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 #include "mru.h"
  27
  28 static const char *pack_usage[] = {
  29         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
  30         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
  31         NULL
  32 };
  33
  34 /*
  35  * Objects we are going to pack are collected in the `to_pack` structure.
  36  * It contains an array (dynamically expanded) of the object data, and a map
  37  * that can resolve SHA1s to their position in the array.
  38  */
  39 static struct packing_data to_pack;
  40
  41 static struct pack_idx_entry **written_list;
  42 static uint32_t nr_result, nr_written;
  43
  44 static int non_empty;
  45 static int reuse_delta = 1, reuse_object = 1;
  46 static int keep_unreachable, unpack_unreachable, include_tag;
  47 static unsigned long unpack_unreachable_expiration;
  48 static int pack_loose_unreachable;
  49 static int local;
  50 static int have_non_local_packs;
  51 static int incremental;
  52 static int ignore_packed_keep;
  53 static int allow_ofs_delta;
  54 static struct pack_idx_option pack_idx_opts;
  55 static const char *base_name;
  56 static int progress = 1;
  57 static int window = 10;
  58 static unsigned long pack_size_limit;
  59 static int depth = 50;
  60 static int delta_search_threads;
  61 static int pack_to_stdout;
  62 static int num_preferred_base;
  63 static struct progress *progress_state;
  64
  65 static struct packed_git *reuse_packfile;
  66 static uint32_t reuse_packfile_objects;
  67 static off_t reuse_packfile_offset;
  68
  69 static int use_bitmap_index_default = 1;
  70 static int use_bitmap_index = -1;
  71 static int write_bitmap_index;
  72 static uint16_t write_bitmap_options;
  73
  74 static unsigned long delta_cache_size = 0;
  75 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
  76 static unsigned long cache_max_small_delta_size = 1000;
  77
  78 static unsigned long window_memory_limit = 0;
  79
  80 /*
  81  * stats
  82  */
  83 static uint32_t written, written_delta;
  84 static uint32_t reused, reused_delta;
  85
  86 /*
  87  * Indexed commits
  88  */
  89 static struct commit **indexed_commits;
  90 static unsigned int indexed_commits_nr;
  91 static unsigned int indexed_commits_alloc;
  92
  93 static void index_commit_for_bitmap(struct commit *commit)
  94 {
  95         if (indexed_commits_nr >= indexed_commits_alloc) {
  96                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
  97                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
  98         }
  99
 100         indexed_commits[indexed_commits_nr++] = commit;
 101 }
 102
 103 static void *get_delta(struct object_entry *entry)
 104 {
 105         unsigned long size, base_size, delta_size;
 106         void *buf, *base_buf, *delta_buf;
 107         enum object_type type;
 108
 109         buf = read_sha1_file(entry->idx.sha1, &type, &size);
 110         if (!buf)
 111                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
 112         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
 113         if (!base_buf)
 114                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
 115         delta_buf = diff_delta(base_buf, base_size,
 116                                buf, size, &delta_size, 0);
 117         if (!delta_buf || delta_size != entry->delta_size)
 118                 die("delta size changed");
 119         free(buf);
 120         free(base_buf);
 121         return delta_buf;
 122 }
 123
 124 static unsigned long do_compress(void **pptr, unsigned long size)
 125 {
 126         git_zstream stream;
 127         void *in, *out;
 128         unsigned long maxsize;
 129
 130         git_deflate_init(&stream, pack_compression_level);
 131         maxsize = git_deflate_bound(&stream, size);
 132
 133         in = *pptr;
 134         out = xmalloc(maxsize);
 135         *pptr = out;
 136
 137         stream.next_in = in;
 138         stream.avail_in = size;
 139         stream.next_out = out;
 140         stream.avail_out = maxsize;
 141         while (git_deflate(&stream, Z_FINISH) == Z_OK)
 142                 ; /* nothing */
 143         git_deflate_end(&stream);
 144
 145         free(in);
 146         return stream.total_out;
 147 }
 148
 149 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
 150                                            const unsigned char *sha1)
 151 {
 152         git_zstream stream;
 153         unsigned char ibuf[1024 * 16];
 154         unsigned char obuf[1024 * 16];
 155         unsigned long olen = 0;
 156
 157         git_deflate_init(&stream, pack_compression_level);
 158
 159         for (;;) {
 160                 ssize_t readlen;
 161                 int zret = Z_OK;
 162                 readlen = read_istream(st, ibuf, sizeof(ibuf));
 163                 if (readlen == -1)
 164                         die(_("unable to read %s"), sha1_to_hex(sha1));
 165
 166                 stream.next_in = ibuf;
 167                 stream.avail_in = readlen;
 168                 while ((stream.avail_in || readlen == 0) &&
 169                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
 170                         stream.next_out = obuf;
 171                         stream.avail_out = sizeof(obuf);
 172                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 173                         sha1write(f, obuf, stream.next_out - obuf);
 174                         olen += stream.next_out - obuf;
 175                 }
 176                 if (stream.avail_in)
 177                         die(_("deflate error (%d)"), zret);
 178                 if (readlen == 0) {
 179                         if (zret != Z_STREAM_END)
 180                                 die(_("deflate error (%d)"), zret);
 181                         break;
 182                 }
 183         }
 184         git_deflate_end(&stream);
 185         return olen;
 186 }
 187
 188 /*
 189  * we are going to reuse the existing object data as is.  make
 190  * sure it is not corrupt.
 191  */
 192 static int check_pack_inflate(struct packed_git *p,
 193                 struct pack_window **w_curs,
 194                 off_t offset,
 195                 off_t len,
 196                 unsigned long expect)
 197 {
 198         git_zstream stream;
 199         unsigned char fakebuf[4096], *in;
 200         int st;
 201
 202         memset(&stream, 0, sizeof(stream));
 203         git_inflate_init(&stream);
 204         do {
 205                 in = use_pack(p, w_curs, offset, &stream.avail_in);
 206                 stream.next_in = in;
 207                 stream.next_out = fakebuf;
 208                 stream.avail_out = sizeof(fakebuf);
 209                 st = git_inflate(&stream, Z_FINISH);
 210                 offset += stream.next_in - in;
 211         } while (st == Z_OK || st == Z_BUF_ERROR);
 212         git_inflate_end(&stream);
 213         return (st == Z_STREAM_END &&
 214                 stream.total_out == expect &&
 215                 stream.total_in == len) ? 0 : -1;
 216 }
 217
 218 static void copy_pack_data(struct sha1file *f,
 219                 struct packed_git *p,
 220                 struct pack_window **w_curs,
 221                 off_t offset,
 222                 off_t len)
 223 {
 224         unsigned char *in;
 225         unsigned long avail;
 226
 227         while (len) {
 228                 in = use_pack(p, w_curs, offset, &avail);
 229                 if (avail > len)
 230                         avail = (unsigned long)len;
 231                 sha1write(f, in, avail);
 232                 offset += avail;
 233                 len -= avail;
 234         }
 235 }
 236
 237 /* Return 0 if we will bust the pack-size limit */
 238 static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
 239                                            unsigned long limit, int usable_delta)
 240 {
 241         unsigned long size, datalen;
 242         unsigned char header[10], dheader[10];
 243         unsigned hdrlen;
 244         enum object_type type;
 245         void *buf;
 246         struct git_istream *st = NULL;
 247
 248         if (!usable_delta) {
 249                 if (entry->type == OBJ_BLOB &&
 250                     entry->size > big_file_threshold &&
 251                     (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
 252                         buf = NULL;
 253                 else {
 254                         buf = read_sha1_file(entry->idx.sha1, &type, &size);
 255                         if (!buf)
 256                                 die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
 257                 }
 258                 /*
 259                  * make sure no cached delta data remains from a
 260                  * previous attempt before a pack split occurred.
 261                  */
 262                 free(entry->delta_data);
 263                 entry->delta_data = NULL;
 264                 entry->z_delta_size = 0;
 265         } else if (entry->delta_data) {
 266                 size = entry->delta_size;
 267                 buf = entry->delta_data;
 268                 entry->delta_data = NULL;
 269                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 270                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 271         } else {
 272                 buf = get_delta(entry);
 273                 size = entry->delta_size;
 274                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 275                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 276         }
 277
 278         if (st) /* large blob case, just assume we don't compress well */
 279                 datalen = size;
 280         else if (entry->z_delta_size)
 281                 datalen = entry->z_delta_size;
 282         else
 283                 datalen = do_compress(&buf, size);
 284
 285         /*
 286          * The object header is a byte of 'type' followed by zero or
 287          * more bytes of length.
 288          */
 289         hdrlen = encode_in_pack_object_header(type, size, header);
 290
 291         if (type == OBJ_OFS_DELTA) {
 292                 /*
 293                  * Deltas with relative base contain an additional
 294                  * encoding of the relative offset for the delta
 295                  * base from this object's position in the pack.
 296                  */
 297                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 298                 unsigned pos = sizeof(dheader) - 1;
 299                 dheader[pos] = ofs & 127;
 300                 while (ofs >>= 7)
 301                         dheader[--pos] = 128 | (--ofs & 127);
 302                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 303                         if (st)
 304                                 close_istream(st);
 305                         free(buf);
 306                         return 0;
 307                 }
 308                 sha1write(f, header, hdrlen);
 309                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 310                 hdrlen += sizeof(dheader) - pos;
 311         } else if (type == OBJ_REF_DELTA) {
 312                 /*
 313                  * Deltas with a base reference contain
 314                  * an additional 20 bytes for the base sha1.
