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

   1 #include "builtin.h"
   2 #include "cache.h"
   3 #include "config.h"
   4 #include "attr.h"
   5 #include "object.h"
   6 #include "blob.h"
   7 #include "commit.h"
   8 #include "tag.h"
   9 #include "tree.h"
  10 #include "delta.h"
  11 #include "pack.h"
  12 #include "pack-revindex.h"
  13 #include "csum-file.h"
  14 #include "tree-walk.h"
  15 #include "diff.h"
  16 #include "revision.h"
  17 #include "list-objects.h"
  18 #include "pack-objects.h"
  19 #include "progress.h"
  20 #include "refs.h"
  21 #include "streaming.h"
  22 #include "thread-utils.h"
  23 #include "pack-bitmap.h"
  24 #include "reachable.h"
  25 #include "sha1-array.h"
  26 #include "argv-array.h"
  27 #include "mru.h"
  28
  29 static const char *pack_usage[] = {
  30         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
  31         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
  32         NULL
  33 };
  34
  35 /*
  36  * Objects we are going to pack are collected in the `to_pack` structure.
  37  * It contains an array (dynamically expanded) of the object data, and a map
  38  * that can resolve SHA1s to their position in the array.
  39  */
  40 static struct packing_data to_pack;
  41
  42 static struct pack_idx_entry **written_list;
  43 static uint32_t nr_result, nr_written;
  44
  45 static int non_empty;
  46 static int reuse_delta = 1, reuse_object = 1;
  47 static int keep_unreachable, unpack_unreachable, include_tag;
  48 static timestamp_t unpack_unreachable_expiration;
  49 static int pack_loose_unreachable;
  50 static int local;
  51 static int have_non_local_packs;
  52 static int incremental;
  53 static int ignore_packed_keep;
  54 static int allow_ofs_delta;
  55 static struct pack_idx_option pack_idx_opts;
  56 static const char *base_name;
  57 static int progress = 1;
  58 static int window = 10;
  59 static unsigned long pack_size_limit;
  60 static int depth = 50;
  61 static int delta_search_threads;
  62 static int pack_to_stdout;
  63 static int num_preferred_base;
  64 static struct progress *progress_state;
  65
  66 static struct packed_git *reuse_packfile;
  67 static uint32_t reuse_packfile_objects;
  68 static off_t reuse_packfile_offset;
  69
  70 static int use_bitmap_index_default = 1;
  71 static int use_bitmap_index = -1;
  72 static int write_bitmap_index;
  73 static uint16_t write_bitmap_options;
  74
  75 static unsigned long delta_cache_size = 0;
  76 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
  77 static unsigned long cache_max_small_delta_size = 1000;
  78
  79 static unsigned long window_memory_limit = 0;
  80
  81 /*
  82  * stats
  83  */
  84 static uint32_t written, written_delta;
  85 static uint32_t reused, reused_delta;
  86
  87 /*
  88  * Indexed commits
  89  */
  90 static struct commit **indexed_commits;
  91 static unsigned int indexed_commits_nr;
  92 static unsigned int indexed_commits_alloc;
  93
  94 static void index_commit_for_bitmap(struct commit *commit)
  95 {
  96         if (indexed_commits_nr >= indexed_commits_alloc) {
  97                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
  98                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
  99         }
 100
 101         indexed_commits[indexed_commits_nr++] = commit;
 102 }
 103
 104 static void *get_delta(struct object_entry *entry)
 105 {
 106         unsigned long size, base_size, delta_size;
 107         void *buf, *base_buf, *delta_buf;
 108         enum object_type type;
 109
 110         buf = read_sha1_file(entry->idx.oid.hash, &type, &size);
 111         if (!buf)
 112                 die("unable to read %s", oid_to_hex(&entry->idx.oid));
 113         base_buf = read_sha1_file(entry->delta->idx.oid.hash, &type,
 114                                   &base_size);
 115         if (!base_buf)
 116                 die("unable to read %s",
 117                     oid_to_hex(&entry->delta->idx.oid));
 118         delta_buf = diff_delta(base_buf, base_size,
 119                                buf, size, &delta_size, 0);
 120         if (!delta_buf || delta_size != entry->delta_size)
 121                 die("delta size changed");
 122         free(buf);
 123         free(base_buf);
 124         return delta_buf;
 125 }
 126
 127 static unsigned long do_compress(void **pptr, unsigned long size)
 128 {
 129         git_zstream stream;
 130         void *in, *out;
 131         unsigned long maxsize;
 132
 133         git_deflate_init(&stream, pack_compression_level);
 134         maxsize = git_deflate_bound(&stream, size);
 135
 136         in = *pptr;
 137         out = xmalloc(maxsize);
 138         *pptr = out;
 139
 140         stream.next_in = in;
 141         stream.avail_in = size;
 142         stream.next_out = out;
 143         stream.avail_out = maxsize;
 144         while (git_deflate(&stream, Z_FINISH) == Z_OK)
 145                 ; /* nothing */
 146         git_deflate_end(&stream);
 147
 148         free(in);
 149         return stream.total_out;
 150 }
 151
 152 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
 153                                            const unsigned char *sha1)
 154 {
 155         git_zstream stream;
 156         unsigned char ibuf[1024 * 16];
 157         unsigned char obuf[1024 * 16];
 158         unsigned long olen = 0;
 159
 160         git_deflate_init(&stream, pack_compression_level);
 161
 162         for (;;) {
 163                 ssize_t readlen;
 164                 int zret = Z_OK;
 165                 readlen = read_istream(st, ibuf, sizeof(ibuf));
 166                 if (readlen == -1)
 167                         die(_("unable to read %s"), sha1_to_hex(sha1));
 168
 169                 stream.next_in = ibuf;
 170                 stream.avail_in = readlen;
 171                 while ((stream.avail_in || readlen == 0) &&
 172                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
 173                         stream.next_out = obuf;
 174                         stream.avail_out = sizeof(obuf);
 175                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 176                         sha1write(f, obuf, stream.next_out - obuf);
 177                         olen += stream.next_out - obuf;
 178                 }
 179                 if (stream.avail_in)
 180                         die(_("deflate error (%d)"), zret);
 181                 if (readlen == 0) {
 182                         if (zret != Z_STREAM_END)
 183                                 die(_("deflate error (%d)"), zret);
 184                         break;
 185                 }
 186         }
 187         git_deflate_end(&stream);
 188         return olen;
 189 }
 190
 191 /*
 192  * we are going to reuse the existing object data as is.  make
 193  * sure it is not corrupt.
 194  */
 195 static int check_pack_inflate(struct packed_git *p,
 196                 struct pack_window **w_curs,
 197                 off_t offset,
 198                 off_t len,
 199                 unsigned long expect)
 200 {
 201         git_zstream stream;
 202         unsigned char fakebuf[4096], *in;
 203         int st;
 204
 205         memset(&stream, 0, sizeof(stream));
 206         git_inflate_init(&stream);
 207         do {
 208                 in = use_pack(p, w_curs, offset, &stream.avail_in);
 209                 stream.next_in = in;
 210                 stream.next_out = fakebuf;
 211                 stream.avail_out = sizeof(fakebuf);
 212                 st = git_inflate(&stream, Z_FINISH);
 213                 offset += stream.next_in - in;
 214         } while (st == Z_OK || st == Z_BUF_ERROR);
 215         git_inflate_end(&stream);
 216         return (st == Z_STREAM_END &&
 217                 stream.total_out == expect &&
 218                 stream.total_in == len) ? 0 : -1;
 219 }
 220
 221 static void copy_pack_data(struct sha1file *f,
 222                 struct packed_git *p,
 223                 struct pack_window **w_curs,
 224                 off_t offset,
 225                 off_t len)
 226 {
 227         unsigned char *in;
 228         unsigned long avail;
 229
 230         while (len) {
 231                 in = use_pack(p, w_curs, offset, &avail);
 232                 if (avail > len)
 233                         avail = (unsigned long)len;
 234                 sha1write(f, in, avail);
 235                 offset += avail;
 236                 len -= avail;
 237         }
 238 }
 239
 240 /* Return 0 if we will bust the pack-size limit */
 241 static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
 242                                            unsigned long limit, int usable_delta)
 243 {
 244         unsigned long size, datalen;
 245         unsigned char header[MAX_PACK_OBJECT_HEADER],
 246                       dheader[MAX_PACK_OBJECT_HEADER];
 247         unsigned hdrlen;
 248         enum object_type type;
 249         void *buf;
 250         struct git_istream *st = NULL;
 251
 252         if (!usable_delta) {
 253                 if (entry->type == OBJ_BLOB &&
 254                     entry->size > big_file_threshold &&
 255                     (st = open_istream(entry->idx.oid.hash, &type, &size, NULL)) != NULL)
 256                         buf = NULL;
 257                 else {
 258                         buf = read_sha1_file(entry->idx.oid.hash, &type,
 259                                              &size);
 260                         if (!buf)
 261                                 die(_("unable to read %s"),
 262                                     oid_to_hex(&entry->idx.oid));
 263                 }
 264                 /*
 265                  * make sure no cached delta data remains from a
 266                  * previous attempt before a pack split occurred.
 267                  */
 268                 free(entry->delta_data);
 269                 entry->delta_data = NULL;
 270                 entry->z_delta_size = 0;
 271         } else if (entry->delta_data) {
 272                 size = entry->delta_size;
 273                 buf = entry->delta_data;
 274                 entry->delta_data = NULL;
 275                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 276                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 277         } else {
 278                 buf = get_delta(entry);
 279                 size = entry->delta_size;
 280                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 281                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 282         }
 283
 284         if (st) /* large blob case, just assume we don't compress well */
 285                 datalen = size;
 286         else if (entry->z_delta_size)
 287                 datalen = entry->z_delta_size;
 288         else
 289                 datalen = do_compress(&buf, size);
 290
 291         /*
 292          * The object header is a byte of 'type' followed by zero or
 293          * more bytes of length.
 294          */
 295         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 296                                               type, size);
 297
 298         if (type == OBJ_OFS_DELTA) {
 299                 /*
 300                  * Deltas with relative base contain an additional
 301                  * encoding of the relative offset for the delta
 302                  * base from this object's position in the pack.
