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_AND_NULL(entry->delta_data);
 269                 entry->z_delta_size = 0;
 270         } else if (entry->delta_data) {
 271                 size = entry->delta_size;
 272                 buf = entry->delta_data;
 273                 entry->delta_data = NULL;
 274                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 275                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 276         } else {
 277                 buf = get_delta(entry);
 278                 size = entry->delta_size;
 279                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 280                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 281         }
 282
 283         if (st) /* large blob case, just assume we don't compress well */
 284                 datalen = size;
 285         else if (entry->z_delta_size)
 286                 datalen = entry->z_delta_size;
 287         else
 288                 datalen = do_compress(&buf, size);
 289
 290         /*
 291          * The object header is a byte of 'type' followed by zero or
 292          * more bytes of length.
 293          */
 294         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 295                                               type, size);
 296
 297         if (type == OBJ_OFS_DELTA) {
 298                 /*
 299                  * Deltas with relative base contain an additional
 300                  * encoding of the relative offset for the delta
 301                  * base from this object's position in the pack.
 302                  */
 303                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 304                 unsigned pos = sizeof(dheader) - 1;
 305                 dheader[pos] = ofs & 127;
 306                 while (ofs >>= 7)
 307                         dheader[--pos] = 128 | (--ofs & 127);
 308                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 309                         if (st)
 310                                 close_istream(st);
 311                         free(buf);
 312                         return 0;
 313                 }
 314                 sha1write(f, header, hdrlen);
 315                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 316                 hdrlen += sizeof(dheader) - pos;
 317         } else if (type == OBJ_REF_DELTA) {
 318                 /*
 319                  * Deltas with a base reference contain
 320                  * an additional 20 bytes for the base sha1.
 321                  */
 322                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 323                         if (st)
 324                                 close_istream(st);
 325                         free(buf);
 326                         return 0;
 327                 }
 328                 sha1write(f, header, hdrlen);
 329                 sha1write(f, entry->delta->idx.oid.hash, 20);
 330                 hdrlen += 20;
 331         } else {
 332                 if (limit && hdrlen + datalen + 20 >= limit) {
 333                         if (st)
 334                                 close_istream(st);
 335                         free(buf);
 336                         return 0;
 337                 }
 338                 sha1write(f, header, hdrlen);
 339         }
 340         if (st) {
 341                 datalen = write_large_blob_data(st, f, entry->idx.oid.hash);
 342                 close_istream(st);
 343         } else {
 344                 sha1write(f, buf, datalen);
 345                 free(buf);
 346         }
 347
 348         return hdrlen + datalen;
 349 }
 350
 351 /* Return 0 if we will bust the pack-size limit */
 352 static off_t write_reuse_object(struct sha1file *f, struct object_entry *entry,
 353                                 unsigned long limit, int usable_delta)
 354 {
 355         struct packed_git *p = entry->in_pack;
 356         struct pack_window *w_curs = NULL;
 357         struct revindex_entry *revidx;
 358         off_t offset;
 359         enum object_type type = entry->type;
 360         off_t datalen;
 361         unsigned char header[MAX_PACK_OBJECT_HEADER],
 362                       dheader[MAX_PACK_OBJECT_HEADER];
 363         unsigned hdrlen;
 364
 365         if (entry->delta)
 366                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 367                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 368         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 369                                               type, entry->size);
 370
 371         offset = entry->in_pack_offset;
 372         revidx = find_pack_revindex(p, offset);
 373         datalen = revidx[1].offset - offset;
 374         if (!pack_to_stdout && p->index_version > 1 &&
 375             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 376                 error("bad packed object CRC for %s",
 377                       oid_to_hex(&entry->idx.oid));
 378                 unuse_pack(&w_curs);
 379                 return write_no_reuse_object(f, entry, limit, usable_delta);
 380         }
 381
 382         offset += entry->in_pack_header_size;
 383         datalen -= entry->in_pack_header_size;
 384
 385         if (!pack_to_stdout && p->index_version == 1 &&
 386             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 387                 error("corrupt packed object for %s",
 388                       oid_to_hex(&entry->idx.oid));
 389                 unuse_pack(&w_curs);
 390                 return write_no_reuse_object(f, entry, limit, usable_delta);
 391         }
 392
 393         if (type == OBJ_OFS_DELTA) {
 394                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 395                 unsigned pos = sizeof(dheader) - 1;
 396                 dheader[pos] = ofs & 127;
 397                 while (ofs >>= 7)
 398                         dheader[--pos] = 128 | (--ofs & 127);
 399                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 400                         unuse_pack(&w_curs);
 401                         return 0;
 402                 }
 403                 sha1write(f, header, hdrlen);
 404                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 405                 hdrlen += sizeof(dheader) - pos;
 406                 reused_delta++;
 407         } else if (type == OBJ_REF_DELTA) {
 408                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 409                         unuse_pack(&w_curs);
 410                         return 0;
 411                 }
 412                 sha1write(f, header, hdrlen);
 413                 sha1write(f, entry->delta->idx.oid.hash, 20);
 414                 hdrlen += 20;
 415                 reused_delta++;
 416         } else {
 417                 if (limit && hdrlen + datalen + 20 >= limit) {
 418                         unuse_pack(&w_curs);
 419                         return 0;
 420                 }
 421                 sha1write(f, header, hdrlen);
 422         }
 423         copy_pack_data(f, p, &w_curs, offset, datalen);
 424         unuse_pack(&w_curs);
 425         reused++;
 426         return hdrlen + datalen;
 427 }
 428
 429 /* Return 0 if we will bust the pack-size limit */
 430 static off_t write_object(struct sha1file *f,
 431                           struct object_entry *entry,
 432                           off_t write_offset)
 433 {
 434         unsigned long limit;
 435         off_t len;
 436         int usable_delta, to_reuse;
 437
 438         if (!pack_to_stdout)
 439                 crc32_begin(f);
 440
 441         /* apply size limit if limited packsize and not first object */
 442         if (!pack_size_limit || !nr_written)
 443                 limit = 0;
 444         else if (pack_size_limit <= write_offset)
 445                 /*
 446                  * the earlier object did not fit the limit; avoid
 447                  * mistaking this with unlimited (i.e. limit = 0).
 448                  */
 449                 limit = 1;
 450         else
 451                 limit = pack_size_limit - write_offset;
 452
 453         if (!entry->delta)
 454                 usable_delta = 0;       /* no delta */
 455         else if (!pack_size_limit)
 456                usable_delta = 1;        /* unlimited packfile */
 457         else if (entry->delta->idx.offset == (off_t)-1)
 458                 usable_delta = 0;       /* base was written to another pack */
 459         else if (entry->delta->idx.offset)
 460                 usable_delta = 1;       /* base already exists in this pack */
 461         else
 462                 usable_delta = 0;       /* base could end up in another pack */
 463
 464         if (!reuse_object)
 465                 to_reuse = 0;   /* explicit */
 466         else if (!entry->in_pack)
 467                 to_reuse = 0;   /* can't reuse what we don't have */
 468         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
 469                                 /* check_object() decided it for us ... */
 470                 to_reuse = usable_delta;
 471                                 /* ... but pack split may override that */
 472         else if (entry->type != entry->in_pack_type)
 473                 to_reuse = 0;   /* pack has delta which is unusable */
 474         else if (entry->delta)
 475                 to_reuse = 0;   /* we want to pack afresh */
 476         else
 477                 to_reuse = 1;   /* we have it in-pack undeltified,
 478                                  * and we do not need to deltify it.
