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 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
  65 static int pack_compression_seen;
  66
  67 static struct packed_git *reuse_packfile;
  68 static uint32_t reuse_packfile_objects;
  69 static off_t reuse_packfile_offset;
  70
  71 static int use_bitmap_index_default = 1;
  72 static int use_bitmap_index = -1;
  73 static int write_bitmap_index;
  74 static uint16_t write_bitmap_options;
  75
  76 static unsigned long delta_cache_size = 0;
  77 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
  78 static unsigned long cache_max_small_delta_size = 1000;
  79
  80 static unsigned long window_memory_limit = 0;
  81
  82 /*
  83  * stats
  84  */
  85 static uint32_t written, written_delta;
  86 static uint32_t reused, reused_delta;
  87
  88 /*
  89  * Indexed commits
  90  */
  91 static struct commit **indexed_commits;
  92 static unsigned int indexed_commits_nr;
  93 static unsigned int indexed_commits_alloc;
  94
  95 static void index_commit_for_bitmap(struct commit *commit)
  96 {
  97         if (indexed_commits_nr >= indexed_commits_alloc) {
  98                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
  99                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
 100         }
 101
 102         indexed_commits[indexed_commits_nr++] = commit;
 103 }
 104
 105 static void *get_delta(struct object_entry *entry)
 106 {
 107         unsigned long size, base_size, delta_size;
 108         void *buf, *base_buf, *delta_buf;
 109         enum object_type type;
 110
 111         buf = read_sha1_file(entry->idx.sha1, &type, &size);
 112         if (!buf)
 113                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
 114         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
 115         if (!base_buf)
 116                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
 117         delta_buf = diff_delta(base_buf, base_size,
 118                                buf, size, &delta_size, 0);
 119         if (!delta_buf || delta_size != entry->delta_size)
 120                 die("delta size changed");
 121         free(buf);
 122         free(base_buf);
 123         return delta_buf;
 124 }
 125
 126 static unsigned long do_compress(void **pptr, unsigned long size)
 127 {
 128         git_zstream stream;
 129         void *in, *out;
 130         unsigned long maxsize;
 131
 132         git_deflate_init(&stream, pack_compression_level);
 133         maxsize = git_deflate_bound(&stream, size);
 134
 135         in = *pptr;
 136         out = xmalloc(maxsize);
 137         *pptr = out;
 138
 139         stream.next_in = in;
 140         stream.avail_in = size;
 141         stream.next_out = out;
 142         stream.avail_out = maxsize;
 143         while (git_deflate(&stream, Z_FINISH) == Z_OK)
 144                 ; /* nothing */
 145         git_deflate_end(&stream);
 146
 147         free(in);
 148         return stream.total_out;
 149 }
 150
 151 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
 152                                            const unsigned char *sha1)
 153 {
 154         git_zstream stream;
 155         unsigned char ibuf[1024 * 16];
 156         unsigned char obuf[1024 * 16];
 157         unsigned long olen = 0;
 158
 159         git_deflate_init(&stream, pack_compression_level);
 160
 161         for (;;) {
 162                 ssize_t readlen;
 163                 int zret = Z_OK;
 164                 readlen = read_istream(st, ibuf, sizeof(ibuf));
 165                 if (readlen == -1)
 166                         die(_("unable to read %s"), sha1_to_hex(sha1));
 167
 168                 stream.next_in = ibuf;
 169                 stream.avail_in = readlen;
 170                 while ((stream.avail_in || readlen == 0) &&
 171                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
 172                         stream.next_out = obuf;
 173                         stream.avail_out = sizeof(obuf);
 174                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 175                         sha1write(f, obuf, stream.next_out - obuf);
 176                         olen += stream.next_out - obuf;
 177                 }
 178                 if (stream.avail_in)
 179                         die(_("deflate error (%d)"), zret);
 180                 if (readlen == 0) {
 181                         if (zret != Z_STREAM_END)
 182                                 die(_("deflate error (%d)"), zret);
 183                         break;
 184                 }
 185         }
 186         git_deflate_end(&stream);
 187         return olen;
 188 }
 189
 190 /*
 191  * we are going to reuse the existing object data as is.  make
 192  * sure it is not corrupt.
 193  */
 194 static int check_pack_inflate(struct packed_git *p,
 195                 struct pack_window **w_curs,
 196                 off_t offset,
 197                 off_t len,
 198                 unsigned long expect)
 199 {
 200         git_zstream stream;
 201         unsigned char fakebuf[4096], *in;
 202         int st;
 203
 204         memset(&stream, 0, sizeof(stream));
 205         git_inflate_init(&stream);
 206         do {
 207                 in = use_pack(p, w_curs, offset, &stream.avail_in);
 208                 stream.next_in = in;
 209                 stream.next_out = fakebuf;
 210                 stream.avail_out = sizeof(fakebuf);
 211                 st = git_inflate(&stream, Z_FINISH);
 212                 offset += stream.next_in - in;
 213         } while (st == Z_OK || st == Z_BUF_ERROR);
 214         git_inflate_end(&stream);
 215         return (st == Z_STREAM_END &&
 216                 stream.total_out == expect &&
 217                 stream.total_in == len) ? 0 : -1;
 218 }
 219
 220 static void copy_pack_data(struct sha1file *f,
 221                 struct packed_git *p,
 222                 struct pack_window **w_curs,
 223                 off_t offset,
 224                 off_t len)
 225 {
 226         unsigned char *in;
 227         unsigned long avail;
 228
 229         while (len) {
 230                 in = use_pack(p, w_curs, offset, &avail);
 231                 if (avail > len)
 232                         avail = (unsigned long)len;
 233                 sha1write(f, in, avail);
 234                 offset += avail;
 235                 len -= avail;
 236         }
 237 }
 238
 239 /* Return 0 if we will bust the pack-size limit */
 240 static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
 241                                            unsigned long limit, int usable_delta)
 242 {
 243         unsigned long size, datalen;
 244         unsigned char header[10], dheader[10];
 245         unsigned hdrlen;
 246         enum object_type type;
 247         void *buf;
 248         struct git_istream *st = NULL;
 249
 250         if (!usable_delta) {
 251                 if (entry->type == OBJ_BLOB &&
 252                     entry->size > big_file_threshold &&
 253                     (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
 254                         buf = NULL;
 255                 else {
 256                         buf = read_sha1_file(entry->idx.sha1, &type, &size);
 257                         if (!buf)
 258                                 die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
 259                 }
 260                 /*
 261                  * make sure no cached delta data remains from a
 262                  * previous attempt before a pack split occurred.
 263                  */
 264                 free(entry->delta_data);
 265                 entry->delta_data = NULL;
 266                 entry->z_delta_size = 0;
 267         } else if (entry->delta_data) {
 268                 size = entry->delta_size;
 269                 buf = entry->delta_data;
 270                 entry->delta_data = NULL;
 271                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 272                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 273         } else {
 274                 buf = get_delta(entry);
 275                 size = entry->delta_size;
 276                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 277                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 278         }
 279
 280         if (st) /* large blob case, just assume we don't compress well */
 281                 datalen = size;
 282         else if (entry->z_delta_size)
 283                 datalen = entry->z_delta_size;
 284         else
 285                 datalen = do_compress(&buf, size);
 286
 287         /*
 288          * The object header is a byte of 'type' followed by zero or
 289          * more bytes of length.
 290          */
 291         hdrlen = encode_in_pack_object_header(type, size, header);
 292
 293         if (type == OBJ_OFS_DELTA) {
 294                 /*
 295                  * Deltas with relative base contain an additional
 296                  * encoding of the relative offset for the delta
 297                  * base from this object's position in the pack.
 298                  */
 299                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 300                 unsigned pos = sizeof(dheader) - 1;
 301                 dheader[pos] = ofs & 127;
 302                 while (ofs >>= 7)
 303                         dheader[--pos] = 128 | (--ofs & 127);
 304                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 305                         if (st)
 306                                 close_istream(st);
 307                         free(buf);
 308                         return 0;
 309                 }
 310                 sha1write(f, header, hdrlen);
 311                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 312                 hdrlen += sizeof(dheader) - pos;
 313         } else if (type == OBJ_REF_DELTA) {
 314                 /*
 315                  * Deltas with a base reference contain
 316                  * an additional 20 bytes for the base sha1.
 317                  */
 318                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 319                         if (st)
 320                                 close_istream(st);
 321                         free(buf);
 322                         return 0;
 323                 }
 324                 sha1write(f, header, hdrlen);
 325                 sha1write(f, entry->delta->idx.sha1, 20);
 326                 hdrlen += 20;
 327         } else {
 328                 if (limit && hdrlen + datalen + 20 >= limit) {
 329                         if (st)
 330                                 close_istream(st);
 331                         free(buf);
 332                         return 0;
 333                 }
 334                 sha1write(f, header, hdrlen);
 335         }
 336         if (st) {
 337                 datalen = write_large_blob_data(st, f, entry->idx.sha1);
 338                 close_istream(st);
 339         } else {
 340                 sha1write(f, buf, datalen);
 341                 free(buf);
 342         }
 343
 344         return hdrlen + datalen;
 345 }
 346
 347 /* Return 0 if we will bust the pack-size limit */
 348 static off_t write_reuse_object(struct sha1file *f, struct object_entry *entry,
 349                                 unsigned long limit, int usable_delta)
 350 {
 351         struct packed_git *p = entry->in_pack;
 352         struct pack_window *w_curs = NULL;
 353         struct revindex_entry *revidx;
 354         off_t offset;
 355         enum object_type type = entry->type;
 356         off_t datalen;
 357         unsigned char header[10], dheader[10];
 358         unsigned hdrlen;
 359
 360         if (entry->delta)
 361                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 362                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 363         hdrlen = encode_in_pack_object_header(type, entry->size, header);
 364
 365         offset = entry->in_pack_offset;
 366         revidx = find_pack_revindex(p, offset);
 367         datalen = revidx[1].offset - offset;
 368         if (!pack_to_stdout && p->index_version > 1 &&
 369             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 370                 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
 371                 unuse_pack(&w_curs);
 372                 return write_no_reuse_object(f, entry, limit, usable_delta);
 373         }
 374
 375         offset += entry->in_pack_header_size;
 376         datalen -= entry->in_pack_header_size;
 377
 378         if (!pack_to_stdout && p->index_version == 1 &&
 379             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 380                 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
 381                 unuse_pack(&w_curs);
 382                 return write_no_reuse_object(f, entry, limit, usable_delta);
 383         }
 384
 385         if (type == OBJ_OFS_DELTA) {
 386                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 387                 unsigned pos = sizeof(dheader) - 1;
 388                 dheader[pos] = ofs & 127;
 389                 while (ofs >>= 7)
 390                         dheader[--pos] = 128 | (--ofs & 127);
 391                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 392                         unuse_pack(&w_curs);
 393                         return 0;
 394                 }
 395                 sha1write(f, header, hdrlen);
 396                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 397                 hdrlen += sizeof(dheader) - pos;
 398                 reused_delta++;
 399         } else if (type == OBJ_REF_DELTA) {
 400                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 401                         unuse_pack(&w_curs);
 402                         return 0;
 403                 }
 404                 sha1write(f, header, hdrlen);
 405                 sha1write(f, entry->delta->idx.sha1, 20);
 406                 hdrlen += 20;
 407                 reused_delta++;
 408         } else {
 409                 if (limit && hdrlen + datalen + 20 >= limit) {
 410                         unuse_pack(&w_curs);
 411                         return 0;
 412                 }
 413                 sha1write(f, header, hdrlen);
 414         }
 415         copy_pack_data(f, p, &w_curs, offset, datalen);
 416         unuse_pack(&w_curs);
 417         reused++;
 418         return hdrlen + datalen;
 419 }
 420
 421 /* Return 0 if we will bust the pack-size limit */
 422 static off_t write_object(struct sha1file *f,
 423                           struct object_entry *entry,
 424                           off_t write_offset)
 425 {
 426         unsigned long limit;
 427         off_t len;
 428         int usable_delta, to_reuse;
 429
 430         if (!pack_to_stdout)
 431                 crc32_begin(f);
 432
 433         /* apply size limit if limited packsize and not first object */
 434         if (!pack_size_limit || !nr_written)
 435                 limit = 0;
 436         else if (pack_size_limit <= write_offset)
 437                 /*
 438                  * the earlier object did not fit the limit; avoid
 439                  * mistaking this with unlimited (i.e. limit = 0).