 303                  */
 304                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 305                 unsigned pos = sizeof(dheader) - 1;
 306                 dheader[pos] = ofs & 127;
 307                 while (ofs >>= 7)
 308                         dheader[--pos] = 128 | (--ofs & 127);
 309                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 310                         if (st)
 311                                 close_istream(st);
 312                         free(buf);
 313                         return 0;
 314                 }
 315                 sha1write(f, header, hdrlen);
 316                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 317                 hdrlen += sizeof(dheader) - pos;
 318         } else if (type == OBJ_REF_DELTA) {
 319                 /*
 320                  * Deltas with a base reference contain
 321                  * an additional 20 bytes for the base sha1.
 322                  */
 323                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 324                         if (st)
 325                                 close_istream(st);
 326                         free(buf);
 327                         return 0;
 328                 }
 329                 sha1write(f, header, hdrlen);
 330                 sha1write(f, entry->delta->idx.oid.hash, 20);
 331                 hdrlen += 20;
 332         } else {
 333                 if (limit && hdrlen + datalen + 20 >= limit) {
 334                         if (st)
 335                                 close_istream(st);
 336                         free(buf);
 337                         return 0;
 338                 }
 339                 sha1write(f, header, hdrlen);
 340         }
 341         if (st) {
 342                 datalen = write_large_blob_data(st, f, entry->idx.oid.hash);
 343                 close_istream(st);
 344         } else {
 345                 sha1write(f, buf, datalen);
 346                 free(buf);
 347         }
 348
 349         return hdrlen + datalen;
 350 }
 351
 352 /* Return 0 if we will bust the pack-size limit */
 353 static off_t write_reuse_object(struct sha1file *f, struct object_entry *entry,
 354                                 unsigned long limit, int usable_delta)
 355 {
 356         struct packed_git *p = entry->in_pack;
 357         struct pack_window *w_curs = NULL;
 358         struct revindex_entry *revidx;
 359         off_t offset;
 360         enum object_type type = entry->type;
 361         off_t datalen;
 362         unsigned char header[MAX_PACK_OBJECT_HEADER],
 363                       dheader[MAX_PACK_OBJECT_HEADER];
 364         unsigned hdrlen;
 365
 366         if (entry->delta)
 367                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 368                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 369         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 370                                               type, entry->size);
 371
 372         offset = entry->in_pack_offset;
 373         revidx = find_pack_revindex(p, offset);
 374         datalen = revidx[1].offset - offset;
 375         if (!pack_to_stdout && p->index_version > 1 &&
 376             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 377                 error("bad packed object CRC for %s",
 378                       oid_to_hex(&entry->idx.oid));
 379                 unuse_pack(&w_curs);
 380                 return write_no_reuse_object(f, entry, limit, usable_delta);
 381         }
 382
 383         offset += entry->in_pack_header_size;
 384         datalen -= entry->in_pack_header_size;
 385
 386         if (!pack_to_stdout && p->index_version == 1 &&
 387             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 388                 error("corrupt packed object for %s",
 389                       oid_to_hex(&entry->idx.oid));
 390                 unuse_pack(&w_curs);
 391                 return write_no_reuse_object(f, entry, limit, usable_delta);
 392         }
 393
 394         if (type == OBJ_OFS_DELTA) {
 395                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 396                 unsigned pos = sizeof(dheader) - 1;
 397                 dheader[pos] = ofs & 127;
 398                 while (ofs >>= 7)
 399                         dheader[--pos] = 128 | (--ofs & 127);
 400                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 401                         unuse_pack(&w_curs);
 402                         return 0;
 403                 }
 404                 sha1write(f, header, hdrlen);
 405                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 406                 hdrlen += sizeof(dheader) - pos;
 407                 reused_delta++;
 408         } else if (type == OBJ_REF_DELTA) {
 409                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 410                         unuse_pack(&w_curs);
 411                         return 0;
 412                 }
 413                 sha1write(f, header, hdrlen);
 414                 sha1write(f, entry->delta->idx.oid.hash, 20);
 415                 hdrlen += 20;
 416                 reused_delta++;
 417         } else {
 418                 if (limit && hdrlen + datalen + 20 >= limit) {
 419                         unuse_pack(&w_curs);
 420                         return 0;
 421                 }
 422                 sha1write(f, header, hdrlen);
 423         }
 424         copy_pack_data(f, p, &w_curs, offset, datalen);
 425         unuse_pack(&w_curs);
 426         reused++;
 427         return hdrlen + datalen;
 428 }
 429
 430 /* Return 0 if we will bust the pack-size limit */
 431 static off_t write_object(struct sha1file *f,
 432                           struct object_entry *entry,
 433                           off_t write_offset)
 434 {
 435         unsigned long limit;
 436         off_t len;
 437         int usable_delta, to_reuse;
 438
 439         if (!pack_to_stdout)
 440                 crc32_begin(f);
 441
 442         /* apply size limit if limited packsize and not first object */
 443         if (!pack_size_limit || !nr_written)
 444                 limit = 0;
 445         else if (pack_size_limit <= write_offset)
 446                 /*
 447                  * the earlier object did not fit the limit; avoid
 448                  * mistaking this with unlimited (i.e. limit = 0).
 449                  */
 450                 limit = 1;
 451         else
 452                 limit = pack_size_limit - write_offset;
 453
 454         if (!entry->delta)
 455                 usable_delta = 0;       /* no delta */
 456         else if (!pack_size_limit)
 457                usable_delta = 1;        /* unlimited packfile */
 458         else if (entry->delta->idx.offset == (off_t)-1)
 459                 usable_delta = 0;       /* base was written to another pack */
 460         else if (entry->delta->idx.offset)
 461                 usable_delta = 1;       /* base already exists in this pack */
 462         else
 463                 usable_delta = 0;       /* base could end up in another pack */
 464
 465         if (!reuse_object)
 466                 to_reuse = 0;   /* explicit */
 467         else if (!entry->in_pack)
 468                 to_reuse = 0;   /* can't reuse what we don't have */
 469         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
 470                                 /* check_object() decided it for us ... */
 471                 to_reuse = usable_delta;
 472                                 /* ... but pack split may override that */
 473         else if (entry->type != entry->in_pack_type)
 474                 to_reuse = 0;   /* pack has delta which is unusable */
 475         else if (entry->delta)
 476                 to_reuse = 0;   /* we want to pack afresh */
 477         else
 478                 to_reuse = 1;   /* we have it in-pack undeltified,
 479                                  * and we do not need to deltify it.
 480                                  */
 481
 482         if (!to_reuse)
 483                 len = write_no_reuse_object(f, entry, limit, usable_delta);
 484         else
 485                 len = write_reuse_object(f, entry, limit, usable_delta);
 486         if (!len)
 487                 return 0;
 488
 489         if (usable_delta)
 490                 written_delta++;
 491         written++;
 492         if (!pack_to_stdout)
 493                 entry->idx.crc32 = crc32_end(f);
 494         return len;
 495 }
 496
 497 enum write_one_status {
 498         WRITE_ONE_SKIP = -1, /* already written */
 499         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 500         WRITE_ONE_WRITTEN = 1, /* normal */
 501         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 502 };
 503
 504 static enum write_one_status write_one(struct sha1file *f,
 505                                        struct object_entry *e,
 506                                        off_t *offset)
 507 {
 508         off_t size;
 509         int recursing;
 510
 511         /*
 512          * we set offset to 1 (which is an impossible value) to mark
 513          * the fact that this object is involved in "write its base
 514          * first before writing a deltified object" recursion.
 515          */
 516         recursing = (e->idx.offset == 1);
 517         if (recursing) {
 518                 warning("recursive delta detected for object %s",
 519                         oid_to_hex(&e->idx.oid));
 520                 return WRITE_ONE_RECURSIVE;
 521         } else if (e->idx.offset || e->preferred_base) {
 522                 /* offset is non zero if object is written already. */
 523                 return WRITE_ONE_SKIP;
 524         }
 525
 526         /* if we are deltified, write out base object first. */
 527         if (e->delta) {
 528                 e->idx.offset = 1; /* now recurse */
 529                 switch (write_one(f, e->delta, offset)) {
 530                 case WRITE_ONE_RECURSIVE:
 531                         /* we cannot depend on this one */
 532                         e->delta = NULL;
 533                         break;
 534                 default:
 535                         break;
 536                 case WRITE_ONE_BREAK:
 537                         e->idx.offset = recursing;
 538                         return WRITE_ONE_BREAK;
 539                 }
 540         }
 541
 542         e->idx.offset = *offset;
 543         size = write_object(f, e, *offset);
 544         if (!size) {
 545                 e->idx.offset = recursing;
 546                 return WRITE_ONE_BREAK;
 547         }
 548         written_list[nr_written++] = &e->idx;
 549
 550         /* make sure off_t is sufficiently large not to wrap */
 551         if (signed_add_overflows(*offset, size))
 552                 die("pack too large for current definition of off_t");
 553         *offset += size;
 554         return WRITE_ONE_WRITTEN;
 555 }
 556
 557 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
 558                        void *cb_data)
 559 {
 560         unsigned char peeled[20];
 561         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
 562
 563         if (entry)
 564                 entry->tagged = 1;
 565         if (!peel_ref(path, peeled)) {
 566                 entry = packlist_find(&to_pack, peeled, NULL);
 567                 if (entry)
 568                         entry->tagged = 1;
 569         }
 570         return 0;
 571 }
 572
 573 static inline void add_to_write_order(struct object_entry **wo,
 574                                unsigned int *endp,
 575                                struct object_entry *e)
 576 {
 577         if (e->filled)
 578                 return;
 579         wo[(*endp)++] = e;
 580         e->filled = 1;
 581 }
 582
 583 static void add_descendants_to_write_order(struct object_entry **wo,
 584                                            unsigned int *endp,
 585                                            struct object_entry *e)
 586 {
 587         int add_to_order = 1;
 588         while (e) {
 589                 if (add_to_order) {
 590                         struct object_entry *s;
 591                         /* add this node... */
 592                         add_to_write_order(wo, endp, e);
 593                         /* all its siblings... */
 594                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
 595                                 add_to_write_order(wo, endp, s);
 596                         }
 597                 }
 598                 /* drop down a level to add left subtree nodes if possible */
 599                 if (e->delta_child) {
 600                         add_to_order = 1;
 601                         e = e->delta_child;
 602                 } else {
 603                         add_to_order = 0;
 604                         /* our sibling might have some children, it is next */
 605                         if (e->delta_sibling) {
 606                                 e = e->delta_sibling;
 607                                 continue;
 608                         }
 609                         /* go back to our parent node */
 610                         e = e->delta;
 611                         while (e && !e->delta_sibling) {
 612                                 /* we're on the right side of a subtree, keep
 613                                  * going up until we can go right again */
 614                                 e = e->delta;
 615                         }
 616                         if (!e) {
 617                                 /* done- we hit our original root node */
 618                                 return;
 619                         }
 620                         /* pass it off to sibling at this level */
 621                         e = e->delta_sibling;
 622                 }
 623         };
 624 }
 625
 626 static void add_family_to_write_order(struct object_entry **wo,
 627                                       unsigned int *endp,
 628                                       struct object_entry *e)
 629 {
 630         struct object_entry *root;
 631
 632         for (root = e; root->delta; root = root->delta)
 633                 ; /* nothing */
 634         add_descendants_to_write_order(wo, endp, root);
 635 }
 636
 637 static struct object_entry **compute_write_order(void)
 638 {
 639         unsigned int i, wo_end, last_untagged;
 640
 641         struct object_entry **wo;
 642         struct object_entry *objects = to_pack.objects;
 643
 644         for (i = 0; i < to_pack.nr_objects; i++) {
 645                 objects[i].tagged = 0;
 646                 objects[i].filled = 0;
 647                 objects[i].delta_child = NULL;
 648                 objects[i].delta_sibling = NULL;
 649         }
 650
 651         /*
 652          * Fully connect delta_child/delta_sibling network.