 479                                  */
 480
 481         if (!to_reuse)
 482                 len = write_no_reuse_object(f, entry, limit, usable_delta);
 483         else
 484                 len = write_reuse_object(f, entry, limit, usable_delta);
 485         if (!len)
 486                 return 0;
 487
 488         if (usable_delta)
 489                 written_delta++;
 490         written++;
 491         if (!pack_to_stdout)
 492                 entry->idx.crc32 = crc32_end(f);
 493         return len;
 494 }
 495
 496 enum write_one_status {
 497         WRITE_ONE_SKIP = -1, /* already written */
 498         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 499         WRITE_ONE_WRITTEN = 1, /* normal */
 500         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 501 };
 502
 503 static enum write_one_status write_one(struct sha1file *f,
 504                                        struct object_entry *e,
 505                                        off_t *offset)
 506 {
 507         off_t size;
 508         int recursing;
 509
 510         /*
 511          * we set offset to 1 (which is an impossible value) to mark
 512          * the fact that this object is involved in "write its base
 513          * first before writing a deltified object" recursion.
 514          */
 515         recursing = (e->idx.offset == 1);
 516         if (recursing) {
 517                 warning("recursive delta detected for object %s",
 518                         oid_to_hex(&e->idx.oid));
 519                 return WRITE_ONE_RECURSIVE;
 520         } else if (e->idx.offset || e->preferred_base) {
 521                 /* offset is non zero if object is written already. */
 522                 return WRITE_ONE_SKIP;
 523         }
 524
 525         /* if we are deltified, write out base object first. */
 526         if (e->delta) {
 527                 e->idx.offset = 1; /* now recurse */
 528                 switch (write_one(f, e->delta, offset)) {
 529                 case WRITE_ONE_RECURSIVE:
 530                         /* we cannot depend on this one */
 531                         e->delta = NULL;
 532                         break;
 533                 default:
 534                         break;
 535                 case WRITE_ONE_BREAK:
 536                         e->idx.offset = recursing;
 537                         return WRITE_ONE_BREAK;
 538                 }
 539         }
 540
 541         e->idx.offset = *offset;
 542         size = write_object(f, e, *offset);
 543         if (!size) {
 544                 e->idx.offset = recursing;
 545                 return WRITE_ONE_BREAK;
 546         }
 547         written_list[nr_written++] = &e->idx;
 548
 549         /* make sure off_t is sufficiently large not to wrap */
 550         if (signed_add_overflows(*offset, size))
 551                 die("pack too large for current definition of off_t");
 552         *offset += size;
 553         return WRITE_ONE_WRITTEN;
 554 }
 555
 556 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
 557                        void *cb_data)
 558 {
 559         unsigned char peeled[20];
 560         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
 561
 562         if (entry)
 563                 entry->tagged = 1;
 564         if (!peel_ref(path, peeled)) {
 565                 entry = packlist_find(&to_pack, peeled, NULL);
 566                 if (entry)
 567                         entry->tagged = 1;
 568         }
 569         return 0;
 570 }
 571
 572 static inline void add_to_write_order(struct object_entry **wo,
 573                                unsigned int *endp,
 574                                struct object_entry *e)
 575 {
 576         if (e->filled)
 577                 return;
 578         wo[(*endp)++] = e;
 579         e->filled = 1;
 580 }
 581
 582 static void add_descendants_to_write_order(struct object_entry **wo,
 583                                            unsigned int *endp,
 584                                            struct object_entry *e)
 585 {
 586         int add_to_order = 1;
 587         while (e) {
 588                 if (add_to_order) {
 589                         struct object_entry *s;
 590                         /* add this node... */
 591                         add_to_write_order(wo, endp, e);
 592                         /* all its siblings... */
 593                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
 594                                 add_to_write_order(wo, endp, s);
 595                         }
 596                 }
 597                 /* drop down a level to add left subtree nodes if possible */
 598                 if (e->delta_child) {
 599                         add_to_order = 1;
 600                         e = e->delta_child;
 601                 } else {
 602                         add_to_order = 0;
 603                         /* our sibling might have some children, it is next */
 604                         if (e->delta_sibling) {
 605                                 e = e->delta_sibling;
 606                                 continue;
 607                         }
 608                         /* go back to our parent node */
 609                         e = e->delta;
 610                         while (e && !e->delta_sibling) {
 611                                 /* we're on the right side of a subtree, keep
 612                                  * going up until we can go right again */
 613                                 e = e->delta;
 614                         }
 615                         if (!e) {
 616                                 /* done- we hit our original root node */
 617                                 return;
 618                         }
 619                         /* pass it off to sibling at this level */
 620                         e = e->delta_sibling;
 621                 }
 622         };
 623 }
 624
 625 static void add_family_to_write_order(struct object_entry **wo,
 626                                       unsigned int *endp,
 627                                       struct object_entry *e)
 628 {
 629         struct object_entry *root;
 630
 631         for (root = e; root->delta; root = root->delta)
 632                 ; /* nothing */
 633         add_descendants_to_write_order(wo, endp, root);
 634 }
 635
 636 static struct object_entry **compute_write_order(void)
 637 {
 638         unsigned int i, wo_end, last_untagged;
 639
 640         struct object_entry **wo;
 641         struct object_entry *objects = to_pack.objects;
 642
 643         for (i = 0; i < to_pack.nr_objects; i++) {
 644                 objects[i].tagged = 0;
 645                 objects[i].filled = 0;
 646                 objects[i].delta_child = NULL;
 647                 objects[i].delta_sibling = NULL;
 648         }
 649
 650         /*
 651          * Fully connect delta_child/delta_sibling network.
 652          * Make sure delta_sibling is sorted in the original
 653          * recency order.
 654          */
 655         for (i = to_pack.nr_objects; i > 0;) {
 656                 struct object_entry *e = &objects[--i];
 657                 if (!e->delta)
 658                         continue;
 659                 /* Mark me as the first child */
 660                 e->delta_sibling = e->delta->delta_child;
 661                 e->delta->delta_child = e;
 662         }
 663
 664         /*
 665          * Mark objects that are at the tip of tags.
 666          */
 667         for_each_tag_ref(mark_tagged, NULL);
 668
 669         /*
 670          * Give the objects in the original recency order until
 671          * we see a tagged tip.
 672          */
 673         ALLOC_ARRAY(wo, to_pack.nr_objects);
 674         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
 675                 if (objects[i].tagged)
 676                         break;
 677                 add_to_write_order(wo, &wo_end, &objects[i]);
 678         }
 679         last_untagged = i;
 680
 681         /*
 682          * Then fill all the tagged tips.
 683          */
 684         for (; i < to_pack.nr_objects; i++) {
 685                 if (objects[i].tagged)
 686                         add_to_write_order(wo, &wo_end, &objects[i]);
 687         }
 688
 689         /*
 690          * And then all remaining commits and tags.
 691          */
 692         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 693                 if (objects[i].type != OBJ_COMMIT &&
 694                     objects[i].type != OBJ_TAG)
 695                         continue;
 696                 add_to_write_order(wo, &wo_end, &objects[i]);
 697         }
 698
 699         /*
 700          * And then all the trees.
 701          */
 702         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 703                 if (objects[i].type != OBJ_TREE)
 704                         continue;
 705                 add_to_write_order(wo, &wo_end, &objects[i]);
 706         }
 707
 708         /*
 709          * Finally all the rest in really tight order
 710          */
 711         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 712                 if (!objects[i].filled)
 713                         add_family_to_write_order(wo, &wo_end, &objects[i]);
 714         }
 715
 716         if (wo_end != to_pack.nr_objects)
 717                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
 718
 719         return wo;
 720 }
 721
 722 static off_t write_reused_pack(struct sha1file *f)
 723 {
 724         unsigned char buffer[8192];
 725         off_t to_write, total;
 726         int fd;
 727
 728         if (!is_pack_valid(reuse_packfile))
 729                 die("packfile is invalid: %s", reuse_packfile->pack_name);
 730
 731         fd = git_open(reuse_packfile->pack_name);
 732         if (fd < 0)
 733                 die_errno("unable to open packfile for reuse: %s",
 734                           reuse_packfile->pack_name);
 735
 736         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
 737                 die_errno("unable to seek in reused packfile");
 738
 739         if (reuse_packfile_offset < 0)
 740                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
 741
 742         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
 743
 744         while (to_write) {
 745                 int read_pack = xread(fd, buffer, sizeof(buffer));
 746
 747                 if (read_pack <= 0)
 748                         die_errno("unable to read from reused packfile");
 749
 750                 if (read_pack > to_write)
 751                         read_pack = to_write;
 752
 753                 sha1write(f, buffer, read_pack);
 754                 to_write -= read_pack;
 755
 756                 /*
 757                  * We don't know the actual number of objects written,
 758                  * only how many bytes written, how many bytes total, and
 759                  * how many objects total. So we can fake it by pretending all
 760                  * objects we are writing are the same size. This gives us a
 761                  * smooth progress meter, and at the end it matches the true
 762                  * answer.