 315                  */
 316                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 317                         if (st)
 318                                 close_istream(st);
 319                         free(buf);
 320                         return 0;
 321                 }
 322                 sha1write(f, header, hdrlen);
 323                 sha1write(f, entry->delta->idx.sha1, 20);
 324                 hdrlen += 20;
 325         } else {
 326                 if (limit && hdrlen + datalen + 20 >= limit) {
 327                         if (st)
 328                                 close_istream(st);
 329                         free(buf);
 330                         return 0;
 331                 }
 332                 sha1write(f, header, hdrlen);
 333         }
 334         if (st) {
 335                 datalen = write_large_blob_data(st, f, entry->idx.sha1);
 336                 close_istream(st);
 337         } else {
 338                 sha1write(f, buf, datalen);
 339                 free(buf);
 340         }
 341
 342         return hdrlen + datalen;
 343 }
 344
 345 /* Return 0 if we will bust the pack-size limit */
 346 static off_t write_reuse_object(struct sha1file *f, struct object_entry *entry,
 347                                 unsigned long limit, int usable_delta)
 348 {
 349         struct packed_git *p = entry->in_pack;
 350         struct pack_window *w_curs = NULL;
 351         struct revindex_entry *revidx;
 352         off_t offset;
 353         enum object_type type = entry->type;
 354         off_t datalen;
 355         unsigned char header[10], dheader[10];
 356         unsigned hdrlen;
 357
 358         if (entry->delta)
 359                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 360                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 361         hdrlen = encode_in_pack_object_header(type, entry->size, header);
 362
 363         offset = entry->in_pack_offset;
 364         revidx = find_pack_revindex(p, offset);
 365         datalen = revidx[1].offset - offset;
 366         if (!pack_to_stdout && p->index_version > 1 &&
 367             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 368                 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
 369                 unuse_pack(&w_curs);
 370                 return write_no_reuse_object(f, entry, limit, usable_delta);
 371         }
 372
 373         offset += entry->in_pack_header_size;
 374         datalen -= entry->in_pack_header_size;
 375
 376         if (!pack_to_stdout && p->index_version == 1 &&
 377             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 378                 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
 379                 unuse_pack(&w_curs);
 380                 return write_no_reuse_object(f, entry, limit, usable_delta);
 381         }
 382
 383         if (type == OBJ_OFS_DELTA) {
 384                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 385                 unsigned pos = sizeof(dheader) - 1;
 386                 dheader[pos] = ofs & 127;
 387                 while (ofs >>= 7)
 388                         dheader[--pos] = 128 | (--ofs & 127);
 389                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 390                         unuse_pack(&w_curs);
 391                         return 0;
 392                 }
 393                 sha1write(f, header, hdrlen);
 394                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 395                 hdrlen += sizeof(dheader) - pos;
 396                 reused_delta++;
 397         } else if (type == OBJ_REF_DELTA) {
 398                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 399                         unuse_pack(&w_curs);
 400                         return 0;
 401                 }
 402                 sha1write(f, header, hdrlen);
 403                 sha1write(f, entry->delta->idx.sha1, 20);
 404                 hdrlen += 20;
 405                 reused_delta++;
 406         } else {
 407                 if (limit && hdrlen + datalen + 20 >= limit) {
 408                         unuse_pack(&w_curs);
 409                         return 0;
 410                 }
 411                 sha1write(f, header, hdrlen);
 412         }
 413         copy_pack_data(f, p, &w_curs, offset, datalen);
 414         unuse_pack(&w_curs);
 415         reused++;
 416         return hdrlen + datalen;
 417 }
 418
 419 /* Return 0 if we will bust the pack-size limit */
 420 static off_t write_object(struct sha1file *f,
 421                           struct object_entry *entry,
 422                           off_t write_offset)
 423 {
 424         unsigned long limit;
 425         off_t 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         off_t 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(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 mru_entry *entry;
 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 (entry = packed_git_mru->head; entry; entry = entry->next) {
1014                 struct packed_git *p = entry->item;
1015                 off_t offset;
1016
1017                 if (p == *found_pack)
1018                         offset = *found_offset;
1019                 else
1020                         offset = find_pack_entry_one(sha1, p);
1021
1022                 if (offset) {
1023                         if (!*found_pack) {
1024                                 if (!is_pack_valid(p))
1025                                         continue;
1026                                 *found_offset = offset;
1027                                 *found_pack = p;
1028                         }
1029                         want = want_found_object(exclude, p);
1030                         if (!exclude && want > 0)
1031                                 mru_mark(packed_git_mru, entry);
1032                         if (want != -1)
1033                                 return want;
1034                 }
1035         }
1036
1037         return 1;
1038 }
1039
1040 static void create_object_entry(const unsigned char *sha1,
1041                                 enum object_type type,
1042                                 uint32_t hash,
1043                                 int exclude,
1044                                 int no_try_delta,
1045                                 uint32_t index_pos,
1046                                 struct packed_git *found_pack,
1047                                 off_t found_offset)
1048 {
1049         struct object_entry *entry;
1050
1051         entry = packlist_alloc(&to_pack, sha1, index_pos);
1052         entry->hash = hash;
1053         if (type)
1054                 entry->type = type;
1055         if (exclude)
1056                 entry->preferred_base = 1;
1057         else
1058                 nr_result++;
1059         if (found_pack) {
1060                 entry->in_pack = found_pack;
1061                 entry->in_pack_offset = found_offset;
1062         }
1063
1064         entry->no_try_delta = no_try_delta;
1065 }
1066
1067 static const char no_closure_warning[] = N_(
1068 "disabling bitmap writing, as some objects are not being packed"
1069 );
1070
1071 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1072                             const char *name, int exclude)
1073 {
1074         struct packed_git *found_pack = NULL;
1075         off_t found_offset = 0;
1076         uint32_t index_pos;
1077
1078         if (have_duplicate_entry(sha1, exclude, &index_pos))
1079                 return 0;
1080
1081         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1082                 /* The pack is missing an object, so it will not have closure */
1083                 if (write_bitmap_index) {
1084                         warning(_(no_closure_warning));
1085                         write_bitmap_index = 0;
1086                 }
1087                 return 0;
1088         }
1089
1090         create_object_entry(sha1, type, pack_name_hash(name),
1091                             exclude, name && no_try_delta(name),
1092                             index_pos, found_pack, found_offset);
1093
1094         display_progress(progress_state, nr_result);
1095         return 1;
1096 }
1097
1098 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1099                                         enum object_type type,
1100                                         int flags, uint32_t name_hash,
1101                                         struct packed_git *pack, off_t offset)
1102 {
1103         uint32_t index_pos;
1104
1105         if (have_duplicate_entry(sha1, 0, &index_pos))
1106                 return 0;
1107
1108         if (!want_object_in_pack(sha1, 0, &pack, &offset))
1109                 return 0;
1110
1111         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1112
1113         display_progress(progress_state, nr_result);
1114         return 1;
1115 }
1116
1117 struct pbase_tree_cache {
1118         unsigned char sha1[20];
1119         int ref;
1120         int temporary;
1121         void *tree_data;
1122         unsigned long tree_size;
1123 };
1124
1125 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1126 static int pbase_tree_cache_ix(const unsigned char *sha1)
1127 {
1128         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1129 }
1130 static int pbase_tree_cache_ix_incr(int ix)
1131 {
1132         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1133 }
1134
1135 static struct pbase_tree {
1136         struct pbase_tree *next;
1137         /* This is a phony "cache" entry; we are not
1138          * going to evict it or find it through _get()
1139          * mechanism -- this is for the toplevel node that
1140          * would almost always change with any commit.
1141          */
1142         struct pbase_tree_cache pcache;
1143 } *pbase_tree;
1144
1145 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1146 {
1147         struct pbase_tree_cache *ent, *nent;
1148         void *data;
1149         unsigned long size;
1150         enum object_type type;
1151         int neigh;
1152         int my_ix = pbase_tree_cache_ix(sha1);
1153         int available_ix = -1;
1154
1155         /* pbase-tree-cache acts as a limited hashtable.
1156          * your object will be found at your index or within a few
1157          * slots after that slot if it is cached.
1158          */
1159         for (neigh = 0; neigh < 8; neigh++) {
1160                 ent = pbase_tree_cache[my_ix];
1161                 if (ent && !hashcmp(ent->sha1, sha1)) {
1162                         ent->ref++;
1163                         return ent;
1164                 }
1165                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1166                          ((0 <= available_ix) &&
1167                           (!ent && pbase_tree_cache[available_ix])))
1168                         available_ix = my_ix;
1169                 if (!ent)
1170                         break;
1171                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1172         }
1173
1174         /* Did not find one.  Either we got a bogus request or
1175          * we need to read and perhaps cache.