 653          * Make sure delta_sibling is sorted in the original
 654          * recency order.
 655          */
 656         for (i = to_pack.nr_objects; i > 0;) {
 657                 struct object_entry *e = &objects[--i];
 658                 if (!e->delta)
 659                         continue;
 660                 /* Mark me as the first child */
 661                 e->delta_sibling = e->delta->delta_child;
 662                 e->delta->delta_child = e;
 663         }
 664
 665         /*
 666          * Mark objects that are at the tip of tags.
 667          */
 668         for_each_tag_ref(mark_tagged, NULL);
 669
 670         /*
 671          * Give the objects in the original recency order until
 672          * we see a tagged tip.
 673          */
 674         ALLOC_ARRAY(wo, to_pack.nr_objects);
 675         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
 676                 if (objects[i].tagged)
 677                         break;
 678                 add_to_write_order(wo, &wo_end, &objects[i]);
 679         }
 680         last_untagged = i;
 681
 682         /*
 683          * Then fill all the tagged tips.
 684          */
 685         for (; i < to_pack.nr_objects; i++) {
 686                 if (objects[i].tagged)
 687                         add_to_write_order(wo, &wo_end, &objects[i]);
 688         }
 689
 690         /*
 691          * And then all remaining commits and tags.
 692          */
 693         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 694                 if (objects[i].type != OBJ_COMMIT &&
 695                     objects[i].type != OBJ_TAG)
 696                         continue;
 697                 add_to_write_order(wo, &wo_end, &objects[i]);
 698         }
 699
 700         /*
 701          * And then all the trees.
 702          */
 703         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 704                 if (objects[i].type != OBJ_TREE)
 705                         continue;
 706                 add_to_write_order(wo, &wo_end, &objects[i]);
 707         }
 708
 709         /*
 710          * Finally all the rest in really tight order
 711          */
 712         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 713                 if (!objects[i].filled)
 714                         add_family_to_write_order(wo, &wo_end, &objects[i]);
 715         }
 716
 717         if (wo_end != to_pack.nr_objects)
 718                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
 719
 720         return wo;
 721 }
 722
 723 static off_t write_reused_pack(struct sha1file *f)
 724 {
 725         unsigned char buffer[8192];
 726         off_t to_write, total;
 727         int fd;
 728
 729         if (!is_pack_valid(reuse_packfile))
 730                 die("packfile is invalid: %s", reuse_packfile->pack_name);
 731
 732         fd = git_open(reuse_packfile->pack_name);
 733         if (fd < 0)
 734                 die_errno("unable to open packfile for reuse: %s",
 735                           reuse_packfile->pack_name);
 736
 737         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
 738                 die_errno("unable to seek in reused packfile");
 739
 740         if (reuse_packfile_offset < 0)
 741                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
 742
 743         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
 744
 745         while (to_write) {
 746                 int read_pack = xread(fd, buffer, sizeof(buffer));
 747
 748                 if (read_pack <= 0)
 749                         die_errno("unable to read from reused packfile");
 750
 751                 if (read_pack > to_write)
 752                         read_pack = to_write;
 753
 754                 sha1write(f, buffer, read_pack);
 755                 to_write -= read_pack;
 756
 757                 /*
 758                  * We don't know the actual number of objects written,
 759                  * only how many bytes written, how many bytes total, and
 760                  * how many objects total. So we can fake it by pretending all
 761                  * objects we are writing are the same size. This gives us a
 762                  * smooth progress meter, and at the end it matches the true
 763                  * answer.
 764                  */
 765                 written = reuse_packfile_objects *
 766                                 (((double)(total - to_write)) / total);
 767                 display_progress(progress_state, written);
 768         }
 769
 770         close(fd);
 771         written = reuse_packfile_objects;
 772         display_progress(progress_state, written);
 773         return reuse_packfile_offset - sizeof(struct pack_header);
 774 }
 775
 776 static const char no_split_warning[] = N_(
 777 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
 778 );
 779
 780 static void write_pack_file(void)
 781 {
 782         uint32_t i = 0, j;
 783         struct sha1file *f;
 784         off_t offset;
 785         uint32_t nr_remaining = nr_result;
 786         time_t last_mtime = 0;
 787         struct object_entry **write_order;
 788
 789         if (progress > pack_to_stdout)
 790                 progress_state = start_progress(_("Writing objects"), nr_result);
 791         ALLOC_ARRAY(written_list, to_pack.nr_objects);
 792         write_order = compute_write_order();
 793
 794         do {
 795                 unsigned char sha1[20];
 796                 char *pack_tmp_name = NULL;
 797
 798                 if (pack_to_stdout)
 799                         f = sha1fd_throughput(1, "<stdout>", progress_state);
 800                 else
 801                         f = create_tmp_packfile(&pack_tmp_name);
 802
 803                 offset = write_pack_header(f, nr_remaining);
 804
 805                 if (reuse_packfile) {
 806                         off_t packfile_size;
 807                         assert(pack_to_stdout);
 808
 809                         packfile_size = write_reused_pack(f);
 810                         offset += packfile_size;
 811                 }
 812
 813                 nr_written = 0;
 814                 for (; i < to_pack.nr_objects; i++) {
 815                         struct object_entry *e = write_order[i];
 816                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 817                                 break;
 818                         display_progress(progress_state, written);
 819                 }
 820
 821                 /*
 822                  * Did we write the wrong # entries in the header?
 823                  * If so, rewrite it like in fast-import
 824                  */
 825                 if (pack_to_stdout) {
 826                         sha1close(f, sha1, CSUM_CLOSE);
 827                 } else if (nr_written == nr_remaining) {
 828                         sha1close(f, sha1, CSUM_FSYNC);
 829                 } else {
 830                         int fd = sha1close(f, sha1, 0);
 831                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
 832                                                  nr_written, sha1, offset);
 833                         close(fd);
 834                         if (write_bitmap_index) {
 835                                 warning(_(no_split_warning));
 836                                 write_bitmap_index = 0;
 837                         }
 838                 }
 839
 840                 if (!pack_to_stdout) {
 841                         struct stat st;
 842                         struct strbuf tmpname = STRBUF_INIT;
 843
 844                         /*
 845                          * Packs are runtime accessed in their mtime
 846                          * order since newer packs are more likely to contain
 847                          * younger objects.  So if we are creating multiple
 848                          * packs then we should modify the mtime of later ones
 849                          * to preserve this property.
 850                          */
 851                         if (stat(pack_tmp_name, &st) < 0) {
 852                                 warning_errno("failed to stat %s", pack_tmp_name);
 853                         } else if (!last_mtime) {
 854                                 last_mtime = st.st_mtime;
 855                         } else {
 856                                 struct utimbuf utb;
 857                                 utb.actime = st.st_atime;
 858                                 utb.modtime = --last_mtime;
 859                                 if (utime(pack_tmp_name, &utb) < 0)
 860                                         warning_errno("failed utime() on %s", pack_tmp_name);
 861                         }
 862
 863                         strbuf_addf(&tmpname, "%s-", base_name);
 864
 865                         if (write_bitmap_index) {
 866                                 bitmap_writer_set_checksum(sha1);
 867                                 bitmap_writer_build_type_index(written_list, nr_written);
 868                         }
 869
 870                         finish_tmp_packfile(&tmpname, pack_tmp_name,
 871                                             written_list, nr_written,
 872                                             &pack_idx_opts, sha1);
 873
 874                         if (write_bitmap_index) {
 875                                 strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
 876
 877                                 stop_progress(&progress_state);
 878
 879                                 bitmap_writer_show_progress(progress);
 880                                 bitmap_writer_reuse_bitmaps(&to_pack);
 881                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
 882                                 bitmap_writer_build(&to_pack);
 883                                 bitmap_writer_finish(written_list, nr_written,
 884                                                      tmpname.buf, write_bitmap_options);
 885                                 write_bitmap_index = 0;
 886                         }
 887
 888                         strbuf_release(&tmpname);
 889                         free(pack_tmp_name);
 890                         puts(sha1_to_hex(sha1));
 891                 }
 892
 893                 /* mark written objects as written to previous pack */
 894                 for (j = 0; j < nr_written; j++) {
 895                         written_list[j]->offset = (off_t)-1;
 896                 }
 897                 nr_remaining -= nr_written;
 898         } while (nr_remaining && i < to_pack.nr_objects);
 899
 900         free(written_list);
 901         free(write_order);
 902         stop_progress(&progress_state);
 903         if (written != nr_result)
 904                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
 905                         written, nr_result);
 906 }
 907
 908 static int no_try_delta(const char *path)
 909 {
 910         static struct attr_check *check;
 911
 912         if (!check)
 913                 check = attr_check_initl("delta", NULL);
 914         if (git_check_attr(path, check))
 915                 return 0;
 916         if (ATTR_FALSE(check->items[0].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                                               oid_to_hex(&entry->idx.oid));
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                                       oid_to_hex(&entry->idx.oid));
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.oid.hash, &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 oidcmp(&a->idx.oid, &b->idx.oid);
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.oid.hash,
1571                                                &entry->size);
1572         }
1573 }
1574
1575 /*
1576  * Follow the chain of deltas from this entry onward, throwing away any links
1577  * that cause us to hit a cycle (as determined by the DFS state flags in
1578  * the entries).
