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         /*
1585          * The actual depth of each object we will write is stored as an int,
1586          * as it cannot exceed our int "depth" limit. But before we break
1587          * changes based no that limit, we may potentially go as deep as the
1588          * number of objects, which is elsewhere bounded to a uint32_t.
1589          */
1590         uint32_t total_depth;
1591         struct object_entry *cur, *next;
1592
1593         for (cur = entry, total_depth = 0;
1594              cur;
1595              cur = cur->delta, total_depth++) {
1596                 if (cur->dfs_state == DFS_DONE) {
1597                         /*
1598                          * We've already seen this object and know it isn't
1599                          * part of a cycle. We do need to append its depth
1600                          * to our count.
1601                          */
1602                         total_depth += cur->depth;
1603                         break;
1604                 }
1605
1606                 /*
1607                  * We break cycles before looping, so an ACTIVE state (or any
1608                  * other cruft which made its way into the state variable)
1609                  * is a bug.
1610                  */
1611                 if (cur->dfs_state != DFS_NONE)
1612                         die("BUG: confusing delta dfs state in first pass: %d",
1613                             cur->dfs_state);
1614
1615                 /*
1616                  * Now we know this is the first time we've seen the object. If
1617                  * it's not a delta, we're done traversing, but we'll mark it
1618                  * done to save time on future traversals.
1619                  */
1620                 if (!cur->delta) {
1621                         cur->dfs_state = DFS_DONE;
1622                         break;
1623                 }
1624
1625                 /*
1626                  * Mark ourselves as active and see if the next step causes
1627                  * us to cycle to another active object. It's important to do
1628                  * this _before_ we loop, because it impacts where we make the
1629                  * cut, and thus how our total_depth counter works.
1630                  * E.g., We may see a partial loop like:
1631                  *
1632                  *   A -> B -> C -> D -> B
1633                  *
1634                  * Cutting B->C breaks the cycle. But now the depth of A is
1635                  * only 1, and our total_depth counter is at 3. The size of the
1636                  * error is always one less than the size of the cycle we
1637                  * broke. Commits C and D were "lost" from A's chain.
1638                  *
1639                  * If we instead cut D->B, then the depth of A is correct at 3.
1640                  * We keep all commits in the chain that we examined.
1641                  */
1642                 cur->dfs_state = DFS_ACTIVE;
1643                 if (cur->delta->dfs_state == DFS_ACTIVE) {
1644                         drop_reused_delta(cur);
1645                         cur->dfs_state = DFS_DONE;
1646                         break;
1647                 }
1648         }
1649
1650         /*
1651          * And now that we've gone all the way to the bottom of the chain, we
1652          * need to clear the active flags and set the depth fields as
1653          * appropriate. Unlike the loop above, which can quit when it drops a
1654          * delta, we need to keep going to look for more depth cuts. So we need
1655          * an extra "next" pointer to keep going after we reset cur->delta.
1656          */
1657         for (cur = entry; cur; cur = next) {
1658                 next = cur->delta;
1659
1660                 /*
1661                  * We should have a chain of zero or more ACTIVE states down to
1662                  * a final DONE. We can quit after the DONE, because either it
1663                  * has no bases, or we've already handled them in a previous
1664                  * call.
1665                  */
1666                 if (cur->dfs_state == DFS_DONE)
1667                         break;
1668                 else if (cur->dfs_state != DFS_ACTIVE)
1669                         die("BUG: confusing delta dfs state in second pass: %d",
1670                             cur->dfs_state);
1671
1672                 /*
1673                  * If the total_depth is more than depth, then we need to snip
1674                  * the chain into two or more smaller chains that don't exceed
1675                  * the maximum depth. Most of the resulting chains will contain
1676                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1677                  * the last chain (i.e., the one containing entry) will contain
1678                  * whatever entries are left over, namely
1679                  * (total_depth % (depth + 1)) of them.
1680                  *
1681                  * Since we are iterating towards decreasing depth, we need to
1682                  * decrement total_depth as we go, and we need to write to the
1683                  * entry what its final depth will be after all of the
1684                  * snipping. Since we're snipping into chains of length (depth
1685                  * + 1) entries, the final depth of an entry will be its
1686                  * original depth modulo (depth + 1). Any time we encounter an
1687                  * entry whose final depth is supposed to be zero, we snip it
1688                  * from its delta base, thereby making it so.
1689                  */
1690                 cur->depth = (total_depth--) % (depth + 1);
1691                 if (!cur->depth)
1692                         drop_reused_delta(cur);
1693
1694                 cur->dfs_state = DFS_DONE;
1695         }
1696 }
1697
1698 static void get_object_details(void)
1699 {
1700         uint32_t i;
1701         struct object_entry **sorted_by_offset;
1702
1703         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1704         for (i = 0; i < to_pack.nr_objects; i++)
1705                 sorted_by_offset[i] = to_pack.objects + i;
1706         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1707
1708         for (i = 0; i < to_pack.nr_objects; i++) {
1709                 struct object_entry *entry = sorted_by_offset[i];
1710                 check_object(entry);
1711                 if (big_file_threshold < entry->size)
1712                         entry->no_try_delta = 1;
1713         }
1714
1715         /*
1716          * This must happen in a second pass, since we rely on the delta
1717          * information for the whole list being completed.
1718          */
1719         for (i = 0; i < to_pack.nr_objects; i++)
1720                 break_delta_chains(&to_pack.objects[i]);
1721
1722         free(sorted_by_offset);
1723 }
1724
1725 /*
1726  * We search for deltas in a list sorted by type, by filename hash, and then
1727  * by size, so that we see progressively smaller and smaller files.
1728  * That's because we prefer deltas to be from the bigger file
1729  * to the smaller -- deletes are potentially cheaper, but perhaps
1730  * more importantly, the bigger file is likely the more recent
1731  * one.  The deepest deltas are therefore the oldest objects which are
1732  * less susceptible to be accessed often.