 763                  */
 764                 written = reuse_packfile_objects *
 765                                 (((double)(total - to_write)) / total);
 766                 display_progress(progress_state, written);
 767         }
 768
 769         close(fd);
 770         written = reuse_packfile_objects;
 771         display_progress(progress_state, written);
 772         return reuse_packfile_offset - sizeof(struct pack_header);
 773 }
 774
 775 static const char no_split_warning[] = N_(
 776 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
 777 );
 778
 779 static void write_pack_file(void)
 780 {
 781         uint32_t i = 0, j;
 782         struct sha1file *f;
 783         off_t offset;
 784         uint32_t nr_remaining = nr_result;
 785         time_t last_mtime = 0;
 786         struct object_entry **write_order;
 787
 788         if (progress > pack_to_stdout)
 789                 progress_state = start_progress(_("Writing objects"), nr_result);
 790         ALLOC_ARRAY(written_list, to_pack.nr_objects);
 791         write_order = compute_write_order();
 792
 793         do {
 794                 unsigned char sha1[20];
 795                 char *pack_tmp_name = NULL;
 796
 797                 if (pack_to_stdout)
 798                         f = sha1fd_throughput(1, "<stdout>", progress_state);
 799                 else
 800                         f = create_tmp_packfile(&pack_tmp_name);
 801
 802                 offset = write_pack_header(f, nr_remaining);
 803
 804                 if (reuse_packfile) {
 805                         off_t packfile_size;
 806                         assert(pack_to_stdout);
 807
 808                         packfile_size = write_reused_pack(f);
 809                         offset += packfile_size;
 810                 }
 811
 812                 nr_written = 0;
 813                 for (; i < to_pack.nr_objects; i++) {
 814                         struct object_entry *e = write_order[i];
 815                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 816                                 break;
 817                         display_progress(progress_state, written);
 818                 }
 819
 820                 /*
 821                  * Did we write the wrong # entries in the header?
 822                  * If so, rewrite it like in fast-import
 823                  */
 824                 if (pack_to_stdout) {
 825                         sha1close(f, sha1, CSUM_CLOSE);
 826                 } else if (nr_written == nr_remaining) {
 827                         sha1close(f, sha1, CSUM_FSYNC);
 828                 } else {
 829                         int fd = sha1close(f, sha1, 0);
 830                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
 831                                                  nr_written, sha1, offset);
 832                         close(fd);
 833                         if (write_bitmap_index) {
 834                                 warning(_(no_split_warning));
 835                                 write_bitmap_index = 0;
 836                         }
 837                 }
 838
 839                 if (!pack_to_stdout) {
 840                         struct stat st;
 841                         struct strbuf tmpname = STRBUF_INIT;
 842
 843                         /*
 844                          * Packs are runtime accessed in their mtime
 845                          * order since newer packs are more likely to contain
 846                          * younger objects.  So if we are creating multiple
 847                          * packs then we should modify the mtime of later ones
 848                          * to preserve this property.
 849                          */
 850                         if (stat(pack_tmp_name, &st) < 0) {
 851                                 warning_errno("failed to stat %s", pack_tmp_name);
 852                         } else if (!last_mtime) {
 853                                 last_mtime = st.st_mtime;
 854                         } else {
 855                                 struct utimbuf utb;
 856                                 utb.actime = st.st_atime;
 857                                 utb.modtime = --last_mtime;
 858                                 if (utime(pack_tmp_name, &utb) < 0)
 859                                         warning_errno("failed utime() on %s", pack_tmp_name);
 860                         }
 861
 862                         strbuf_addf(&tmpname, "%s-", base_name);
 863
 864                         if (write_bitmap_index) {
 865                                 bitmap_writer_set_checksum(sha1);
 866                                 bitmap_writer_build_type_index(written_list, nr_written);
 867                         }
 868
 869                         finish_tmp_packfile(&tmpname, pack_tmp_name,
 870                                             written_list, nr_written,
 871                                             &pack_idx_opts, sha1);
 872
 873                         if (write_bitmap_index) {
 874                                 strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
 875
 876                                 stop_progress(&progress_state);
 877
 878                                 bitmap_writer_show_progress(progress);
 879                                 bitmap_writer_reuse_bitmaps(&to_pack);
 880                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
 881                                 bitmap_writer_build(&to_pack);
 882                                 bitmap_writer_finish(written_list, nr_written,
 883                                                      tmpname.buf, write_bitmap_options);
 884                                 write_bitmap_index = 0;
 885                         }
 886
 887                         strbuf_release(&tmpname);
 888                         free(pack_tmp_name);
 889                         puts(sha1_to_hex(sha1));
 890                 }
 891
 892                 /* mark written objects as written to previous pack */
 893                 for (j = 0; j < nr_written; j++) {
 894                         written_list[j]->offset = (off_t)-1;
 895                 }
 896                 nr_remaining -= nr_written;
 897         } while (nr_remaining && i < to_pack.nr_objects);
 898
 899         free(written_list);
 900         free(write_order);
 901         stop_progress(&progress_state);
 902         if (written != nr_result)
 903                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
 904                         written, nr_result);
 905 }
 906
 907 static int no_try_delta(const char *path)
 908 {
 909         static struct attr_check *check;
 910
 911         if (!check)
 912                 check = attr_check_initl("delta", NULL);
 913         if (git_check_attr(path, check))
 914                 return 0;
 915         if (ATTR_FALSE(check->items[0].value))
 916                 return 1;
 917         return 0;
 918 }
 919
 920 /*
 921  * When adding an object, check whether we have already added it
 922  * to our packing list. If so, we can skip. However, if we are
 923  * being asked to excludei t, but the previous mention was to include
 924  * it, make sure to adjust its flags and tweak our numbers accordingly.
 925  *
 926  * As an optimization, we pass out the index position where we would have
 927  * found the item, since that saves us from having to look it up again a
 928  * few lines later when we want to add the new entry.
 929  */
 930 static int have_duplicate_entry(const unsigned char *sha1,
 931                                 int exclude,
 932                                 uint32_t *index_pos)
 933 {
 934         struct object_entry *entry;
 935
 936         entry = packlist_find(&to_pack, sha1, index_pos);
 937         if (!entry)
 938                 return 0;
 939
 940         if (exclude) {
 941                 if (!entry->preferred_base)
 942                         nr_result--;
 943                 entry->preferred_base = 1;
 944         }
 945
 946         return 1;
 947 }
 948
 949 static int want_found_object(int exclude, struct packed_git *p)
 950 {
 951         if (exclude)
 952                 return 1;
 953         if (incremental)
 954                 return 0;
 955
 956         /*
 957          * When asked to do --local (do not include an object that appears in a
 958          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
 959          * an object that appears in a pack marked with .keep), finding a pack
 960          * that matches the criteria is sufficient for us to decide to omit it.
 961          * However, even if this pack does not satisfy the criteria, we need to
 962          * make sure no copy of this object appears in _any_ pack that makes us
 963          * to omit the object, so we need to check all the packs.
 964          *
 965          * We can however first check whether these options can possible matter;
 966          * if they do not matter we know we want the object in generated pack.
 967          * Otherwise, we signal "-1" at the end to tell the caller that we do
 968          * not know either way, and it needs to check more packs.