 440                  */
 441                 limit = 1;
 442         else
 443                 limit = pack_size_limit - write_offset;
 444
 445         if (!entry->delta)
 446                 usable_delta = 0;       /* no delta */
 447         else if (!pack_size_limit)
 448                usable_delta = 1;        /* unlimited packfile */
 449         else if (entry->delta->idx.offset == (off_t)-1)
 450                 usable_delta = 0;       /* base was written to another pack */
 451         else if (entry->delta->idx.offset)
 452                 usable_delta = 1;       /* base already exists in this pack */
 453         else
 454                 usable_delta = 0;       /* base could end up in another pack */
 455
 456         if (!reuse_object)
 457                 to_reuse = 0;   /* explicit */
 458         else if (!entry->in_pack)
 459                 to_reuse = 0;   /* can't reuse what we don't have */
 460         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
 461                                 /* check_object() decided it for us ... */
 462                 to_reuse = usable_delta;
 463                                 /* ... but pack split may override that */
 464         else if (entry->type != entry->in_pack_type)
 465                 to_reuse = 0;   /* pack has delta which is unusable */
 466         else if (entry->delta)
 467                 to_reuse = 0;   /* we want to pack afresh */
 468         else
 469                 to_reuse = 1;   /* we have it in-pack undeltified,
 470                                  * and we do not need to deltify it.
 471                                  */
 472
 473         if (!to_reuse)
 474                 len = write_no_reuse_object(f, entry, limit, usable_delta);
 475         else
 476                 len = write_reuse_object(f, entry, limit, usable_delta);
 477         if (!len)
 478                 return 0;
 479
 480         if (usable_delta)
 481                 written_delta++;
 482         written++;
 483         if (!pack_to_stdout)
 484                 entry->idx.crc32 = crc32_end(f);
 485         return len;
 486 }
 487
 488 enum write_one_status {
 489         WRITE_ONE_SKIP = -1, /* already written */
 490         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 491         WRITE_ONE_WRITTEN = 1, /* normal */
 492         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 493 };
 494
 495 static enum write_one_status write_one(struct sha1file *f,
 496                                        struct object_entry *e,
 497                                        off_t *offset)
 498 {
 499         off_t size;
 500         int recursing;
 501
 502         /*
 503          * we set offset to 1 (which is an impossible value) to mark
 504          * the fact that this object is involved in "write its base
 505          * first before writing a deltified object" recursion.
 506          */
 507         recursing = (e->idx.offset == 1);
 508         if (recursing) {
 509                 warning("recursive delta detected for object %s",
 510                         sha1_to_hex(e->idx.sha1));
 511                 return WRITE_ONE_RECURSIVE;
 512         } else if (e->idx.offset || e->preferred_base) {
 513                 /* offset is non zero if object is written already. */
 514                 return WRITE_ONE_SKIP;
 515         }
 516
 517         /* if we are deltified, write out base object first. */
 518         if (e->delta) {
 519                 e->idx.offset = 1; /* now recurse */
 520                 switch (write_one(f, e->delta, offset)) {
 521                 case WRITE_ONE_RECURSIVE:
 522                         /* we cannot depend on this one */
 523                         e->delta = NULL;
 524                         break;
 525                 default:
 526                         break;
 527                 case WRITE_ONE_BREAK:
 528                         e->idx.offset = recursing;
 529                         return WRITE_ONE_BREAK;
 530                 }
 531         }
 532
 533         e->idx.offset = *offset;
 534         size = write_object(f, e, *offset);
 535         if (!size) {
 536                 e->idx.offset = recursing;
 537                 return WRITE_ONE_BREAK;
 538         }
 539         written_list[nr_written++] = &e->idx;
 540
 541         /* make sure off_t is sufficiently large not to wrap */
 542         if (signed_add_overflows(*offset, size))
 543                 die("pack too large for current definition of off_t");
 544         *offset += size;
 545         return WRITE_ONE_WRITTEN;
 546 }
 547
 548 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
 549                        void *cb_data)
 550 {
 551         unsigned char peeled[20];
 552         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
 553
 554         if (entry)
 555                 entry->tagged = 1;
 556         if (!peel_ref(path, peeled)) {
 557                 entry = packlist_find(&to_pack, peeled, NULL);
 558                 if (entry)
 559                         entry->tagged = 1;
 560         }
 561         return 0;
 562 }
 563
 564 static inline void add_to_write_order(struct object_entry **wo,
 565                                unsigned int *endp,
 566                                struct object_entry *e)
 567 {
 568         if (e->filled)
 569                 return;
 570         wo[(*endp)++] = e;
 571         e->filled = 1;
 572 }
 573
 574 static void add_descendants_to_write_order(struct object_entry **wo,
 575                                            unsigned int *endp,
 576                                            struct object_entry *e)
 577 {
 578         int add_to_order = 1;
 579         while (e) {
 580                 if (add_to_order) {
 581                         struct object_entry *s;
 582                         /* add this node... */
 583                         add_to_write_order(wo, endp, e);
 584                         /* all its siblings... */
 585                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
 586                                 add_to_write_order(wo, endp, s);
 587                         }
 588                 }
 589                 /* drop down a level to add left subtree nodes if possible */
 590                 if (e->delta_child) {
 591                         add_to_order = 1;
 592                         e = e->delta_child;
 593                 } else {
 594                         add_to_order = 0;
 595                         /* our sibling might have some children, it is next */
 596                         if (e->delta_sibling) {
 597                                 e = e->delta_sibling;
 598                                 continue;
 599                         }
 600                         /* go back to our parent node */
 601                         e = e->delta;
 602                         while (e && !e->delta_sibling) {
 603                                 /* we're on the right side of a subtree, keep
 604                                  * going up until we can go right again */
 605                                 e = e->delta;
 606                         }
 607                         if (!e) {
 608                                 /* done- we hit our original root node */
 609                                 return;
 610                         }
 611                         /* pass it off to sibling at this level */
 612                         e = e->delta_sibling;
 613                 }
 614         };
 615 }
 616
 617 static void add_family_to_write_order(struct object_entry **wo,
 618                                       unsigned int *endp,
 619                                       struct object_entry *e)
 620 {
 621         struct object_entry *root;
 622
 623         for (root = e; root->delta; root = root->delta)
 624                 ; /* nothing */
 625         add_descendants_to_write_order(wo, endp, root);
 626 }
 627
 628 static struct object_entry **compute_write_order(void)
 629 {
 630         unsigned int i, wo_end, last_untagged;
 631
 632         struct object_entry **wo;
 633         struct object_entry *objects = to_pack.objects;
 634
 635         for (i = 0; i < to_pack.nr_objects; i++) {
 636                 objects[i].tagged = 0;
 637                 objects[i].filled = 0;
 638                 objects[i].delta_child = NULL;
 639                 objects[i].delta_sibling = NULL;
 640         }
 641
 642         /*
 643          * Fully connect delta_child/delta_sibling network.
 644          * Make sure delta_sibling is sorted in the original
 645          * recency order.
 646          */
 647         for (i = to_pack.nr_objects; i > 0;) {
 648                 struct object_entry *e = &objects[--i];
 649                 if (!e->delta)
 650                         continue;
 651                 /* Mark me as the first child */
 652                 e->delta_sibling = e->delta->delta_child;
 653                 e->delta->delta_child = e;
 654         }
 655
 656         /*
 657          * Mark objects that are at the tip of tags.
 658          */
 659         for_each_tag_ref(mark_tagged, NULL);
 660
 661         /*
 662          * Give the objects in the original recency order until
 663          * we see a tagged tip.
 664          */
 665         ALLOC_ARRAY(wo, to_pack.nr_objects);
 666         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
 667                 if (objects[i].tagged)
 668                         break;
 669                 add_to_write_order(wo, &wo_end, &objects[i]);
 670         }
 671         last_untagged = i;
 672
 673         /*
 674          * Then fill all the tagged tips.
 675          */
 676         for (; i < to_pack.nr_objects; i++) {
 677                 if (objects[i].tagged)
 678                         add_to_write_order(wo, &wo_end, &objects[i]);
 679         }
 680
 681         /*
 682          * And then all remaining commits and tags.
 683          */
 684         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 685                 if (objects[i].type != OBJ_COMMIT &&
 686                     objects[i].type != OBJ_TAG)
 687                         continue;
 688                 add_to_write_order(wo, &wo_end, &objects[i]);
 689         }
 690
 691         /*
 692          * And then all the trees.
 693          */
 694         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 695                 if (objects[i].type != OBJ_TREE)
 696                         continue;
 697                 add_to_write_order(wo, &wo_end, &objects[i]);
 698         }
 699
 700         /*
 701          * Finally all the rest in really tight order
 702          */
 703         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 704                 if (!objects[i].filled)
 705                         add_family_to_write_order(wo, &wo_end, &objects[i]);
 706         }
 707
 708         if (wo_end != to_pack.nr_objects)
 709                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
 710
 711         return wo;
 712 }
 713
 714 static off_t write_reused_pack(struct sha1file *f)
 715 {
 716         unsigned char buffer[8192];
 717         off_t to_write, total;
 718         int fd;
 719
 720         if (!is_pack_valid(reuse_packfile))
 721                 die("packfile is invalid: %s", reuse_packfile->pack_name);
 722
 723         fd = git_open(reuse_packfile->pack_name);
 724         if (fd < 0)
 725                 die_errno("unable to open packfile for reuse: %s",
 726                           reuse_packfile->pack_name);
 727
 728         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
 729                 die_errno("unable to seek in reused packfile");
 730
 731         if (reuse_packfile_offset < 0)
 732                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
 733
 734         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
 735
 736         while (to_write) {
 737                 int read_pack = xread(fd, buffer, sizeof(buffer));
 738
 739                 if (read_pack <= 0)
 740                         die_errno("unable to read from reused packfile");
 741
 742                 if (read_pack > to_write)
 743                         read_pack = to_write;
 744
 745                 sha1write(f, buffer, read_pack);
 746                 to_write -= read_pack;
 747
 748                 /*
 749                  * We don't know the actual number of objects written,
 750                  * only how many bytes written, how many bytes total, and
 751                  * how many objects total. So we can fake it by pretending all
 752                  * objects we are writing are the same size. This gives us a
 753                  * smooth progress meter, and at the end it matches the true
 754                  * answer.
 755                  */
 756                 written = reuse_packfile_objects *
 757                                 (((double)(total - to_write)) / total);
 758                 display_progress(progress_state, written);
 759         }
 760
 761         close(fd);
 762         written = reuse_packfile_objects;
 763         display_progress(progress_state, written);
 764         return reuse_packfile_offset - sizeof(struct pack_header);
 765 }
 766
 767 static const char no_split_warning[] = N_(
 768 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
 769 );
 770
 771 static void write_pack_file(void)
 772 {
 773         uint32_t i = 0, j;
 774         struct sha1file *f;
 775         off_t offset;
 776         uint32_t nr_remaining = nr_result;
 777         time_t last_mtime = 0;
 778         struct object_entry **write_order;
 779
 780         if (progress > pack_to_stdout)
 781                 progress_state = start_progress(_("Writing objects"), nr_result);
 782         ALLOC_ARRAY(written_list, to_pack.nr_objects);
 783         write_order = compute_write_order();
 784
 785         do {
 786                 unsigned char sha1[20];
 787                 char *pack_tmp_name = NULL;
 788
 789                 if (pack_to_stdout)
 790                         f = sha1fd_throughput(1, "<stdout>", progress_state);
 791                 else
 792                         f = create_tmp_packfile(&pack_tmp_name);
 793
 794                 offset = write_pack_header(f, nr_remaining);
 795
 796                 if (reuse_packfile) {
 797                         off_t packfile_size;
 798                         assert(pack_to_stdout);
 799
 800                         packfile_size = write_reused_pack(f);
 801                         offset += packfile_size;
 802                 }
 803
 804                 nr_written = 0;
 805                 for (; i < to_pack.nr_objects; i++) {
 806                         struct object_entry *e = write_order[i];
 807                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 808                                 break;
 809                         display_progress(progress_state, written);
 810                 }
 811
 812                 /*
 813                  * Did we write the wrong # entries in the header?