1733  */
1734 static int type_size_sort(const void *_a, const void *_b)
1735 {
1736         const struct object_entry *a = *(struct object_entry **)_a;
1737         const struct object_entry *b = *(struct object_entry **)_b;
1738
1739         if (a->type > b->type)
1740                 return -1;
1741         if (a->type < b->type)
1742                 return 1;
1743         if (a->hash > b->hash)
1744                 return -1;
1745         if (a->hash < b->hash)
1746                 return 1;
1747         if (a->preferred_base > b->preferred_base)
1748                 return -1;
1749         if (a->preferred_base < b->preferred_base)
1750                 return 1;
1751         if (a->size > b->size)
1752                 return -1;
1753         if (a->size < b->size)
1754                 return 1;
1755         return a < b ? -1 : (a > b);  /* newest first */
1756 }
1757
1758 struct unpacked {
1759         struct object_entry *entry;
1760         void *data;
1761         struct delta_index *index;
1762         unsigned depth;
1763 };
1764
1765 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1766                            unsigned long delta_size)
1767 {
1768         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1769                 return 0;
1770
1771         if (delta_size < cache_max_small_delta_size)
1772                 return 1;
1773
1774         /* cache delta, if objects are large enough compared to delta size */
1775         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1776                 return 1;
1777
1778         return 0;
1779 }
1780
1781 #ifndef NO_PTHREADS
1782
1783 static pthread_mutex_t read_mutex;
1784 #define read_lock()             pthread_mutex_lock(&read_mutex)
1785 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1786
1787 static pthread_mutex_t cache_mutex;
1788 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1789 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1790
1791 static pthread_mutex_t progress_mutex;
1792 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1793 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1794
1795 #else
1796
1797 #define read_lock()             (void)0
1798 #define read_unlock()           (void)0
1799 #define cache_lock()            (void)0
1800 #define cache_unlock()          (void)0
1801 #define progress_lock()         (void)0
1802 #define progress_unlock()       (void)0
1803
1804 #endif
1805
1806 static int try_delta(struct unpacked *trg, struct unpacked *src,
1807                      unsigned max_depth, unsigned long *mem_usage)
1808 {
1809         struct object_entry *trg_entry = trg->entry;
1810         struct object_entry *src_entry = src->entry;
1811         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1812         unsigned ref_depth;
1813         enum object_type type;
1814         void *delta_buf;
1815
1816         /* Don't bother doing diffs between different types */
1817         if (trg_entry->type != src_entry->type)
1818                 return -1;
1819
1820         /*
1821          * We do not bother to try a delta that we discarded on an
1822          * earlier try, but only when reusing delta data.  Note that
1823          * src_entry that is marked as the preferred_base should always
1824          * be considered, as even if we produce a suboptimal delta against
1825          * it, we will still save the transfer cost, as we already know
1826          * the other side has it and we won't send src_entry at all.
1827          */
1828         if (reuse_delta && trg_entry->in_pack &&
1829             trg_entry->in_pack == src_entry->in_pack &&
1830             !src_entry->preferred_base &&
1831             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1832             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1833                 return 0;
1834
1835         /* Let's not bust the allowed depth. */
1836         if (src->depth >= max_depth)
1837                 return 0;
1838
1839         /* Now some size filtering heuristics. */
1840         trg_size = trg_entry->size;
1841         if (!trg_entry->delta) {
1842                 max_size = trg_size/2 - 20;
1843                 ref_depth = 1;
1844         } else {
1845                 max_size = trg_entry->delta_size;
1846                 ref_depth = trg->depth;
1847         }
1848         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1849                                                 (max_depth - ref_depth + 1);
1850         if (max_size == 0)
1851                 return 0;
1852         src_size = src_entry->size;
1853         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1854         if (sizediff >= max_size)
1855                 return 0;
1856         if (trg_size < src_size / 32)
1857                 return 0;
1858
1859         /* Load data if not already done */
1860         if (!trg->data) {
1861                 read_lock();
1862                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1863                 read_unlock();
1864                 if (!trg->data)
1865                         die("object %s cannot be read",
1866                             sha1_to_hex(trg_entry->idx.sha1));
1867                 if (sz != trg_size)
1868                         die("object %s inconsistent object length (%lu vs %lu)",
1869                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1870                 *mem_usage += sz;
1871         }
1872         if (!src->data) {
1873                 read_lock();
1874                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1875                 read_unlock();
1876                 if (!src->data) {
1877                         if (src_entry->preferred_base) {
1878                                 static int warned = 0;
1879                                 if (!warned++)
1880                                         warning("object %s cannot be read",
1881                                                 sha1_to_hex(src_entry->idx.sha1));
1882                                 /*
1883                                  * Those objects are not included in the
1884                                  * resulting pack.  Be resilient and ignore
1885                                  * them if they can't be read, in case the
1886                                  * pack could be created nevertheless.
1887                                  */
1888                                 return 0;
1889                         }
1890                         die("object %s cannot be read",
1891                             sha1_to_hex(src_entry->idx.sha1));
1892                 }
1893                 if (sz != src_size)
1894                         die("object %s inconsistent object length (%lu vs %lu)",
1895                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1896                 *mem_usage += sz;
1897         }
1898         if (!src->index) {
1899                 src->index = create_delta_index(src->data, src_size);
1900                 if (!src->index) {
1901                         static int warned = 0;
1902                         if (!warned++)
1903                                 warning("suboptimal pack - out of memory");
1904                         return 0;
1905                 }
1906                 *mem_usage += sizeof_delta_index(src->index);
1907         }
1908
1909         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1910         if (!delta_buf)
1911                 return 0;
1912
1913         if (trg_entry->delta) {
1914                 /* Prefer only shallower same-sized deltas. */
1915                 if (delta_size == trg_entry->delta_size &&
1916                     src->depth + 1 >= trg->depth) {
1917                         free(delta_buf);
1918                         return 0;
1919                 }
1920         }
1921
1922         /*
1923          * Handle memory allocation outside of the cache
1924          * accounting lock.  Compiler will optimize the strangeness
1925          * away when NO_PTHREADS is defined.