1176          */
1177         data = read_sha1_file(sha1, &type, &size);
1178         if (!data)
1179                 return NULL;
1180         if (type != OBJ_TREE) {
1181                 free(data);
1182                 return NULL;
1183         }
1184
1185         /* We need to either cache or return a throwaway copy */
1186
1187         if (available_ix < 0)
1188                 ent = NULL;
1189         else {
1190                 ent = pbase_tree_cache[available_ix];
1191                 my_ix = available_ix;
1192         }
1193
1194         if (!ent) {
1195                 nent = xmalloc(sizeof(*nent));
1196                 nent->temporary = (available_ix < 0);
1197         }
1198         else {
1199                 /* evict and reuse */
1200                 free(ent->tree_data);
1201                 nent = ent;
1202         }
1203         hashcpy(nent->sha1, sha1);
1204         nent->tree_data = data;
1205         nent->tree_size = size;
1206         nent->ref = 1;
1207         if (!nent->temporary)
1208                 pbase_tree_cache[my_ix] = nent;
1209         return nent;
1210 }
1211
1212 static void pbase_tree_put(struct pbase_tree_cache *cache)
1213 {
1214         if (!cache->temporary) {
1215                 cache->ref--;
1216                 return;
1217         }
1218         free(cache->tree_data);
1219         free(cache);
1220 }
1221
1222 static int name_cmp_len(const char *name)
1223 {
1224         int i;
1225         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1226                 ;
1227         return i;
1228 }
1229
1230 static void add_pbase_object(struct tree_desc *tree,
1231                              const char *name,
1232                              int cmplen,
1233                              const char *fullname)
1234 {
1235         struct name_entry entry;
1236         int cmp;
1237
1238         while (tree_entry(tree,&entry)) {
1239                 if (S_ISGITLINK(entry.mode))
1240                         continue;
1241                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1242                       memcmp(name, entry.path, cmplen);
1243                 if (cmp > 0)
1244                         continue;
1245                 if (cmp < 0)
1246                         return;
1247                 if (name[cmplen] != '/') {
1248                         add_object_entry(entry.oid->hash,
1249                                          object_type(entry.mode),
1250                                          fullname, 1);
1251                         return;
1252                 }
1253                 if (S_ISDIR(entry.mode)) {
1254                         struct tree_desc sub;
1255                         struct pbase_tree_cache *tree;
1256                         const char *down = name+cmplen+1;
1257                         int downlen = name_cmp_len(down);
1258
1259                         tree = pbase_tree_get(entry.oid->hash);
1260                         if (!tree)
1261                                 return;
1262                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1263
1264                         add_pbase_object(&sub, down, downlen, fullname);
1265                         pbase_tree_put(tree);
1266                 }
1267         }
1268 }
1269
1270 static unsigned *done_pbase_paths;
1271 static int done_pbase_paths_num;
1272 static int done_pbase_paths_alloc;
1273 static int done_pbase_path_pos(unsigned hash)
1274 {
1275         int lo = 0;
1276         int hi = done_pbase_paths_num;
1277         while (lo < hi) {
1278                 int mi = (hi + lo) / 2;
1279                 if (done_pbase_paths[mi] == hash)
1280                         return mi;
1281                 if (done_pbase_paths[mi] < hash)
1282                         hi = mi;
1283                 else
1284                         lo = mi + 1;
1285         }
1286         return -lo-1;
1287 }
1288
1289 static int check_pbase_path(unsigned hash)
1290 {
1291         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1292         if (0 <= pos)
1293                 return 1;
1294         pos = -pos - 1;
1295         ALLOC_GROW(done_pbase_paths,
1296                    done_pbase_paths_num + 1,
1297                    done_pbase_paths_alloc);
1298         done_pbase_paths_num++;
1299         if (pos < done_pbase_paths_num)
1300                 memmove(done_pbase_paths + pos + 1,
1301                         done_pbase_paths + pos,
1302                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1303         done_pbase_paths[pos] = hash;
1304         return 0;
1305 }
1306
1307 static void add_preferred_base_object(const char *name)
1308 {
1309         struct pbase_tree *it;
1310         int cmplen;
1311         unsigned hash = pack_name_hash(name);
1312
1313         if (!num_preferred_base || check_pbase_path(hash))
1314                 return;
1315
1316         cmplen = name_cmp_len(name);
1317         for (it = pbase_tree; it; it = it->next) {
1318                 if (cmplen == 0) {
1319                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1320                 }
1321                 else {
1322                         struct tree_desc tree;
1323                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1324                         add_pbase_object(&tree, name, cmplen, name);
1325                 }
1326         }
1327 }
1328
1329 static void add_preferred_base(unsigned char *sha1)
1330 {
1331         struct pbase_tree *it;
1332         void *data;
1333         unsigned long size;
1334         unsigned char tree_sha1[20];
1335
1336         if (window <= num_preferred_base++)
1337                 return;
1338
1339         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1340         if (!data)
1341                 return;
1342
1343         for (it = pbase_tree; it; it = it->next) {
1344                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1345                         free(data);
1346                         return;
1347                 }
1348         }
1349
1350         it = xcalloc(1, sizeof(*it));
1351         it->next = pbase_tree;
1352         pbase_tree = it;
1353
1354         hashcpy(it->pcache.sha1, tree_sha1);
1355         it->pcache.tree_data = data;
1356         it->pcache.tree_size = size;
1357 }
1358
1359 static void cleanup_preferred_base(void)
1360 {
1361         struct pbase_tree *it;
1362         unsigned i;
1363
1364         it = pbase_tree;
1365         pbase_tree = NULL;
1366         while (it) {
1367                 struct pbase_tree *this = it;
1368                 it = this->next;
1369                 free(this->pcache.tree_data);
1370                 free(this);
1371         }
1372
1373         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1374                 if (!pbase_tree_cache[i])
1375                         continue;
1376                 free(pbase_tree_cache[i]->tree_data);
1377                 free(pbase_tree_cache[i]);
1378                 pbase_tree_cache[i] = NULL;
1379         }
1380
1381         free(done_pbase_paths);
1382         done_pbase_paths = NULL;
1383         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1384 }
1385
1386 static void check_object(struct object_entry *entry)
1387 {
1388         if (entry->in_pack) {
1389                 struct packed_git *p = entry->in_pack;
1390                 struct pack_window *w_curs = NULL;
1391                 const unsigned char *base_ref = NULL;
1392                 struct object_entry *base_entry;
1393                 unsigned long used, used_0;
1394                 unsigned long avail;
1395                 off_t ofs;
1396                 unsigned char *buf, c;
1397
1398                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1399
1400                 /*
1401                  * We want in_pack_type even if we do not reuse delta
1402                  * since non-delta representations could still be reused.
1403                  */
1404                 used = unpack_object_header_buffer(buf, avail,
1405                                                    &entry->in_pack_type,
1406                                                    &entry->size);
1407                 if (used == 0)
1408                         goto give_up;
1409
1410                 /*
1411                  * Determine if this is a delta and if so whether we can
1412                  * reuse it or not.  Otherwise let's find out as cheaply as
1413                  * possible what the actual type and size for this object is.
1414                  */
1415                 switch (entry->in_pack_type) {
1416                 default:
1417                         /* Not a delta hence we've already got all we need. */
1418                         entry->type = entry->in_pack_type;
1419                         entry->in_pack_header_size = used;
1420                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1421                                 goto give_up;
1422                         unuse_pack(&w_curs);
1423                         return;
1424                 case OBJ_REF_DELTA:
1425                         if (reuse_delta && !entry->preferred_base)
1426                                 base_ref = use_pack(p, &w_curs,
1427                                                 entry->in_pack_offset + used, NULL);
1428                         entry->in_pack_header_size = used + 20;
1429                         break;
1430                 case OBJ_OFS_DELTA:
1431                         buf = use_pack(p, &w_curs,
1432                                        entry->in_pack_offset + used, NULL);
1433                         used_0 = 0;
1434                         c = buf[used_0++];
1435                         ofs = c & 127;
1436                         while (c & 128) {
1437                                 ofs += 1;
1438                                 if (!ofs || MSB(ofs, 7)) {
1439                                         error("delta base offset overflow in pack for %s",
1440                                               sha1_to_hex(entry->idx.sha1));
1441                                         goto give_up;
1442                                 }
1443                                 c = buf[used_0++];
1444                                 ofs = (ofs << 7) + (c & 127);
1445                         }
1446                         ofs = entry->in_pack_offset - ofs;
1447                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1448                                 error("delta base offset out of bound for %s",
1449                                       sha1_to_hex(entry->idx.sha1));
1450                                 goto give_up;
1451                         }
1452                         if (reuse_delta && !entry->preferred_base) {
1453                                 struct revindex_entry *revidx;
1454                                 revidx = find_pack_revindex(p, ofs);
1455                                 if (!revidx)
1456                                         goto give_up;
1457                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1458                         }
1459                         entry->in_pack_header_size = used + used_0;
1460                         break;
1461                 }
1462
1463                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1464                         /*
1465                          * If base_ref was set above that means we wish to
1466                          * reuse delta data, and we even found that base
1467                          * in the list of objects we want to pack. Goodie!
1468                          *
1469                          * Depth value does not matter - find_deltas() will
1470                          * never consider reused delta as the base object to
1471                          * deltify other objects against, in order to avoid
1472                          * circular deltas.
1473                          */
1474                         entry->type = entry->in_pack_type;
1475                         entry->delta = base_entry;
1476                         entry->delta_size = entry->size;
1477                         entry->delta_sibling = base_entry->delta_child;
1478                         base_entry->delta_child = entry;
1479                         unuse_pack(&w_curs);
1480                         return;
1481                 }
1482
1483                 if (entry->type) {
1484                         /*
1485                          * This must be a delta and we already know what the
1486                          * final object type is.  Let's extract the actual
1487                          * object size from the delta header.
1488                          */
1489                         entry->size = get_size_from_delta(p, &w_curs,
1490                                         entry->in_pack_offset + entry->in_pack_header_size);
1491                         if (entry->size == 0)
1492                                 goto give_up;
1493                         unuse_pack(&w_curs);
1494                         return;
1495                 }
1496
1497                 /*
1498                  * No choice but to fall back to the recursive delta walk
1499                  * with sha1_object_info() to find about the object type
1500                  * at this point...
1501                  */
1502                 give_up:
1503                 unuse_pack(&w_curs);
1504         }
1505
1506         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1507         /*
1508          * The error condition is checked in prepare_pack().  This is
1509          * to permit a missing preferred base object to be ignored
1510          * as a preferred base.  Doing so can result in a larger
1511          * pack file, but the transfer will still take place.
1512          */
1513 }
1514
1515 static int pack_offset_sort(const void *_a, const void *_b)
1516 {
1517         const struct object_entry *a = *(struct object_entry **)_a;
1518         const struct object_entry *b = *(struct object_entry **)_b;
1519
1520         /* avoid filesystem trashing with loose objects */
1521         if (!a->in_pack && !b->in_pack)
1522                 return hashcmp(a->idx.sha1, b->idx.sha1);
1523
1524         if (a->in_pack < b->in_pack)
1525                 return -1;
1526         if (a->in_pack > b->in_pack)
1527                 return 1;
1528         return a->in_pack_offset < b->in_pack_offset ? -1 :
1529                         (a->in_pack_offset > b->in_pack_offset);
1530 }
1531
1532 /*
1533  * Drop an on-disk delta we were planning to reuse. Naively, this would
1534  * just involve blanking out the "delta" field, but we have to deal
1535  * with some extra book-keeping:
1536  *
1537  *   1. Removing ourselves from the delta_sibling linked list.