1579  *
1580  * We also detect too-long reused chains that would violate our --depth
1581  * limit.
1582  */
1583 static void break_delta_chains(struct object_entry *entry)
1584 {
1585         /*
1586          * The actual depth of each object we will write is stored as an int,
1587          * as it cannot exceed our int "depth" limit. But before we break
1588          * changes based no that limit, we may potentially go as deep as the
1589          * number of objects, which is elsewhere bounded to a uint32_t.
1590          */
1591         uint32_t total_depth;
1592         struct object_entry *cur, *next;
1593
1594         for (cur = entry, total_depth = 0;
1595              cur;
1596              cur = cur->delta, total_depth++) {
1597                 if (cur->dfs_state == DFS_DONE) {
1598                         /*
1599                          * We've already seen this object and know it isn't
1600                          * part of a cycle. We do need to append its depth
1601                          * to our count.
1602                          */
1603                         total_depth += cur->depth;
1604                         break;
1605                 }
1606
1607                 /*
1608                  * We break cycles before looping, so an ACTIVE state (or any
1609                  * other cruft which made its way into the state variable)
1610                  * is a bug.
1611                  */
1612                 if (cur->dfs_state != DFS_NONE)
1613                         die("BUG: confusing delta dfs state in first pass: %d",
1614                             cur->dfs_state);
1615
1616                 /*
1617                  * Now we know this is the first time we've seen the object. If
1618                  * it's not a delta, we're done traversing, but we'll mark it
1619                  * done to save time on future traversals.
1620                  */
1621                 if (!cur->delta) {
1622                         cur->dfs_state = DFS_DONE;
1623                         break;
1624                 }
1625
1626                 /*
1627                  * Mark ourselves as active and see if the next step causes
1628                  * us to cycle to another active object. It's important to do
1629                  * this _before_ we loop, because it impacts where we make the
1630                  * cut, and thus how our total_depth counter works.
1631                  * E.g., We may see a partial loop like:
1632                  *
1633                  *   A -> B -> C -> D -> B
1634                  *
1635                  * Cutting B->C breaks the cycle. But now the depth of A is
1636                  * only 1, and our total_depth counter is at 3. The size of the
1637                  * error is always one less than the size of the cycle we
1638                  * broke. Commits C and D were "lost" from A's chain.
1639                  *
1640                  * If we instead cut D->B, then the depth of A is correct at 3.
1641                  * We keep all commits in the chain that we examined.
1642                  */
1643                 cur->dfs_state = DFS_ACTIVE;
1644                 if (cur->delta->dfs_state == DFS_ACTIVE) {
1645                         drop_reused_delta(cur);
1646                         cur->dfs_state = DFS_DONE;
1647                         break;
1648                 }
1649         }
1650
1651         /*
1652          * And now that we've gone all the way to the bottom of the chain, we
1653          * need to clear the active flags and set the depth fields as
1654          * appropriate. Unlike the loop above, which can quit when it drops a
1655          * delta, we need to keep going to look for more depth cuts. So we need
1656          * an extra "next" pointer to keep going after we reset cur->delta.
1657          */
1658         for (cur = entry; cur; cur = next) {
1659                 next = cur->delta;
1660
1661                 /*
1662                  * We should have a chain of zero or more ACTIVE states down to
1663                  * a final DONE. We can quit after the DONE, because either it
1664                  * has no bases, or we've already handled them in a previous
1665                  * call.
1666                  */
1667                 if (cur->dfs_state == DFS_DONE)
1668                         break;
1669                 else if (cur->dfs_state != DFS_ACTIVE)
1670                         die("BUG: confusing delta dfs state in second pass: %d",
1671                             cur->dfs_state);
1672
1673                 /*
1674                  * If the total_depth is more than depth, then we need to snip
1675                  * the chain into two or more smaller chains that don't exceed
1676                  * the maximum depth. Most of the resulting chains will contain
1677                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1678                  * the last chain (i.e., the one containing entry) will contain
1679                  * whatever entries are left over, namely
1680                  * (total_depth % (depth + 1)) of them.
1681                  *
1682                  * Since we are iterating towards decreasing depth, we need to
1683                  * decrement total_depth as we go, and we need to write to the
1684                  * entry what its final depth will be after all of the
1685                  * snipping. Since we're snipping into chains of length (depth
1686                  * + 1) entries, the final depth of an entry will be its
1687                  * original depth modulo (depth + 1). Any time we encounter an
1688                  * entry whose final depth is supposed to be zero, we snip it
1689                  * from its delta base, thereby making it so.
1690                  */
1691                 cur->depth = (total_depth--) % (depth + 1);
1692                 if (!cur->depth)
1693                         drop_reused_delta(cur);
1694
1695                 cur->dfs_state = DFS_DONE;
1696         }
1697 }
1698
1699 static void get_object_details(void)
1700 {
1701         uint32_t i;
1702         struct object_entry **sorted_by_offset;
1703
1704         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1705         for (i = 0; i < to_pack.nr_objects; i++)
1706                 sorted_by_offset[i] = to_pack.objects + i;
1707         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1708
1709         for (i = 0; i < to_pack.nr_objects; i++) {
1710                 struct object_entry *entry = sorted_by_offset[i];
1711                 check_object(entry);
1712                 if (big_file_threshold < entry->size)
1713                         entry->no_try_delta = 1;
1714         }
1715
1716         /*
1717          * This must happen in a second pass, since we rely on the delta
1718          * information for the whole list being completed.
1719          */
1720         for (i = 0; i < to_pack.nr_objects; i++)
1721                 break_delta_chains(&to_pack.objects[i]);
1722
1723         free(sorted_by_offset);
1724 }
1725
1726 /*
1727  * We search for deltas in a list sorted by type, by filename hash, and then
1728  * by size, so that we see progressively smaller and smaller files.
1729  * That's because we prefer deltas to be from the bigger file
1730  * to the smaller -- deletes are potentially cheaper, but perhaps
1731  * more importantly, the bigger file is likely the more recent
1732  * one.  The deepest deltas are therefore the oldest objects which are
1733  * less susceptible to be accessed often.
1734  */
1735 static int type_size_sort(const void *_a, const void *_b)
1736 {
1737         const struct object_entry *a = *(struct object_entry **)_a;
1738         const struct object_entry *b = *(struct object_entry **)_b;
1739
1740         if (a->type > b->type)
1741                 return -1;
1742         if (a->type < b->type)
1743                 return 1;
1744         if (a->hash > b->hash)
1745                 return -1;
1746         if (a->hash < b->hash)
1747                 return 1;
1748         if (a->preferred_base > b->preferred_base)
1749                 return -1;
1750         if (a->preferred_base < b->preferred_base)
1751                 return 1;
1752         if (a->size > b->size)
1753                 return -1;
1754         if (a->size < b->size)
1755                 return 1;
1756         return a < b ? -1 : (a > b);  /* newest first */
1757 }
1758
1759 struct unpacked {
1760         struct object_entry *entry;
1761         void *data;
1762         struct delta_index *index;
1763         unsigned depth;
1764 };
1765
1766 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1767                            unsigned long delta_size)
1768 {
1769         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1770                 return 0;
1771
1772         if (delta_size < cache_max_small_delta_size)
1773                 return 1;
1774
1775         /* cache delta, if objects are large enough compared to delta size */
1776         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1777                 return 1;
1778
1779         return 0;
1780 }
1781
1782 #ifndef NO_PTHREADS
1783
1784 static pthread_mutex_t read_mutex;
1785 #define read_lock()             pthread_mutex_lock(&read_mutex)
1786 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1787
1788 static pthread_mutex_t cache_mutex;
1789 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1790 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1791
1792 static pthread_mutex_t progress_mutex;
1793 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1794 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1795
1796 #else
1797
1798 #define read_lock()             (void)0
1799 #define read_unlock()           (void)0
1800 #define cache_lock()            (void)0
1801 #define cache_unlock()          (void)0
1802 #define progress_lock()         (void)0
1803 #define progress_unlock()       (void)0
1804
1805 #endif
1806
1807 static int try_delta(struct unpacked *trg, struct unpacked *src,
1808                      unsigned max_depth, unsigned long *mem_usage)
1809 {
1810         struct object_entry *trg_entry = trg->entry;
1811         struct object_entry *src_entry = src->entry;
1812         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1813         unsigned ref_depth;
1814         enum object_type type;
1815         void *delta_buf;
1816
1817         /* Don't bother doing diffs between different types */
1818         if (trg_entry->type != src_entry->type)
1819                 return -1;
1820
1821         /*
1822          * We do not bother to try a delta that we discarded on an
1823          * earlier try, but only when reusing delta data.  Note that
1824          * src_entry that is marked as the preferred_base should always
1825          * be considered, as even if we produce a suboptimal delta against
1826          * it, we will still save the transfer cost, as we already know
1827          * the other side has it and we won't send src_entry at all.
1828          */
1829         if (reuse_delta && trg_entry->in_pack &&
1830             trg_entry->in_pack == src_entry->in_pack &&
1831             !src_entry->preferred_base &&
1832             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1833             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1834                 return 0;
1835
1836         /* Let's not bust the allowed depth. */
1837         if (src->depth >= max_depth)
1838                 return 0;
1839
1840         /* Now some size filtering heuristics. */
1841         trg_size = trg_entry->size;
1842         if (!trg_entry->delta) {
1843                 max_size = trg_size/2 - 20;
1844                 ref_depth = 1;
1845         } else {
1846                 max_size = trg_entry->delta_size;
1847                 ref_depth = trg->depth;
1848         }
1849         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1850                                                 (max_depth - ref_depth + 1);
1851         if (max_size == 0)
1852                 return 0;
1853         src_size = src_entry->size;
1854         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1855         if (sizediff >= max_size)
1856                 return 0;
1857         if (trg_size < src_size / 32)
1858                 return 0;
1859
1860         /* Load data if not already done */
1861         if (!trg->data) {
1862                 read_lock();
1863                 trg->data = read_sha1_file(trg_entry->idx.oid.hash, &type,
1864                                            &sz);
1865                 read_unlock();
1866                 if (!trg->data)
1867                         die("object %s cannot be read",
1868                             oid_to_hex(&trg_entry->idx.oid));
1869                 if (sz != trg_size)
1870                         die("object %s inconsistent object length (%lu vs %lu)",
1871                             oid_to_hex(&trg_entry->idx.oid), sz,
1872                             trg_size);
1873                 *mem_usage += sz;
1874         }
1875         if (!src->data) {
1876                 read_lock();
1877                 src->data = read_sha1_file(src_entry->idx.oid.hash, &type,
1878                                            &sz);
1879                 read_unlock();
1880                 if (!src->data) {
1881                         if (src_entry->preferred_base) {
1882                                 static int warned = 0;
1883                                 if (!warned++)
1884                                         warning("object %s cannot be read",
1885                                                 oid_to_hex(&src_entry->idx.oid));
1886                                 /*
1887                                  * Those objects are not included in the
1888                                  * resulting pack.  Be resilient and ignore
1889                                  * them if they can't be read, in case the
1890                                  * pack could be created nevertheless.