1926          */
1927         free(trg_entry->delta_data);
1928         cache_lock();
1929         if (trg_entry->delta_data) {
1930                 delta_cache_size -= trg_entry->delta_size;
1931                 trg_entry->delta_data = NULL;
1932         }
1933         if (delta_cacheable(src_size, trg_size, delta_size)) {
1934                 delta_cache_size += delta_size;
1935                 cache_unlock();
1936                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1937         } else {
1938                 cache_unlock();
1939                 free(delta_buf);
1940         }
1941
1942         trg_entry->delta = src_entry;
1943         trg_entry->delta_size = delta_size;
1944         trg->depth = src->depth + 1;
1945
1946         return 1;
1947 }
1948
1949 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1950 {
1951         struct object_entry *child = me->delta_child;
1952         unsigned int m = n;
1953         while (child) {
1954                 unsigned int c = check_delta_limit(child, n + 1);
1955                 if (m < c)
1956                         m = c;
1957                 child = child->delta_sibling;
1958         }
1959         return m;
1960 }
1961
1962 static unsigned long free_unpacked(struct unpacked *n)
1963 {
1964         unsigned long freed_mem = sizeof_delta_index(n->index);
1965         free_delta_index(n->index);
1966         n->index = NULL;
1967         if (n->data) {
1968                 freed_mem += n->entry->size;
1969                 free(n->data);
1970                 n->data = NULL;
1971         }
1972         n->entry = NULL;
1973         n->depth = 0;
1974         return freed_mem;
1975 }
1976
1977 static void find_deltas(struct object_entry **list, unsigned *list_size,
1978                         int window, int depth, unsigned *processed)
1979 {
1980         uint32_t i, idx = 0, count = 0;
1981         struct unpacked *array;
1982         unsigned long mem_usage = 0;
1983
1984         array = xcalloc(window, sizeof(struct unpacked));
1985
1986         for (;;) {
1987                 struct object_entry *entry;
1988                 struct unpacked *n = array + idx;
1989                 int j, max_depth, best_base = -1;
1990
1991                 progress_lock();
1992                 if (!*list_size) {
1993                         progress_unlock();
1994                         break;
1995                 }
1996                 entry = *list++;
1997                 (*list_size)--;
1998                 if (!entry->preferred_base) {
1999                         (*processed)++;
2000                         display_progress(progress_state, *processed);
2001                 }
2002                 progress_unlock();
2003
2004                 mem_usage -= free_unpacked(n);
2005                 n->entry = entry;
2006
2007                 while (window_memory_limit &&
2008                        mem_usage > window_memory_limit &&
2009                        count > 1) {
2010                         uint32_t tail = (idx + window - count) % window;
2011                         mem_usage -= free_unpacked(array + tail);
2012                         count--;
2013                 }
2014
2015                 /* We do not compute delta to *create* objects we are not
2016                  * going to pack.
2017                  */
2018                 if (entry->preferred_base)
2019                         goto next;
2020
2021                 /*
2022                  * If the current object is at pack edge, take the depth the
2023                  * objects that depend on the current object into account
2024                  * otherwise they would become too deep.
2025                  */
2026                 max_depth = depth;
2027                 if (entry->delta_child) {
2028                         max_depth -= check_delta_limit(entry, 0);
2029                         if (max_depth <= 0)
2030                                 goto next;
2031                 }
2032
2033                 j = window;
2034                 while (--j > 0) {
2035                         int ret;
2036                         uint32_t other_idx = idx + j;
2037                         struct unpacked *m;
2038                         if (other_idx >= window)
2039                                 other_idx -= window;
2040                         m = array + other_idx;
2041                         if (!m->entry)
2042                                 break;
2043                         ret = try_delta(n, m, max_depth, &mem_usage);
2044                         if (ret < 0)
2045                                 break;
2046                         else if (ret > 0)
2047                                 best_base = other_idx;
2048                 }
2049
2050                 /*
2051                  * If we decided to cache the delta data, then it is best
2052                  * to compress it right away.  First because we have to do
2053                  * it anyway, and doing it here while we're threaded will
2054                  * save a lot of time in the non threaded write phase,
2055                  * as well as allow for caching more deltas within
2056                  * the same cache size limit.
2057                  * ...
2058                  * But only if not writing to stdout, since in that case
2059                  * the network is most likely throttling writes anyway,
2060                  * and therefore it is best to go to the write phase ASAP
2061                  * instead, as we can afford spending more time compressing
2062                  * between writes at that moment.
2063                  */
2064                 if (entry->delta_data && !pack_to_stdout) {
2065                         entry->z_delta_size = do_compress(&entry->delta_data,
2066                                                           entry->delta_size);
2067                         cache_lock();
2068                         delta_cache_size -= entry->delta_size;
2069                         delta_cache_size += entry->z_delta_size;
2070                         cache_unlock();
2071                 }
2072
2073                 /* if we made n a delta, and if n is already at max
2074                  * depth, leaving it in the window is pointless.  we
2075                  * should evict it first.
2076                  */
2077                 if (entry->delta && max_depth <= n->depth)
2078                         continue;
2079
2080                 /*
2081                  * Move the best delta base up in the window, after the
2082                  * currently deltified object, to keep it longer.  It will
2083                  * be the first base object to be attempted next.
2084                  */
2085                 if (entry->delta) {
2086                         struct unpacked swap = array[best_base];
2087                         int dist = (window + idx - best_base) % window;
2088                         int dst = best_base;
2089                         while (dist--) {
2090                                 int src = (dst + 1) % window;
2091                                 array[dst] = array[src];
2092                                 dst = src;
2093                         }
2094                         array[dst] = swap;
2095                 }
2096
2097                 next:
2098                 idx++;
2099                 if (count + 1 < window)
2100                         count++;
2101                 if (idx >= window)
2102                         idx = 0;
2103         }
2104
2105         for (i = 0; i < window; ++i) {
2106                 free_delta_index(array[i].index);
2107                 free(array[i].data);
2108         }
2109         free(array);
2110 }
2111
2112 #ifndef NO_PTHREADS
2113
2114 static void try_to_free_from_threads(size_t size)
2115 {
2116         read_lock();
2117         release_pack_memory(size);
2118         read_unlock();
2119 }
2120
2121 static try_to_free_t old_try_to_free_routine;
2122
2123 /*
2124  * The main thread waits on the condition that (at least) one of the workers
2125  * has stopped working (which is indicated in the .working member of
2126  * struct thread_params).
2127  * When a work thread has completed its work, it sets .working to 0 and
2128  * signals the main thread and waits on the condition that .data_ready
2129  * becomes 1.
2130  */
2131
2132 struct thread_params {
2133         pthread_t thread;
2134         struct object_entry **list;
2135         unsigned list_size;
2136         unsigned remaining;
2137         int window;
2138         int depth;
2139         int working;
2140         int data_ready;
2141         pthread_mutex_t mutex;
2142         pthread_cond_t cond;
2143         unsigned *processed;
2144 };
2145
2146 static pthread_cond_t progress_cond;
2147
2148 /*
2149  * Mutex and conditional variable can't be statically-initialized on Windows.