1538  *
1539  *   2. Updating our size/type to the non-delta representation. These were
1540  *      either not recorded initially (size) or overwritten with the delta type
1541  *      (type) when check_object() decided to reuse the delta.
1542  */
1543 static void drop_reused_delta(struct object_entry *entry)
1544 {
1545         struct object_entry **p = &entry->delta->delta_child;
1546         struct object_info oi = OBJECT_INFO_INIT;
1547
1548         while (*p) {
1549                 if (*p == entry)
1550                         *p = (*p)->delta_sibling;
1551                 else
1552                         p = &(*p)->delta_sibling;
1553         }
1554         entry->delta = NULL;
1555
1556         oi.sizep = &entry->size;
1557         oi.typep = &entry->type;
1558         if (packed_object_info(entry->in_pack, entry->in_pack_offset, &oi) < 0) {
1559                 /*
1560                  * We failed to get the info from this pack for some reason;
1561                  * fall back to sha1_object_info, which may find another copy.
1562                  * And if that fails, the error will be recorded in entry->type
1563                  * and dealt with in prepare_pack().
1564                  */
1565                 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1566         }
1567 }
1568
1569 /*
1570  * Follow the chain of deltas from this entry onward, throwing away any links
1571  * that cause us to hit a cycle (as determined by the DFS state flags in
1572  * the entries).
1573  */
1574 static void break_delta_chains(struct object_entry *entry)
1575 {
1576         /* If it's not a delta, it can't be part of a cycle. */
1577         if (!entry->delta) {
1578                 entry->dfs_state = DFS_DONE;
1579                 return;
1580         }
1581
1582         switch (entry->dfs_state) {
1583         case DFS_NONE:
1584                 /*
1585                  * This is the first time we've seen the object. We mark it as
1586                  * part of the active potential cycle and recurse.
1587                  */
1588                 entry->dfs_state = DFS_ACTIVE;
1589                 break_delta_chains(entry->delta);
1590                 entry->dfs_state = DFS_DONE;
1591                 break;
1592
1593         case DFS_DONE:
1594                 /* object already examined, and not part of a cycle */
1595                 break;
1596
1597         case DFS_ACTIVE:
1598                 /*
1599                  * We found a cycle that needs broken. It would be correct to
1600                  * break any link in the chain, but it's convenient to
1601                  * break this one.
1602                  */
1603                 drop_reused_delta(entry);
1604                 entry->dfs_state = DFS_DONE;
1605                 break;
1606         }
1607 }
1608
1609 static void get_object_details(void)
1610 {
1611         uint32_t i;
1612         struct object_entry **sorted_by_offset;
1613
1614         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1615         for (i = 0; i < to_pack.nr_objects; i++)
1616                 sorted_by_offset[i] = to_pack.objects + i;
1617         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1618
1619         for (i = 0; i < to_pack.nr_objects; i++) {
1620                 struct object_entry *entry = sorted_by_offset[i];
1621                 check_object(entry);
1622                 if (big_file_threshold < entry->size)
1623                         entry->no_try_delta = 1;
1624         }
1625
1626         /*
1627          * This must happen in a second pass, since we rely on the delta
1628          * information for the whole list being completed.
1629          */
1630         for (i = 0; i < to_pack.nr_objects; i++)
1631                 break_delta_chains(&to_pack.objects[i]);
1632
1633         free(sorted_by_offset);
1634 }
1635
1636 /*
1637  * We search for deltas in a list sorted by type, by filename hash, and then
1638  * by size, so that we see progressively smaller and smaller files.
1639  * That's because we prefer deltas to be from the bigger file
1640  * to the smaller -- deletes are potentially cheaper, but perhaps
1641  * more importantly, the bigger file is likely the more recent
1642  * one.  The deepest deltas are therefore the oldest objects which are
1643  * less susceptible to be accessed often.
1644  */
1645 static int type_size_sort(const void *_a, const void *_b)
1646 {
1647         const struct object_entry *a = *(struct object_entry **)_a;
1648         const struct object_entry *b = *(struct object_entry **)_b;
1649
1650         if (a->type > b->type)
1651                 return -1;
1652         if (a->type < b->type)
1653                 return 1;
1654         if (a->hash > b->hash)
1655                 return -1;
1656         if (a->hash < b->hash)
1657                 return 1;
1658         if (a->preferred_base > b->preferred_base)
1659                 return -1;
1660         if (a->preferred_base < b->preferred_base)
1661                 return 1;
1662         if (a->size > b->size)
1663                 return -1;
1664         if (a->size < b->size)
1665                 return 1;
1666         return a < b ? -1 : (a > b);  /* newest first */
1667 }
1668
1669 struct unpacked {
1670         struct object_entry *entry;
1671         void *data;
1672         struct delta_index *index;
1673         unsigned depth;
1674 };
1675
1676 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1677                            unsigned long delta_size)
1678 {
1679         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1680                 return 0;
1681
1682         if (delta_size < cache_max_small_delta_size)
1683                 return 1;
1684
1685         /* cache delta, if objects are large enough compared to delta size */
1686         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1687                 return 1;
1688
1689         return 0;
1690 }
1691
1692 #ifndef NO_PTHREADS
1693
1694 static pthread_mutex_t read_mutex;
1695 #define read_lock()             pthread_mutex_lock(&read_mutex)
1696 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1697
1698 static pthread_mutex_t cache_mutex;
1699 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1700 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1701
1702 static pthread_mutex_t progress_mutex;
1703 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1704 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1705
1706 #else
1707
1708 #define read_lock()             (void)0
1709 #define read_unlock()           (void)0
1710 #define cache_lock()            (void)0
1711 #define cache_unlock()          (void)0
1712 #define progress_lock()         (void)0
1713 #define progress_unlock()       (void)0
1714
1715 #endif
1716
1717 static int try_delta(struct unpacked *trg, struct unpacked *src,
1718                      unsigned max_depth, unsigned long *mem_usage)
1719 {
1720         struct object_entry *trg_entry = trg->entry;
1721         struct object_entry *src_entry = src->entry;
1722         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1723         unsigned ref_depth;
1724         enum object_type type;
1725         void *delta_buf;
1726
1727         /* Don't bother doing diffs between different types */
1728         if (trg_entry->type != src_entry->type)
1729                 return -1;
1730
1731         /*
1732          * We do not bother to try a delta that we discarded on an
1733          * earlier try, but only when reusing delta data.  Note that
1734          * src_entry that is marked as the preferred_base should always
1735          * be considered, as even if we produce a suboptimal delta against
1736          * it, we will still save the transfer cost, as we already know
1737          * the other side has it and we won't send src_entry at all.
1738          */
1739         if (reuse_delta && trg_entry->in_pack &&
1740             trg_entry->in_pack == src_entry->in_pack &&
1741             !src_entry->preferred_base &&
1742             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1743             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1744                 return 0;
1745
1746         /* Let's not bust the allowed depth. */
1747         if (src->depth >= max_depth)
1748                 return 0;
1749
1750         /* Now some size filtering heuristics. */
1751         trg_size = trg_entry->size;
1752         if (!trg_entry->delta) {
1753                 max_size = trg_size/2 - 20;
1754                 ref_depth = 1;
1755         } else {
1756                 max_size = trg_entry->delta_size;
1757                 ref_depth = trg->depth;
1758         }
1759         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1760                                                 (max_depth - ref_depth + 1);
1761         if (max_size == 0)
1762                 return 0;
1763         src_size = src_entry->size;
1764         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1765         if (sizediff >= max_size)
1766                 return 0;
1767         if (trg_size < src_size / 32)
1768                 return 0;
1769
1770         /* Load data if not already done */
1771         if (!trg->data) {
1772                 read_lock();
1773                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1774                 read_unlock();
1775                 if (!trg->data)
1776                         die("object %s cannot be read",
1777                             sha1_to_hex(trg_entry->idx.sha1));
1778                 if (sz != trg_size)
1779                         die("object %s inconsistent object length (%lu vs %lu)",
1780                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1781                 *mem_usage += sz;
1782         }
1783         if (!src->data) {
1784                 read_lock();
1785                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1786                 read_unlock();
1787                 if (!src->data) {
1788                         if (src_entry->preferred_base) {
1789                                 static int warned = 0;
1790                                 if (!warned++)
1791                                         warning("object %s cannot be read",
1792                                                 sha1_to_hex(src_entry->idx.sha1));
1793                                 /*
1794                                  * Those objects are not included in the
1795                                  * resulting pack.  Be resilient and ignore
1796                                  * them if they can't be read, in case the
1797                                  * pack could be created nevertheless.
1798                                  */
1799                                 return 0;
1800                         }
1801                         die("object %s cannot be read",
1802                             sha1_to_hex(src_entry->idx.sha1));
1803                 }
1804                 if (sz != src_size)
1805                         die("object %s inconsistent object length (%lu vs %lu)",
1806                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1807                 *mem_usage += sz;
1808         }
1809         if (!src->index) {
1810                 src->index = create_delta_index(src->data, src_size);
1811                 if (!src->index) {
1812                         static int warned = 0;
1813                         if (!warned++)
1814                                 warning("suboptimal pack - out of memory");
1815                         return 0;
1816                 }
1817                 *mem_usage += sizeof_delta_index(src->index);
1818         }
1819
1820         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1821         if (!delta_buf)
1822                 return 0;
1823
1824         if (trg_entry->delta) {
1825                 /* Prefer only shallower same-sized deltas. */
1826                 if (delta_size == trg_entry->delta_size &&
1827                     src->depth + 1 >= trg->depth) {
1828                         free(delta_buf);
1829                         return 0;
1830                 }
1831         }
1832
1833         /*
1834          * Handle memory allocation outside of the cache
1835          * accounting lock.  Compiler will optimize the strangeness
1836          * away when NO_PTHREADS is defined.