1891                                  */
1892                                 return 0;
1893                         }
1894                         die("object %s cannot be read",
1895                             oid_to_hex(&src_entry->idx.oid));
1896                 }
1897                 if (sz != src_size)
1898                         die("object %s inconsistent object length (%lu vs %lu)",
1899                             oid_to_hex(&src_entry->idx.oid), sz,
1900                             src_size);
1901                 *mem_usage += sz;
1902         }
1903         if (!src->index) {
1904                 src->index = create_delta_index(src->data, src_size);
1905                 if (!src->index) {
1906                         static int warned = 0;
1907                         if (!warned++)
1908                                 warning("suboptimal pack - out of memory");
1909                         return 0;
1910                 }
1911                 *mem_usage += sizeof_delta_index(src->index);
1912         }
1913
1914         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1915         if (!delta_buf)
1916                 return 0;
1917
1918         if (trg_entry->delta) {
1919                 /* Prefer only shallower same-sized deltas. */
1920                 if (delta_size == trg_entry->delta_size &&
1921                     src->depth + 1 >= trg->depth) {
1922                         free(delta_buf);
1923                         return 0;
1924                 }
1925         }
1926
1927         /*
1928          * Handle memory allocation outside of the cache
1929          * accounting lock.  Compiler will optimize the strangeness
1930          * away when NO_PTHREADS is defined.
1931          */
1932         free(trg_entry->delta_data);
1933         cache_lock();
1934         if (trg_entry->delta_data) {
1935                 delta_cache_size -= trg_entry->delta_size;
1936                 trg_entry->delta_data = NULL;
1937         }
1938         if (delta_cacheable(src_size, trg_size, delta_size)) {
1939                 delta_cache_size += delta_size;
1940                 cache_unlock();
1941                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1942         } else {
1943                 cache_unlock();
1944                 free(delta_buf);
1945         }
1946
1947         trg_entry->delta = src_entry;
1948         trg_entry->delta_size = delta_size;
1949         trg->depth = src->depth + 1;
1950
1951         return 1;
1952 }
1953
1954 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1955 {
1956         struct object_entry *child = me->delta_child;
1957         unsigned int m = n;
1958         while (child) {
1959                 unsigned int c = check_delta_limit(child, n + 1);
1960                 if (m < c)
1961                         m = c;
1962                 child = child->delta_sibling;
1963         }
1964         return m;
1965 }
1966
1967 static unsigned long free_unpacked(struct unpacked *n)
1968 {
1969         unsigned long freed_mem = sizeof_delta_index(n->index);
1970         free_delta_index(n->index);
1971         n->index = NULL;
1972         if (n->data) {
1973                 freed_mem += n->entry->size;
1974                 free(n->data);
1975                 n->data = NULL;
1976         }
1977         n->entry = NULL;
1978         n->depth = 0;
1979         return freed_mem;
1980 }
1981
1982 static void find_deltas(struct object_entry **list, unsigned *list_size,
1983                         int window, int depth, unsigned *processed)
1984 {
1985         uint32_t i, idx = 0, count = 0;
1986         struct unpacked *array;
1987         unsigned long mem_usage = 0;
1988
1989         array = xcalloc(window, sizeof(struct unpacked));
1990
1991         for (;;) {
1992                 struct object_entry *entry;
1993                 struct unpacked *n = array + idx;
1994                 int j, max_depth, best_base = -1;
1995
1996                 progress_lock();
1997                 if (!*list_size) {
1998                         progress_unlock();
1999                         break;
2000                 }
2001                 entry = *list++;
2002                 (*list_size)--;
2003                 if (!entry->preferred_base) {
2004                         (*processed)++;
2005                         display_progress(progress_state, *processed);
2006                 }
2007                 progress_unlock();
2008
2009                 mem_usage -= free_unpacked(n);
2010                 n->entry = entry;
2011
2012                 while (window_memory_limit &&
2013                        mem_usage > window_memory_limit &&
2014                        count > 1) {
2015                         uint32_t tail = (idx + window - count) % window;
2016                         mem_usage -= free_unpacked(array + tail);
2017                         count--;
2018                 }
2019
2020                 /* We do not compute delta to *create* objects we are not
2021                  * going to pack.
2022                  */
2023                 if (entry->preferred_base)
2024                         goto next;
2025
2026                 /*
2027                  * If the current object is at pack edge, take the depth the
2028                  * objects that depend on the current object into account
2029                  * otherwise they would become too deep.
2030                  */
2031                 max_depth = depth;
2032                 if (entry->delta_child) {
2033                         max_depth -= check_delta_limit(entry, 0);
2034                         if (max_depth <= 0)
2035                                 goto next;
2036                 }
2037
2038                 j = window;
2039                 while (--j > 0) {
2040                         int ret;
2041                         uint32_t other_idx = idx + j;
2042                         struct unpacked *m;
2043                         if (other_idx >= window)
2044                                 other_idx -= window;
2045                         m = array + other_idx;
2046                         if (!m->entry)
2047                                 break;
2048                         ret = try_delta(n, m, max_depth, &mem_usage);
2049                         if (ret < 0)
2050                                 break;
2051                         else if (ret > 0)
2052                                 best_base = other_idx;
2053                 }
2054
2055                 /*
2056                  * If we decided to cache the delta data, then it is best
2057                  * to compress it right away.  First because we have to do
2058                  * it anyway, and doing it here while we're threaded will
2059                  * save a lot of time in the non threaded write phase,
2060                  * as well as allow for caching more deltas within
2061                  * the same cache size limit.
2062                  * ...
2063                  * But only if not writing to stdout, since in that case
2064                  * the network is most likely throttling writes anyway,
2065                  * and therefore it is best to go to the write phase ASAP
2066                  * instead, as we can afford spending more time compressing
2067                  * between writes at that moment.
2068                  */
2069                 if (entry->delta_data && !pack_to_stdout) {
2070                         entry->z_delta_size = do_compress(&entry->delta_data,
2071                                                           entry->delta_size);
2072                         cache_lock();
2073                         delta_cache_size -= entry->delta_size;
2074                         delta_cache_size += entry->z_delta_size;
2075                         cache_unlock();
2076                 }
2077
2078                 /* if we made n a delta, and if n is already at max
2079                  * depth, leaving it in the window is pointless.  we
2080                  * should evict it first.
2081                  */
2082                 if (entry->delta && max_depth <= n->depth)
2083                         continue;
2084
2085                 /*
2086                  * Move the best delta base up in the window, after the
2087                  * currently deltified object, to keep it longer.  It will
2088                  * be the first base object to be attempted next.
2089                  */
2090                 if (entry->delta) {
2091                         struct unpacked swap = array[best_base];
2092                         int dist = (window + idx - best_base) % window;
2093                         int dst = best_base;
2094                         while (dist--) {
2095                                 int src = (dst + 1) % window;
2096                                 array[dst] = array[src];
2097                                 dst = src;
2098                         }
2099                         array[dst] = swap;
2100                 }
2101
2102                 next:
2103                 idx++;
2104                 if (count + 1 < window)
2105                         count++;
2106                 if (idx >= window)
2107                         idx = 0;
2108         }
2109
2110         for (i = 0; i < window; ++i) {
2111                 free_delta_index(array[i].index);
2112                 free(array[i].data);
2113         }
2114         free(array);
2115 }
2116
2117 #ifndef NO_PTHREADS
2118
2119 static void try_to_free_from_threads(size_t size)
2120 {
2121         read_lock();
2122         release_pack_memory(size);
2123         read_unlock();
2124 }
2125
2126 static try_to_free_t old_try_to_free_routine;
2127
2128 /*
2129  * The main thread waits on the condition that (at least) one of the workers
2130  * has stopped working (which is indicated in the .working member of
2131  * struct thread_params).
2132  * When a work thread has completed its work, it sets .working to 0 and
2133  * signals the main thread and waits on the condition that .data_ready
2134  * becomes 1.
2135  */
2136
2137 struct thread_params {
2138         pthread_t thread;
2139         struct object_entry **list;
2140         unsigned list_size;
2141         unsigned remaining;
2142         int window;
2143         int depth;
2144         int working;
2145         int data_ready;
2146         pthread_mutex_t mutex;
2147         pthread_cond_t cond;
2148         unsigned *processed;
2149 };
2150
2151 static pthread_cond_t progress_cond;
2152
2153 /*
2154  * Mutex and conditional variable can't be statically-initialized on Windows.
2155  */
2156 static void init_threaded_search(void)
2157 {
2158         init_recursive_mutex(&read_mutex);
2159         pthread_mutex_init(&cache_mutex, NULL);
2160         pthread_mutex_init(&progress_mutex, NULL);
2161         pthread_cond_init(&progress_cond, NULL);
2162         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2163 }
2164
2165 static void cleanup_threaded_search(void)
2166 {
2167         set_try_to_free_routine(old_try_to_free_routine);
2168         pthread_cond_destroy(&progress_cond);
2169         pthread_mutex_destroy(&read_mutex);
2170         pthread_mutex_destroy(&cache_mutex);
2171         pthread_mutex_destroy(&progress_mutex);
2172 }
2173
2174 static void *threaded_find_deltas(void *arg)
2175 {
2176         struct thread_params *me = arg;
2177
2178         while (me->remaining) {
2179                 find_deltas(me->list, &me->remaining,
2180                             me->window, me->depth, me->processed);
2181
2182                 progress_lock();
2183                 me->working = 0;
2184                 pthread_cond_signal(&progress_cond);
2185                 progress_unlock();
2186
2187                 /*
2188                  * We must not set ->data_ready before we wait on the
2189                  * condition because the main thread may have set it to 1
2190                  * before we get here. In order to be sure that new
2191                  * work is available if we see 1 in ->data_ready, it
2192                  * was initialized to 0 before this thread was spawned
2193                  * and we reset it to 0 right away.