2150  */
2151 static void init_threaded_search(void)
2152 {
2153         init_recursive_mutex(&read_mutex);
2154         pthread_mutex_init(&cache_mutex, NULL);
2155         pthread_mutex_init(&progress_mutex, NULL);
2156         pthread_cond_init(&progress_cond, NULL);
2157         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2158 }
2159
2160 static void cleanup_threaded_search(void)
2161 {
2162         set_try_to_free_routine(old_try_to_free_routine);
2163         pthread_cond_destroy(&progress_cond);
2164         pthread_mutex_destroy(&read_mutex);
2165         pthread_mutex_destroy(&cache_mutex);
2166         pthread_mutex_destroy(&progress_mutex);
2167 }
2168
2169 static void *threaded_find_deltas(void *arg)
2170 {
2171         struct thread_params *me = arg;
2172
2173         while (me->remaining) {
2174                 find_deltas(me->list, &me->remaining,
2175                             me->window, me->depth, me->processed);
2176
2177                 progress_lock();
2178                 me->working = 0;
2179                 pthread_cond_signal(&progress_cond);
2180                 progress_unlock();
2181
2182                 /*
2183                  * We must not set ->data_ready before we wait on the
2184                  * condition because the main thread may have set it to 1
2185                  * before we get here. In order to be sure that new
2186                  * work is available if we see 1 in ->data_ready, it
2187                  * was initialized to 0 before this thread was spawned
2188                  * and we reset it to 0 right away.
2189                  */
2190                 pthread_mutex_lock(&me->mutex);
2191                 while (!me->data_ready)
2192                         pthread_cond_wait(&me->cond, &me->mutex);
2193                 me->data_ready = 0;
2194                 pthread_mutex_unlock(&me->mutex);
2195         }
2196         /* leave ->working 1 so that this doesn't get more work assigned */
2197         return NULL;
2198 }
2199
2200 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2201                            int window, int depth, unsigned *processed)
2202 {
2203         struct thread_params *p;
2204         int i, ret, active_threads = 0;
2205
2206         init_threaded_search();
2207
2208         if (delta_search_threads <= 1) {
2209                 find_deltas(list, &list_size, window, depth, processed);
2210                 cleanup_threaded_search();
2211                 return;
2212         }
2213         if (progress > pack_to_stdout)
2214                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2215                                 delta_search_threads);
2216         p = xcalloc(delta_search_threads, sizeof(*p));
2217
2218         /* Partition the work amongst work threads. */
2219         for (i = 0; i < delta_search_threads; i++) {
2220                 unsigned sub_size = list_size / (delta_search_threads - i);
2221
2222                 /* don't use too small segments or no deltas will be found */
2223                 if (sub_size < 2*window && i+1 < delta_search_threads)
2224                         sub_size = 0;
2225
2226                 p[i].window = window;
2227                 p[i].depth = depth;
2228                 p[i].processed = processed;
2229                 p[i].working = 1;
2230                 p[i].data_ready = 0;
2231
2232                 /* try to split chunks on "path" boundaries */
2233                 while (sub_size && sub_size < list_size &&
2234                        list[sub_size]->hash &&
2235                        list[sub_size]->hash == list[sub_size-1]->hash)
2236                         sub_size++;
2237
2238                 p[i].list = list;
2239                 p[i].list_size = sub_size;
2240                 p[i].remaining = sub_size;
2241
2242                 list += sub_size;
2243                 list_size -= sub_size;
2244         }
2245
2246         /* Start work threads. */
2247         for (i = 0; i < delta_search_threads; i++) {
2248                 if (!p[i].list_size)
2249                         continue;
2250                 pthread_mutex_init(&p[i].mutex, NULL);
2251                 pthread_cond_init(&p[i].cond, NULL);
2252                 ret = pthread_create(&p[i].thread, NULL,
2253                                      threaded_find_deltas, &p[i]);
2254                 if (ret)
2255                         die("unable to create thread: %s", strerror(ret));
2256                 active_threads++;
2257         }
2258
2259         /*
2260          * Now let's wait for work completion.  Each time a thread is done
2261          * with its work, we steal half of the remaining work from the
2262          * thread with the largest number of unprocessed objects and give
2263          * it to that newly idle thread.  This ensure good load balancing
2264          * until the remaining object list segments are simply too short
2265          * to be worth splitting anymore.
2266          */
2267         while (active_threads) {
2268                 struct thread_params *target = NULL;
2269                 struct thread_params *victim = NULL;
2270                 unsigned sub_size = 0;
2271
2272                 progress_lock();
2273                 for (;;) {
2274                         for (i = 0; !target && i < delta_search_threads; i++)
2275                                 if (!p[i].working)
2276                                         target = &p[i];
2277                         if (target)
2278                                 break;
2279                         pthread_cond_wait(&progress_cond, &progress_mutex);
2280                 }
2281
2282                 for (i = 0; i < delta_search_threads; i++)
2283                         if (p[i].remaining > 2*window &&
2284                             (!victim || victim->remaining < p[i].remaining))
2285                                 victim = &p[i];
2286                 if (victim) {
2287                         sub_size = victim->remaining / 2;
2288                         list = victim->list + victim->list_size - sub_size;
2289                         while (sub_size && list[0]->hash &&
2290                                list[0]->hash == list[-1]->hash) {
2291                                 list++;
2292                                 sub_size--;
2293                         }
2294                         if (!sub_size) {
2295                                 /*
2296                                  * It is possible for some "paths" to have
2297                                  * so many objects that no hash boundary
2298                                  * might be found.  Let's just steal the
2299                                  * exact half in that case.
2300                                  */
2301                                 sub_size = victim->remaining / 2;
2302                                 list -= sub_size;
2303                         }
2304                         target->list = list;
2305                         victim->list_size -= sub_size;
2306                         victim->remaining -= sub_size;
2307                 }
2308                 target->list_size = sub_size;
2309                 target->remaining = sub_size;
2310                 target->working = 1;
2311                 progress_unlock();
2312
2313                 pthread_mutex_lock(&target->mutex);
2314                 target->data_ready = 1;
2315                 pthread_cond_signal(&target->cond);
2316                 pthread_mutex_unlock(&target->mutex);
2317
2318                 if (!sub_size) {
2319                         pthread_join(target->thread, NULL);
2320                         pthread_cond_destroy(&target->cond);
2321                         pthread_mutex_destroy(&target->mutex);
2322                         active_threads--;
2323                 }
2324         }
2325         cleanup_threaded_search();
2326         free(p);
2327 }
2328
2329 #else
2330 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2331 #endif
2332
2333 static void add_tag_chain(const struct object_id *oid)
2334 {
2335         struct tag *tag;
2336
2337         /*
2338          * We catch duplicates already in add_object_entry(), but we'd
2339          * prefer to do this extra check to avoid having to parse the
2340          * tag at all if we already know that it's being packed (e.g., if
2341          * it was included via bitmaps, we would not have parsed it
2342          * previously).