1837          */
1838         free(trg_entry->delta_data);
1839         cache_lock();
1840         if (trg_entry->delta_data) {
1841                 delta_cache_size -= trg_entry->delta_size;
1842                 trg_entry->delta_data = NULL;
1843         }
1844         if (delta_cacheable(src_size, trg_size, delta_size)) {
1845                 delta_cache_size += delta_size;
1846                 cache_unlock();
1847                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1848         } else {
1849                 cache_unlock();
1850                 free(delta_buf);
1851         }
1852
1853         trg_entry->delta = src_entry;
1854         trg_entry->delta_size = delta_size;
1855         trg->depth = src->depth + 1;
1856
1857         return 1;
1858 }
1859
1860 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1861 {
1862         struct object_entry *child = me->delta_child;
1863         unsigned int m = n;
1864         while (child) {
1865                 unsigned int c = check_delta_limit(child, n + 1);
1866                 if (m < c)
1867                         m = c;
1868                 child = child->delta_sibling;
1869         }
1870         return m;
1871 }
1872
1873 static unsigned long free_unpacked(struct unpacked *n)
1874 {
1875         unsigned long freed_mem = sizeof_delta_index(n->index);
1876         free_delta_index(n->index);
1877         n->index = NULL;
1878         if (n->data) {
1879                 freed_mem += n->entry->size;
1880                 free(n->data);
1881                 n->data = NULL;
1882         }
1883         n->entry = NULL;
1884         n->depth = 0;
1885         return freed_mem;
1886 }
1887
1888 static void find_deltas(struct object_entry **list, unsigned *list_size,
1889                         int window, int depth, unsigned *processed)
1890 {
1891         uint32_t i, idx = 0, count = 0;
1892         struct unpacked *array;
1893         unsigned long mem_usage = 0;
1894
1895         array = xcalloc(window, sizeof(struct unpacked));
1896
1897         for (;;) {
1898                 struct object_entry *entry;
1899                 struct unpacked *n = array + idx;
1900                 int j, max_depth, best_base = -1;
1901
1902                 progress_lock();
1903                 if (!*list_size) {
1904                         progress_unlock();
1905                         break;
1906                 }
1907                 entry = *list++;
1908                 (*list_size)--;
1909                 if (!entry->preferred_base) {
1910                         (*processed)++;
1911                         display_progress(progress_state, *processed);
1912                 }
1913                 progress_unlock();
1914
1915                 mem_usage -= free_unpacked(n);
1916                 n->entry = entry;
1917
1918                 while (window_memory_limit &&
1919                        mem_usage > window_memory_limit &&
1920                        count > 1) {
1921                         uint32_t tail = (idx + window - count) % window;
1922                         mem_usage -= free_unpacked(array + tail);
1923                         count--;
1924                 }
1925
1926                 /* We do not compute delta to *create* objects we are not
1927                  * going to pack.
1928                  */
1929                 if (entry->preferred_base)
1930                         goto next;
1931
1932                 /*
1933                  * If the current object is at pack edge, take the depth the
1934                  * objects that depend on the current object into account
1935                  * otherwise they would become too deep.
1936                  */
1937                 max_depth = depth;
1938                 if (entry->delta_child) {
1939                         max_depth -= check_delta_limit(entry, 0);
1940                         if (max_depth <= 0)
1941                                 goto next;
1942                 }
1943
1944                 j = window;
1945                 while (--j > 0) {
1946                         int ret;
1947                         uint32_t other_idx = idx + j;
1948                         struct unpacked *m;
1949                         if (other_idx >= window)
1950                                 other_idx -= window;
1951                         m = array + other_idx;
1952                         if (!m->entry)
1953                                 break;
1954                         ret = try_delta(n, m, max_depth, &mem_usage);
1955                         if (ret < 0)
1956                                 break;
1957                         else if (ret > 0)
1958                                 best_base = other_idx;
1959                 }
1960
1961                 /*
1962                  * If we decided to cache the delta data, then it is best
1963                  * to compress it right away.  First because we have to do
1964                  * it anyway, and doing it here while we're threaded will
1965                  * save a lot of time in the non threaded write phase,
1966                  * as well as allow for caching more deltas within
1967                  * the same cache size limit.
1968                  * ...
1969                  * But only if not writing to stdout, since in that case
1970                  * the network is most likely throttling writes anyway,
1971                  * and therefore it is best to go to the write phase ASAP
1972                  * instead, as we can afford spending more time compressing
1973                  * between writes at that moment.
1974                  */
1975                 if (entry->delta_data && !pack_to_stdout) {
1976                         entry->z_delta_size = do_compress(&entry->delta_data,
1977                                                           entry->delta_size);
1978                         cache_lock();
1979                         delta_cache_size -= entry->delta_size;
1980                         delta_cache_size += entry->z_delta_size;
1981                         cache_unlock();
1982                 }
1983
1984                 /* if we made n a delta, and if n is already at max
1985                  * depth, leaving it in the window is pointless.  we
1986                  * should evict it first.
1987                  */
1988                 if (entry->delta && max_depth <= n->depth)
1989                         continue;
1990
1991                 /*
1992                  * Move the best delta base up in the window, after the
1993                  * currently deltified object, to keep it longer.  It will
1994                  * be the first base object to be attempted next.
1995                  */
1996                 if (entry->delta) {
1997                         struct unpacked swap = array[best_base];
1998                         int dist = (window + idx - best_base) % window;
1999                         int dst = best_base;
2000                         while (dist--) {
2001                                 int src = (dst + 1) % window;
2002                                 array[dst] = array[src];
2003                                 dst = src;
2004                         }
2005                         array[dst] = swap;
2006                 }
2007
2008                 next:
2009                 idx++;
2010                 if (count + 1 < window)
2011                         count++;
2012                 if (idx >= window)
2013                         idx = 0;
2014         }
2015
2016         for (i = 0; i < window; ++i) {
2017                 free_delta_index(array[i].index);
2018                 free(array[i].data);
2019         }
2020         free(array);
2021 }
2022
2023 #ifndef NO_PTHREADS
2024
2025 static void try_to_free_from_threads(size_t size)
2026 {
2027         read_lock();
2028         release_pack_memory(size);
2029         read_unlock();
2030 }
2031
2032 static try_to_free_t old_try_to_free_routine;
2033
2034 /*
2035  * The main thread waits on the condition that (at least) one of the workers
2036  * has stopped working (which is indicated in the .working member of
2037  * struct thread_params).
2038  * When a work thread has completed its work, it sets .working to 0 and
2039  * signals the main thread and waits on the condition that .data_ready
2040  * becomes 1.
2041  */
2042
2043 struct thread_params {
2044         pthread_t thread;
2045         struct object_entry **list;
2046         unsigned list_size;
2047         unsigned remaining;
2048         int window;
2049         int depth;
2050         int working;
2051         int data_ready;
2052         pthread_mutex_t mutex;
2053         pthread_cond_t cond;
2054         unsigned *processed;
2055 };
2056
2057 static pthread_cond_t progress_cond;
2058
2059 /*
2060  * Mutex and conditional variable can't be statically-initialized on Windows.
2061  */
2062 static void init_threaded_search(void)
2063 {
2064         init_recursive_mutex(&read_mutex);
2065         pthread_mutex_init(&cache_mutex, NULL);
2066         pthread_mutex_init(&progress_mutex, NULL);
2067         pthread_cond_init(&progress_cond, NULL);
2068         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2069 }
2070
2071 static void cleanup_threaded_search(void)
2072 {
2073         set_try_to_free_routine(old_try_to_free_routine);
2074         pthread_cond_destroy(&progress_cond);
2075         pthread_mutex_destroy(&read_mutex);
2076         pthread_mutex_destroy(&cache_mutex);
2077         pthread_mutex_destroy(&progress_mutex);
2078 }
2079
2080 static void *threaded_find_deltas(void *arg)
2081 {
2082         struct thread_params *me = arg;
2083
2084         while (me->remaining) {
2085                 find_deltas(me->list, &me->remaining,
2086                             me->window, me->depth, me->processed);
2087
2088                 progress_lock();
2089                 me->working = 0;
2090                 pthread_cond_signal(&progress_cond);
2091                 progress_unlock();
2092
2093                 /*
2094                  * We must not set ->data_ready before we wait on the
2095                  * condition because the main thread may have set it to 1
2096                  * before we get here. In order to be sure that new
2097                  * work is available if we see 1 in ->data_ready, it
2098                  * was initialized to 0 before this thread was spawned
2099                  * and we reset it to 0 right away.