2194                  */
2195                 pthread_mutex_lock(&me->mutex);
2196                 while (!me->data_ready)
2197                         pthread_cond_wait(&me->cond, &me->mutex);
2198                 me->data_ready = 0;
2199                 pthread_mutex_unlock(&me->mutex);
2200         }
2201         /* leave ->working 1 so that this doesn't get more work assigned */
2202         return NULL;
2203 }
2204
2205 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2206                            int window, int depth, unsigned *processed)
2207 {
2208         struct thread_params *p;
2209         int i, ret, active_threads = 0;
2210
2211         init_threaded_search();
2212
2213         if (delta_search_threads <= 1) {
2214                 find_deltas(list, &list_size, window, depth, processed);
2215                 cleanup_threaded_search();
2216                 return;
2217         }
2218         if (progress > pack_to_stdout)
2219                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2220                                 delta_search_threads);
2221         p = xcalloc(delta_search_threads, sizeof(*p));
2222
2223         /* Partition the work amongst work threads. */
2224         for (i = 0; i < delta_search_threads; i++) {
2225                 unsigned sub_size = list_size / (delta_search_threads - i);
2226
2227                 /* don't use too small segments or no deltas will be found */
2228                 if (sub_size < 2*window && i+1 < delta_search_threads)
2229                         sub_size = 0;
2230
2231                 p[i].window = window;
2232                 p[i].depth = depth;
2233                 p[i].processed = processed;
2234                 p[i].working = 1;
2235                 p[i].data_ready = 0;
2236
2237                 /* try to split chunks on "path" boundaries */
2238                 while (sub_size && sub_size < list_size &&
2239                        list[sub_size]->hash &&
2240                        list[sub_size]->hash == list[sub_size-1]->hash)
2241                         sub_size++;
2242
2243                 p[i].list = list;
2244                 p[i].list_size = sub_size;
2245                 p[i].remaining = sub_size;
2246
2247                 list += sub_size;
2248                 list_size -= sub_size;
2249         }
2250
2251         /* Start work threads. */
2252         for (i = 0; i < delta_search_threads; i++) {
2253                 if (!p[i].list_size)
2254                         continue;
2255                 pthread_mutex_init(&p[i].mutex, NULL);
2256                 pthread_cond_init(&p[i].cond, NULL);
2257                 ret = pthread_create(&p[i].thread, NULL,
2258                                      threaded_find_deltas, &p[i]);
2259                 if (ret)
2260                         die("unable to create thread: %s", strerror(ret));
2261                 active_threads++;
2262         }
2263
2264         /*
2265          * Now let's wait for work completion.  Each time a thread is done
2266          * with its work, we steal half of the remaining work from the
2267          * thread with the largest number of unprocessed objects and give
2268          * it to that newly idle thread.  This ensure good load balancing
2269          * until the remaining object list segments are simply too short
2270          * to be worth splitting anymore.
2271          */
2272         while (active_threads) {
2273                 struct thread_params *target = NULL;
2274                 struct thread_params *victim = NULL;
2275                 unsigned sub_size = 0;
2276
2277                 progress_lock();
2278                 for (;;) {
2279                         for (i = 0; !target && i < delta_search_threads; i++)
2280                                 if (!p[i].working)
2281                                         target = &p[i];
2282                         if (target)
2283                                 break;
2284                         pthread_cond_wait(&progress_cond, &progress_mutex);
2285                 }
2286
2287                 for (i = 0; i < delta_search_threads; i++)
2288                         if (p[i].remaining > 2*window &&
2289                             (!victim || victim->remaining < p[i].remaining))
2290                                 victim = &p[i];
2291                 if (victim) {
2292                         sub_size = victim->remaining / 2;
2293                         list = victim->list + victim->list_size - sub_size;
2294                         while (sub_size && list[0]->hash &&
2295                                list[0]->hash == list[-1]->hash) {
2296                                 list++;
2297                                 sub_size--;
2298                         }
2299                         if (!sub_size) {
2300                                 /*
2301                                  * It is possible for some "paths" to have
2302                                  * so many objects that no hash boundary
2303                                  * might be found.  Let's just steal the
2304                                  * exact half in that case.
2305                                  */
2306                                 sub_size = victim->remaining / 2;
2307                                 list -= sub_size;
2308                         }
2309                         target->list = list;
2310                         victim->list_size -= sub_size;
2311                         victim->remaining -= sub_size;
2312                 }
2313                 target->list_size = sub_size;
2314                 target->remaining = sub_size;
2315                 target->working = 1;
2316                 progress_unlock();
2317
2318                 pthread_mutex_lock(&target->mutex);
2319                 target->data_ready = 1;
2320                 pthread_cond_signal(&target->cond);
2321                 pthread_mutex_unlock(&target->mutex);
2322
2323                 if (!sub_size) {
2324                         pthread_join(target->thread, NULL);
2325                         pthread_cond_destroy(&target->cond);
2326                         pthread_mutex_destroy(&target->mutex);
2327                         active_threads--;
2328                 }
2329         }
2330         cleanup_threaded_search();
2331         free(p);
2332 }
2333
2334 #else
2335 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2336 #endif
2337
2338 static void add_tag_chain(const struct object_id *oid)
2339 {
2340         struct tag *tag;
2341
2342         /*
2343          * We catch duplicates already in add_object_entry(), but we'd
2344          * prefer to do this extra check to avoid having to parse the
2345          * tag at all if we already know that it's being packed (e.g., if
2346          * it was included via bitmaps, we would not have parsed it
2347          * previously).
2348          */
2349         if (packlist_find(&to_pack, oid->hash, NULL))
2350                 return;
2351
2352         tag = lookup_tag(oid);
2353         while (1) {
2354                 if (!tag || parse_tag(tag) || !tag->tagged)
2355                         die("unable to pack objects reachable from tag %s",
2356                             oid_to_hex(oid));
2357
2358                 add_object_entry(tag->object.oid.hash, OBJ_TAG, NULL, 0);
2359
2360                 if (tag->tagged->type != OBJ_TAG)
2361                         return;
2362
2363                 tag = (struct tag *)tag->tagged;
2364         }
2365 }
2366
2367 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2368 {
2369         struct object_id peeled;
2370
2371         if (starts_with(path, "refs/tags/") && /* is a tag? */
2372             !peel_ref(path, peeled.hash)    && /* peelable? */
2373             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2374                 add_tag_chain(oid);
2375         return 0;
2376 }
2377
2378 static void prepare_pack(int window, int depth)
2379 {
2380         struct object_entry **delta_list;
2381         uint32_t i, nr_deltas;
2382         unsigned n;
2383
2384         get_object_details();
2385
2386         /*
2387          * If we're locally repacking then we need to be doubly careful
2388          * from now on in order to make sure no stealth corruption gets
2389          * propagated to the new pack.  Clients receiving streamed packs
2390          * should validate everything they get anyway so no need to incur
2391          * the additional cost here in that case.
2392          */
2393         if (!pack_to_stdout)
2394                 do_check_packed_object_crc = 1;
2395
2396         if (!to_pack.nr_objects || !window || !depth)
2397                 return;
2398
2399         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2400         nr_deltas = n = 0;
2401
2402         for (i = 0; i < to_pack.nr_objects; i++) {
2403                 struct object_entry *entry = to_pack.objects + i;
2404
2405                 if (entry->delta)
2406                         /* This happens if we decided to reuse existing
2407                          * delta from a pack.  "reuse_delta &&" is implied.
2408                          */
2409                         continue;
2410
2411                 if (entry->size < 50)
2412                         continue;
2413
2414                 if (entry->no_try_delta)
2415                         continue;
2416
2417                 if (!entry->preferred_base) {
2418                         nr_deltas++;
2419                         if (entry->type < 0)
2420                                 die("unable to get type of object %s",
2421                                     oid_to_hex(&entry->idx.oid));
2422                 } else {
2423                         if (entry->type < 0) {
2424                                 /*
2425                                  * This object is not found, but we
2426                                  * don't have to include it anyway.