2343          */
2344         if (packlist_find(&to_pack, oid->hash, NULL))
2345                 return;
2346
2347         tag = lookup_tag(oid->hash);
2348         while (1) {
2349                 if (!tag || parse_tag(tag) || !tag->tagged)
2350                         die("unable to pack objects reachable from tag %s",
2351                             oid_to_hex(oid));
2352
2353                 add_object_entry(tag->object.oid.hash, OBJ_TAG, NULL, 0);
2354
2355                 if (tag->tagged->type != OBJ_TAG)
2356                         return;
2357
2358                 tag = (struct tag *)tag->tagged;
2359         }
2360 }
2361
2362 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2363 {
2364         struct object_id peeled;
2365
2366         if (starts_with(path, "refs/tags/") && /* is a tag? */
2367             !peel_ref(path, peeled.hash)    && /* peelable? */
2368             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2369                 add_tag_chain(oid);
2370         return 0;
2371 }
2372
2373 static void prepare_pack(int window, int depth)
2374 {
2375         struct object_entry **delta_list;
2376         uint32_t i, nr_deltas;
2377         unsigned n;
2378
2379         get_object_details();
2380
2381         /*
2382          * If we're locally repacking then we need to be doubly careful
2383          * from now on in order to make sure no stealth corruption gets
2384          * propagated to the new pack.  Clients receiving streamed packs
2385          * should validate everything they get anyway so no need to incur
2386          * the additional cost here in that case.
2387          */
2388         if (!pack_to_stdout)
2389                 do_check_packed_object_crc = 1;
2390
2391         if (!to_pack.nr_objects || !window || !depth)
2392                 return;
2393
2394         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2395         nr_deltas = n = 0;
2396
2397         for (i = 0; i < to_pack.nr_objects; i++) {
2398                 struct object_entry *entry = to_pack.objects + i;
2399
2400                 if (entry->delta)
2401                         /* This happens if we decided to reuse existing
2402                          * delta from a pack.  "reuse_delta &&" is implied.
2403                          */
2404                         continue;
2405
2406                 if (entry->size < 50)
2407                         continue;
2408
2409                 if (entry->no_try_delta)
2410                         continue;
2411
2412                 if (!entry->preferred_base) {
2413                         nr_deltas++;
2414                         if (entry->type < 0)
2415                                 die("unable to get type of object %s",
2416                                     sha1_to_hex(entry->idx.sha1));
2417                 } else {
2418                         if (entry->type < 0) {
2419                                 /*
2420                                  * This object is not found, but we
2421                                  * don't have to include it anyway.
2422                                  */
2423                                 continue;
2424                         }
2425                 }
2426
2427                 delta_list[n++] = entry;
2428         }
2429
2430         if (nr_deltas && n > 1) {
2431                 unsigned nr_done = 0;
2432                 if (progress)
2433                         progress_state = start_progress(_("Compressing objects"),
2434                                                         nr_deltas);
2435                 QSORT(delta_list, n, type_size_sort);
2436                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2437                 stop_progress(&progress_state);
2438                 if (nr_done != nr_deltas)
2439                         die("inconsistency with delta count");
2440         }
2441         free(delta_list);
2442 }
2443
2444 static int git_pack_config(const char *k, const char *v, void *cb)
2445 {
2446         if (!strcmp(k, "pack.window")) {
2447                 window = git_config_int(k, v);
2448                 return 0;
2449         }
2450         if (!strcmp(k, "pack.windowmemory")) {
2451                 window_memory_limit = git_config_ulong(k, v);
2452                 return 0;
2453         }
2454         if (!strcmp(k, "pack.depth")) {
2455                 depth = git_config_int(k, v);
2456                 return 0;
2457         }
2458         if (!strcmp(k, "pack.compression")) {
2459                 int level = git_config_int(k, v);
2460                 if (level == -1)
2461                         level = Z_DEFAULT_COMPRESSION;
2462                 else if (level < 0 || level > Z_BEST_COMPRESSION)
2463                         die("bad pack compression level %d", level);
2464                 pack_compression_level = level;
2465                 pack_compression_seen = 1;
2466                 return 0;
2467         }
2468         if (!strcmp(k, "pack.deltacachesize")) {
2469                 max_delta_cache_size = git_config_int(k, v);
2470                 return 0;
2471         }
2472         if (!strcmp(k, "pack.deltacachelimit")) {
2473                 cache_max_small_delta_size = git_config_int(k, v);
2474                 return 0;
2475         }
2476         if (!strcmp(k, "pack.writebitmaphashcache")) {
2477                 if (git_config_bool(k, v))
2478                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2479                 else
2480                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2481         }
2482         if (!strcmp(k, "pack.usebitmaps")) {
2483                 use_bitmap_index_default = git_config_bool(k, v);
2484                 return 0;
2485         }
2486         if (!strcmp(k, "pack.threads")) {
2487                 delta_search_threads = git_config_int(k, v);
2488                 if (delta_search_threads < 0)
2489                         die("invalid number of threads specified (%d)",
2490                             delta_search_threads);
2491 #ifdef NO_PTHREADS
2492                 if (delta_search_threads != 1)
2493                         warning("no threads support, ignoring %s", k);
2494 #endif
2495                 return 0;
2496         }
2497         if (!strcmp(k, "pack.indexversion")) {
2498                 pack_idx_opts.version = git_config_int(k, v);
2499                 if (pack_idx_opts.version > 2)
2500                         die("bad pack.indexversion=%"PRIu32,
2501                             pack_idx_opts.version);
2502                 return 0;
2503         }
2504         return git_default_config(k, v, cb);
2505 }
2506
2507 static void read_object_list_from_stdin(void)
2508 {
2509         char line[40 + 1 + PATH_MAX + 2];
2510         unsigned char sha1[20];
2511
2512         for (;;) {
2513                 if (!fgets(line, sizeof(line), stdin)) {
2514                         if (feof(stdin))
2515                                 break;
2516                         if (!ferror(stdin))
2517                                 die("fgets returned NULL, not EOF, not error!");
2518                         if (errno != EINTR)
2519                                 die_errno("fgets");
2520                         clearerr(stdin);
2521                         continue;
2522                 }
2523                 if (line[0] == '-') {
2524                         if (get_sha1_hex(line+1, sha1))
2525                                 die("expected edge sha1, got garbage:\n %s",
2526                                     line);
2527                         add_preferred_base(sha1);
2528                         continue;
2529                 }
2530                 if (get_sha1_hex(line, sha1))
2531                         die("expected sha1, got garbage:\n %s", line);
2532
2533                 add_preferred_base_object(line+41);
2534                 add_object_entry(sha1, 0, line+41, 0);
2535         }
2536 }
2537
2538 #define OBJECT_ADDED (1u<<20)
2539
2540 static void show_commit(struct commit *commit, void *data)
2541 {
2542         add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2543         commit->object.flags |= OBJECT_ADDED;
2544
2545         if (write_bitmap_index)
2546                 index_commit_for_bitmap(commit);
2547 }
2548
2549 static void show_object(struct object *obj, const char *name, void *data)
2550 {
2551         add_preferred_base_object(name);
2552         add_object_entry(obj->oid.hash, obj->type, name, 0);
2553         obj->flags |= OBJECT_ADDED;
2554 }
2555
2556 static void show_edge(struct commit *commit)
2557 {
2558         add_preferred_base(commit->object.oid.hash);
2559 }
2560
2561 struct in_pack_object {
2562         off_t offset;
2563         struct object *object;
2564 };
2565
2566 struct in_pack {
2567         int alloc;
2568         int nr;
2569         struct in_pack_object *array;
2570 };
2571
2572 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2573 {
2574         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2575         in_pack->array[in_pack->nr].object = object;
2576         in_pack->nr++;
2577 }
2578
2579 /*
2580  * Compare the objects in the offset order, in order to emulate the
2581  * "git rev-list --objects" output that produced the pack originally.