 814                  * If so, rewrite it like in fast-import
 815                  */
 816                 if (pack_to_stdout) {
 817                         sha1close(f, sha1, CSUM_CLOSE);
 818                 } else if (nr_written == nr_remaining) {
 819                         sha1close(f, sha1, CSUM_FSYNC);
 820                 } else {
 821                         int fd = sha1close(f, sha1, 0);
 822                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
 823                                                  nr_written, sha1, offset);
 824                         close(fd);
 825                         if (write_bitmap_index) {
 826                                 warning(_(no_split_warning));
 827                                 write_bitmap_index = 0;
 828                         }
 829                 }
 830
 831                 if (!pack_to_stdout) {
 832                         struct stat st;
 833                         struct strbuf tmpname = STRBUF_INIT;
 834
 835                         /*
 836                          * Packs are runtime accessed in their mtime
 837                          * order since newer packs are more likely to contain
 838                          * younger objects.  So if we are creating multiple
 839                          * packs then we should modify the mtime of later ones
 840                          * to preserve this property.
 841                          */
 842                         if (stat(pack_tmp_name, &st) < 0) {
 843                                 warning_errno("failed to stat %s", pack_tmp_name);
 844                         } else if (!last_mtime) {
 845                                 last_mtime = st.st_mtime;
 846                         } else {
 847                                 struct utimbuf utb;
 848                                 utb.actime = st.st_atime;
 849                                 utb.modtime = --last_mtime;
 850                                 if (utime(pack_tmp_name, &utb) < 0)
 851                                         warning_errno("failed utime() on %s", pack_tmp_name);
 852                         }
 853
 854                         strbuf_addf(&tmpname, "%s-", base_name);
 855
 856                         if (write_bitmap_index) {
 857                                 bitmap_writer_set_checksum(sha1);
 858                                 bitmap_writer_build_type_index(written_list, nr_written);
 859                         }
 860
 861                         finish_tmp_packfile(&tmpname, pack_tmp_name,
 862                                             written_list, nr_written,
 863                                             &pack_idx_opts, sha1);
 864
 865                         if (write_bitmap_index) {
 866                                 strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
 867
 868                                 stop_progress(&progress_state);
 869
 870                                 bitmap_writer_show_progress(progress);
 871                                 bitmap_writer_reuse_bitmaps(&to_pack);
 872                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
 873                                 bitmap_writer_build(&to_pack);
 874                                 bitmap_writer_finish(written_list, nr_written,
 875                                                      tmpname.buf, write_bitmap_options);
 876                                 write_bitmap_index = 0;
 877                         }
 878
 879                         strbuf_release(&tmpname);
 880                         free(pack_tmp_name);
 881                         puts(sha1_to_hex(sha1));
 882                 }
 883
 884                 /* mark written objects as written to previous pack */
 885                 for (j = 0; j < nr_written; j++) {
 886                         written_list[j]->offset = (off_t)-1;
 887                 }
 888                 nr_remaining -= nr_written;
 889         } while (nr_remaining && i < to_pack.nr_objects);
 890
 891         free(written_list);
 892         free(write_order);
 893         stop_progress(&progress_state);
 894         if (written != nr_result)
 895                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
 896                         written, nr_result);
 897 }
 898
 899 static void setup_delta_attr_check(struct git_attr_check *check)
 900 {
 901         static struct git_attr *attr_delta;
 902
 903         if (!attr_delta)
 904                 attr_delta = git_attr("delta");
 905
 906         check[0].attr = attr_delta;
 907 }
 908
 909 static int no_try_delta(const char *path)
 910 {
 911         struct git_attr_check check[1];
 912
 913         setup_delta_attr_check(check);
 914         if (git_check_attr(path, ARRAY_SIZE(check), check))
 915                 return 0;
 916         if (ATTR_FALSE(check->value))
 917                 return 1;
 918         return 0;
 919 }
 920
 921 /*
 922  * When adding an object, check whether we have already added it
 923  * to our packing list. If so, we can skip. However, if we are
 924  * being asked to excludei t, but the previous mention was to include
 925  * it, make sure to adjust its flags and tweak our numbers accordingly.
 926  *
 927  * As an optimization, we pass out the index position where we would have
 928  * found the item, since that saves us from having to look it up again a
 929  * few lines later when we want to add the new entry.
 930  */
 931 static int have_duplicate_entry(const unsigned char *sha1,
 932                                 int exclude,
 933                                 uint32_t *index_pos)
 934 {
 935         struct object_entry *entry;
 936
 937         entry = packlist_find(&to_pack, sha1, index_pos);
 938         if (!entry)
 939                 return 0;
 940
 941         if (exclude) {
 942                 if (!entry->preferred_base)
 943                         nr_result--;
 944                 entry->preferred_base = 1;
 945         }
 946
 947         return 1;
 948 }
 949
 950 static int want_found_object(int exclude, struct packed_git *p)
 951 {
 952         if (exclude)
 953                 return 1;
 954         if (incremental)
 955                 return 0;
 956
 957         /*
 958          * When asked to do --local (do not include an object that appears in a
 959          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
 960          * an object that appears in a pack marked with .keep), finding a pack
 961          * that matches the criteria is sufficient for us to decide to omit it.
 962          * However, even if this pack does not satisfy the criteria, we need to
 963          * make sure no copy of this object appears in _any_ pack that makes us
 964          * to omit the object, so we need to check all the packs.
 965          *
 966          * We can however first check whether these options can possible matter;
 967          * if they do not matter we know we want the object in generated pack.
 968          * Otherwise, we signal "-1" at the end to tell the caller that we do
 969          * not know either way, and it needs to check more packs.
 970          */
 971         if (!ignore_packed_keep &&
 972             (!local || !have_non_local_packs))
 973                 return 1;
 974
 975         if (local && !p->pack_local)
 976                 return 0;
 977         if (ignore_packed_keep && p->pack_local && p->pack_keep)
 978                 return 0;
 979
 980         /* we don't know yet; keep looking for more packs */
 981         return -1;
 982 }
 983
 984 /*
 985  * Check whether we want the object in the pack (e.g., we do not want
 986  * objects found in non-local stores if the "--local" option was used).
 987  *
 988  * If the caller already knows an existing pack it wants to take the object
 989  * from, that is passed in *found_pack and *found_offset; otherwise this
 990  * function finds if there is any pack that has the object and returns the pack
 991  * and its offset in these variables.
 992  */
 993 static int want_object_in_pack(const unsigned char *sha1,
 994                                int exclude,
 995                                struct packed_git **found_pack,
 996                                off_t *found_offset)
 997 {
 998         struct mru_entry *entry;
 999         int want;
1000
1001         if (!exclude && local && has_loose_object_nonlocal(sha1))
1002                 return 0;
1003
1004         /*
1005          * If we already know the pack object lives in, start checks from that
1006          * pack - in the usual case when neither --local was given nor .keep files
1007          * are present we will determine the answer right now.
1008          */
1009         if (*found_pack) {
1010                 want = want_found_object(exclude, *found_pack);
1011                 if (want != -1)
1012                         return want;
1013         }
1014
1015         for (entry = packed_git_mru->head; entry; entry = entry->next) {
1016                 struct packed_git *p = entry->item;
1017                 off_t offset;
1018
1019                 if (p == *found_pack)
1020                         offset = *found_offset;
1021                 else
1022                         offset = find_pack_entry_one(sha1, p);
1023
1024                 if (offset) {
1025                         if (!*found_pack) {
1026                                 if (!is_pack_valid(p))
1027                                         continue;
1028                                 *found_offset = offset;
1029                                 *found_pack = p;
1030                         }
1031                         want = want_found_object(exclude, p);
1032                         if (!exclude && want > 0)
1033                                 mru_mark(packed_git_mru, entry);
1034                         if (want != -1)
1035                                 return want;
1036                 }
1037         }
1038
1039         return 1;
1040 }
1041
1042 static void create_object_entry(const unsigned char *sha1,
1043                                 enum object_type type,
1044                                 uint32_t hash,
1045                                 int exclude,
1046                                 int no_try_delta,
1047                                 uint32_t index_pos,
1048                                 struct packed_git *found_pack,
1049                                 off_t found_offset)
1050 {
1051         struct object_entry *entry;
1052
1053         entry = packlist_alloc(&to_pack, sha1, index_pos);
1054         entry->hash = hash;
1055         if (type)
1056                 entry->type = type;
1057         if (exclude)
1058                 entry->preferred_base = 1;
1059         else
1060                 nr_result++;
1061         if (found_pack) {
1062                 entry->in_pack = found_pack;
1063                 entry->in_pack_offset = found_offset;
1064         }
1065
1066         entry->no_try_delta = no_try_delta;
1067 }
1068
1069 static const char no_closure_warning[] = N_(
1070 "disabling bitmap writing, as some objects are not being packed"
1071 );
1072
1073 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1074                             const char *name, int exclude)
1075 {
1076         struct packed_git *found_pack = NULL;
1077         off_t found_offset = 0;
1078         uint32_t index_pos;
1079
1080         if (have_duplicate_entry(sha1, exclude, &index_pos))
1081                 return 0;
1082
1083         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1084                 /* The pack is missing an object, so it will not have closure */
1085                 if (write_bitmap_index) {
1086                         warning(_(no_closure_warning));
1087                         write_bitmap_index = 0;
1088                 }
1089                 return 0;
1090         }
1091
1092         create_object_entry(sha1, type, pack_name_hash(name),
1093                             exclude, name && no_try_delta(name),
1094                             index_pos, found_pack, found_offset);
1095
1096         display_progress(progress_state, nr_result);
1097         return 1;
1098 }
1099
1100 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1101                                         enum object_type type,
1102                                         int flags, uint32_t name_hash,
1103                                         struct packed_git *pack, off_t offset)
1104 {
1105         uint32_t index_pos;
1106
1107         if (have_duplicate_entry(sha1, 0, &index_pos))
1108                 return 0;
1109
1110         if (!want_object_in_pack(sha1, 0, &pack, &offset))
1111                 return 0;
1112
1113         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1114
1115         display_progress(progress_state, nr_result);
1116         return 1;
1117 }
1118
1119 struct pbase_tree_cache {
1120         unsigned char sha1[20];
1121         int ref;
1122         int temporary;
1123         void *tree_data;
1124         unsigned long tree_size;
1125 };
1126
1127 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1128 static int pbase_tree_cache_ix(const unsigned char *sha1)
1129 {
1130         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1131 }
1132 static int pbase_tree_cache_ix_incr(int ix)
1133 {
1134         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1135 }
1136
1137 static struct pbase_tree {
1138         struct pbase_tree *next;
1139         /* This is a phony "cache" entry; we are not
1140          * going to evict it or find it through _get()
1141          * mechanism -- this is for the toplevel node that
1142          * would almost always change with any commit.
1143          */
1144         struct pbase_tree_cache pcache;
1145 } *pbase_tree;
1146
1147 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1148 {
1149         struct pbase_tree_cache *ent, *nent;
1150         void *data;
1151         unsigned long size;
1152         enum object_type type;
1153         int neigh;
1154         int my_ix = pbase_tree_cache_ix(sha1);
1155         int available_ix = -1;
1156
1157         /* pbase-tree-cache acts as a limited hashtable.
1158          * your object will be found at your index or within a few
1159          * slots after that slot if it is cached.
1160          */
1161         for (neigh = 0; neigh < 8; neigh++) {
1162                 ent = pbase_tree_cache[my_ix];
1163                 if (ent && !hashcmp(ent->sha1, sha1)) {
1164                         ent->ref++;
1165                         return ent;
1166                 }
1167                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1168                          ((0 <= available_ix) &&
1169                           (!ent && pbase_tree_cache[available_ix])))
1170                         available_ix = my_ix;
1171                 if (!ent)
1172                         break;
1173                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1174         }
1175
1176         /* Did not find one.  Either we got a bogus request or
1177          * we need to read and perhaps cache.