2100                  */
2101                 pthread_mutex_lock(&me->mutex);
2102                 while (!me->data_ready)
2103                         pthread_cond_wait(&me->cond, &me->mutex);
2104                 me->data_ready = 0;
2105                 pthread_mutex_unlock(&me->mutex);
2106         }
2107         /* leave ->working 1 so that this doesn't get more work assigned */
2108         return NULL;
2109 }
2110
2111 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2112                            int window, int depth, unsigned *processed)
2113 {
2114         struct thread_params *p;
2115         int i, ret, active_threads = 0;
2116
2117         init_threaded_search();
2118
2119         if (delta_search_threads <= 1) {
2120                 find_deltas(list, &list_size, window, depth, processed);
2121                 cleanup_threaded_search();
2122                 return;
2123         }
2124         if (progress > pack_to_stdout)
2125                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2126                                 delta_search_threads);
2127         p = xcalloc(delta_search_threads, sizeof(*p));
2128
2129         /* Partition the work amongst work threads. */
2130         for (i = 0; i < delta_search_threads; i++) {
2131                 unsigned sub_size = list_size / (delta_search_threads - i);
2132
2133                 /* don't use too small segments or no deltas will be found */
2134                 if (sub_size < 2*window && i+1 < delta_search_threads)
2135                         sub_size = 0;
2136
2137                 p[i].window = window;
2138                 p[i].depth = depth;
2139                 p[i].processed = processed;
2140                 p[i].working = 1;
2141                 p[i].data_ready = 0;
2142
2143                 /* try to split chunks on "path" boundaries */
2144                 while (sub_size && sub_size < list_size &&
2145                        list[sub_size]->hash &&
2146                        list[sub_size]->hash == list[sub_size-1]->hash)
2147                         sub_size++;
2148
2149                 p[i].list = list;
2150                 p[i].list_size = sub_size;
2151                 p[i].remaining = sub_size;
2152
2153                 list += sub_size;
2154                 list_size -= sub_size;
2155         }
2156
2157         /* Start work threads. */
2158         for (i = 0; i < delta_search_threads; i++) {
2159                 if (!p[i].list_size)
2160                         continue;
2161                 pthread_mutex_init(&p[i].mutex, NULL);
2162                 pthread_cond_init(&p[i].cond, NULL);
2163                 ret = pthread_create(&p[i].thread, NULL,
2164                                      threaded_find_deltas, &p[i]);
2165                 if (ret)
2166                         die("unable to create thread: %s", strerror(ret));
2167                 active_threads++;
2168         }
2169
2170         /*
2171          * Now let's wait for work completion.  Each time a thread is done
2172          * with its work, we steal half of the remaining work from the
2173          * thread with the largest number of unprocessed objects and give
2174          * it to that newly idle thread.  This ensure good load balancing
2175          * until the remaining object list segments are simply too short
2176          * to be worth splitting anymore.
2177          */
2178         while (active_threads) {
2179                 struct thread_params *target = NULL;
2180                 struct thread_params *victim = NULL;
2181                 unsigned sub_size = 0;
2182
2183                 progress_lock();
2184                 for (;;) {
2185                         for (i = 0; !target && i < delta_search_threads; i++)
2186                                 if (!p[i].working)
2187                                         target = &p[i];
2188                         if (target)
2189                                 break;
2190                         pthread_cond_wait(&progress_cond, &progress_mutex);
2191                 }
2192
2193                 for (i = 0; i < delta_search_threads; i++)
2194                         if (p[i].remaining > 2*window &&
2195                             (!victim || victim->remaining < p[i].remaining))
2196                                 victim = &p[i];
2197                 if (victim) {
2198                         sub_size = victim->remaining / 2;
2199                         list = victim->list + victim->list_size - sub_size;
2200                         while (sub_size && list[0]->hash &&
2201                                list[0]->hash == list[-1]->hash) {
2202                                 list++;
2203                                 sub_size--;
2204                         }
2205                         if (!sub_size) {
2206                                 /*
2207                                  * It is possible for some "paths" to have
2208                                  * so many objects that no hash boundary
2209                                  * might be found.  Let's just steal the
2210                                  * exact half in that case.
2211                                  */
2212                                 sub_size = victim->remaining / 2;
2213                                 list -= sub_size;
2214                         }
2215                         target->list = list;
2216                         victim->list_size -= sub_size;
2217                         victim->remaining -= sub_size;
2218                 }
2219                 target->list_size = sub_size;
2220                 target->remaining = sub_size;
2221                 target->working = 1;
2222                 progress_unlock();
2223
2224                 pthread_mutex_lock(&target->mutex);
2225                 target->data_ready = 1;
2226                 pthread_cond_signal(&target->cond);
2227                 pthread_mutex_unlock(&target->mutex);
2228
2229                 if (!sub_size) {
2230                         pthread_join(target->thread, NULL);
2231                         pthread_cond_destroy(&target->cond);
2232                         pthread_mutex_destroy(&target->mutex);
2233                         active_threads--;
2234                 }
2235         }
2236         cleanup_threaded_search();
2237         free(p);
2238 }
2239
2240 #else
2241 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2242 #endif
2243
2244 static void add_tag_chain(const struct object_id *oid)
2245 {
2246         struct tag *tag;
2247
2248         /*
2249          * We catch duplicates already in add_object_entry(), but we'd
2250          * prefer to do this extra check to avoid having to parse the
2251          * tag at all if we already know that it's being packed (e.g., if
2252          * it was included via bitmaps, we would not have parsed it
2253          * previously).
2254          */
2255         if (packlist_find(&to_pack, oid->hash, NULL))
2256                 return;
2257
2258         tag = lookup_tag(oid->hash);
2259         while (1) {
2260                 if (!tag || parse_tag(tag) || !tag->tagged)
2261                         die("unable to pack objects reachable from tag %s",
2262                             oid_to_hex(oid));
2263
2264                 add_object_entry(tag->object.oid.hash, OBJ_TAG, NULL, 0);
2265
2266                 if (tag->tagged->type != OBJ_TAG)
2267                         return;
2268
2269                 tag = (struct tag *)tag->tagged;
2270         }
2271 }
2272
2273 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2274 {
2275         struct object_id peeled;
2276
2277         if (starts_with(path, "refs/tags/") && /* is a tag? */
2278             !peel_ref(path, peeled.hash)    && /* peelable? */
2279             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2280                 add_tag_chain(oid);
2281         return 0;
2282 }
2283
2284 static void prepare_pack(int window, int depth)
2285 {
2286         struct object_entry **delta_list;
2287         uint32_t i, nr_deltas;
2288         unsigned n;
2289
2290         get_object_details();
2291
2292         /*
2293          * If we're locally repacking then we need to be doubly careful
2294          * from now on in order to make sure no stealth corruption gets
2295          * propagated to the new pack.  Clients receiving streamed packs
2296          * should validate everything they get anyway so no need to incur
2297          * the additional cost here in that case.
2298          */
2299         if (!pack_to_stdout)
2300                 do_check_packed_object_crc = 1;
2301
2302         if (!to_pack.nr_objects || !window || !depth)
2303                 return;
2304
2305         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2306         nr_deltas = n = 0;
2307
2308         for (i = 0; i < to_pack.nr_objects; i++) {
2309                 struct object_entry *entry = to_pack.objects + i;
2310
2311                 if (entry->delta)
2312                         /* This happens if we decided to reuse existing
2313                          * delta from a pack.  "reuse_delta &&" is implied.
2314                          */
2315                         continue;
2316
2317                 if (entry->size < 50)
2318                         continue;
2319
2320                 if (entry->no_try_delta)
2321                         continue;
2322
2323                 if (!entry->preferred_base) {
2324                         nr_deltas++;
2325                         if (entry->type < 0)
2326                                 die("unable to get type of object %s",
2327                                     sha1_to_hex(entry->idx.sha1));
2328                 } else {
2329                         if (entry->type < 0) {
2330                                 /*
2331                                  * This object is not found, but we
2332                                  * don't have to include it anyway.
2333                                  */
2334                                 continue;
2335                         }
2336                 }
2337
2338                 delta_list[n++] = entry;
2339         }
2340
2341         if (nr_deltas && n > 1) {
2342                 unsigned nr_done = 0;
2343                 if (progress)
2344                         progress_state = start_progress(_("Compressing objects"),
2345                                                         nr_deltas);
2346                 QSORT(delta_list, n, type_size_sort);
2347                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2348                 stop_progress(&progress_state);
2349                 if (nr_done != nr_deltas)
2350                         die("inconsistency with delta count");
2351         }
2352         free(delta_list);
2353 }
2354
2355 static int git_pack_config(const char *k, const char *v, void *cb)
2356 {
2357         if (!strcmp(k, "pack.window")) {
2358                 window = git_config_int(k, v);
2359                 return 0;
2360         }
2361         if (!strcmp(k, "pack.windowmemory")) {
2362                 window_memory_limit = git_config_ulong(k, v);
2363                 return 0;
2364         }
2365         if (!strcmp(k, "pack.depth")) {
2366                 depth = git_config_int(k, v);
2367                 return 0;
2368         }
2369         if (!strcmp(k, "pack.deltacachesize")) {
2370                 max_delta_cache_size = git_config_int(k, v);
2371                 return 0;
2372         }
2373         if (!strcmp(k, "pack.deltacachelimit")) {
2374                 cache_max_small_delta_size = git_config_int(k, v);
2375                 return 0;
2376         }
2377         if (!strcmp(k, "pack.writebitmaphashcache")) {
2378                 if (git_config_bool(k, v))
2379                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2380                 else
2381                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2382         }
2383         if (!strcmp(k, "pack.usebitmaps")) {
2384                 use_bitmap_index_default = git_config_bool(k, v);
2385                 return 0;
2386         }
2387         if (!strcmp(k, "pack.threads")) {
2388                 delta_search_threads = git_config_int(k, v);
2389                 if (delta_search_threads < 0)
2390                         die("invalid number of threads specified (%d)",
2391                             delta_search_threads);
2392 #ifdef NO_PTHREADS
2393                 if (delta_search_threads != 1)
2394                         warning("no threads support, ignoring %s", k);
2395 #endif
2396                 return 0;
2397         }
2398         if (!strcmp(k, "pack.indexversion")) {
2399                 pack_idx_opts.version = git_config_int(k, v);
2400                 if (pack_idx_opts.version > 2)
2401                         die("bad pack.indexversion=%"PRIu32,
2402                             pack_idx_opts.version);
2403                 return 0;
2404         }
2405         return git_default_config(k, v, cb);
2406 }
2407
2408 static void read_object_list_from_stdin(void)
2409 {
2410         char line[40 + 1 + PATH_MAX + 2];
2411         unsigned char sha1[20];
2412
2413         for (;;) {
2414                 if (!fgets(line, sizeof(line), stdin)) {
2415                         if (feof(stdin))
2416                                 break;
2417                         if (!ferror(stdin))
2418                                 die("fgets returned NULL, not EOF, not error!");
2419                         if (errno != EINTR)
2420                                 die_errno("fgets");
2421                         clearerr(stdin);
2422                         continue;
2423                 }
2424                 if (line[0] == '-') {
2425                         if (get_sha1_hex(line+1, sha1))
2426                                 die("expected edge sha1, got garbage:\n %s",
2427                                     line);
2428                         add_preferred_base(sha1);
2429                         continue;
2430                 }
2431                 if (get_sha1_hex(line, sha1))
2432                         die("expected sha1, got garbage:\n %s", line);
2433
2434                 add_preferred_base_object(line+41);
2435                 add_object_entry(sha1, 0, line+41, 0);
2436         }
2437 }
2438
2439 #define OBJECT_ADDED (1u<<20)
2440
2441 static void show_commit(struct commit *commit, void *data)
2442 {
2443         add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2444         commit->object.flags |= OBJECT_ADDED;
2445
2446         if (write_bitmap_index)
2447                 index_commit_for_bitmap(commit);
2448 }
2449
2450 static void show_object(struct object *obj, const char *name, void *data)
2451 {
2452         add_preferred_base_object(name);
2453         add_object_entry(obj->oid.hash, obj->type, name, 0);
2454         obj->flags |= OBJECT_ADDED;
2455 }
2456
2457 static void show_edge(struct commit *commit)
2458 {
2459         add_preferred_base(commit->object.oid.hash);
2460 }
2461
2462 struct in_pack_object {
2463         off_t offset;
2464         struct object *object;
2465 };
2466
2467 struct in_pack {
2468         int alloc;
2469         int nr;
2470         struct in_pack_object *array;
2471 };
2472
2473 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2474 {
2475         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2476         in_pack->array[in_pack->nr].object = object;
2477         in_pack->nr++;
2478 }
2479
2480 /*
2481  * Compare the objects in the offset order, in order to emulate the
2482  * "git rev-list --objects" output that produced the pack originally.