2582  */
2583 static int ofscmp(const void *a_, const void *b_)
2584 {
2585         struct in_pack_object *a = (struct in_pack_object *)a_;
2586         struct in_pack_object *b = (struct in_pack_object *)b_;
2587
2588         if (a->offset < b->offset)
2589                 return -1;
2590         else if (a->offset > b->offset)
2591                 return 1;
2592         else
2593                 return oidcmp(&a->object->oid, &b->object->oid);
2594 }
2595
2596 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2597 {
2598         struct packed_git *p;
2599         struct in_pack in_pack;
2600         uint32_t i;
2601
2602         memset(&in_pack, 0, sizeof(in_pack));
2603
2604         for (p = packed_git; p; p = p->next) {
2605                 const unsigned char *sha1;
2606                 struct object *o;
2607
2608                 if (!p->pack_local || p->pack_keep)
2609                         continue;
2610                 if (open_pack_index(p))
2611                         die("cannot open pack index");
2612
2613                 ALLOC_GROW(in_pack.array,
2614                            in_pack.nr + p->num_objects,
2615                            in_pack.alloc);
2616
2617                 for (i = 0; i < p->num_objects; i++) {
2618                         sha1 = nth_packed_object_sha1(p, i);
2619                         o = lookup_unknown_object(sha1);
2620                         if (!(o->flags & OBJECT_ADDED))
2621                                 mark_in_pack_object(o, p, &in_pack);
2622                         o->flags |= OBJECT_ADDED;
2623                 }
2624         }
2625
2626         if (in_pack.nr) {
2627                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2628                 for (i = 0; i < in_pack.nr; i++) {
2629                         struct object *o = in_pack.array[i].object;
2630                         add_object_entry(o->oid.hash, o->type, "", 0);
2631                 }
2632         }
2633         free(in_pack.array);
2634 }
2635
2636 static int add_loose_object(const unsigned char *sha1, const char *path,
2637                             void *data)
2638 {
2639         enum object_type type = sha1_object_info(sha1, NULL);
2640
2641         if (type < 0) {
2642                 warning("loose object at %s could not be examined", path);
2643                 return 0;
2644         }
2645
2646         add_object_entry(sha1, type, "", 0);
2647         return 0;
2648 }
2649
2650 /*
2651  * We actually don't even have to worry about reachability here.
2652  * add_object_entry will weed out duplicates, so we just add every
2653  * loose object we find.
2654  */
2655 static void add_unreachable_loose_objects(void)
2656 {
2657         for_each_loose_file_in_objdir(get_object_directory(),
2658                                       add_loose_object,
2659                                       NULL, NULL, NULL);
2660 }
2661
2662 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2663 {
2664         static struct packed_git *last_found = (void *)1;
2665         struct packed_git *p;
2666
2667         p = (last_found != (void *)1) ? last_found : packed_git;
2668
2669         while (p) {
2670                 if ((!p->pack_local || p->pack_keep) &&
2671                         find_pack_entry_one(sha1, p)) {
2672                         last_found = p;
2673                         return 1;
2674                 }
2675                 if (p == last_found)
2676                         p = packed_git;
2677                 else
2678                         p = p->next;
2679                 if (p == last_found)
2680                         p = p->next;
2681         }
2682         return 0;
2683 }
2684
2685 /*
2686  * Store a list of sha1s that are should not be discarded
2687  * because they are either written too recently, or are
2688  * reachable from another object that was.
2689  *
2690  * This is filled by get_object_list.
2691  */
2692 static struct sha1_array recent_objects;
2693
2694 static int loosened_object_can_be_discarded(const unsigned char *sha1,
2695                                             unsigned long mtime)
2696 {
2697         if (!unpack_unreachable_expiration)
2698                 return 0;
2699         if (mtime > unpack_unreachable_expiration)
2700                 return 0;
2701         if (sha1_array_lookup(&recent_objects, sha1) >= 0)
2702                 return 0;
2703         return 1;
2704 }
2705
2706 static void loosen_unused_packed_objects(struct rev_info *revs)
2707 {
2708         struct packed_git *p;
2709         uint32_t i;
2710         const unsigned char *sha1;
2711
2712         for (p = packed_git; p; p = p->next) {
2713                 if (!p->pack_local || p->pack_keep)
2714                         continue;
2715
2716                 if (open_pack_index(p))
2717                         die("cannot open pack index");
2718
2719                 for (i = 0; i < p->num_objects; i++) {
2720                         sha1 = nth_packed_object_sha1(p, i);
2721                         if (!packlist_find(&to_pack, sha1, NULL) &&
2722                             !has_sha1_pack_kept_or_nonlocal(sha1) &&
2723                             !loosened_object_can_be_discarded(sha1, p->mtime))
2724                                 if (force_object_loose(sha1, p->mtime))
2725                                         die("unable to force loose object");
2726                 }
2727         }
2728 }
2729
2730 /*
2731  * This tracks any options which pack-reuse code expects to be on, or which a
2732  * reader of the pack might not understand, and which would therefore prevent
2733  * blind reuse of what we have on disk.