1178          */
1179         data = read_sha1_file(sha1, &type, &size);
1180         if (!data)
1181                 return NULL;
1182         if (type != OBJ_TREE) {
1183                 free(data);
1184                 return NULL;
1185         }
1186
1187         /* We need to either cache or return a throwaway copy */
1188
1189         if (available_ix < 0)
1190                 ent = NULL;
1191         else {
1192                 ent = pbase_tree_cache[available_ix];
1193                 my_ix = available_ix;
1194         }
1195
1196         if (!ent) {
1197                 nent = xmalloc(sizeof(*nent));
1198                 nent->temporary = (available_ix < 0);
1199         }
1200         else {
1201                 /* evict and reuse */
1202                 free(ent->tree_data);
1203                 nent = ent;
1204         }
1205         hashcpy(nent->sha1, sha1);
1206         nent->tree_data = data;
1207         nent->tree_size = size;
1208         nent->ref = 1;
1209         if (!nent->temporary)
1210                 pbase_tree_cache[my_ix] = nent;
1211         return nent;
1212 }
1213
1214 static void pbase_tree_put(struct pbase_tree_cache *cache)
1215 {
1216         if (!cache->temporary) {
1217                 cache->ref--;
1218                 return;
1219         }
1220         free(cache->tree_data);
1221         free(cache);
1222 }
1223
1224 static int name_cmp_len(const char *name)
1225 {
1226         int i;
1227         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1228                 ;
1229         return i;
1230 }
1231
1232 static void add_pbase_object(struct tree_desc *tree,
1233                              const char *name,
1234                              int cmplen,
1235                              const char *fullname)
1236 {
1237         struct name_entry entry;
1238         int cmp;
1239
1240         while (tree_entry(tree,&entry)) {
1241                 if (S_ISGITLINK(entry.mode))
1242                         continue;
1243                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1244                       memcmp(name, entry.path, cmplen);
1245                 if (cmp > 0)
1246                         continue;
1247                 if (cmp < 0)
1248                         return;
1249                 if (name[cmplen] != '/') {
1250                         add_object_entry(entry.oid->hash,
1251                                          object_type(entry.mode),
1252                                          fullname, 1);
1253                         return;
1254                 }
1255                 if (S_ISDIR(entry.mode)) {
1256                         struct tree_desc sub;
1257                         struct pbase_tree_cache *tree;
1258                         const char *down = name+cmplen+1;
1259                         int downlen = name_cmp_len(down);
1260
1261                         tree = pbase_tree_get(entry.oid->hash);
1262                         if (!tree)
1263                                 return;
1264                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1265
1266                         add_pbase_object(&sub, down, downlen, fullname);
1267                         pbase_tree_put(tree);
1268                 }
1269         }
1270 }
1271
1272 static unsigned *done_pbase_paths;
1273 static int done_pbase_paths_num;
1274 static int done_pbase_paths_alloc;
1275 static int done_pbase_path_pos(unsigned hash)
1276 {
1277         int lo = 0;
1278         int hi = done_pbase_paths_num;
1279         while (lo < hi) {
1280                 int mi = (hi + lo) / 2;
1281                 if (done_pbase_paths[mi] == hash)
1282                         return mi;
1283                 if (done_pbase_paths[mi] < hash)
1284                         hi = mi;
1285                 else
1286                         lo = mi + 1;
1287         }
1288         return -lo-1;
1289 }
1290
1291 static int check_pbase_path(unsigned hash)
1292 {
1293         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1294         if (0 <= pos)
1295                 return 1;
1296         pos = -pos - 1;
1297         ALLOC_GROW(done_pbase_paths,
1298                    done_pbase_paths_num + 1,
1299                    done_pbase_paths_alloc);
1300         done_pbase_paths_num++;
1301         if (pos < done_pbase_paths_num)
1302                 memmove(done_pbase_paths + pos + 1,
1303                         done_pbase_paths + pos,
1304                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1305         done_pbase_paths[pos] = hash;
1306         return 0;
1307 }
1308
1309 static void add_preferred_base_object(const char *name)
1310 {
1311         struct pbase_tree *it;
1312         int cmplen;
1313         unsigned hash = pack_name_hash(name);
1314
1315         if (!num_preferred_base || check_pbase_path(hash))
1316                 return;
1317
1318         cmplen = name_cmp_len(name);
1319         for (it = pbase_tree; it; it = it->next) {
1320                 if (cmplen == 0) {
1321                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1322                 }
1323                 else {
1324                         struct tree_desc tree;
1325                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1326                         add_pbase_object(&tree, name, cmplen, name);
1327                 }
1328         }
1329 }
1330
1331 static void add_preferred_base(unsigned char *sha1)
1332 {
1333         struct pbase_tree *it;
1334         void *data;
1335         unsigned long size;
1336         unsigned char tree_sha1[20];
1337
1338         if (window <= num_preferred_base++)
1339                 return;
1340
1341         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1342         if (!data)
1343                 return;
1344
1345         for (it = pbase_tree; it; it = it->next) {
1346                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1347                         free(data);
1348                         return;
1349                 }
1350         }
1351
1352         it = xcalloc(1, sizeof(*it));
1353         it->next = pbase_tree;
1354         pbase_tree = it;
1355
1356         hashcpy(it->pcache.sha1, tree_sha1);
1357         it->pcache.tree_data = data;
1358         it->pcache.tree_size = size;
1359 }
1360
1361 static void cleanup_preferred_base(void)
1362 {
1363         struct pbase_tree *it;
1364         unsigned i;
1365
1366         it = pbase_tree;
1367         pbase_tree = NULL;
1368         while (it) {
1369                 struct pbase_tree *this = it;
1370                 it = this->next;
1371                 free(this->pcache.tree_data);
1372                 free(this);
1373         }
1374
1375         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1376                 if (!pbase_tree_cache[i])
1377                         continue;
1378                 free(pbase_tree_cache[i]->tree_data);
1379                 free(pbase_tree_cache[i]);
1380                 pbase_tree_cache[i] = NULL;
1381         }
1382
1383         free(done_pbase_paths);
1384         done_pbase_paths = NULL;
1385         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1386 }
1387
1388 static void check_object(struct object_entry *entry)
1389 {
1390         if (entry->in_pack) {
1391                 struct packed_git *p = entry->in_pack;
1392                 struct pack_window *w_curs = NULL;
1393                 const unsigned char *base_ref = NULL;
1394                 struct object_entry *base_entry;
1395                 unsigned long used, used_0;
1396                 unsigned long avail;
1397                 off_t ofs;
1398                 unsigned char *buf, c;
1399
1400                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1401
1402                 /*
1403                  * We want in_pack_type even if we do not reuse delta
1404                  * since non-delta representations could still be reused.
1405                  */
1406                 used = unpack_object_header_buffer(buf, avail,
1407                                                    &entry->in_pack_type,
1408                                                    &entry->size);
1409                 if (used == 0)
1410                         goto give_up;
1411
1412                 /*
1413                  * Determine if this is a delta and if so whether we can
1414                  * reuse it or not.  Otherwise let's find out as cheaply as
1415                  * possible what the actual type and size for this object is.
1416                  */
1417                 switch (entry->in_pack_type) {
1418                 default:
1419                         /* Not a delta hence we've already got all we need. */
1420                         entry->type = entry->in_pack_type;
1421                         entry->in_pack_header_size = used;
1422                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1423                                 goto give_up;
1424                         unuse_pack(&w_curs);
1425                         return;
1426                 case OBJ_REF_DELTA:
1427                         if (reuse_delta && !entry->preferred_base)
1428                                 base_ref = use_pack(p, &w_curs,
1429                                                 entry->in_pack_offset + used, NULL);
1430                         entry->in_pack_header_size = used + 20;
1431                         break;
1432                 case OBJ_OFS_DELTA:
1433                         buf = use_pack(p, &w_curs,
1434                                        entry->in_pack_offset + used, NULL);
1435                         used_0 = 0;
1436                         c = buf[used_0++];
1437                         ofs = c & 127;
1438                         while (c & 128) {
1439                                 ofs += 1;
1440                                 if (!ofs || MSB(ofs, 7)) {
1441                                         error("delta base offset overflow in pack for %s",
1442                                               sha1_to_hex(entry->idx.sha1));
1443                                         goto give_up;
1444                                 }
1445                                 c = buf[used_0++];
1446                                 ofs = (ofs << 7) + (c & 127);
1447                         }
1448                         ofs = entry->in_pack_offset - ofs;
1449                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1450                                 error("delta base offset out of bound for %s",
1451                                       sha1_to_hex(entry->idx.sha1));
1452                                 goto give_up;
1453                         }
1454                         if (reuse_delta && !entry->preferred_base) {
1455                                 struct revindex_entry *revidx;
1456                                 revidx = find_pack_revindex(p, ofs);
1457                                 if (!revidx)
1458                                         goto give_up;
1459                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1460                         }
1461                         entry->in_pack_header_size = used + used_0;
1462                         break;
1463                 }
1464
1465                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1466                         /*
1467                          * If base_ref was set above that means we wish to
1468                          * reuse delta data, and we even found that base
1469                          * in the list of objects we want to pack. Goodie!
1470                          *
1471                          * Depth value does not matter - find_deltas() will
1472                          * never consider reused delta as the base object to
1473                          * deltify other objects against, in order to avoid
1474                          * circular deltas.
1475                          */
1476                         entry->type = entry->in_pack_type;
1477                         entry->delta = base_entry;
1478                         entry->delta_size = entry->size;
1479                         entry->delta_sibling = base_entry->delta_child;
1480                         base_entry->delta_child = entry;
1481                         unuse_pack(&w_curs);
1482                         return;
1483                 }
1484
1485                 if (entry->type) {
1486                         /*
1487                          * This must be a delta and we already know what the
1488                          * final object type is.  Let's extract the actual
1489                          * object size from the delta header.
1490                          */
1491                         entry->size = get_size_from_delta(p, &w_curs,
1492                                         entry->in_pack_offset + entry->in_pack_header_size);
1493                         if (entry->size == 0)
1494                                 goto give_up;
1495                         unuse_pack(&w_curs);
1496                         return;
1497                 }
1498
1499                 /*
1500                  * No choice but to fall back to the recursive delta walk
1501                  * with sha1_object_info() to find about the object type
1502                  * at this point...
1503                  */
1504                 give_up:
1505                 unuse_pack(&w_curs);
1506         }
1507
1508         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1509         /*
1510          * The error condition is checked in prepare_pack().  This is
1511          * to permit a missing preferred base object to be ignored
1512          * as a preferred base.  Doing so can result in a larger
1513          * pack file, but the transfer will still take place.
1514          */
1515 }
1516
1517 static int pack_offset_sort(const void *_a, const void *_b)
1518 {
1519         const struct object_entry *a = *(struct object_entry **)_a;
1520         const struct object_entry *b = *(struct object_entry **)_b;
1521
1522         /* avoid filesystem trashing with loose objects */
1523         if (!a->in_pack && !b->in_pack)
1524                 return hashcmp(a->idx.sha1, b->idx.sha1);
1525
1526         if (a->in_pack < b->in_pack)
1527                 return -1;
1528         if (a->in_pack > b->in_pack)
1529                 return 1;
1530         return a->in_pack_offset < b->in_pack_offset ? -1 :
1531                         (a->in_pack_offset > b->in_pack_offset);
1532 }
1533
1534 /*
1535  * Drop an on-disk delta we were planning to reuse. Naively, this would
1536  * just involve blanking out the "delta" field, but we have to deal
1537  * with some extra book-keeping:
1538  *
1539  *   1. Removing ourselves from the delta_sibling linked list.
1540  *
1541  *   2. Updating our size/type to the non-delta representation. These were
1542  *      either not recorded initially (size) or overwritten with the delta type
1543  *      (type) when check_object() decided to reuse the delta.
1544  *
1545  *   3. Resetting our delta depth, as we are now a base object.
1546  */
1547 static void drop_reused_delta(struct object_entry *entry)
1548 {
1549         struct object_entry **p = &entry->delta->delta_child;
1550         struct object_info oi = OBJECT_INFO_INIT;
1551
1552         while (*p) {
1553                 if (*p == entry)
1554                         *p = (*p)->delta_sibling;
1555                 else
1556                         p = &(*p)->delta_sibling;
1557         }
1558         entry->delta = NULL;
1559         entry->depth = 0;
1560
1561         oi.sizep = &entry->size;
1562         oi.typep = &entry->type;
1563         if (packed_object_info(entry->in_pack, entry->in_pack_offset, &oi) < 0) {
1564                 /*
1565                  * We failed to get the info from this pack for some reason;
1566                  * fall back to sha1_object_info, which may find another copy.