 969          */
 970         if (!ignore_packed_keep &&
 971             (!local || !have_non_local_packs))
 972                 return 1;
 973
 974         if (local && !p->pack_local)
 975                 return 0;
 976         if (ignore_packed_keep && p->pack_local && p->pack_keep)
 977                 return 0;
 978
 979         /* we don't know yet; keep looking for more packs */
 980         return -1;
 981 }
 982
 983 /*
 984  * Check whether we want the object in the pack (e.g., we do not want
 985  * objects found in non-local stores if the "--local" option was used).
 986  *
 987  * If the caller already knows an existing pack it wants to take the object
 988  * from, that is passed in *found_pack and *found_offset; otherwise this
 989  * function finds if there is any pack that has the object and returns the pack
 990  * and its offset in these variables.
 991  */
 992 static int want_object_in_pack(const unsigned char *sha1,
 993                                int exclude,
 994                                struct packed_git **found_pack,
 995                                off_t *found_offset)
 996 {
 997         struct mru_entry *entry;
 998         int want;
 999
1000         if (!exclude && local && has_loose_object_nonlocal(sha1))
1001                 return 0;
1002
1003         /*
1004          * If we already know the pack object lives in, start checks from that
1005          * pack - in the usual case when neither --local was given nor .keep files
1006          * are present we will determine the answer right now.
1007          */
1008         if (*found_pack) {
1009                 want = want_found_object(exclude, *found_pack);
1010                 if (want != -1)
1011                         return want;
1012         }
1013
1014         for (entry = packed_git_mru->head; entry; entry = entry->next) {
1015                 struct packed_git *p = entry->item;
1016                 off_t offset;
1017
1018                 if (p == *found_pack)
1019                         offset = *found_offset;
1020                 else
1021                         offset = find_pack_entry_one(sha1, p);
1022
1023                 if (offset) {
1024                         if (!*found_pack) {
1025                                 if (!is_pack_valid(p))
1026                                         continue;
1027                                 *found_offset = offset;
1028                                 *found_pack = p;
1029                         }
1030                         want = want_found_object(exclude, p);
1031                         if (!exclude && want > 0)
1032                                 mru_mark(packed_git_mru, entry);
1033                         if (want != -1)
1034                                 return want;
1035                 }
1036         }
1037
1038         return 1;
1039 }
1040
1041 static void create_object_entry(const unsigned char *sha1,
1042                                 enum object_type type,
1043                                 uint32_t hash,
1044                                 int exclude,
1045                                 int no_try_delta,
1046                                 uint32_t index_pos,
1047                                 struct packed_git *found_pack,
1048                                 off_t found_offset)
1049 {
1050         struct object_entry *entry;
1051
1052         entry = packlist_alloc(&to_pack, sha1, index_pos);
1053         entry->hash = hash;
1054         if (type)
1055                 entry->type = type;
1056         if (exclude)
1057                 entry->preferred_base = 1;
1058         else
1059                 nr_result++;
1060         if (found_pack) {
1061                 entry->in_pack = found_pack;
1062                 entry->in_pack_offset = found_offset;
1063         }
1064
1065         entry->no_try_delta = no_try_delta;
1066 }
1067
1068 static const char no_closure_warning[] = N_(
1069 "disabling bitmap writing, as some objects are not being packed"
1070 );
1071
1072 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1073                             const char *name, int exclude)
1074 {
1075         struct packed_git *found_pack = NULL;
1076         off_t found_offset = 0;
1077         uint32_t index_pos;
1078
1079         if (have_duplicate_entry(sha1, exclude, &index_pos))
1080                 return 0;
1081
1082         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1083                 /* The pack is missing an object, so it will not have closure */
1084                 if (write_bitmap_index) {
1085                         warning(_(no_closure_warning));
1086                         write_bitmap_index = 0;
1087                 }
1088                 return 0;
1089         }
1090
1091         create_object_entry(sha1, type, pack_name_hash(name),
1092                             exclude, name && no_try_delta(name),
1093                             index_pos, found_pack, found_offset);
1094
1095         display_progress(progress_state, nr_result);
1096         return 1;
1097 }
1098
1099 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1100                                         enum object_type type,
1101                                         int flags, uint32_t name_hash,
1102                                         struct packed_git *pack, off_t offset)
1103 {
1104         uint32_t index_pos;
1105
1106         if (have_duplicate_entry(sha1, 0, &index_pos))
1107                 return 0;
1108
1109         if (!want_object_in_pack(sha1, 0, &pack, &offset))
1110                 return 0;
1111
1112         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1113
1114         display_progress(progress_state, nr_result);
1115         return 1;
1116 }
1117
1118 struct pbase_tree_cache {
1119         unsigned char sha1[20];
1120         int ref;
1121         int temporary;
1122         void *tree_data;
1123         unsigned long tree_size;
1124 };
1125
1126 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1127 static int pbase_tree_cache_ix(const unsigned char *sha1)
1128 {
1129         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1130 }
1131 static int pbase_tree_cache_ix_incr(int ix)
1132 {
1133         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1134 }
1135
1136 static struct pbase_tree {
1137         struct pbase_tree *next;
1138         /* This is a phony "cache" entry; we are not
1139          * going to evict it or find it through _get()
1140          * mechanism -- this is for the toplevel node that
1141          * would almost always change with any commit.
1142          */
1143         struct pbase_tree_cache pcache;
1144 } *pbase_tree;
1145
1146 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1147 {
1148         struct pbase_tree_cache *ent, *nent;
1149         void *data;
1150         unsigned long size;
1151         enum object_type type;
1152         int neigh;
1153         int my_ix = pbase_tree_cache_ix(sha1);
1154         int available_ix = -1;
1155
1156         /* pbase-tree-cache acts as a limited hashtable.
1157          * your object will be found at your index or within a few
1158          * slots after that slot if it is cached.
1159          */
1160         for (neigh = 0; neigh < 8; neigh++) {
1161                 ent = pbase_tree_cache[my_ix];
1162                 if (ent && !hashcmp(ent->sha1, sha1)) {
1163                         ent->ref++;
1164                         return ent;
1165                 }
1166                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1167                          ((0 <= available_ix) &&
1168                           (!ent && pbase_tree_cache[available_ix])))
1169                         available_ix = my_ix;
1170                 if (!ent)
1171                         break;
1172                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1173         }
1174
1175         /* Did not find one.  Either we got a bogus request or
1176          * we need to read and perhaps cache.