2427                                  */
2428                                 continue;
2429                         }
2430                 }
2431
2432                 delta_list[n++] = entry;
2433         }
2434
2435         if (nr_deltas && n > 1) {
2436                 unsigned nr_done = 0;
2437                 if (progress)
2438                         progress_state = start_progress(_("Compressing objects"),
2439                                                         nr_deltas);
2440                 QSORT(delta_list, n, type_size_sort);
2441                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2442                 stop_progress(&progress_state);
2443                 if (nr_done != nr_deltas)
2444                         die("inconsistency with delta count");
2445         }
2446         free(delta_list);
2447 }
2448
2449 static int git_pack_config(const char *k, const char *v, void *cb)
2450 {
2451         if (!strcmp(k, "pack.window")) {
2452                 window = git_config_int(k, v);
2453                 return 0;
2454         }
2455         if (!strcmp(k, "pack.windowmemory")) {
2456                 window_memory_limit = git_config_ulong(k, v);
2457                 return 0;
2458         }
2459         if (!strcmp(k, "pack.depth")) {
2460                 depth = git_config_int(k, v);
2461                 return 0;
2462         }
2463         if (!strcmp(k, "pack.deltacachesize")) {
2464                 max_delta_cache_size = git_config_int(k, v);
2465                 return 0;
2466         }
2467         if (!strcmp(k, "pack.deltacachelimit")) {
2468                 cache_max_small_delta_size = git_config_int(k, v);
2469                 return 0;
2470         }
2471         if (!strcmp(k, "pack.writebitmaphashcache")) {
2472                 if (git_config_bool(k, v))
2473                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2474                 else
2475                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2476         }
2477         if (!strcmp(k, "pack.usebitmaps")) {
2478                 use_bitmap_index_default = git_config_bool(k, v);
2479                 return 0;
2480         }
2481         if (!strcmp(k, "pack.threads")) {
2482                 delta_search_threads = git_config_int(k, v);
2483                 if (delta_search_threads < 0)
2484                         die("invalid number of threads specified (%d)",
2485                             delta_search_threads);
2486 #ifdef NO_PTHREADS
2487                 if (delta_search_threads != 1) {
2488                         warning("no threads support, ignoring %s", k);
2489                         delta_search_threads = 0;
2490                 }
2491 #endif
2492                 return 0;
2493         }
2494         if (!strcmp(k, "pack.indexversion")) {
2495                 pack_idx_opts.version = git_config_int(k, v);
2496                 if (pack_idx_opts.version > 2)
2497                         die("bad pack.indexversion=%"PRIu32,
2498                             pack_idx_opts.version);
2499                 return 0;
2500         }
2501         return git_default_config(k, v, cb);
2502 }
2503
2504 static void read_object_list_from_stdin(void)
2505 {
2506         char line[40 + 1 + PATH_MAX + 2];
2507         unsigned char sha1[20];
2508
2509         for (;;) {
2510                 if (!fgets(line, sizeof(line), stdin)) {
2511                         if (feof(stdin))
2512                                 break;
2513                         if (!ferror(stdin))
2514                                 die("fgets returned NULL, not EOF, not error!");
2515                         if (errno != EINTR)
2516                                 die_errno("fgets");
2517                         clearerr(stdin);
2518                         continue;
2519                 }
2520                 if (line[0] == '-') {
2521                         if (get_sha1_hex(line+1, sha1))
2522                                 die("expected edge sha1, got garbage:\n %s",
2523                                     line);
2524                         add_preferred_base(sha1);
2525                         continue;
2526                 }
2527                 if (get_sha1_hex(line, sha1))
2528                         die("expected sha1, got garbage:\n %s", line);
2529
2530                 add_preferred_base_object(line+41);
2531                 add_object_entry(sha1, 0, line+41, 0);
2532         }
2533 }
2534
2535 #define OBJECT_ADDED (1u<<20)
2536
2537 static void show_commit(struct commit *commit, void *data)
2538 {
2539         add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2540         commit->object.flags |= OBJECT_ADDED;
2541
2542         if (write_bitmap_index)
2543                 index_commit_for_bitmap(commit);
2544 }
2545
2546 static void show_object(struct object *obj, const char *name, void *data)
2547 {
2548         add_preferred_base_object(name);
2549         add_object_entry(obj->oid.hash, obj->type, name, 0);
2550         obj->flags |= OBJECT_ADDED;
2551 }
2552
2553 static void show_edge(struct commit *commit)
2554 {
2555         add_preferred_base(commit->object.oid.hash);
2556 }
2557
2558 struct in_pack_object {
2559         off_t offset;
2560         struct object *object;
2561 };
2562
2563 struct in_pack {
2564         int alloc;
2565         int nr;
2566         struct in_pack_object *array;
2567 };
2568
2569 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2570 {
2571         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2572         in_pack->array[in_pack->nr].object = object;
2573         in_pack->nr++;
2574 }
2575
2576 /*
2577  * Compare the objects in the offset order, in order to emulate the
2578  * "git rev-list --objects" output that produced the pack originally.
2579  */
2580 static int ofscmp(const void *a_, const void *b_)
2581 {
2582         struct in_pack_object *a = (struct in_pack_object *)a_;
2583         struct in_pack_object *b = (struct in_pack_object *)b_;
2584
2585         if (a->offset < b->offset)
2586                 return -1;
2587         else if (a->offset > b->offset)
2588                 return 1;
2589         else
2590                 return oidcmp(&a->object->oid, &b->object->oid);
2591 }
2592
2593 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2594 {
2595         struct packed_git *p;
2596         struct in_pack in_pack;
2597         uint32_t i;
2598
2599         memset(&in_pack, 0, sizeof(in_pack));
2600
2601         for (p = packed_git; p; p = p->next) {
2602                 const unsigned char *sha1;
2603                 struct object *o;
2604
2605                 if (!p->pack_local || p->pack_keep)
2606                         continue;
2607                 if (open_pack_index(p))
2608                         die("cannot open pack index");
2609
2610                 ALLOC_GROW(in_pack.array,
2611                            in_pack.nr + p->num_objects,
2612                            in_pack.alloc);
2613
2614                 for (i = 0; i < p->num_objects; i++) {
2615                         sha1 = nth_packed_object_sha1(p, i);
2616                         o = lookup_unknown_object(sha1);
2617                         if (!(o->flags & OBJECT_ADDED))
2618                                 mark_in_pack_object(o, p, &in_pack);
2619                         o->flags |= OBJECT_ADDED;
2620                 }
2621         }
2622
2623         if (in_pack.nr) {
2624                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2625                 for (i = 0; i < in_pack.nr; i++) {
2626                         struct object *o = in_pack.array[i].object;
2627                         add_object_entry(o->oid.hash, o->type, "", 0);
2628                 }
2629         }
2630         free(in_pack.array);
2631 }
2632
2633 static int add_loose_object(const struct object_id *oid, const char *path,
2634                             void *data)
2635 {
2636         enum object_type type = sha1_object_info(oid->hash, NULL);
2637
2638         if (type < 0) {
2639                 warning("loose object at %s could not be examined", path);
2640                 return 0;
2641         }
2642
2643         add_object_entry(oid->hash, type, "", 0);
2644         return 0;
2645 }
2646
2647 /*
2648  * We actually don't even have to worry about reachability here.
2649  * add_object_entry will weed out duplicates, so we just add every
2650  * loose object we find.
2651  */
2652 static void add_unreachable_loose_objects(void)
2653 {
2654         for_each_loose_file_in_objdir(get_object_directory(),
2655                                       add_loose_object,
2656                                       NULL, NULL, NULL);
2657 }
2658
2659 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2660 {
2661         static struct packed_git *last_found = (void *)1;
2662         struct packed_git *p;
2663
2664         p = (last_found != (void *)1) ? last_found : packed_git;
2665
2666         while (p) {
2667                 if ((!p->pack_local || p->pack_keep) &&
2668                         find_pack_entry_one(sha1, p)) {
2669                         last_found = p;
2670                         return 1;
2671                 }
2672                 if (p == last_found)
2673                         p = packed_git;
2674                 else
2675                         p = p->next;
2676                 if (p == last_found)
2677                         p = p->next;
2678         }
2679         return 0;
2680 }
2681
2682 /*
2683  * Store a list of sha1s that are should not be discarded
2684  * because they are either written too recently, or are
2685  * reachable from another object that was.
2686  *
2687  * This is filled by get_object_list.
2688  */
2689 static struct oid_array recent_objects;
2690
2691 static int loosened_object_can_be_discarded(const struct object_id *oid,
2692                                             timestamp_t mtime)
2693 {
2694         if (!unpack_unreachable_expiration)
2695                 return 0;
2696         if (mtime > unpack_unreachable_expiration)
2697                 return 0;
2698         if (oid_array_lookup(&recent_objects, oid) >= 0)
2699                 return 0;
2700         return 1;
2701 }
2702
2703 static void loosen_unused_packed_objects(struct rev_info *revs)
2704 {
2705         struct packed_git *p;
2706         uint32_t i;
2707         struct object_id oid;
2708
2709         for (p = packed_git; p; p = p->next) {
2710                 if (!p->pack_local || p->pack_keep)
2711                         continue;
2712
2713                 if (open_pack_index(p))
2714                         die("cannot open pack index");
2715
2716                 for (i = 0; i < p->num_objects; i++) {
2717                         nth_packed_object_oid(&oid, p, i);
2718                         if (!packlist_find(&to_pack, oid.hash, NULL) &&
2719                             !has_sha1_pack_kept_or_nonlocal(oid.hash) &&
2720                             !loosened_object_can_be_discarded(&oid, p->mtime))
2721                                 if (force_object_loose(oid.hash, p->mtime))
2722                                         die("unable to force loose object");
2723                 }
2724         }
2725 }
2726
2727 /*
2728  * This tracks any options which pack-reuse code expects to be on, or which a
2729  * reader of the pack might not understand, and which would therefore prevent
2730  * blind reuse of what we have on disk.