1567                  * And if that fails, the error will be recorded in entry->type
1568                  * and dealt with in prepare_pack().
1569                  */
1570                 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1571         }
1572 }
1573
1574 /*
1575  * Follow the chain of deltas from this entry onward, throwing away any links
1576  * that cause us to hit a cycle (as determined by the DFS state flags in
1577  * the entries).
1578  *
1579  * We also detect too-long reused chains that would violate our --depth
1580  * limit.
1581  */
1582 static void break_delta_chains(struct object_entry *entry)
1583 {
1584         /* If it's not a delta, it can't be part of a cycle. */
1585         if (!entry->delta) {
1586                 entry->dfs_state = DFS_DONE;
1587                 return;
1588         }
1589
1590         switch (entry->dfs_state) {
1591         case DFS_NONE:
1592                 /*
1593                  * This is the first time we've seen the object. We mark it as
1594                  * part of the active potential cycle and recurse.
1595                  */
1596                 entry->dfs_state = DFS_ACTIVE;
1597                 break_delta_chains(entry->delta);
1598
1599                 /*
1600                  * Once we've recursed, our base (if we still have one) knows
1601                  * its depth, so we can compute ours (and check it against
1602                  * the limit).
1603                  */
1604                 if (entry->delta) {
1605                         entry->depth = entry->delta->depth + 1;
1606                         if (entry->depth > depth)
1607                                 drop_reused_delta(entry);
1608                 }
1609
1610                 entry->dfs_state = DFS_DONE;
1611                 break;
1612
1613         case DFS_DONE:
1614                 /* object already examined, and not part of a cycle */
1615                 break;
1616
1617         case DFS_ACTIVE:
1618                 /*
1619                  * We found a cycle that needs broken. It would be correct to
1620                  * break any link in the chain, but it's convenient to
1621                  * break this one.
1622                  */
1623                 drop_reused_delta(entry);
1624                 entry->dfs_state = DFS_DONE;
1625                 break;
1626         }
1627 }
1628
1629 static void get_object_details(void)
1630 {
1631         uint32_t i;
1632         struct object_entry **sorted_by_offset;
1633
1634         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1635         for (i = 0; i < to_pack.nr_objects; i++)
1636                 sorted_by_offset[i] = to_pack.objects + i;
1637         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1638
1639         for (i = 0; i < to_pack.nr_objects; i++) {
1640                 struct object_entry *entry = sorted_by_offset[i];
1641                 check_object(entry);
1642                 if (big_file_threshold < entry->size)
1643                         entry->no_try_delta = 1;
1644         }
1645
1646         /*
1647          * This must happen in a second pass, since we rely on the delta
1648          * information for the whole list being completed.
1649          */
1650         for (i = 0; i < to_pack.nr_objects; i++)
1651                 break_delta_chains(&to_pack.objects[i]);
1652
1653         free(sorted_by_offset);
1654 }
1655
1656 /*
1657  * We search for deltas in a list sorted by type, by filename hash, and then
1658  * by size, so that we see progressively smaller and smaller files.
1659  * That's because we prefer deltas to be from the bigger file
1660  * to the smaller -- deletes are potentially cheaper, but perhaps
1661  * more importantly, the bigger file is likely the more recent
1662  * one.  The deepest deltas are therefore the oldest objects which are
1663  * less susceptible to be accessed often.
1664  */
1665 static int type_size_sort(const void *_a, const void *_b)
1666 {
1667         const struct object_entry *a = *(struct object_entry **)_a;
1668         const struct object_entry *b = *(struct object_entry **)_b;
1669
1670         if (a->type > b->type)
1671                 return -1;
1672         if (a->type < b->type)
1673                 return 1;
1674         if (a->hash > b->hash)
1675                 return -1;
1676         if (a->hash < b->hash)
1677                 return 1;
1678         if (a->preferred_base > b->preferred_base)
1679                 return -1;
1680         if (a->preferred_base < b->preferred_base)
1681                 return 1;
1682         if (a->size > b->size)
1683                 return -1;
1684         if (a->size < b->size)
1685                 return 1;
1686         return a < b ? -1 : (a > b);  /* newest first */
1687 }
1688
1689 struct unpacked {
1690         struct object_entry *entry;
1691         void *data;
1692         struct delta_index *index;
1693         unsigned depth;
1694 };
1695
1696 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1697                            unsigned long delta_size)
1698 {
1699         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1700                 return 0;
1701
1702         if (delta_size < cache_max_small_delta_size)
1703                 return 1;
1704
1705         /* cache delta, if objects are large enough compared to delta size */
1706         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1707                 return 1;
1708
1709         return 0;
1710 }
1711
1712 #ifndef NO_PTHREADS
1713
1714 static pthread_mutex_t read_mutex;
1715 #define read_lock()             pthread_mutex_lock(&read_mutex)
1716 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1717
1718 static pthread_mutex_t cache_mutex;
1719 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1720 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1721
1722 static pthread_mutex_t progress_mutex;
1723 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1724 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1725
1726 #else
1727
1728 #define read_lock()             (void)0
1729 #define read_unlock()           (void)0
1730 #define cache_lock()            (void)0
1731 #define cache_unlock()          (void)0
1732 #define progress_lock()         (void)0
1733 #define progress_unlock()       (void)0
1734
1735 #endif
1736
1737 static int try_delta(struct unpacked *trg, struct unpacked *src,
1738                      unsigned max_depth, unsigned long *mem_usage)
1739 {
1740         struct object_entry *trg_entry = trg->entry;
1741         struct object_entry *src_entry = src->entry;
1742         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1743         unsigned ref_depth;
1744         enum object_type type;
1745         void *delta_buf;
1746
1747         /* Don't bother doing diffs between different types */
1748         if (trg_entry->type != src_entry->type)
1749                 return -1;
1750
1751         /*
1752          * We do not bother to try a delta that we discarded on an
1753          * earlier try, but only when reusing delta data.  Note that
1754          * src_entry that is marked as the preferred_base should always
1755          * be considered, as even if we produce a suboptimal delta against
1756          * it, we will still save the transfer cost, as we already know
1757          * the other side has it and we won't send src_entry at all.
1758          */
1759         if (reuse_delta && trg_entry->in_pack &&
1760             trg_entry->in_pack == src_entry->in_pack &&
1761             !src_entry->preferred_base &&
1762             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1763             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1764                 return 0;
1765
1766         /* Let's not bust the allowed depth. */
1767         if (src->depth >= max_depth)
1768                 return 0;
1769
1770         /* Now some size filtering heuristics. */
1771         trg_size = trg_entry->size;
1772         if (!trg_entry->delta) {
1773                 max_size = trg_size/2 - 20;
1774                 ref_depth = 1;
1775         } else {
1776                 max_size = trg_entry->delta_size;
1777                 ref_depth = trg->depth;
1778         }
1779         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1780                                                 (max_depth - ref_depth + 1);
1781         if (max_size == 0)
1782                 return 0;
1783         src_size = src_entry->size;
1784         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1785         if (sizediff >= max_size)
1786                 return 0;
1787         if (trg_size < src_size / 32)
1788                 return 0;
1789
1790         /* Load data if not already done */
1791         if (!trg->data) {
1792                 read_lock();
1793                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1794                 read_unlock();
1795                 if (!trg->data)
1796                         die("object %s cannot be read",
1797                             sha1_to_hex(trg_entry->idx.sha1));
1798                 if (sz != trg_size)
1799                         die("object %s inconsistent object length (%lu vs %lu)",
1800                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1801                 *mem_usage += sz;
1802         }
1803         if (!src->data) {
1804                 read_lock();
1805                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1806                 read_unlock();
1807                 if (!src->data) {
1808                         if (src_entry->preferred_base) {
1809                                 static int warned = 0;
1810                                 if (!warned++)
1811                                         warning("object %s cannot be read",
1812                                                 sha1_to_hex(src_entry->idx.sha1));
1813                                 /*
1814                                  * Those objects are not included in the
1815                                  * resulting pack.  Be resilient and ignore
1816                                  * them if they can't be read, in case the
1817                                  * pack could be created nevertheless.
1818                                  */
1819                                 return 0;
1820                         }
1821                         die("object %s cannot be read",
1822                             sha1_to_hex(src_entry->idx.sha1));
1823                 }
1824                 if (sz != src_size)
1825                         die("object %s inconsistent object length (%lu vs %lu)",
1826                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1827                 *mem_usage += sz;
1828         }
1829         if (!src->index) {
1830                 src->index = create_delta_index(src->data, src_size);
1831                 if (!src->index) {
1832                         static int warned = 0;
1833                         if (!warned++)
1834                                 warning("suboptimal pack - out of memory");
1835                         return 0;
1836                 }
1837                 *mem_usage += sizeof_delta_index(src->index);
1838         }
1839
1840         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1841         if (!delta_buf)
1842                 return 0;
1843
1844         if (trg_entry->delta) {
1845                 /* Prefer only shallower same-sized deltas. */
1846                 if (delta_size == trg_entry->delta_size &&
1847                     src->depth + 1 >= trg->depth) {
1848                         free(delta_buf);
1849                         return 0;
1850                 }
1851         }
1852
1853         /*
1854          * Handle memory allocation outside of the cache
1855          * accounting lock.  Compiler will optimize the strangeness
1856          * away when NO_PTHREADS is defined.
1857          */
1858         free(trg_entry->delta_data);
1859         cache_lock();
1860         if (trg_entry->delta_data) {
1861                 delta_cache_size -= trg_entry->delta_size;
1862                 trg_entry->delta_data = NULL;
1863         }
1864         if (delta_cacheable(src_size, trg_size, delta_size)) {
1865                 delta_cache_size += delta_size;
1866                 cache_unlock();
1867                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1868         } else {
1869                 cache_unlock();
1870                 free(delta_buf);
1871         }
1872
1873         trg_entry->delta = src_entry;
1874         trg_entry->delta_size = delta_size;
1875         trg->depth = src->depth + 1;
1876
1877         return 1;
1878 }
1879
1880 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1881 {
1882         struct object_entry *child = me->delta_child;
1883         unsigned int m = n;
1884         while (child) {
1885                 unsigned int c = check_delta_limit(child, n + 1);
1886                 if (m < c)
1887                         m = c;
1888                 child = child->delta_sibling;
1889         }
1890         return m;
1891 }
1892
1893 static unsigned long free_unpacked(struct unpacked *n)
1894 {
1895         unsigned long freed_mem = sizeof_delta_index(n->index);
1896         free_delta_index(n->index);
1897         n->index = NULL;
1898         if (n->data) {
1899                 freed_mem += n->entry->size;
1900                 free(n->data);
1901                 n->data = NULL;
1902         }
1903         n->entry = NULL;
1904         n->depth = 0;
1905         return freed_mem;
1906 }
1907
1908 static void find_deltas(struct object_entry **list, unsigned *list_size,
1909                         int window, int depth, unsigned *processed)
1910 {
1911         uint32_t i, idx = 0, count = 0;
1912         struct unpacked *array;
1913         unsigned long mem_usage = 0;
1914
1915         array = xcalloc(window, sizeof(struct unpacked));
1916
1917         for (;;) {
1918                 struct object_entry *entry;
1919                 struct unpacked *n = array + idx;
1920                 int j, max_depth, best_base = -1;
1921
1922                 progress_lock();
1923                 if (!*list_size) {
1924                         progress_unlock();
1925                         break;
1926                 }
1927                 entry = *list++;
1928                 (*list_size)--;
1929                 if (!entry->preferred_base) {
1930                         (*processed)++;
1931                         display_progress(progress_state, *processed);
1932                 }
1933                 progress_unlock();
1934
1935                 mem_usage -= free_unpacked(n);
1936                 n->entry = entry;
1937
1938                 while (window_memory_limit &&
1939                        mem_usage > window_memory_limit &&
1940                        count > 1) {
1941                         uint32_t tail = (idx + window - count) % window;
1942                         mem_usage -= free_unpacked(array + tail);
1943                         count--;
1944                 }
1945
1946                 /* We do not compute delta to *create* objects we are not
1947                  * going to pack.
1948                  */
1949                 if (entry->preferred_base)
1950                         goto next;
1951
1952                 /*
1953                  * If the current object is at pack edge, take the depth the
1954                  * objects that depend on the current object into account
1955                  * otherwise they would become too deep.