1177          */
1178         data = read_sha1_file(sha1, &type, &size);
1179         if (!data)
1180                 return NULL;
1181         if (type != OBJ_TREE) {
1182                 free(data);
1183                 return NULL;
1184         }
1185
1186         /* We need to either cache or return a throwaway copy */
1187
1188         if (available_ix < 0)
1189                 ent = NULL;
1190         else {
1191                 ent = pbase_tree_cache[available_ix];
1192                 my_ix = available_ix;
1193         }
1194
1195         if (!ent) {
1196                 nent = xmalloc(sizeof(*nent));
1197                 nent->temporary = (available_ix < 0);
1198         }
1199         else {
1200                 /* evict and reuse */
1201                 free(ent->tree_data);
1202                 nent = ent;
1203         }
1204         hashcpy(nent->sha1, sha1);
1205         nent->tree_data = data;
1206         nent->tree_size = size;
1207         nent->ref = 1;
1208         if (!nent->temporary)
1209                 pbase_tree_cache[my_ix] = nent;
1210         return nent;
1211 }
1212
1213 static void pbase_tree_put(struct pbase_tree_cache *cache)
1214 {
1215         if (!cache->temporary) {
1216                 cache->ref--;
1217                 return;
1218         }
1219         free(cache->tree_data);
1220         free(cache);
1221 }
1222
1223 static int name_cmp_len(const char *name)
1224 {
1225         int i;
1226         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1227                 ;
1228         return i;
1229 }
1230
1231 static void add_pbase_object(struct tree_desc *tree,
1232                              const char *name,
1233                              int cmplen,
1234                              const char *fullname)
1235 {
1236         struct name_entry entry;
1237         int cmp;
1238
1239         while (tree_entry(tree,&entry)) {
1240                 if (S_ISGITLINK(entry.mode))
1241                         continue;
1242                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1243                       memcmp(name, entry.path, cmplen);
1244                 if (cmp > 0)
1245                         continue;
1246                 if (cmp < 0)
1247                         return;
1248                 if (name[cmplen] != '/') {
1249                         add_object_entry(entry.oid->hash,
1250                                          object_type(entry.mode),
1251                                          fullname, 1);
1252                         return;
1253                 }
1254                 if (S_ISDIR(entry.mode)) {
1255                         struct tree_desc sub;
1256                         struct pbase_tree_cache *tree;
1257                         const char *down = name+cmplen+1;
1258                         int downlen = name_cmp_len(down);
1259
1260                         tree = pbase_tree_get(entry.oid->hash);
1261                         if (!tree)
1262                                 return;
1263                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1264
1265                         add_pbase_object(&sub, down, downlen, fullname);
1266                         pbase_tree_put(tree);
1267                 }
1268         }
1269 }
1270
1271 static unsigned *done_pbase_paths;
1272 static int done_pbase_paths_num;
1273 static int done_pbase_paths_alloc;
1274 static int done_pbase_path_pos(unsigned hash)
1275 {
1276         int lo = 0;
1277         int hi = done_pbase_paths_num;
1278         while (lo < hi) {
1279                 int mi = (hi + lo) / 2;
1280                 if (done_pbase_paths[mi] == hash)
1281                         return mi;
1282                 if (done_pbase_paths[mi] < hash)
1283                         hi = mi;
1284                 else
1285                         lo = mi + 1;
1286         }
1287         return -lo-1;
1288 }
1289
1290 static int check_pbase_path(unsigned hash)
1291 {
1292         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1293         if (0 <= pos)
1294                 return 1;
1295         pos = -pos - 1;
1296         ALLOC_GROW(done_pbase_paths,
1297                    done_pbase_paths_num + 1,
1298                    done_pbase_paths_alloc);
1299         done_pbase_paths_num++;
1300         if (pos < done_pbase_paths_num)
1301                 memmove(done_pbase_paths + pos + 1,
1302                         done_pbase_paths + pos,
1303                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1304         done_pbase_paths[pos] = hash;
1305         return 0;
1306 }
1307
1308 static void add_preferred_base_object(const char *name)
1309 {
1310         struct pbase_tree *it;
1311         int cmplen;
1312         unsigned hash = pack_name_hash(name);
1313
1314         if (!num_preferred_base || check_pbase_path(hash))
1315                 return;
1316
1317         cmplen = name_cmp_len(name);
1318         for (it = pbase_tree; it; it = it->next) {
1319                 if (cmplen == 0) {
1320                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1321                 }
1322                 else {
1323                         struct tree_desc tree;
1324                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1325                         add_pbase_object(&tree, name, cmplen, name);
1326                 }
1327         }
1328 }
1329
1330 static void add_preferred_base(unsigned char *sha1)
1331 {
1332         struct pbase_tree *it;
1333         void *data;
1334         unsigned long size;
1335         unsigned char tree_sha1[20];
1336
1337         if (window <= num_preferred_base++)
1338                 return;
1339
1340         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1341         if (!data)
1342                 return;
1343
1344         for (it = pbase_tree; it; it = it->next) {
1345                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1346                         free(data);
1347                         return;
1348                 }
1349         }
1350
1351         it = xcalloc(1, sizeof(*it));
1352         it->next = pbase_tree;
1353         pbase_tree = it;
1354
1355         hashcpy(it->pcache.sha1, tree_sha1);
1356         it->pcache.tree_data = data;
1357         it->pcache.tree_size = size;
1358 }
1359
1360 static void cleanup_preferred_base(void)
1361 {
1362         struct pbase_tree *it;
1363         unsigned i;
1364
1365         it = pbase_tree;
1366         pbase_tree = NULL;
1367         while (it) {
1368                 struct pbase_tree *this = it;
1369                 it = this->next;
1370                 free(this->pcache.tree_data);
1371                 free(this);
1372         }
1373
1374         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1375                 if (!pbase_tree_cache[i])
1376                         continue;
1377                 free(pbase_tree_cache[i]->tree_data);
1378                 FREE_AND_NULL(pbase_tree_cache[i]);
1379         }
1380
1381         FREE_AND_NULL(done_pbase_paths);
1382         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1383 }
1384
1385 static void check_object(struct object_entry *entry)
1386 {
1387         if (entry->in_pack) {
1388                 struct packed_git *p = entry->in_pack;
1389                 struct pack_window *w_curs = NULL;
1390                 const unsigned char *base_ref = NULL;
1391                 struct object_entry *base_entry;
1392                 unsigned long used, used_0;
1393                 unsigned long avail;
1394                 off_t ofs;
1395                 unsigned char *buf, c;
1396
1397                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1398
1399                 /*
1400                  * We want in_pack_type even if we do not reuse delta
1401                  * since non-delta representations could still be reused.
1402                  */
1403                 used = unpack_object_header_buffer(buf, avail,
1404                                                    &entry->in_pack_type,
1405                                                    &entry->size);
1406                 if (used == 0)
1407                         goto give_up;
1408
1409                 /*
1410                  * Determine if this is a delta and if so whether we can
1411                  * reuse it or not.  Otherwise let's find out as cheaply as
1412                  * possible what the actual type and size for this object is.
1413                  */
1414                 switch (entry->in_pack_type) {
1415                 default:
1416                         /* Not a delta hence we've already got all we need. */
1417                         entry->type = entry->in_pack_type;
1418                         entry->in_pack_header_size = used;
1419                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1420                                 goto give_up;
1421                         unuse_pack(&w_curs);
1422                         return;
1423                 case OBJ_REF_DELTA:
1424                         if (reuse_delta && !entry->preferred_base)
1425                                 base_ref = use_pack(p, &w_curs,
1426                                                 entry->in_pack_offset + used, NULL);
1427                         entry->in_pack_header_size = used + 20;
1428                         break;
1429                 case OBJ_OFS_DELTA:
1430                         buf = use_pack(p, &w_curs,
1431                                        entry->in_pack_offset + used, NULL);
1432                         used_0 = 0;
1433                         c = buf[used_0++];
1434                         ofs = c & 127;
1435                         while (c & 128) {
1436                                 ofs += 1;
1437                                 if (!ofs || MSB(ofs, 7)) {
1438                                         error("delta base offset overflow in pack for %s",
1439                                               oid_to_hex(&entry->idx.oid));
1440                                         goto give_up;
1441                                 }
1442                                 c = buf[used_0++];
1443                                 ofs = (ofs << 7) + (c & 127);
1444                         }
1445                         ofs = entry->in_pack_offset - ofs;
1446                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1447                                 error("delta base offset out of bound for %s",
1448                                       oid_to_hex(&entry->idx.oid));
1449                                 goto give_up;
1450                         }
1451                         if (reuse_delta && !entry->preferred_base) {
1452                                 struct revindex_entry *revidx;
1453                                 revidx = find_pack_revindex(p, ofs);
1454                                 if (!revidx)
1455                                         goto give_up;
1456                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1457                         }
1458                         entry->in_pack_header_size = used + used_0;
1459                         break;
1460                 }
1461
1462                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1463                         /*
1464                          * If base_ref was set above that means we wish to
1465                          * reuse delta data, and we even found that base
1466                          * in the list of objects we want to pack. Goodie!
1467                          *
1468                          * Depth value does not matter - find_deltas() will
1469                          * never consider reused delta as the base object to
1470                          * deltify other objects against, in order to avoid
1471                          * circular deltas.