2731  */
2732 static int pack_options_allow_reuse(void)
2733 {
2734         return pack_to_stdout &&
2735                allow_ofs_delta &&
2736                !ignore_packed_keep &&
2737                (!local || !have_non_local_packs) &&
2738                !incremental;
2739 }
2740
2741 static int get_object_list_from_bitmap(struct rev_info *revs)
2742 {
2743         if (prepare_bitmap_walk(revs) < 0)
2744                 return -1;
2745
2746         if (pack_options_allow_reuse() &&
2747             !reuse_partial_packfile_from_bitmap(
2748                         &reuse_packfile,
2749                         &reuse_packfile_objects,
2750                         &reuse_packfile_offset)) {
2751                 assert(reuse_packfile_objects);
2752                 nr_result += reuse_packfile_objects;
2753                 display_progress(progress_state, nr_result);
2754         }
2755
2756         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2757         return 0;
2758 }
2759
2760 static void record_recent_object(struct object *obj,
2761                                  const char *name,
2762                                  void *data)
2763 {
2764         oid_array_append(&recent_objects, &obj->oid);
2765 }
2766
2767 static void record_recent_commit(struct commit *commit, void *data)
2768 {
2769         oid_array_append(&recent_objects, &commit->object.oid);
2770 }
2771
2772 static void get_object_list(int ac, const char **av)
2773 {
2774         struct rev_info revs;
2775         char line[1000];
2776         int flags = 0;
2777
2778         init_revisions(&revs, NULL);
2779         save_commit_buffer = 0;
2780         setup_revisions(ac, av, &revs, NULL);
2781
2782         /* make sure shallows are read */
2783         is_repository_shallow();
2784
2785         while (fgets(line, sizeof(line), stdin) != NULL) {
2786                 int len = strlen(line);
2787                 if (len && line[len - 1] == '\n')
2788                         line[--len] = 0;
2789                 if (!len)
2790                         break;
2791                 if (*line == '-') {
2792                         if (!strcmp(line, "--not")) {
2793                                 flags ^= UNINTERESTING;
2794                                 write_bitmap_index = 0;
2795                                 continue;
2796                         }
2797                         if (starts_with(line, "--shallow ")) {
2798                                 struct object_id oid;
2799                                 if (get_oid_hex(line + 10, &oid))
2800                                         die("not an SHA-1 '%s'", line + 10);
2801                                 register_shallow(&oid);
2802                                 use_bitmap_index = 0;
2803                                 continue;
2804                         }
2805                         die("not a rev '%s'", line);
2806                 }
2807                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2808                         die("bad revision '%s'", line);
2809         }
2810
2811         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2812                 return;
2813
2814         if (prepare_revision_walk(&revs))
2815                 die("revision walk setup failed");
2816         mark_edges_uninteresting(&revs, show_edge);
2817         traverse_commit_list(&revs, show_commit, show_object, NULL);
2818
2819         if (unpack_unreachable_expiration) {
2820                 revs.ignore_missing_links = 1;
2821                 if (add_unseen_recent_objects_to_traversal(&revs,
2822                                 unpack_unreachable_expiration))
2823                         die("unable to add recent objects");
2824                 if (prepare_revision_walk(&revs))
2825                         die("revision walk setup failed");
2826                 traverse_commit_list(&revs, record_recent_commit,
2827                                      record_recent_object, NULL);
2828         }
2829
2830         if (keep_unreachable)
2831                 add_objects_in_unpacked_packs(&revs);
2832         if (pack_loose_unreachable)
2833                 add_unreachable_loose_objects();
2834         if (unpack_unreachable)
2835                 loosen_unused_packed_objects(&revs);
2836
2837         oid_array_clear(&recent_objects);
2838 }
2839
2840 static int option_parse_index_version(const struct option *opt,
2841                                       const char *arg, int unset)
2842 {
2843         char *c;
2844         const char *val = arg;
2845         pack_idx_opts.version = strtoul(val, &c, 10);
2846         if (pack_idx_opts.version > 2)
2847                 die(_("unsupported index version %s"), val);
2848         if (*c == ',' && c[1])
2849                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2850         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2851                 die(_("bad index version '%s'"), val);
2852         return 0;
2853 }
2854
2855 static int option_parse_unpack_unreachable(const struct option *opt,
2856                                            const char *arg, int unset)
2857 {
2858         if (unset) {
2859                 unpack_unreachable = 0;
2860                 unpack_unreachable_expiration = 0;
2861         }
2862         else {
2863                 unpack_unreachable = 1;
2864                 if (arg)
2865                         unpack_unreachable_expiration = approxidate(arg);
2866         }
2867         return 0;
2868 }
2869
2870 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2871 {
2872         int use_internal_rev_list = 0;
2873         int thin = 0;
2874         int shallow = 0;
2875         int all_progress_implied = 0;
2876         struct argv_array rp = ARGV_ARRAY_INIT;
2877         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2878         int rev_list_index = 0;
2879         struct option pack_objects_options[] = {
2880                 OPT_SET_INT('q', "quiet", &progress,
2881                             N_("do not show progress meter"), 0),
2882                 OPT_SET_INT(0, "progress", &progress,
2883                             N_("show progress meter"), 1),
2884                 OPT_SET_INT(0, "all-progress", &progress,
2885                             N_("show progress meter during object writing phase"), 2),
2886                 OPT_BOOL(0, "all-progress-implied",
2887                          &all_progress_implied,
2888                          N_("similar to --all-progress when progress meter is shown")),
2889                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2890                   N_("write the pack index file in the specified idx format version"),
2891                   0, option_parse_index_version },
2892                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2893                               N_("maximum size of each output pack file")),
2894                 OPT_BOOL(0, "local", &local,
2895                          N_("ignore borrowed objects from alternate object store")),
2896                 OPT_BOOL(0, "incremental", &incremental,
2897                          N_("ignore packed objects")),
2898                 OPT_INTEGER(0, "window", &window,
2899                             N_("limit pack window by objects")),
2900                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2901                               N_("limit pack window by memory in addition to object limit")),
2902                 OPT_INTEGER(0, "depth", &depth,
2903                             N_("maximum length of delta chain allowed in the resulting pack")),
2904                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2905                          N_("reuse existing deltas")),
2906                 OPT_BOOL(0, "reuse-object", &reuse_object,
2907                          N_("reuse existing objects")),
2908                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2909                          N_("use OFS_DELTA objects")),
2910                 OPT_INTEGER(0, "threads", &delta_search_threads,
2911                             N_("use threads when searching for best delta matches")),
2912                 OPT_BOOL(0, "non-empty", &non_empty,
2913                          N_("do not create an empty pack output")),
2914                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2915                          N_("read revision arguments from standard input")),
2916                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2917                   N_("limit the objects to those that are not yet packed"),
2918                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2919                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2920                   N_("include objects reachable from any reference"),
2921                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2922                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2923                   N_("include objects referred by reflog entries"),
2924                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2925                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2926                   N_("include objects referred to by the index"),
2927                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2928                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2929                          N_("output pack to stdout")),
2930                 OPT_BOOL(0, "include-tag", &include_tag,
2931                          N_("include tag objects that refer to objects to be packed")),
2932                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2933                          N_("keep unreachable objects")),
2934                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
2935                          N_("pack loose unreachable objects")),
2936                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2937                   N_("unpack unreachable objects newer than <time>"),
2938                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2939                 OPT_BOOL(0, "thin", &thin,
2940                          N_("create thin packs")),
2941                 OPT_BOOL(0, "shallow", &shallow,
2942                          N_("create packs suitable for shallow fetches")),
2943                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2944                          N_("ignore packs that have companion .keep file")),
2945                 OPT_INTEGER(0, "compression", &pack_compression_level,
2946                             N_("pack compression level")),
2947                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2948                             N_("do not hide commits by grafts"), 0),
2949                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2950                          N_("use a bitmap index if available to speed up counting objects")),
2951                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2952                          N_("write a bitmap index together with the pack index")),
2953                 OPT_END(),
2954         };
2955
2956         check_replace_refs = 0;
2957
2958         reset_pack_idx_option(&pack_idx_opts);
2959         git_config(git_pack_config, NULL);
2960
2961         progress = isatty(2);
2962         argc = parse_options(argc, argv, prefix, pack_objects_options,
2963                              pack_usage, 0);
2964
2965         if (argc) {
2966                 base_name = argv[0];
2967                 argc--;
2968         }
2969         if (pack_to_stdout != !base_name || argc)
2970                 usage_with_options(pack_usage, pack_objects_options);
2971
2972         argv_array_push(&rp, "pack-objects");
2973         if (thin) {
2974                 use_internal_rev_list = 1;
2975                 argv_array_push(&rp, shallow
2976                                 ? "--objects-edge-aggressive"
2977                                 : "--objects-edge");
2978         } else
2979                 argv_array_push(&rp, "--objects");
2980
2981         if (rev_list_all) {
2982                 use_internal_rev_list = 1;
2983                 argv_array_push(&rp, "--all");
2984         }
2985         if (rev_list_reflog) {
2986                 use_internal_rev_list = 1;
2987                 argv_array_push(&rp, "--reflog");
2988         }
2989         if (rev_list_index) {
2990                 use_internal_rev_list = 1;
2991                 argv_array_push(&rp, "--indexed-objects");
2992         }
2993         if (rev_list_unpacked) {
2994                 use_internal_rev_list = 1;
2995                 argv_array_push(&rp, "--unpacked");
2996         }
2997
2998         if (!reuse_object)
2999                 reuse_delta = 0;
3000         if (pack_compression_level == -1)
3001                 pack_compression_level = Z_DEFAULT_COMPRESSION;
3002         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
3003                 die("bad pack compression level %d", pack_compression_level);
3004
3005         if (!delta_search_threads)      /* --threads=0 means autodetect */
3006                 delta_search_threads = online_cpus();
3007
3008 #ifdef NO_PTHREADS
3009         if (delta_search_threads != 1)
3010                 warning("no threads support, ignoring --threads");
3011 #endif
3012         if (!pack_to_stdout && !pack_size_limit)
3013                 pack_size_limit = pack_size_limit_cfg;
3014         if (pack_to_stdout && pack_size_limit)
3015                 die("--max-pack-size cannot be used to build a pack for transfer.");
3016         if (pack_size_limit && pack_size_limit < 1024*1024) {
3017                 warning("minimum pack size limit is 1 MiB");
3018                 pack_size_limit = 1024*1024;
3019         }
3020
3021         if (!pack_to_stdout && thin)
3022                 die("--thin cannot be used to build an indexable pack.");
3023
3024         if (keep_unreachable && unpack_unreachable)
3025                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
3026         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3027                 unpack_unreachable_expiration = 0;
3028
3029         /*
3030          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3031          *
3032          * - to produce good pack (with bitmap index not-yet-packed objects are
3033          *   packed in suboptimal order).
3034          *
3035          * - to use more robust pack-generation codepath (avoiding possible
3036          *   bugs in bitmap code and possible bitmap index corruption).
3037          */
3038         if (!pack_to_stdout)
3039                 use_bitmap_index_default = 0;
3040
3041         if (use_bitmap_index < 0)
3042                 use_bitmap_index = use_bitmap_index_default;
3043
3044         /* "hard" reasons not to use bitmaps; these just won't work at all */
3045         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
3046                 use_bitmap_index = 0;
3047
3048         if (pack_to_stdout || !rev_list_all)
3049                 write_bitmap_index = 0;
3050
3051         if (progress && all_progress_implied)
3052                 progress = 2;
3053
3054         prepare_packed_git();
3055         if (ignore_packed_keep) {
3056                 struct packed_git *p;
3057                 for (p = packed_git; p; p = p->next)
3058                         if (p->pack_local && p->pack_keep)
3059                                 break;
3060                 if (!p) /* no keep-able packs found */
3061                         ignore_packed_keep = 0;
3062         }
3063         if (local) {
3064                 /*
3065                  * unlike ignore_packed_keep above, we do not want to
3066                  * unset "local" based on looking at packs, as it
3067                  * also covers non-local objects
3068                  */
3069                 struct packed_git *p;
3070                 for (p = packed_git; p; p = p->next) {
3071                         if (!p->pack_local) {
3072                                 have_non_local_packs = 1;
3073                                 break;
3074                         }
3075                 }
3076         }
3077
3078         if (progress)
3079                 progress_state = start_progress(_("Counting objects"), 0);
3080         if (!use_internal_rev_list)
3081                 read_object_list_from_stdin();
3082         else {
3083                 get_object_list(rp.argc, rp.argv);
3084                 argv_array_clear(&rp);
3085         }
3086         cleanup_preferred_base();
3087         if (include_tag && nr_result)
3088                 for_each_ref(add_ref_tag, NULL);
3089         stop_progress(&progress_state);
3090
3091         if (non_empty && !nr_result)
3092                 return 0;
3093         if (nr_result)
3094                 prepare_pack(window, depth);
3095         write_pack_file();
3096         if (progress)
3097                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
3098                         " reused %"PRIu32" (delta %"PRIu32")\n",
3099                         written, written_delta, reused, reused_delta);
3100         return 0;
3101 }