2483  */
2484 static int ofscmp(const void *a_, const void *b_)
2485 {
2486         struct in_pack_object *a = (struct in_pack_object *)a_;
2487         struct in_pack_object *b = (struct in_pack_object *)b_;
2488
2489         if (a->offset < b->offset)
2490                 return -1;
2491         else if (a->offset > b->offset)
2492                 return 1;
2493         else
2494                 return oidcmp(&a->object->oid, &b->object->oid);
2495 }
2496
2497 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2498 {
2499         struct packed_git *p;
2500         struct in_pack in_pack;
2501         uint32_t i;
2502
2503         memset(&in_pack, 0, sizeof(in_pack));
2504
2505         for (p = packed_git; p; p = p->next) {
2506                 const unsigned char *sha1;
2507                 struct object *o;
2508
2509                 if (!p->pack_local || p->pack_keep)
2510                         continue;
2511                 if (open_pack_index(p))
2512                         die("cannot open pack index");
2513
2514                 ALLOC_GROW(in_pack.array,
2515                            in_pack.nr + p->num_objects,
2516                            in_pack.alloc);
2517
2518                 for (i = 0; i < p->num_objects; i++) {
2519                         sha1 = nth_packed_object_sha1(p, i);
2520                         o = lookup_unknown_object(sha1);
2521                         if (!(o->flags & OBJECT_ADDED))
2522                                 mark_in_pack_object(o, p, &in_pack);
2523                         o->flags |= OBJECT_ADDED;
2524                 }
2525         }
2526
2527         if (in_pack.nr) {
2528                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2529                 for (i = 0; i < in_pack.nr; i++) {
2530                         struct object *o = in_pack.array[i].object;
2531                         add_object_entry(o->oid.hash, o->type, "", 0);
2532                 }
2533         }
2534         free(in_pack.array);
2535 }
2536
2537 static int add_loose_object(const unsigned char *sha1, const char *path,
2538                             void *data)
2539 {
2540         enum object_type type = sha1_object_info(sha1, NULL);
2541
2542         if (type < 0) {
2543                 warning("loose object at %s could not be examined", path);
2544                 return 0;
2545         }
2546
2547         add_object_entry(sha1, type, "", 0);
2548         return 0;
2549 }
2550
2551 /*
2552  * We actually don't even have to worry about reachability here.
2553  * add_object_entry will weed out duplicates, so we just add every
2554  * loose object we find.
2555  */
2556 static void add_unreachable_loose_objects(void)
2557 {
2558         for_each_loose_file_in_objdir(get_object_directory(),
2559                                       add_loose_object,
2560                                       NULL, NULL, NULL);
2561 }
2562
2563 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2564 {
2565         static struct packed_git *last_found = (void *)1;
2566         struct packed_git *p;
2567
2568         p = (last_found != (void *)1) ? last_found : packed_git;
2569
2570         while (p) {
2571                 if ((!p->pack_local || p->pack_keep) &&
2572                         find_pack_entry_one(sha1, p)) {
2573                         last_found = p;
2574                         return 1;
2575                 }
2576                 if (p == last_found)
2577                         p = packed_git;
2578                 else
2579                         p = p->next;
2580                 if (p == last_found)
2581                         p = p->next;
2582         }
2583         return 0;
2584 }
2585
2586 /*
2587  * Store a list of sha1s that are should not be discarded
2588  * because they are either written too recently, or are
2589  * reachable from another object that was.
2590  *
2591  * This is filled by get_object_list.
2592  */
2593 static struct sha1_array recent_objects;
2594
2595 static int loosened_object_can_be_discarded(const unsigned char *sha1,
2596                                             unsigned long mtime)
2597 {
2598         if (!unpack_unreachable_expiration)
2599                 return 0;
2600         if (mtime > unpack_unreachable_expiration)
2601                 return 0;
2602         if (sha1_array_lookup(&recent_objects, sha1) >= 0)
2603                 return 0;
2604         return 1;
2605 }
2606
2607 static void loosen_unused_packed_objects(struct rev_info *revs)
2608 {
2609         struct packed_git *p;
2610         uint32_t i;
2611         const unsigned char *sha1;
2612
2613         for (p = packed_git; p; p = p->next) {
2614                 if (!p->pack_local || p->pack_keep)
2615                         continue;
2616
2617                 if (open_pack_index(p))
2618                         die("cannot open pack index");
2619
2620                 for (i = 0; i < p->num_objects; i++) {
2621                         sha1 = nth_packed_object_sha1(p, i);
2622                         if (!packlist_find(&to_pack, sha1, NULL) &&
2623                             !has_sha1_pack_kept_or_nonlocal(sha1) &&
2624                             !loosened_object_can_be_discarded(sha1, p->mtime))
2625                                 if (force_object_loose(sha1, p->mtime))
2626                                         die("unable to force loose object");
2627                 }
2628         }
2629 }
2630
2631 /*
2632  * This tracks any options which pack-reuse code expects to be on, or which a
2633  * reader of the pack might not understand, and which would therefore prevent
2634  * blind reuse of what we have on disk.
2635  */
2636 static int pack_options_allow_reuse(void)
2637 {
2638         return pack_to_stdout && allow_ofs_delta;
2639 }
2640
2641 static int get_object_list_from_bitmap(struct rev_info *revs)
2642 {
2643         if (prepare_bitmap_walk(revs) < 0)
2644                 return -1;
2645
2646         if (pack_options_allow_reuse() &&
2647             !reuse_partial_packfile_from_bitmap(
2648                         &reuse_packfile,
2649                         &reuse_packfile_objects,
2650                         &reuse_packfile_offset)) {
2651                 assert(reuse_packfile_objects);
2652                 nr_result += reuse_packfile_objects;
2653                 display_progress(progress_state, nr_result);
2654         }
2655
2656         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2657         return 0;
2658 }
2659
2660 static void record_recent_object(struct object *obj,
2661                                  const char *name,
2662                                  void *data)
2663 {
2664         sha1_array_append(&recent_objects, obj->oid.hash);
2665 }
2666
2667 static void record_recent_commit(struct commit *commit, void *data)
2668 {
2669         sha1_array_append(&recent_objects, commit->object.oid.hash);
2670 }
2671
2672 static void get_object_list(int ac, const char **av)
2673 {
2674         struct rev_info revs;
2675         char line[1000];
2676         int flags = 0;
2677
2678         init_revisions(&revs, NULL);
2679         save_commit_buffer = 0;
2680         setup_revisions(ac, av, &revs, NULL);
2681
2682         /* make sure shallows are read */
2683         is_repository_shallow();
2684
2685         while (fgets(line, sizeof(line), stdin) != NULL) {
2686                 int len = strlen(line);
2687                 if (len && line[len - 1] == '\n')
2688                         line[--len] = 0;
2689                 if (!len)
2690                         break;
2691                 if (*line == '-') {
2692                         if (!strcmp(line, "--not")) {
2693                                 flags ^= UNINTERESTING;
2694                                 write_bitmap_index = 0;
2695                                 continue;
2696                         }
2697                         if (starts_with(line, "--shallow ")) {
2698                                 unsigned char sha1[20];
2699                                 if (get_sha1_hex(line + 10, sha1))
2700                                         die("not an SHA-1 '%s'", line + 10);
2701                                 register_shallow(sha1);
2702                                 use_bitmap_index = 0;
2703                                 continue;
2704                         }
2705                         die("not a rev '%s'", line);
2706                 }
2707                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2708                         die("bad revision '%s'", line);
2709         }
2710
2711         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2712                 return;
2713
2714         if (prepare_revision_walk(&revs))
2715                 die("revision walk setup failed");
2716         mark_edges_uninteresting(&revs, show_edge);
2717         traverse_commit_list(&revs, show_commit, show_object, NULL);
2718
2719         if (unpack_unreachable_expiration) {
2720                 revs.ignore_missing_links = 1;
2721                 if (add_unseen_recent_objects_to_traversal(&revs,
2722                                 unpack_unreachable_expiration))
2723                         die("unable to add recent objects");
2724                 if (prepare_revision_walk(&revs))
2725                         die("revision walk setup failed");
2726                 traverse_commit_list(&revs, record_recent_commit,
2727                                      record_recent_object, NULL);
2728         }
2729
2730         if (keep_unreachable)
2731                 add_objects_in_unpacked_packs(&revs);
2732         if (pack_loose_unreachable)
2733                 add_unreachable_loose_objects();
2734         if (unpack_unreachable)
2735                 loosen_unused_packed_objects(&revs);
2736
2737         sha1_array_clear(&recent_objects);
2738 }
2739
2740 static int option_parse_index_version(const struct option *opt,
2741                                       const char *arg, int unset)
2742 {
2743         char *c;
2744         const char *val = arg;
2745         pack_idx_opts.version = strtoul(val, &c, 10);
2746         if (pack_idx_opts.version > 2)
2747                 die(_("unsupported index version %s"), val);
2748         if (*c == ',' && c[1])
2749                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2750         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2751                 die(_("bad index version '%s'"), val);
2752         return 0;
2753 }
2754
2755 static int option_parse_unpack_unreachable(const struct option *opt,
2756                                            const char *arg, int unset)
2757 {
2758         if (unset) {
2759                 unpack_unreachable = 0;
2760                 unpack_unreachable_expiration = 0;
2761         }
2762         else {
2763                 unpack_unreachable = 1;
2764                 if (arg)
2765                         unpack_unreachable_expiration = approxidate(arg);
2766         }
2767         return 0;
2768 }
2769
2770 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2771 {
2772         int use_internal_rev_list = 0;
2773         int thin = 0;
2774         int shallow = 0;
2775         int all_progress_implied = 0;
2776         struct argv_array rp = ARGV_ARRAY_INIT;
2777         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2778         int rev_list_index = 0;