1472                          */
1473                         entry->type = entry->in_pack_type;
1474                         entry->delta = base_entry;
1475                         entry->delta_size = entry->size;
1476                         entry->delta_sibling = base_entry->delta_child;
1477                         base_entry->delta_child = entry;
1478                         unuse_pack(&w_curs);
1479                         return;
1480                 }
1481
1482                 if (entry->type) {
1483                         /*
1484                          * This must be a delta and we already know what the
1485                          * final object type is.  Let's extract the actual
1486                          * object size from the delta header.
1487                          */
1488                         entry->size = get_size_from_delta(p, &w_curs,
1489                                         entry->in_pack_offset + entry->in_pack_header_size);
1490                         if (entry->size == 0)
1491                                 goto give_up;
1492                         unuse_pack(&w_curs);
1493                         return;
1494                 }
1495
1496                 /*
1497                  * No choice but to fall back to the recursive delta walk
1498                  * with sha1_object_info() to find about the object type
1499                  * at this point...
1500                  */
1501                 give_up:
1502                 unuse_pack(&w_curs);
1503         }
1504
1505         entry->type = sha1_object_info(entry->idx.oid.hash, &entry->size);
1506         /*
1507          * The error condition is checked in prepare_pack().  This is
1508          * to permit a missing preferred base object to be ignored
1509          * as a preferred base.  Doing so can result in a larger
1510          * pack file, but the transfer will still take place.
1511          */
1512 }
1513
1514 static int pack_offset_sort(const void *_a, const void *_b)
1515 {
1516         const struct object_entry *a = *(struct object_entry **)_a;
1517         const struct object_entry *b = *(struct object_entry **)_b;
1518
1519         /* avoid filesystem trashing with loose objects */
1520         if (!a->in_pack && !b->in_pack)
1521                 return oidcmp(&a->idx.oid, &b->idx.oid);
1522
1523         if (a->in_pack < b->in_pack)
1524                 return -1;
1525         if (a->in_pack > b->in_pack)
1526                 return 1;
1527         return a->in_pack_offset < b->in_pack_offset ? -1 :
1528                         (a->in_pack_offset > b->in_pack_offset);
1529 }
1530
1531 /*
1532  * Drop an on-disk delta we were planning to reuse. Naively, this would
1533  * just involve blanking out the "delta" field, but we have to deal
1534  * with some extra book-keeping:
1535  *
1536  *   1. Removing ourselves from the delta_sibling linked list.
1537  *
1538  *   2. Updating our size/type to the non-delta representation. These were
1539  *      either not recorded initially (size) or overwritten with the delta type
1540  *      (type) when check_object() decided to reuse the delta.
1541  *
1542  *   3. Resetting our delta depth, as we are now a base object.
1543  */
1544 static void drop_reused_delta(struct object_entry *entry)
1545 {
1546         struct object_entry **p = &entry->delta->delta_child;
1547         struct object_info oi = OBJECT_INFO_INIT;
1548
1549         while (*p) {
1550                 if (*p == entry)
1551                         *p = (*p)->delta_sibling;
1552                 else
1553                         p = &(*p)->delta_sibling;
1554         }
1555         entry->delta = NULL;
1556         entry->depth = 0;
1557
1558         oi.sizep = &entry->size;
1559         oi.typep = &entry->type;
1560         if (packed_object_info(entry->in_pack, entry->in_pack_offset, &oi) < 0) {
1561                 /*
1562                  * We failed to get the info from this pack for some reason;
1563                  * fall back to sha1_object_info, which may find another copy.
1564                  * And if that fails, the error will be recorded in entry->type
1565                  * and dealt with in prepare_pack().
1566                  */
1567                 entry->type = sha1_object_info(entry->idx.oid.hash,
1568                                                &entry->size);
1569         }
1570 }
1571
1572 /*
1573  * Follow the chain of deltas from this entry onward, throwing away any links
1574  * that cause us to hit a cycle (as determined by the DFS state flags in
1575  * the entries).
1576  *
1577  * We also detect too-long reused chains that would violate our --depth
1578  * limit.
1579  */
1580 static void break_delta_chains(struct object_entry *entry)
1581 {
1582         /*
1583          * The actual depth of each object we will write is stored as an int,
1584          * as it cannot exceed our int "depth" limit. But before we break
1585          * changes based no that limit, we may potentially go as deep as the
1586          * number of objects, which is elsewhere bounded to a uint32_t.
1587          */
1588         uint32_t total_depth;
1589         struct object_entry *cur, *next;
1590
1591         for (cur = entry, total_depth = 0;
1592              cur;
1593              cur = cur->delta, total_depth++) {
1594                 if (cur->dfs_state == DFS_DONE) {
1595                         /*
1596                          * We've already seen this object and know it isn't
1597                          * part of a cycle. We do need to append its depth
1598                          * to our count.
1599                          */
1600                         total_depth += cur->depth;
1601                         break;
1602                 }
1603
1604                 /*
1605                  * We break cycles before looping, so an ACTIVE state (or any
1606                  * other cruft which made its way into the state variable)
1607                  * is a bug.
1608                  */
1609                 if (cur->dfs_state != DFS_NONE)
1610                         die("BUG: confusing delta dfs state in first pass: %d",
1611                             cur->dfs_state);
1612
1613                 /*
1614                  * Now we know this is the first time we've seen the object. If
1615                  * it's not a delta, we're done traversing, but we'll mark it
1616                  * done to save time on future traversals.
1617                  */
1618                 if (!cur->delta) {
1619                         cur->dfs_state = DFS_DONE;
1620                         break;
1621                 }
1622
1623                 /*
1624                  * Mark ourselves as active and see if the next step causes
1625                  * us to cycle to another active object. It's important to do
1626                  * this _before_ we loop, because it impacts where we make the
1627                  * cut, and thus how our total_depth counter works.
1628                  * E.g., We may see a partial loop like:
1629                  *
1630                  *   A -> B -> C -> D -> B
1631                  *
1632                  * Cutting B->C breaks the cycle. But now the depth of A is
1633                  * only 1, and our total_depth counter is at 3. The size of the
1634                  * error is always one less than the size of the cycle we
1635                  * broke. Commits C and D were "lost" from A's chain.
1636                  *
1637                  * If we instead cut D->B, then the depth of A is correct at 3.
1638                  * We keep all commits in the chain that we examined.
1639                  */
1640                 cur->dfs_state = DFS_ACTIVE;
1641                 if (cur->delta->dfs_state == DFS_ACTIVE) {
1642                         drop_reused_delta(cur);
1643                         cur->dfs_state = DFS_DONE;
1644                         break;
1645                 }
1646         }
1647
1648         /*
1649          * And now that we've gone all the way to the bottom of the chain, we
1650          * need to clear the active flags and set the depth fields as
1651          * appropriate. Unlike the loop above, which can quit when it drops a
1652          * delta, we need to keep going to look for more depth cuts. So we need
1653          * an extra "next" pointer to keep going after we reset cur->delta.
1654          */
1655         for (cur = entry; cur; cur = next) {
1656                 next = cur->delta;
1657
1658                 /*
1659                  * We should have a chain of zero or more ACTIVE states down to
1660                  * a final DONE. We can quit after the DONE, because either it
1661                  * has no bases, or we've already handled them in a previous
1662                  * call.
1663                  */
1664                 if (cur->dfs_state == DFS_DONE)
1665                         break;
1666                 else if (cur->dfs_state != DFS_ACTIVE)
1667                         die("BUG: confusing delta dfs state in second pass: %d",
1668                             cur->dfs_state);
1669
1670                 /*
1671                  * If the total_depth is more than depth, then we need to snip
1672                  * the chain into two or more smaller chains that don't exceed
1673                  * the maximum depth. Most of the resulting chains will contain
1674                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1675                  * the last chain (i.e., the one containing entry) will contain
1676                  * whatever entries are left over, namely
1677                  * (total_depth % (depth + 1)) of them.