1956                  */
1957                 max_depth = depth;
1958                 if (entry->delta_child) {
1959                         max_depth -= check_delta_limit(entry, 0);
1960                         if (max_depth <= 0)
1961                                 goto next;
1962                 }
1963
1964                 j = window;
1965                 while (--j > 0) {
1966                         int ret;
1967                         uint32_t other_idx = idx + j;
1968                         struct unpacked *m;
1969                         if (other_idx >= window)
1970                                 other_idx -= window;
1971                         m = array + other_idx;
1972                         if (!m->entry)
1973                                 break;
1974                         ret = try_delta(n, m, max_depth, &mem_usage);
1975                         if (ret < 0)
1976                                 break;
1977                         else if (ret > 0)
1978                                 best_base = other_idx;
1979                 }
1980
1981                 /*
1982                  * If we decided to cache the delta data, then it is best
1983                  * to compress it right away.  First because we have to do
1984                  * it anyway, and doing it here while we're threaded will
1985                  * save a lot of time in the non threaded write phase,
1986                  * as well as allow for caching more deltas within
1987                  * the same cache size limit.
1988                  * ...
1989                  * But only if not writing to stdout, since in that case
1990                  * the network is most likely throttling writes anyway,
1991                  * and therefore it is best to go to the write phase ASAP
1992                  * instead, as we can afford spending more time compressing
1993                  * between writes at that moment.
1994                  */
1995                 if (entry->delta_data && !pack_to_stdout) {
1996                         entry->z_delta_size = do_compress(&entry->delta_data,
1997                                                           entry->delta_size);
1998                         cache_lock();
1999                         delta_cache_size -= entry->delta_size;
2000                         delta_cache_size += entry->z_delta_size;
2001                         cache_unlock();
2002                 }
2003
2004                 /* if we made n a delta, and if n is already at max
2005                  * depth, leaving it in the window is pointless.  we
2006                  * should evict it first.
2007                  */
2008                 if (entry->delta && max_depth <= n->depth)
2009                         continue;
2010
2011                 /*
2012                  * Move the best delta base up in the window, after the
2013                  * currently deltified object, to keep it longer.  It will
2014                  * be the first base object to be attempted next.
2015                  */
2016                 if (entry->delta) {
2017                         struct unpacked swap = array[best_base];
2018                         int dist = (window + idx - best_base) % window;
2019                         int dst = best_base;
2020                         while (dist--) {
2021                                 int src = (dst + 1) % window;
2022                                 array[dst] = array[src];
2023                                 dst = src;
2024                         }
2025                         array[dst] = swap;
2026                 }
2027
2028                 next:
2029                 idx++;
2030                 if (count + 1 < window)
2031                         count++;
2032                 if (idx >= window)
2033                         idx = 0;
2034         }
2035
2036         for (i = 0; i < window; ++i) {
2037                 free_delta_index(array[i].index);
2038                 free(array[i].data);
2039         }
2040         free(array);
2041 }
2042
2043 #ifndef NO_PTHREADS
2044
2045 static void try_to_free_from_threads(size_t size)
2046 {
2047         read_lock();
2048         release_pack_memory(size);
2049         read_unlock();
2050 }
2051
2052 static try_to_free_t old_try_to_free_routine;
2053
2054 /*
2055  * The main thread waits on the condition that (at least) one of the workers
2056  * has stopped working (which is indicated in the .working member of
2057  * struct thread_params).
2058  * When a work thread has completed its work, it sets .working to 0 and
2059  * signals the main thread and waits on the condition that .data_ready
2060  * becomes 1.
2061  */
2062
2063 struct thread_params {
2064         pthread_t thread;
2065         struct object_entry **list;
2066         unsigned list_size;
2067         unsigned remaining;
2068         int window;
2069         int depth;
2070         int working;
2071         int data_ready;
2072         pthread_mutex_t mutex;
2073         pthread_cond_t cond;
2074         unsigned *processed;
2075 };
2076
2077 static pthread_cond_t progress_cond;
2078
2079 /*
2080  * Mutex and conditional variable can't be statically-initialized on Windows.
2081  */
2082 static void init_threaded_search(void)
2083 {
2084         init_recursive_mutex(&read_mutex);
2085         pthread_mutex_init(&cache_mutex, NULL);
2086         pthread_mutex_init(&progress_mutex, NULL);
2087         pthread_cond_init(&progress_cond, NULL);
2088         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2089 }
2090
2091 static void cleanup_threaded_search(void)
2092 {
2093         set_try_to_free_routine(old_try_to_free_routine);
2094         pthread_cond_destroy(&progress_cond);
2095         pthread_mutex_destroy(&read_mutex);
2096         pthread_mutex_destroy(&cache_mutex);
2097         pthread_mutex_destroy(&progress_mutex);
2098 }
2099
2100 static void *threaded_find_deltas(void *arg)
2101 {
2102         struct thread_params *me = arg;
2103
2104         while (me->remaining) {
2105                 find_deltas(me->list, &me->remaining,
2106                             me->window, me->depth, me->processed);
2107
2108                 progress_lock();
2109                 me->working = 0;
2110                 pthread_cond_signal(&progress_cond);
2111                 progress_unlock();
2112
2113                 /*
2114                  * We must not set ->data_ready before we wait on the
2115                  * condition because the main thread may have set it to 1
2116                  * before we get here. In order to be sure that new
2117                  * work is available if we see 1 in ->data_ready, it
2118                  * was initialized to 0 before this thread was spawned
2119                  * and we reset it to 0 right away.
2120                  */
2121                 pthread_mutex_lock(&me->mutex);
2122                 while (!me->data_ready)
2123                         pthread_cond_wait(&me->cond, &me->mutex);
2124                 me->data_ready = 0;
2125                 pthread_mutex_unlock(&me->mutex);
2126         }
2127         /* leave ->working 1 so that this doesn't get more work assigned */
2128         return NULL;
2129 }
2130
2131 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2132                            int window, int depth, unsigned *processed)
2133 {
2134         struct thread_params *p;
2135         int i, ret, active_threads = 0;
2136
2137         init_threaded_search();
2138
2139         if (delta_search_threads <= 1) {
2140                 find_deltas(list, &list_size, window, depth, processed);
2141                 cleanup_threaded_search();
2142                 return;
2143         }
2144         if (progress > pack_to_stdout)
2145                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2146                                 delta_search_threads);
2147         p = xcalloc(delta_search_threads, sizeof(*p));
2148
2149         /* Partition the work amongst work threads. */
2150         for (i = 0; i < delta_search_threads; i++) {
2151                 unsigned sub_size = list_size / (delta_search_threads - i);
2152
2153                 /* don't use too small segments or no deltas will be found */
2154                 if (sub_size < 2*window && i+1 < delta_search_threads)
2155                         sub_size = 0;
2156
2157                 p[i].window = window;
2158                 p[i].depth = depth;
2159                 p[i].processed = processed;
2160                 p[i].working = 1;
2161                 p[i].data_ready = 0;
2162
2163                 /* try to split chunks on "path" boundaries */
2164                 while (sub_size && sub_size < list_size &&
2165                        list[sub_size]->hash &&
2166                        list[sub_size]->hash == list[sub_size-1]->hash)
2167                         sub_size++;
2168
2169                 p[i].list = list;
2170                 p[i].list_size = sub_size;
2171                 p[i].remaining = sub_size;
2172
2173                 list += sub_size;
2174                 list_size -= sub_size;
2175         }
2176
2177         /* Start work threads. */
2178         for (i = 0; i < delta_search_threads; i++) {
2179                 if (!p[i].list_size)
2180                         continue;
2181                 pthread_mutex_init(&p[i].mutex, NULL);
2182                 pthread_cond_init(&p[i].cond, NULL);
2183                 ret = pthread_create(&p[i].thread, NULL,
2184                                      threaded_find_deltas, &p[i]);
2185                 if (ret)
2186                         die("unable to create thread: %s", strerror(ret));
2187                 active_threads++;
2188         }
2189
2190         /*
2191          * Now let's wait for work completion.  Each time a thread is done
2192          * with its work, we steal half of the remaining work from the
2193          * thread with the largest number of unprocessed objects and give
2194          * it to that newly idle thread.  This ensure good load balancing
2195          * until the remaining object list segments are simply too short
2196          * to be worth splitting anymore.
2197          */
2198         while (active_threads) {
2199                 struct thread_params *target = NULL;
2200                 struct thread_params *victim = NULL;
2201                 unsigned sub_size = 0;
2202
2203                 progress_lock();
2204                 for (;;) {
2205                         for (i = 0; !target && i < delta_search_threads; i++)
2206                                 if (!p[i].working)
2207                                         target = &p[i];
2208                         if (target)
2209                                 break;
2210                         pthread_cond_wait(&progress_cond, &progress_mutex);
2211                 }
2212
2213                 for (i = 0; i < delta_search_threads; i++)
2214                         if (p[i].remaining > 2*window &&
2215                             (!victim || victim->remaining < p[i].remaining))
2216                                 victim = &p[i];
2217                 if (victim) {
2218                         sub_size = victim->remaining / 2;
2219                         list = victim->list + victim->list_size - sub_size;
2220                         while (sub_size && list[0]->hash &&
2221                                list[0]->hash == list[-1]->hash) {
2222                                 list++;
2223                                 sub_size--;
2224                         }
2225                         if (!sub_size) {
2226                                 /*
2227                                  * It is possible for some "paths" to have
2228                                  * so many objects that no hash boundary
2229                                  * might be found.  Let's just steal the
2230                                  * exact half in that case.
2231                                  */
2232                                 sub_size = victim->remaining / 2;
2233                                 list -= sub_size;
2234                         }
2235                         target->list = list;
2236                         victim->list_size -= sub_size;
2237                         victim->remaining -= sub_size;
2238                 }
2239                 target->list_size = sub_size;
2240                 target->remaining = sub_size;
2241                 target->working = 1;
2242                 progress_unlock();
2243
2244                 pthread_mutex_lock(&target->mutex);
2245                 target->data_ready = 1;
2246                 pthread_cond_signal(&target->cond);
2247                 pthread_mutex_unlock(&target->mutex);
2248
2249                 if (!sub_size) {
2250                         pthread_join(target->thread, NULL);
2251                         pthread_cond_destroy(&target->cond);
2252                         pthread_mutex_destroy(&target->mutex);
2253                         active_threads--;
2254                 }
2255         }
2256         cleanup_threaded_search();
2257         free(p);
2258 }
2259
2260 #else
2261 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2262 #endif
2263
2264 static void add_tag_chain(const struct object_id *oid)
2265 {
2266         struct tag *tag;
2267
2268         /*
2269          * We catch duplicates already in add_object_entry(), but we'd
2270          * prefer to do this extra check to avoid having to parse the
2271          * tag at all if we already know that it's being packed (e.g., if
2272          * it was included via bitmaps, we would not have parsed it
2273          * previously).
2274          */
2275         if (packlist_find(&to_pack, oid->hash, NULL))
2276                 return;
2277
2278         tag = lookup_tag(oid->hash);
2279         while (1) {
2280                 if (!tag || parse_tag(tag) || !tag->tagged)
2281                         die("unable to pack objects reachable from tag %s",
2282                             oid_to_hex(oid));
2283
2284                 add_object_entry(tag->object.oid.hash, OBJ_TAG, NULL, 0);
2285
2286                 if (tag->tagged->type != OBJ_TAG)
2287                         return;
2288
2289                 tag = (struct tag *)tag->tagged;
2290         }
2291 }
2292
2293 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2294 {
2295         struct object_id peeled;
2296
2297         if (starts_with(path, "refs/tags/") && /* is a tag? */
2298             !peel_ref(path, peeled.hash)    && /* peelable? */
2299             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2300                 add_tag_chain(oid);
2301         return 0;
2302 }
2303
2304 static void prepare_pack(int window, int depth)
2305 {
2306         struct object_entry **delta_list;
2307         uint32_t i, nr_deltas;
2308         unsigned n;
2309
2310         get_object_details();
2311
2312         /*
2313          * If we're locally repacking then we need to be doubly careful
2314          * from now on in order to make sure no stealth corruption gets
2315          * propagated to the new pack.  Clients receiving streamed packs
2316          * should validate everything they get anyway so no need to incur
2317          * the additional cost here in that case.
