
2707                 OPT_BOOL(0, "non-empty", &non_empty,
2708                          N_("do not create an empty pack output")),
2709                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2710                          N_("read revision arguments from standard input")),
2711                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2712                   N_("limit the objects to those that are not yet packed"),
2713                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2714                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2715                   N_("include objects reachable from any reference"),
2716                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2717                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2718                   N_("include objects referred by reflog entries"),
2719                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2720                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2721                   N_("include objects referred to by the index"),
2722                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2723                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2724                          N_("output pack to stdout")),
2725                 OPT_BOOL(0, "include-tag", &include_tag,
2726                          N_("include tag objects that refer to objects to be packed")),
2727                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2728                          N_("keep unreachable objects")),
2729                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
2730                          N_("pack loose unreachable objects")),
2731                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2732                   N_("unpack unreachable objects newer than <time>"),
2733                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2734                 OPT_BOOL(0, "thin", &thin,
2735                          N_("create thin packs")),
2736                 OPT_BOOL(0, "shallow", &shallow,
2737                          N_("create packs suitable for shallow fetches")),
2738                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2739                          N_("ignore packs that have companion .keep file")),
2740                 OPT_INTEGER(0, "compression", &pack_compression_level,
2741                             N_("pack compression level")),
2742                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2743                             N_("do not hide commits by grafts"), 0),
2744                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2745                          N_("use a bitmap index if available to speed up counting objects")),
2746                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2747                          N_("write a bitmap index together with the pack index")),
2748                 OPT_END(),
2749         };
2750
2751         check_replace_refs = 0;
2752
2753         reset_pack_idx_option(&pack_idx_opts);
2754         git_config(git_pack_config, NULL);
2755         if (!pack_compression_seen && core_compression_seen)
2756                 pack_compression_level = core_compression_level;
2757
2758         progress = isatty(2);
2759         argc = parse_options(argc, argv, prefix, pack_objects_options,
2760                              pack_usage, 0);
2761
2762         if (argc) {
2763                 base_name = argv[0];
2764                 argc--;
2765         }
2766         if (pack_to_stdout != !base_name || argc)
2767                 usage_with_options(pack_usage, pack_objects_options);
2768
2769         argv_array_push(&rp, "pack-objects");
2770         if (thin) {
2771                 use_internal_rev_list = 1;
2772                 argv_array_push(&rp, shallow
2773                                 ? "--objects-edge-aggressive"
2774                                 : "--objects-edge");
2775         } else
2776                 argv_array_push(&rp, "--objects");
2777
2778         if (rev_list_all) {
2779                 use_internal_rev_list = 1;
2780                 argv_array_push(&rp, "--all");
2781         }
2782         if (rev_list_reflog) {
2783                 use_internal_rev_list = 1;
2784                 argv_array_push(&rp, "--reflog");
2785         }
2786         if (rev_list_index) {
2787                 use_internal_rev_list = 1;
2788                 argv_array_push(&rp, "--indexed-objects");
2789         }
2790         if (rev_list_unpacked) {
2791                 use_internal_rev_list = 1;
2792                 argv_array_push(&rp, "--unpacked");
2793         }
2794
2795         if (!reuse_object)
2796                 reuse_delta = 0;
2797         if (pack_compression_level == -1)
2798                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2799         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2800                 die("bad pack compression level %d", pack_compression_level);
2801
2802         if (!delta_search_threads)      /* --threads=0 means autodetect */
2803                 delta_search_threads = online_cpus();
2804
2805 #ifdef NO_PTHREADS
2806         if (delta_search_threads != 1)
2807                 warning("no threads support, ignoring --threads");
2808 #endif
2809         if (!pack_to_stdout && !pack_size_limit)
2810                 pack_size_limit = pack_size_limit_cfg;
2811         if (pack_to_stdout && pack_size_limit)
2812                 die("--max-pack-size cannot be used to build a pack for transfer.");
2813         if (pack_size_limit && pack_size_limit < 1024*1024) {
2814                 warning("minimum pack size limit is 1 MiB");
2815                 pack_size_limit = 1024*1024;
2816         }
2817
2818         if (!pack_to_stdout && thin)
2819                 die("--thin cannot be used to build an indexable pack.");
2820
2821         if (keep_unreachable && unpack_unreachable)
2822                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2823         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
2824                 unpack_unreachable_expiration = 0;
2825
2826         /*
2827          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
2828          *
2829          * - to produce good pack (with bitmap index not-yet-packed objects are
2830          *   packed in suboptimal order).
2831          *
2832          * - to use more robust pack-generation codepath (avoiding possible
2833          *   bugs in bitmap code and possible bitmap index corruption).
2834          */
2835         if (!pack_to_stdout)
2836                 use_bitmap_index_default = 0;
2837
2838         if (use_bitmap_index < 0)
2839                 use_bitmap_index = use_bitmap_index_default;
2840
2841         /* "hard" reasons not to use bitmaps; these just won't work at all */
2842         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
2843                 use_bitmap_index = 0;
2844
2845         if (pack_to_stdout || !rev_list_all)
2846                 write_bitmap_index = 0;
2847
2848         if (progress && all_progress_implied)
2849                 progress = 2;
2850
2851         prepare_packed_git();
2852         if (ignore_packed_keep) {
2853                 struct packed_git *p;
2854                 for (p = packed_git; p; p = p->next)
2855                         if (p->pack_local && p->pack_keep)
2856                                 break;
2857                 if (!p) /* no keep-able packs found */
2858                         ignore_packed_keep = 0;
2859         }
2860         if (local) {
2861                 /*
2862                  * unlike ignore_packed_keep above, we do not want to
2863                  * unset "local" based on looking at packs, as it
2864                  * also covers non-local objects
2865                  */
2866                 struct packed_git *p;
2867                 for (p = packed_git; p; p = p->next) {
2868                         if (!p->pack_local) {
2869                                 have_non_local_packs = 1;
2870                                 break;
2871                         }
2872                 }
2873         }
2874
2875         if (progress)
2876                 progress_state = start_progress(_("Counting objects"), 0);
2877         if (!use_internal_rev_list)
2878                 read_object_list_from_stdin();
2879         else {
2880                 get_object_list(rp.argc, rp.argv);
2881                 argv_array_clear(&rp);
2882         }
2883         cleanup_preferred_base();
2884         if (include_tag && nr_result)
2885                 for_each_ref(add_ref_tag, NULL);
2886         stop_progress(&progress_state);
2887
2888         if (non_empty && !nr_result)
2889                 return 0;
2890         if (nr_result)
2891                 prepare_pack(window, depth);
2892         write_pack_file();
2893         if (progress)
2894                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2895                         " reused %"PRIu32" (delta %"PRIu32")\n",
2896                         written, written_delta, reused, reused_delta);
2897         return 0;
2898 }