2779         struct option pack_objects_options[] = {
2780                 OPT_SET_INT('q', "quiet", &progress,
2781                             N_("do not show progress meter"), 0),
2782                 OPT_SET_INT(0, "progress", &progress,
2783                             N_("show progress meter"), 1),
2784                 OPT_SET_INT(0, "all-progress", &progress,
2785                             N_("show progress meter during object writing phase"), 2),
2786                 OPT_BOOL(0, "all-progress-implied",
2787                          &all_progress_implied,
2788                          N_("similar to --all-progress when progress meter is shown")),
2789                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2790                   N_("write the pack index file in the specified idx format version"),
2791                   0, option_parse_index_version },
2792                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2793                               N_("maximum size of each output pack file")),
2794                 OPT_BOOL(0, "local", &local,
2795                          N_("ignore borrowed objects from alternate object store")),
2796                 OPT_BOOL(0, "incremental", &incremental,
2797                          N_("ignore packed objects")),
2798                 OPT_INTEGER(0, "window", &window,
2799                             N_("limit pack window by objects")),
2800                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2801                               N_("limit pack window by memory in addition to object limit")),
2802                 OPT_INTEGER(0, "depth", &depth,
2803                             N_("maximum length of delta chain allowed in the resulting pack")),
2804                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2805                          N_("reuse existing deltas")),
2806                 OPT_BOOL(0, "reuse-object", &reuse_object,
2807                          N_("reuse existing objects")),
2808                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2809                          N_("use OFS_DELTA objects")),
2810                 OPT_INTEGER(0, "threads", &delta_search_threads,
2811                             N_("use threads when searching for best delta matches")),
2812                 OPT_BOOL(0, "non-empty", &non_empty,
2813                          N_("do not create an empty pack output")),
2814                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2815                          N_("read revision arguments from standard input")),
2816                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2817                   N_("limit the objects to those that are not yet packed"),
2818                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2819                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2820                   N_("include objects reachable from any reference"),
2821                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2822                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2823                   N_("include objects referred by reflog entries"),
2824                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2825                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2826                   N_("include objects referred to by the index"),
2827                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2828                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2829                          N_("output pack to stdout")),
2830                 OPT_BOOL(0, "include-tag", &include_tag,
2831                          N_("include tag objects that refer to objects to be packed")),
2832                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2833                          N_("keep unreachable objects")),
2834                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
2835                          N_("pack loose unreachable objects")),
2836                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2837                   N_("unpack unreachable objects newer than <time>"),
2838                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2839                 OPT_BOOL(0, "thin", &thin,
2840                          N_("create thin packs")),
2841                 OPT_BOOL(0, "shallow", &shallow,
2842                          N_("create packs suitable for shallow fetches")),
2843                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2844                          N_("ignore packs that have companion .keep file")),
2845                 OPT_INTEGER(0, "compression", &pack_compression_level,
2846                             N_("pack compression level")),
2847                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2848                             N_("do not hide commits by grafts"), 0),
2849                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2850                          N_("use a bitmap index if available to speed up counting objects")),
2851                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2852                          N_("write a bitmap index together with the pack index")),
2853                 OPT_END(),
2854         };
2855
2856         check_replace_refs = 0;
2857
2858         reset_pack_idx_option(&pack_idx_opts);
2859         git_config(git_pack_config, NULL);
2860
2861         progress = isatty(2);
2862         argc = parse_options(argc, argv, prefix, pack_objects_options,
2863                              pack_usage, 0);
2864
2865         if (argc) {
2866                 base_name = argv[0];
2867                 argc--;
2868         }
2869         if (pack_to_stdout != !base_name || argc)
2870                 usage_with_options(pack_usage, pack_objects_options);
2871
2872         argv_array_push(&rp, "pack-objects");
2873         if (thin) {
2874                 use_internal_rev_list = 1;
2875                 argv_array_push(&rp, shallow
2876                                 ? "--objects-edge-aggressive"
2877                                 : "--objects-edge");
2878         } else
2879                 argv_array_push(&rp, "--objects");
2880
2881         if (rev_list_all) {
2882                 use_internal_rev_list = 1;
2883                 argv_array_push(&rp, "--all");
2884         }
2885         if (rev_list_reflog) {
2886                 use_internal_rev_list = 1;
2887                 argv_array_push(&rp, "--reflog");
2888         }
2889         if (rev_list_index) {
2890                 use_internal_rev_list = 1;
2891                 argv_array_push(&rp, "--indexed-objects");
2892         }
2893         if (rev_list_unpacked) {
2894                 use_internal_rev_list = 1;
2895                 argv_array_push(&rp, "--unpacked");
2896         }
2897
2898         if (!reuse_object)
2899                 reuse_delta = 0;
2900         if (pack_compression_level == -1)
2901                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2902         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2903                 die("bad pack compression level %d", pack_compression_level);
2904
2905         if (!delta_search_threads)      /* --threads=0 means autodetect */
2906                 delta_search_threads = online_cpus();
2907
2908 #ifdef NO_PTHREADS
2909         if (delta_search_threads != 1)
2910                 warning("no threads support, ignoring --threads");
2911 #endif
2912         if (!pack_to_stdout && !pack_size_limit)
2913                 pack_size_limit = pack_size_limit_cfg;
2914         if (pack_to_stdout && pack_size_limit)
2915                 die("--max-pack-size cannot be used to build a pack for transfer.");
2916         if (pack_size_limit && pack_size_limit < 1024*1024) {
2917                 warning("minimum pack size limit is 1 MiB");
2918                 pack_size_limit = 1024*1024;
2919         }
2920
2921         if (!pack_to_stdout && thin)
2922                 die("--thin cannot be used to build an indexable pack.");
2923
2924         if (keep_unreachable && unpack_unreachable)
2925                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2926         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
2927                 unpack_unreachable_expiration = 0;
2928
2929         /*
2930          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
2931          *
2932          * - to produce good pack (with bitmap index not-yet-packed objects are
2933          *   packed in suboptimal order).
2934          *
2935          * - to use more robust pack-generation codepath (avoiding possible
2936          *   bugs in bitmap code and possible bitmap index corruption).
2937          */
2938         if (!pack_to_stdout)
2939                 use_bitmap_index_default = 0;
2940
2941         if (use_bitmap_index < 0)
2942                 use_bitmap_index = use_bitmap_index_default;
2943
2944         /* "hard" reasons not to use bitmaps; these just won't work at all */
2945         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
2946                 use_bitmap_index = 0;
2947
2948         if (pack_to_stdout || !rev_list_all)
2949                 write_bitmap_index = 0;
2950
2951         if (progress && all_progress_implied)
2952                 progress = 2;
2953
2954         prepare_packed_git();
2955         if (ignore_packed_keep) {
2956                 struct packed_git *p;
2957                 for (p = packed_git; p; p = p->next)
2958                         if (p->pack_local && p->pack_keep)
2959                                 break;
2960                 if (!p) /* no keep-able packs found */
2961                         ignore_packed_keep = 0;
2962         }
2963         if (local) {
2964                 /*
2965                  * unlike ignore_packed_keep above, we do not want to
2966                  * unset "local" based on looking at packs, as it
2967                  * also covers non-local objects
2968                  */
2969                 struct packed_git *p;
2970                 for (p = packed_git; p; p = p->next) {
2971                         if (!p->pack_local) {
2972                                 have_non_local_packs = 1;
2973                                 break;
2974                         }
2975                 }
2976         }
2977
2978         if (progress)
2979                 progress_state = start_progress(_("Counting objects"), 0);
2980         if (!use_internal_rev_list)
2981                 read_object_list_from_stdin();
2982         else {
2983                 get_object_list(rp.argc, rp.argv);
2984                 argv_array_clear(&rp);
2985         }
2986         cleanup_preferred_base();
2987         if (include_tag && nr_result)
2988                 for_each_ref(add_ref_tag, NULL);
2989         stop_progress(&progress_state);
2990
2991         if (non_empty && !nr_result)
2992                 return 0;
2993         if (nr_result)
2994                 prepare_pack(window, depth);
2995         write_pack_file();
2996         if (progress)
2997                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2998                         " reused %"PRIu32" (delta %"PRIu32")\n",
2999                         written, written_delta, reused, reused_delta);
3000         return 0;
3001 }