2318          */
2319         if (!pack_to_stdout)
2320                 do_check_packed_object_crc = 1;
2321
2322         if (!to_pack.nr_objects || !window || !depth)
2323                 return;
2324
2325         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2326         nr_deltas = n = 0;
2327
2328         for (i = 0; i < to_pack.nr_objects; i++) {
2329                 struct object_entry *entry = to_pack.objects + i;
2330
2331                 if (entry->delta)
2332                         /* This happens if we decided to reuse existing
2333                          * delta from a pack.  "reuse_delta &&" is implied.
2334                          */
2335                         continue;
2336
2337                 if (entry->size < 50)
2338                         continue;
2339
2340                 if (entry->no_try_delta)
2341                         continue;
2342
2343                 if (!entry->preferred_base) {
2344                         nr_deltas++;
2345                         if (entry->type < 0)
2346                                 die("unable to get type of object %s",
2347                                     sha1_to_hex(entry->idx.sha1));
2348                 } else {
2349                         if (entry->type < 0) {
2350                                 /*
2351                                  * This object is not found, but we
2352                                  * don't have to include it anyway.
2353                                  */
2354                                 continue;
2355                         }
2356                 }
2357
2358                 delta_list[n++] = entry;
2359         }
2360
2361         if (nr_deltas && n > 1) {
2362                 unsigned nr_done = 0;
2363                 if (progress)
2364                         progress_state = start_progress(_("Compressing objects"),
2365                                                         nr_deltas);
2366                 QSORT(delta_list, n, type_size_sort);
2367                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2368                 stop_progress(&progress_state);
2369                 if (nr_done != nr_deltas)
2370                         die("inconsistency with delta count");
2371         }
2372         free(delta_list);
2373 }
2374
2375 static int git_pack_config(const char *k, const char *v, void *cb)
2376 {
2377         if (!strcmp(k, "pack.window")) {
2378                 window = git_config_int(k, v);
2379                 return 0;
2380         }
2381         if (!strcmp(k, "pack.windowmemory")) {
2382                 window_memory_limit = git_config_ulong(k, v);
2383                 return 0;
2384         }
2385         if (!strcmp(k, "pack.depth")) {
2386                 depth = git_config_int(k, v);
2387                 return 0;
2388         }
2389         if (!strcmp(k, "pack.compression")) {
2390                 int level = git_config_int(k, v);
2391                 if (level == -1)
2392                         level = Z_DEFAULT_COMPRESSION;
2393                 else if (level < 0 || level > Z_BEST_COMPRESSION)
2394                         die("bad pack compression level %d", level);
2395                 pack_compression_level = level;
2396                 pack_compression_seen = 1;
2397                 return 0;
2398         }
2399         if (!strcmp(k, "pack.deltacachesize")) {
2400                 max_delta_cache_size = git_config_int(k, v);
2401                 return 0;
2402         }
2403         if (!strcmp(k, "pack.deltacachelimit")) {
2404                 cache_max_small_delta_size = git_config_int(k, v);
2405                 return 0;
2406         }
2407         if (!strcmp(k, "pack.writebitmaphashcache")) {
2408                 if (git_config_bool(k, v))
2409                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2410                 else
2411                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2412         }
2413         if (!strcmp(k, "pack.usebitmaps")) {
2414                 use_bitmap_index_default = git_config_bool(k, v);
2415                 return 0;
2416         }
2417         if (!strcmp(k, "pack.threads")) {
2418                 delta_search_threads = git_config_int(k, v);
2419                 if (delta_search_threads < 0)
2420                         die("invalid number of threads specified (%d)",
2421                             delta_search_threads);
2422 #ifdef NO_PTHREADS
2423                 if (delta_search_threads != 1)
2424                         warning("no threads support, ignoring %s", k);
2425 #endif
2426                 return 0;
2427         }
2428         if (!strcmp(k, "pack.indexversion")) {
2429                 pack_idx_opts.version = git_config_int(k, v);
2430                 if (pack_idx_opts.version > 2)
2431                         die("bad pack.indexversion=%"PRIu32,
2432                             pack_idx_opts.version);
2433                 return 0;
2434         }
2435         return git_default_config(k, v, cb);
2436 }
2437
2438 static void read_object_list_from_stdin(void)
2439 {
2440         char line[40 + 1 + PATH_MAX + 2];
2441         unsigned char sha1[20];
2442
2443         for (;;) {
2444                 if (!fgets(line, sizeof(line), stdin)) {
2445                         if (feof(stdin))
2446                                 break;
2447                         if (!ferror(stdin))
2448                                 die("fgets returned NULL, not EOF, not error!");
2449                         if (errno != EINTR)
2450                                 die_errno("fgets");
2451                         clearerr(stdin);
2452                         continue;
2453                 }
2454                 if (line[0] == '-') {
2455                         if (get_sha1_hex(line+1, sha1))
2456                                 die("expected edge sha1, got garbage:\n %s",
2457                                     line);
2458                         add_preferred_base(sha1);
2459                         continue;
2460                 }
2461                 if (get_sha1_hex(line, sha1))
2462                         die("expected sha1, got garbage:\n %s", line);
2463
2464                 add_preferred_base_object(line+41);
2465                 add_object_entry(sha1, 0, line+41, 0);
2466         }
2467 }
2468
2469 #define OBJECT_ADDED (1u<<20)
2470
2471 static void show_commit(struct commit *commit, void *data)
2472 {
2473         add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2474         commit->object.flags |= OBJECT_ADDED;
2475
2476         if (write_bitmap_index)
2477                 index_commit_for_bitmap(commit);
2478 }
2479
2480 static void show_object(struct object *obj, const char *name, void *data)
2481 {
2482         add_preferred_base_object(name);
2483         add_object_entry(obj->oid.hash, obj->type, name, 0);
2484         obj->flags |= OBJECT_ADDED;
2485 }
2486
2487 static void show_edge(struct commit *commit)
2488 {
2489         add_preferred_base(commit->object.oid.hash);
2490 }
2491
2492 struct in_pack_object {
2493         off_t offset;
2494         struct object *object;
2495 };
2496
2497 struct in_pack {
2498         int alloc;
2499         int nr;
2500         struct in_pack_object *array;
2501 };
2502
2503 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2504 {
2505         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2506         in_pack->array[in_pack->nr].object = object;
2507         in_pack->nr++;
2508 }
2509
2510 /*
2511  * Compare the objects in the offset order, in order to emulate the
2512  * "git rev-list --objects" output that produced the pack originally.
2513  */
2514 static int ofscmp(const void *a_, const void *b_)
2515 {
2516         struct in_pack_object *a = (struct in_pack_object *)a_;
2517         struct in_pack_object *b = (struct in_pack_object *)b_;
2518
2519         if (a->offset < b->offset)
2520                 return -1;
2521         else if (a->offset > b->offset)
2522                 return 1;
2523         else
2524                 return oidcmp(&a->object->oid, &b->object->oid);
2525 }
2526
2527 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2528 {
2529         struct packed_git *p;
2530         struct in_pack in_pack;
2531         uint32_t i;
2532
2533         memset(&in_pack, 0, sizeof(in_pack));
2534
2535         for (p = packed_git; p; p = p->next) {
2536                 const unsigned char *sha1;
2537                 struct object *o;
2538
2539                 if (!p->pack_local || p->pack_keep)
2540                         continue;
2541                 if (open_pack_index(p))
2542                         die("cannot open pack index");
2543
2544                 ALLOC_GROW(in_pack.array,
2545                            in_pack.nr + p->num_objects,
2546                            in_pack.alloc);
2547
2548                 for (i = 0; i < p->num_objects; i++) {
2549                         sha1 = nth_packed_object_sha1(p, i);
2550                         o = lookup_unknown_object(sha1);
2551                         if (!(o->flags & OBJECT_ADDED))
2552                                 mark_in_pack_object(o, p, &in_pack);
2553                         o->flags |= OBJECT_ADDED;
2554                 }
2555         }
2556
2557         if (in_pack.nr) {
2558                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2559                 for (i = 0; i < in_pack.nr; i++) {
2560                         struct object *o = in_pack.array[i].object;
2561                         add_object_entry(o->oid.hash, o->type, "", 0);
2562                 }
2563         }
2564         free(in_pack.array);
2565 }
2566
2567 static int add_loose_object(const unsigned char *sha1, const char *path,
2568                             void *data)
2569 {
2570         enum object_type type = sha1_object_info(sha1, NULL);
2571
2572         if (type < 0) {
2573                 warning("loose object at %s could not be examined", path);
2574                 return 0;
2575         }
2576
2577         add_object_entry(sha1, type, "", 0);
2578         return 0;
2579 }
2580
2581 /*
2582  * We actually don't even have to worry about reachability here.
2583  * add_object_entry will weed out duplicates, so we just add every
2584  * loose object we find.
2585  */
2586 static void add_unreachable_loose_objects(void)
2587 {
2588         for_each_loose_file_in_objdir(get_object_directory(),
2589                                       add_loose_object,
2590                                       NULL, NULL, NULL);
2591 }
2592
2593 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2594 {
2595         static struct packed_git *last_found = (void *)1;
2596         struct packed_git *p;
2597
2598         p = (last_found != (void *)1) ? last_found : packed_git;
2599
2600         while (p) {
2601                 if ((!p->pack_local || p->pack_keep) &&
2602                         find_pack_entry_one(sha1, p)) {
2603                         last_found = p;
2604                         return 1;
2605                 }
2606                 if (p == last_found)
2607                         p = packed_git;
2608                 else
2609                         p = p->next;
2610                 if (p == last_found)
2611                         p = p->next;
2612         }
2613         return 0;
2614 }
2615
2616 /*
2617  * Store a list of sha1s that are should not be discarded
2618  * because they are either written too recently, or are
2619  * reachable from another object that was.
2620  *
2621  * This is filled by get_object_list.
2622  */
2623 static struct sha1_array recent_objects;
2624
2625 static int loosened_object_can_be_discarded(const unsigned char *sha1,
2626                                             unsigned long mtime)
2627 {
2628         if (!unpack_unreachable_expiration)
2629                 return 0;
2630         if (mtime > unpack_unreachable_expiration)
2631                 return 0;
2632         if (sha1_array_lookup(&recent_objects, sha1) >= 0)
2633                 return 0;
2634         return 1;
2635 }
2636
2637 static void loosen_unused_packed_objects(struct rev_info *revs)
2638 {
2639         struct packed_git *p;
2640         uint32_t i;
2641         const unsigned char *sha1;
2642
2643         for (p = packed_git; p; p = p->next) {
2644                 if (!p->pack_local || p->pack_keep)
2645                         continue;
2646
2647                 if (open_pack_index(p))
2648                         die("cannot open pack index");
2649
2650                 for (i = 0; i < p->num_objects; i++) {
2651                         sha1 = nth_packed_object_sha1(p, i);
2652                         if (!packlist_find(&to_pack, sha1, NULL) &&
2653                             !has_sha1_pack_kept_or_nonlocal(sha1) &&
2654                             !loosened_object_can_be_discarded(sha1, p->mtime))
2655                                 if (force_object_loose(sha1, p->mtime))
2656                                         die("unable to force loose object");
2657                 }
2658         }
2659 }
2660
2661 /*
2662  * This tracks any options which pack-reuse code expects to be on, or which a
2663  * reader of the pack might not understand, and which would therefore prevent
2664  * blind reuse of what we have on disk.