2734  */
2735 static int pack_options_allow_reuse(void)
2736 {
2737         return pack_to_stdout && allow_ofs_delta;
2738 }
2739
2740 static int get_object_list_from_bitmap(struct rev_info *revs)
2741 {
2742         if (prepare_bitmap_walk(revs) < 0)
2743                 return -1;
2744
2745         if (pack_options_allow_reuse() &&
2746             !reuse_partial_packfile_from_bitmap(
2747                         &reuse_packfile,
2748                         &reuse_packfile_objects,
2749                         &reuse_packfile_offset)) {
2750                 assert(reuse_packfile_objects);
2751                 nr_result += reuse_packfile_objects;
2752                 display_progress(progress_state, nr_result);
2753         }
2754
2755         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2756         return 0;
2757 }
2758
2759 static void record_recent_object(struct object *obj,
2760                                  const char *name,
2761                                  void *data)
2762 {
2763         sha1_array_append(&recent_objects, obj->oid.hash);
2764 }
2765
2766 static void record_recent_commit(struct commit *commit, void *data)
2767 {
2768         sha1_array_append(&recent_objects, commit->object.oid.hash);
2769 }
2770
2771 static void get_object_list(int ac, const char **av)
2772 {
2773         struct rev_info revs;
2774         char line[1000];
2775         int flags = 0;
2776
2777         init_revisions(&revs, NULL);
2778         save_commit_buffer = 0;
2779         setup_revisions(ac, av, &revs, NULL);
2780
2781         /* make sure shallows are read */
2782         is_repository_shallow();
2783
2784         while (fgets(line, sizeof(line), stdin) != NULL) {
2785                 int len = strlen(line);
2786                 if (len && line[len - 1] == '\n')
2787                         line[--len] = 0;
2788                 if (!len)
2789                         break;
2790                 if (*line == '-') {
2791                         if (!strcmp(line, "--not")) {
2792                                 flags ^= UNINTERESTING;
2793                                 write_bitmap_index = 0;
2794                                 continue;
2795                         }
2796                         if (starts_with(line, "--shallow ")) {
2797                                 unsigned char sha1[20];
2798                                 if (get_sha1_hex(line + 10, sha1))
2799                                         die("not an SHA-1 '%s'", line + 10);
2800                                 register_shallow(sha1);
2801                                 use_bitmap_index = 0;
2802                                 continue;
2803                         }
2804                         die("not a rev '%s'", line);
2805                 }
2806                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2807                         die("bad revision '%s'", line);
2808         }
2809
2810         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2811                 return;
2812
2813         if (prepare_revision_walk(&revs))
2814                 die("revision walk setup failed");
2815         mark_edges_uninteresting(&revs, show_edge);
2816         traverse_commit_list(&revs, show_commit, show_object, NULL);
2817
2818         if (unpack_unreachable_expiration) {
2819                 revs.ignore_missing_links = 1;
2820                 if (add_unseen_recent_objects_to_traversal(&revs,
2821                                 unpack_unreachable_expiration))
2822                         die("unable to add recent objects");
2823                 if (prepare_revision_walk(&revs))
2824                         die("revision walk setup failed");
2825                 traverse_commit_list(&revs, record_recent_commit,
2826                                      record_recent_object, NULL);
2827         }
2828
2829         if (keep_unreachable)
2830                 add_objects_in_unpacked_packs(&revs);
2831         if (pack_loose_unreachable)
2832                 add_unreachable_loose_objects();
2833         if (unpack_unreachable)
2834                 loosen_unused_packed_objects(&revs);
2835
2836         sha1_array_clear(&recent_objects);
2837 }
2838
2839 static int option_parse_index_version(const struct option *opt,
2840                                       const char *arg, int unset)
2841 {
2842         char *c;
2843         const char *val = arg;
2844         pack_idx_opts.version = strtoul(val, &c, 10);
2845         if (pack_idx_opts.version > 2)
2846                 die(_("unsupported index version %s"), val);
2847         if (*c == ',' && c[1])
2848                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2849         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2850                 die(_("bad index version '%s'"), val);
2851         return 0;
2852 }
2853
2854 static int option_parse_unpack_unreachable(const struct option *opt,
2855                                            const char *arg, int unset)
2856 {
2857         if (unset) {
2858                 unpack_unreachable = 0;
2859                 unpack_unreachable_expiration = 0;
2860         }
2861         else {
2862                 unpack_unreachable = 1;
2863                 if (arg)
2864                         unpack_unreachable_expiration = approxidate(arg);
2865         }
2866         return 0;
2867 }
2868
2869 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2870 {
2871         int use_internal_rev_list = 0;
2872         int thin = 0;
2873         int shallow = 0;
2874         int all_progress_implied = 0;
2875         struct argv_array rp = ARGV_ARRAY_INIT;
2876         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2877         int rev_list_index = 0;
2878         struct option pack_objects_options[] = {
2879                 OPT_SET_INT('q', "quiet", &progress,
2880                             N_("do not show progress meter"), 0),
2881                 OPT_SET_INT(0, "progress", &progress,
2882                             N_("show progress meter"), 1),
2883                 OPT_SET_INT(0, "all-progress", &progress,
2884                             N_("show progress meter during object writing phase"), 2),
2885                 OPT_BOOL(0, "all-progress-implied",
2886                          &all_progress_implied,
2887                          N_("similar to --all-progress when progress meter is shown")),
2888                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2889                   N_("write the pack index file in the specified idx format version"),
2890                   0, option_parse_index_version },
2891                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2892                               N_("maximum size of each output pack file")),
2893                 OPT_BOOL(0, "local", &local,
2894                          N_("ignore borrowed objects from alternate object store")),
2895                 OPT_BOOL(0, "incremental", &incremental,
2896                          N_("ignore packed objects")),
2897                 OPT_INTEGER(0, "window", &window,
2898                             N_("limit pack window by objects")),
2899                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2900                               N_("limit pack window by memory in addition to object limit")),
2901                 OPT_INTEGER(0, "depth", &depth,
2902                             N_("maximum length of delta chain allowed in the resulting pack")),
2903                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2904                          N_("reuse existing deltas")),
2905                 OPT_BOOL(0, "reuse-object", &reuse_object,
2906                          N_("reuse existing objects")),
2907                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2908                          N_("use OFS_DELTA objects")),
2909                 OPT_INTEGER(0, "threads", &delta_search_threads,
2910                             N_("use threads when searching for best delta matches")),
2911                 OPT_BOOL(0, "non-empty", &non_empty,
2912                          N_("do not create an empty pack output")),
2913                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2914                          N_("read revision arguments from standard input")),
2915                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2916                   N_("limit the objects to those that are not yet packed"),
2917                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2918                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2919                   N_("include objects reachable from any reference"),