1678                  *
1679                  * Since we are iterating towards decreasing depth, we need to
1680                  * decrement total_depth as we go, and we need to write to the
1681                  * entry what its final depth will be after all of the
1682                  * snipping. Since we're snipping into chains of length (depth
1683                  * + 1) entries, the final depth of an entry will be its
1684                  * original depth modulo (depth + 1). Any time we encounter an
1685                  * entry whose final depth is supposed to be zero, we snip it
1686                  * from its delta base, thereby making it so.
1687                  */
1688                 cur->depth = (total_depth--) % (depth + 1);
1689                 if (!cur->depth)
1690                         drop_reused_delta(cur);
1691
1692                 cur->dfs_state = DFS_DONE;
1693         }
1694 }
1695
1696 static void get_object_details(void)
1697 {
1698         uint32_t i;
1699         struct object_entry **sorted_by_offset;
1700
1701         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1702         for (i = 0; i < to_pack.nr_objects; i++)
1703                 sorted_by_offset[i] = to_pack.objects + i;
1704         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1705
1706         for (i = 0; i < to_pack.nr_objects; i++) {
1707                 struct object_entry *entry = sorted_by_offset[i];
1708                 check_object(entry);
1709                 if (big_file_threshold < entry->size)
1710                         entry->no_try_delta = 1;
1711         }
1712
1713         /*
1714          * This must happen in a second pass, since we rely on the delta
1715          * information for the whole list being completed.
1716          */
1717         for (i = 0; i < to_pack.nr_objects; i++)
1718                 break_delta_chains(&to_pack.objects[i]);
1719
1720         free(sorted_by_offset);
1721 }
1722
1723 /*
1724  * We search for deltas in a list sorted by type, by filename hash, and then
1725  * by size, so that we see progressively smaller and smaller files.
1726  * That's because we prefer deltas to be from the bigger file
1727  * to the smaller -- deletes are potentially cheaper, but perhaps
1728  * more importantly, the bigger file is likely the more recent
1729  * one.  The deepest deltas are therefore the oldest objects which are
1730  * less susceptible to be accessed often.
1731  */
1732 static int type_size_sort(const void *_a, const void *_b)
1733 {
1734         const struct object_entry *a = *(struct object_entry **)_a;
1735         const struct object_entry *b = *(struct object_entry **)_b;
1736
1737         if (a->type > b->type)
1738                 return -1;
1739         if (a->type < b->type)
1740                 return 1;
1741         if (a->hash > b->hash)
1742                 return -1;
1743         if (a->hash < b->hash)
1744                 return 1;
1745         if (a->preferred_base > b->preferred_base)
1746                 return -1;
1747         if (a->preferred_base < b->preferred_base)
1748                 return 1;
1749         if (a->size > b->size)
1750                 return -1;
1751         if (a->size < b->size)
1752                 return 1;
1753         return a < b ? -1 : (a > b);  /* newest first */
1754 }
1755
1756 struct unpacked {
1757         struct object_entry *entry;
1758         void *data;
1759         struct delta_index *index;
1760         unsigned depth;
1761 };
1762
1763 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1764                            unsigned long delta_size)
1765 {
1766         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1767                 return 0;
1768
1769         if (delta_size < cache_max_small_delta_size)
1770                 return 1;
1771
1772         /* cache delta, if objects are large enough compared to delta size */
1773         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1774                 return 1;
1775
1776         return 0;
1777 }
1778
1779 #ifndef NO_PTHREADS
1780
1781 static pthread_mutex_t read_mutex;
1782 #define read_lock()             pthread_mutex_lock(&read_mutex)
1783 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1784
1785 static pthread_mutex_t cache_mutex;
1786 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1787 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1788
1789 static pthread_mutex_t progress_mutex;
1790 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1791 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1792
1793 #else
1794
1795 #define read_lock()             (void)0
1796 #define read_unlock()           (void)0
1797 #define cache_lock()            (void)0
1798 #define cache_unlock()          (void)0
1799 #define progress_lock()         (void)0
1800 #define progress_unlock()       (void)0
1801
1802 #endif
1803
1804 static int try_delta(struct unpacked *trg, struct unpacked *src,
1805                      unsigned max_depth, unsigned long *mem_usage)
1806 {
1807         struct object_entry *trg_entry = trg->entry;
1808         struct object_entry *src_entry = src->entry;
1809         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1810         unsigned ref_depth;
1811         enum object_type type;
1812         void *delta_buf;
1813
1814         /* Don't bother doing diffs between different types */
1815         if (trg_entry->type != src_entry->type)
1816                 return -1;
1817
1818         /*
1819          * We do not bother to try a delta that we discarded on an
1820          * earlier try, but only when reusing delta data.  Note that
1821          * src_entry that is marked as the preferred_base should always
1822          * be considered, as even if we produce a suboptimal delta against
1823          * it, we will still save the transfer cost, as we already know
1824          * the other side has it and we won't send src_entry at all.
1825          */
1826         if (reuse_delta && trg_entry->in_pack &&
1827             trg_entry->in_pack == src_entry->in_pack &&
1828             !src_entry->preferred_base &&
1829             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1830             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1831                 return 0;
1832
1833         /* Let's not bust the allowed depth. */
1834         if (src->depth >= max_depth)
1835                 return 0;
1836
1837         /* Now some size filtering heuristics. */
1838         trg_size = trg_entry->size;
1839         if (!trg_entry->delta) {
1840                 max_size = trg_size/2 - 20;
1841                 ref_depth = 1;
1842         } else {
1843                 max_size = trg_entry->delta_size;
1844                 ref_depth = trg->depth;
1845         }
1846         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1847                                                 (max_depth - ref_depth + 1);
1848         if (max_size == 0)
1849                 return 0;
1850         src_size = src_entry->size;
1851         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1852         if (sizediff >= max_size)
1853                 return 0;
1854         if (trg_size < src_size / 32)
1855                 return 0;
1856
1857         /* Load data if not already done */
1858         if (!trg->data) {
1859                 read_lock();
1860                 trg->data = read_sha1_file(trg_entry->idx.oid.hash, &type,
1861                                            &sz);
1862                 read_unlock();
1863                 if (!trg->data)
1864                         die("object %s cannot be read",
1865                             oid_to_hex(&trg_entry->idx.oid));
1866                 if (sz != trg_size)
1867                         die("object %s inconsistent object length (%lu vs %lu)",
1868                             oid_to_hex(&trg_entry->idx.oid), sz,
1869                             trg_size);
1870                 *mem_usage += sz;
1871         }
1872         if (!src->data) {
1873                 read_lock();
1874                 src->data = read_sha1_file(src_entry->idx.oid.hash, &type,
1875                                            &sz);
1876                 read_unlock();
1877                 if (!src->data) {
1878                         if (src_entry->preferred_base) {
1879                                 static int warned = 0;
1880                                 if (!warned++)
1881                                         warning("object %s cannot be read",
1882                                                 oid_to_hex(&src_entry->idx.oid));
1883                                 /*
1884                                  * Those objects are not included in the
1885                                  * resulting pack.  Be resilient and ignore
1886                                  * them if they can't be read, in case the
1887                                  * pack could be created nevertheless.
1888                                  */
1889                                 return 0;
1890                         }
1891                         die("object %s cannot be read",
1892                             oid_to_hex(&src_entry->idx.oid));
1893                 }
1894                 if (sz != src_size)
1895                         die("object %s inconsistent object length (%lu vs %lu)",
1896                             oid_to_hex(&src_entry->idx.oid), sz,
1897                             src_size);
1898                 *mem_usage += sz;
1899         }
1900         if (!src->index) {
1901                 src->index = create_delta_index(src->data, src_size);
1902                 if (!src->index) {
1903                         static int warned = 0;
1904                         if (!warned++)
1905                                 warning("suboptimal pack - out of memory");
1906                         return 0;
1907                 }
1908                 *mem_usage += sizeof_delta_index(src->index);
1909         }
1910
1911         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1912         if (!delta_buf)
1913                 return 0;
1914
1915         if (trg_entry->delta) {
1916                 /* Prefer only shallower same-sized deltas. */
1917                 if (delta_size == trg_entry->delta_size &&
1918                     src->depth + 1 >= trg->depth) {
1919                         free(delta_buf);
1920                         return 0;
1921                 }
1922         }
1923
1924         /*
1925          * Handle memory allocation outside of the cache
1926          * accounting lock.  Compiler will optimize the strangeness
1927          * away when NO_PTHREADS is defined.