2665  */
2666 static int pack_options_allow_reuse(void)
2667 {
2668         return pack_to_stdout && allow_ofs_delta;
2669 }
2670
2671 static int get_object_list_from_bitmap(struct rev_info *revs)
2672 {
2673         if (prepare_bitmap_walk(revs) < 0)
2674                 return -1;
2675
2676         if (pack_options_allow_reuse() &&
2677             !reuse_partial_packfile_from_bitmap(
2678                         &reuse_packfile,
2679                         &reuse_packfile_objects,
2680                         &reuse_packfile_offset)) {
2681                 assert(reuse_packfile_objects);
2682                 nr_result += reuse_packfile_objects;
2683                 display_progress(progress_state, nr_result);
2684         }
2685
2686         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2687         return 0;
2688 }
2689
2690 static void record_recent_object(struct object *obj,
2691                                  const char *name,
2692                                  void *data)
2693 {
2694         sha1_array_append(&recent_objects, obj->oid.hash);
2695 }
2696
2697 static void record_recent_commit(struct commit *commit, void *data)
2698 {
2699         sha1_array_append(&recent_objects, commit->object.oid.hash);
2700 }
2701
2702 static void get_object_list(int ac, const char **av)
2703 {
2704         struct rev_info revs;
2705         char line[1000];
2706         int flags = 0;
2707
2708         init_revisions(&revs, NULL);
2709         save_commit_buffer = 0;
2710         setup_revisions(ac, av, &revs, NULL);
2711
2712         /* make sure shallows are read */
2713         is_repository_shallow();
2714
2715         while (fgets(line, sizeof(line), stdin) != NULL) {
2716                 int len = strlen(line);
2717                 if (len && line[len - 1] == '\n')
2718                         line[--len] = 0;
2719                 if (!len)
2720                         break;
2721                 if (*line == '-') {
2722                         if (!strcmp(line, "--not")) {
2723                                 flags ^= UNINTERESTING;
2724                                 write_bitmap_index = 0;
2725                                 continue;
2726                         }
2727                         if (starts_with(line, "--shallow ")) {
2728                                 unsigned char sha1[20];
2729                                 if (get_sha1_hex(line + 10, sha1))
2730                                         die("not an SHA-1 '%s'", line + 10);
2731                                 register_shallow(sha1);
2732                                 use_bitmap_index = 0;
2733                                 continue;
2734                         }
2735                         die("not a rev '%s'", line);
2736                 }
2737                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2738                         die("bad revision '%s'", line);
2739         }
2740
2741         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2742                 return;
2743
2744         if (prepare_revision_walk(&revs))
2745                 die("revision walk setup failed");
2746         mark_edges_uninteresting(&revs, show_edge);
2747         traverse_commit_list(&revs, show_commit, show_object, NULL);
2748
2749         if (unpack_unreachable_expiration) {
2750                 revs.ignore_missing_links = 1;
2751                 if (add_unseen_recent_objects_to_traversal(&revs,
2752                                 unpack_unreachable_expiration))
2753                         die("unable to add recent objects");
2754                 if (prepare_revision_walk(&revs))
2755                         die("revision walk setup failed");
2756                 traverse_commit_list(&revs, record_recent_commit,
2757                                      record_recent_object, NULL);
2758         }
2759
2760         if (keep_unreachable)
2761                 add_objects_in_unpacked_packs(&revs);
2762         if (pack_loose_unreachable)
2763                 add_unreachable_loose_objects();
2764         if (unpack_unreachable)
2765                 loosen_unused_packed_objects(&revs);
2766
2767         sha1_array_clear(&recent_objects);
2768 }
2769
2770 static int option_parse_index_version(const struct option *opt,
2771                                       const char *arg, int unset)
2772 {
2773         char *c;
2774         const char *val = arg;
2775         pack_idx_opts.version = strtoul(val, &c, 10);
2776         if (pack_idx_opts.version > 2)
2777                 die(_("unsupported index version %s"), val);
2778         if (*c == ',' && c[1])
2779                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2780         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2781                 die(_("bad index version '%s'"), val);
2782         return 0;
2783 }
2784
2785 static int option_parse_unpack_unreachable(const struct option *opt,
2786                                            const char *arg, int unset)
2787 {
2788         if (unset) {
2789                 unpack_unreachable = 0;
2790                 unpack_unreachable_expiration = 0;
2791         }
2792         else {
2793                 unpack_unreachable = 1;
2794                 if (arg)
2795                         unpack_unreachable_expiration = approxidate(arg);
2796         }
2797         return 0;
2798 }
2799
2800 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2801 {
2802         int use_internal_rev_list = 0;
2803         int thin = 0;
2804         int shallow = 0;
2805         int all_progress_implied = 0;
2806         struct argv_array rp = ARGV_ARRAY_INIT;
2807         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2808         int rev_list_index = 0;
2809         struct option pack_objects_options[] = {
2810                 OPT_SET_INT('q', "quiet", &progress,
2811                             N_("do not show progress meter"), 0),
2812                 OPT_SET_INT(0, "progress", &progress,
2813                             N_("show progress meter"), 1),
2814                 OPT_SET_INT(0, "all-progress", &progress,
2815                             N_("show progress meter during object writing phase"), 2),
2816                 OPT_BOOL(0, "all-progress-implied",
2817                          &all_progress_implied,
2818                          N_("similar to --all-progress when progress meter is shown")),
2819                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2820                   N_("write the pack index file in the specified idx format version"),
2821                   0, option_parse_index_version },
2822                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2823                               N_("maximum size of each output pack file")),
2824                 OPT_BOOL(0, "local", &local,
2825                          N_("ignore borrowed objects from alternate object store")),
2826                 OPT_BOOL(0, "incremental", &incremental,
2827                          N_("ignore packed objects")),
2828                 OPT_INTEGER(0, "window", &window,
2829                             N_("limit pack window by objects")),
2830                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2831                               N_("limit pack window by memory in addition to object limit")),
2832                 OPT_INTEGER(0, "depth", &depth,
2833                             N_("maximum length of delta chain allowed in the resulting pack")),
2834                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2835                          N_("reuse existing deltas")),
2836                 OPT_BOOL(0, "reuse-object", &reuse_object,
2837                          N_("reuse existing objects")),
2838                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2839                          N_("use OFS_DELTA objects")),
2840                 OPT_INTEGER(0, "threads", &delta_search_threads,
2841                             N_("use threads when searching for best delta matches")),
2842                 OPT_BOOL(0, "non-empty", &non_empty,
2843                          N_("do not create an empty pack output")),
2844                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2845                          N_("read revision arguments from standard input")),
2846                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2847                   N_("limit the objects to those that are not yet packed"),
2848                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2849                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2850                   N_("include objects reachable from any reference"),
2851                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2852                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2853                   N_("include objects referred by reflog entries"),
2854                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2855                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2856                   N_("include objects referred to by the index"),
2857                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2858                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2859                          N_("output pack to stdout")),
2860                 OPT_BOOL(0, "include-tag", &include_tag,
2861                          N_("include tag objects that refer to objects to be packed")),
2862                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2863                          N_("keep unreachable objects")),
2864                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
2865                          N_("pack loose unreachable objects")),
2866                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2867                   N_("unpack unreachable objects newer than <time>"),
2868                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2869                 OPT_BOOL(0, "thin", &thin,
2870                          N_("create thin packs")),
2871                 OPT_BOOL(0, "shallow", &shallow,
2872                          N_("create packs suitable for shallow fetches")),
2873                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2874                          N_("ignore packs that have companion .keep file")),
2875                 OPT_INTEGER(0, "compression", &pack_compression_level,
2876                             N_("pack compression level")),
2877                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2878                             N_("do not hide commits by grafts"), 0),
2879                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2880                          N_("use a bitmap index if available to speed up counting objects")),
2881                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2882                          N_("write a bitmap index together with the pack index")),
2883                 OPT_END(),
2884         };
2885
2886         check_replace_refs = 0;
2887
2888         reset_pack_idx_option(&pack_idx_opts);
2889         git_config(git_pack_config, NULL);
2890         if (!pack_compression_seen && core_compression_seen)
2891                 pack_compression_level = core_compression_level;
2892
2893         progress = isatty(2);
2894         argc = parse_options(argc, argv, prefix, pack_objects_options,
2895                              pack_usage, 0);
2896
2897         if (argc) {
2898                 base_name = argv[0];
2899                 argc--;
2900         }
2901         if (pack_to_stdout != !base_name || argc)
2902                 usage_with_options(pack_usage, pack_objects_options);
2903
2904         argv_array_push(&rp, "pack-objects");
2905         if (thin) {
2906                 use_internal_rev_list = 1;
2907                 argv_array_push(&rp, shallow
2908                                 ? "--objects-edge-aggressive"
2909                                 : "--objects-edge");
2910         } else
2911                 argv_array_push(&rp, "--objects");
2912
2913         if (rev_list_all) {
2914                 use_internal_rev_list = 1;
2915                 argv_array_push(&rp, "--all");
2916         }
2917         if (rev_list_reflog) {
2918                 use_internal_rev_list = 1;
2919                 argv_array_push(&rp, "--reflog");
2920         }
2921         if (rev_list_index) {
2922                 use_internal_rev_list = 1;
2923                 argv_array_push(&rp, "--indexed-objects");
2924         }
2925         if (rev_list_unpacked) {
2926                 use_internal_rev_list = 1;
2927                 argv_array_push(&rp, "--unpacked");
2928         }
2929
2930         if (!reuse_object)
2931                 reuse_delta = 0;
2932         if (pack_compression_level == -1)
2933                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2934         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2935                 die("bad pack compression level %d", pack_compression_level);
2936
2937         if (!delta_search_threads)      /* --threads=0 means autodetect */
2938                 delta_search_threads = online_cpus();
2939
2940 #ifdef NO_PTHREADS
2941         if (delta_search_threads != 1)
2942                 warning("no threads support, ignoring --threads");
2943 #endif
2944         if (!pack_to_stdout && !pack_size_limit)
2945                 pack_size_limit = pack_size_limit_cfg;
2946         if (pack_to_stdout && pack_size_limit)
2947                 die("--max-pack-size cannot be used to build a pack for transfer.");
2948         if (pack_size_limit && pack_size_limit < 1024*1024) {
2949                 warning("minimum pack size limit is 1 MiB");
2950                 pack_size_limit = 1024*1024;
2951         }
2952
2953         if (!pack_to_stdout && thin)
2954                 die("--thin cannot be used to build an indexable pack.");
2955
2956         if (keep_unreachable && unpack_unreachable)
2957                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2958         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
2959                 unpack_unreachable_expiration = 0;
2960
2961         /*
2962          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
2963          *
2964          * - to produce good pack (with bitmap index not-yet-packed objects are
2965          *   packed in suboptimal order).
2966          *
2967          * - to use more robust pack-generation codepath (avoiding possible
2968          *   bugs in bitmap code and possible bitmap index corruption).
2969          */
2970         if (!pack_to_stdout)
2971                 use_bitmap_index_default = 0;
2972
2973         if (use_bitmap_index < 0)
2974                 use_bitmap_index = use_bitmap_index_default;
2975
2976         /* "hard" reasons not to use bitmaps; these just won't work at all */
2977         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
2978                 use_bitmap_index = 0;
2979
2980         if (pack_to_stdout || !rev_list_all)
2981                 write_bitmap_index = 0;
2982
2983         if (progress && all_progress_implied)
2984                 progress = 2;
2985
2986         prepare_packed_git();
2987         if (ignore_packed_keep) {
2988                 struct packed_git *p;
2989                 for (p = packed_git; p; p = p->next)
2990                         if (p->pack_local && p->pack_keep)
2991                                 break;
2992                 if (!p) /* no keep-able packs found */
2993                         ignore_packed_keep = 0;
2994         }
2995         if (local) {
2996                 /*
2997                  * unlike ignore_packed_keep above, we do not want to
2998                  * unset "local" based on looking at packs, as it
2999                  * also covers non-local objects
3000                  */
3001                 struct packed_git *p;
3002                 for (p = packed_git; p; p = p->next) {
3003                         if (!p->pack_local) {
3004                                 have_non_local_packs = 1;
3005                                 break;
3006                         }
3007                 }
3008         }
3009
3010         if (progress)
3011                 progress_state = start_progress(_("Counting objects"), 0);
3012         if (!use_internal_rev_list)
3013                 read_object_list_from_stdin();
3014         else {
3015                 get_object_list(rp.argc, rp.argv);
3016                 argv_array_clear(&rp);
3017         }
3018         cleanup_preferred_base();
3019         if (include_tag && nr_result)
3020                 for_each_ref(add_ref_tag, NULL);
3021         stop_progress(&progress_state);
3022
3023         if (non_empty && !nr_result)
3024                 return 0;
3025         if (nr_result)
3026                 prepare_pack(window, depth);
3027         write_pack_file();
3028         if (progress)
3029                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
3030                         " reused %"PRIu32" (delta %"PRIu32")\n",
3031                         written, written_delta, reused, reused_delta);
3032         return 0;
3033 }