2920                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2921                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2922                   N_("include objects referred by reflog entries"),
2923                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2924                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2925                   N_("include objects referred to by the index"),
2926                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2927                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2928                          N_("output pack to stdout")),
2929                 OPT_BOOL(0, "include-tag", &include_tag,
2930                          N_("include tag objects that refer to objects to be packed")),
2931                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2932                          N_("keep unreachable objects")),
2933                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
2934                          N_("pack loose unreachable objects")),
2935                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2936                   N_("unpack unreachable objects newer than <time>"),
2937                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2938                 OPT_BOOL(0, "thin", &thin,
2939                          N_("create thin packs")),
2940                 OPT_BOOL(0, "shallow", &shallow,
2941                          N_("create packs suitable for shallow fetches")),
2942                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2943                          N_("ignore packs that have companion .keep file")),
2944                 OPT_INTEGER(0, "compression", &pack_compression_level,
2945                             N_("pack compression level")),
2946                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2947                             N_("do not hide commits by grafts"), 0),
2948                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2949                          N_("use a bitmap index if available to speed up counting objects")),
2950                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2951                          N_("write a bitmap index together with the pack index")),
2952                 OPT_END(),
2953         };
2954
2955         check_replace_refs = 0;
2956
2957         reset_pack_idx_option(&pack_idx_opts);
2958         git_config(git_pack_config, NULL);
2959         if (!pack_compression_seen && core_compression_seen)
2960                 pack_compression_level = core_compression_level;
2961
2962         progress = isatty(2);
2963         argc = parse_options(argc, argv, prefix, pack_objects_options,
2964                              pack_usage, 0);
2965
2966         if (argc) {
2967                 base_name = argv[0];
2968                 argc--;
2969         }
2970         if (pack_to_stdout != !base_name || argc)
2971                 usage_with_options(pack_usage, pack_objects_options);
2972
2973         argv_array_push(&rp, "pack-objects");
2974         if (thin) {
2975                 use_internal_rev_list = 1;
2976                 argv_array_push(&rp, shallow
2977                                 ? "--objects-edge-aggressive"
2978                                 : "--objects-edge");
2979         } else
2980                 argv_array_push(&rp, "--objects");
2981
2982         if (rev_list_all) {
2983                 use_internal_rev_list = 1;
2984                 argv_array_push(&rp, "--all");
2985         }
2986         if (rev_list_reflog) {
2987                 use_internal_rev_list = 1;
2988                 argv_array_push(&rp, "--reflog");
2989         }
2990         if (rev_list_index) {
2991                 use_internal_rev_list = 1;
2992                 argv_array_push(&rp, "--indexed-objects");
2993         }
2994         if (rev_list_unpacked) {
2995                 use_internal_rev_list = 1;
2996                 argv_array_push(&rp, "--unpacked");
2997         }
2998
2999         if (!reuse_object)
3000                 reuse_delta = 0;
3001         if (pack_compression_level == -1)
3002                 pack_compression_level = Z_DEFAULT_COMPRESSION;
3003         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
3004                 die("bad pack compression level %d", pack_compression_level);
3005
3006         if (!delta_search_threads)      /* --threads=0 means autodetect */
3007                 delta_search_threads = online_cpus();
3008
3009 #ifdef NO_PTHREADS
3010         if (delta_search_threads != 1)
3011                 warning("no threads support, ignoring --threads");
3012 #endif
3013         if (!pack_to_stdout && !pack_size_limit)
3014                 pack_size_limit = pack_size_limit_cfg;
3015         if (pack_to_stdout && pack_size_limit)
3016                 die("--max-pack-size cannot be used to build a pack for transfer.");
3017         if (pack_size_limit && pack_size_limit < 1024*1024) {
3018                 warning("minimum pack size limit is 1 MiB");
3019                 pack_size_limit = 1024*1024;
3020         }
3021
3022         if (!pack_to_stdout && thin)
3023                 die("--thin cannot be used to build an indexable pack.");
3024
3025         if (keep_unreachable && unpack_unreachable)
3026                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
3027         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3028                 unpack_unreachable_expiration = 0;
3029
3030         /*
3031          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3032          *
3033          * - to produce good pack (with bitmap index not-yet-packed objects are
3034          *   packed in suboptimal order).
3035          *
3036          * - to use more robust pack-generation codepath (avoiding possible
3037          *   bugs in bitmap code and possible bitmap index corruption).
3038          */
3039         if (!pack_to_stdout)
3040                 use_bitmap_index_default = 0;
3041
3042         if (use_bitmap_index < 0)
3043                 use_bitmap_index = use_bitmap_index_default;
3044
3045         /* "hard" reasons not to use bitmaps; these just won't work at all */
3046         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
3047                 use_bitmap_index = 0;
3048
3049         if (pack_to_stdout || !rev_list_all)
3050                 write_bitmap_index = 0;
3051
3052         if (progress && all_progress_implied)
3053                 progress = 2;
3054
3055         prepare_packed_git();
3056         if (ignore_packed_keep) {
3057                 struct packed_git *p;
3058                 for (p = packed_git; p; p = p->next)
3059                         if (p->pack_local && p->pack_keep)
3060                                 break;
3061                 if (!p) /* no keep-able packs found */
3062                         ignore_packed_keep = 0;
3063         }
3064         if (local) {
3065                 /*
3066                  * unlike ignore_packed_keep above, we do not want to
3067                  * unset "local" based on looking at packs, as it
3068                  * also covers non-local objects
3069                  */
3070                 struct packed_git *p;
3071                 for (p = packed_git; p; p = p->next) {
3072                         if (!p->pack_local) {
3073                                 have_non_local_packs = 1;
3074                                 break;
3075                         }
3076                 }
3077         }
3078
3079         if (progress)
3080                 progress_state = start_progress(_("Counting objects"), 0);
3081         if (!use_internal_rev_list)
3082                 read_object_list_from_stdin();
3083         else {
3084                 get_object_list(rp.argc, rp.argv);
3085                 argv_array_clear(&rp);
3086         }
3087         cleanup_preferred_base();
3088         if (include_tag && nr_result)
3089                 for_each_ref(add_ref_tag, NULL);
3090         stop_progress(&progress_state);
3091
3092         if (non_empty && !nr_result)
3093                 return 0;
3094         if (nr_result)
3095                 prepare_pack(window, depth);
3096         write_pack_file();
3097         if (progress)
3098                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
3099                         " reused %"PRIu32" (delta %"PRIu32")\n",
3100                         written, written_delta, reused, reused_delta);
3101         return 0;
3102 }