1928          */
1929         free(trg_entry->delta_data);
1930         cache_lock();
1931         if (trg_entry->delta_data) {
1932                 delta_cache_size -= trg_entry->delta_size;
1933                 trg_entry->delta_data = NULL;
1934         }
1935         if (delta_cacheable(src_size, trg_size, delta_size)) {
1936                 delta_cache_size += delta_size;
1937                 cache_unlock();
1938                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1939         } else {
1940                 cache_unlock();
1941                 free(delta_buf);
1942         }
1943
1944         trg_entry->delta = src_entry;
1945         trg_entry->delta_size = delta_size;
1946         trg->depth = src->depth + 1;
1947
1948         return 1;
1949 }
1950
1951 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1952 {
1953         struct object_entry *child = me->delta_child;
1954         unsigned int m = n;
1955         while (child) {
1956                 unsigned int c = check_delta_limit(child, n + 1);
1957                 if (m < c)
1958                         m = c;
1959                 child = child->delta_sibling;
1960         }
1961         return m;
1962 }
1963
1964 static unsigned long free_unpacked(struct unpacked *n)
1965 {
1966         unsigned long freed_mem = sizeof_delta_index(n->index);
1967         free_delta_index(n->index);
1968         n->index = NULL;
1969         if (n->data) {
1970                 freed_mem += n->entry->size;
1971                 FREE_AND_NULL(n->data);
1972         }
1973         n->entry = NULL;
1974         n->depth = 0;
1975         return freed_mem;
1976 }
1977
1978 static void find_deltas(struct object_entry **list, unsigned *list_size,
1979                         int window, int depth, unsigned *processed)
1980 {
1981         uint32_t i, idx = 0, count = 0;
1982         struct unpacked *array;
1983         unsigned long mem_usage = 0;
1984
1985         array = xcalloc(window, sizeof(struct unpacked));
1986
1987         for (;;) {
1988                 struct object_entry *entry;
1989                 struct unpacked *n = array + idx;
1990                 int j, max_depth, best_base = -1;
1991
1992                 progress_lock();
1993                 if (!*list_size) {
1994                         progress_unlock();
1995                         break;
1996                 }
1997                 entry = *list++;
1998                 (*list_size)--;
1999                 if (!entry->preferred_base) {
2000                         (*processed)++;
2001                         display_progress(progress_state, *processed);
2002                 }
2003                 progress_unlock();
2004
2005                 mem_usage -= free_unpacked(n);
2006                 n->entry = entry;
2007
2008                 while (window_memory_limit &&
2009                        mem_usage > window_memory_limit &&
2010                        count > 1) {
2011                         uint32_t tail = (idx + window - count) % window;
2012                         mem_usage -= free_unpacked(array + tail);
2013                         count--;
2014                 }
2015
2016                 /* We do not compute delta to *create* objects we are not
2017                  * going to pack.
2018                  */
2019                 if (entry->preferred_base)
2020                         goto next;
2021
2022                 /*
2023                  * If the current object is at pack edge, take the depth the
2024                  * objects that depend on the current object into account
2025                  * otherwise they would become too deep.
2026                  */
2027                 max_depth = depth;
2028                 if (entry->delta_child) {
2029                         max_depth -= check_delta_limit(entry, 0);
2030                         if (max_depth <= 0)
2031                                 goto next;
2032                 }
2033
2034                 j = window;
2035                 while (--j > 0) {
2036                         int ret;
2037                         uint32_t other_idx = idx + j;
2038                         struct unpacked *m;
2039                         if (other_idx >= window)
2040                                 other_idx -= window;
2041                         m = array + other_idx;
2042                         if (!m->entry)
2043                                 break;
2044                         ret = try_delta(n, m, max_depth, &mem_usage);
2045                         if (ret < 0)
2046                                 break;
2047                         else if (ret > 0)
2048                                 best_base = other_idx;
2049                 }
2050
2051                 /*
2052                  * If we decided to cache the delta data, then it is best
2053                  * to compress it right away.  First because we have to do
2054                  * it anyway, and doing it here while we're threaded will
2055                  * save a lot of time in the non threaded write phase,
2056                  * as well as allow for caching more deltas within
2057                  * the same cache size limit.
2058                  * ...
2059                  * But only if not writing to stdout, since in that case
2060                  * the network is most likely throttling writes anyway,
2061                  * and therefore it is best to go to the write phase ASAP
2062                  * instead, as we can afford spending more time compressing
2063                  * between writes at that moment.
2064                  */
2065                 if (entry->delta_data && !pack_to_stdout) {
2066                         entry->z_delta_size = do_compress(&entry->delta_data,
2067                                                           entry->delta_size);
2068                         cache_lock();
2069                         delta_cache_size -= entry->delta_size;
2070                         delta_cache_size += entry->z_delta_size;
2071                         cache_unlock();
2072                 }
2073
2074                 /* if we made n a delta, and if n is already at max
2075                  * depth, leaving it in the window is pointless.  we
2076                  * should evict it first.
2077                  */
2078                 if (entry->delta && max_depth <= n->depth)
2079                         continue;
2080
2081                 /*
2082                  * Move the best delta base up in the window, after the
2083                  * currently deltified object, to keep it longer.  It will
2084                  * be the first base object to be attempted next.
2085                  */
2086                 if (entry->delta) {
2087                         struct unpacked swap = array[best_base];
2088                         int dist = (window + idx - best_base) % window;
2089                         int dst = best_base;
2090                         while (dist--) {
2091                                 int src = (dst + 1) % window;
2092                                 array[dst] = array[src];
2093                                 dst = src;
2094                         }
2095                         array[dst] = swap;
2096                 }
2097
2098                 next:
2099                 idx++;
2100                 if (count + 1 < window)
2101                         count++;
2102                 if (idx >= window)
2103                         idx = 0;
2104         }
2105
2106         for (i = 0; i < window; ++i) {
2107                 free_delta_index(array[i].index);
2108                 free(array[i].data);
2109         }
2110         free(array);
2111 }
2112
2113 #ifndef NO_PTHREADS
2114
2115 static void try_to_free_from_threads(size_t size)
2116 {
2117         read_lock();
2118         release_pack_memory(size);
2119         read_unlock();
2120 }
2121
2122 static try_to_free_t old_try_to_free_routine;
2123
2124 /*
2125  * The main thread waits on the condition that (at least) one of the workers
2126  * has stopped working (which is indicated in the .working member of
2127  * struct thread_params).
2128  * When a work thread has completed its work, it sets .working to 0 and
2129  * signals the main thread and waits on the condition that .data_ready
2130  * becomes 1.
2131  */
2132
2133 struct thread_params {
2134         pthread_t thread;
2135         struct object_entry **list;
2136         unsigned list_size;
2137         unsigned remaining;
2138         int window;
2139         int depth;
2140         int working;
2141         int data_ready;
2142         pthread_mutex_t mutex;
2143         pthread_cond_t cond;
2144         unsigned *processed;
2145 };
2146
2147 static pthread_cond_t progress_cond;
2148
2149 /*
2150  * Mutex and conditional variable can't be statically-initialized on Windows.
2151  */
2152 static void init_threaded_search(void)
2153 {
2154         init_recursive_mutex(&read_mutex);
2155         pthread_mutex_init(&cache_mutex, NULL);
2156         pthread_mutex_init(&progress_mutex, NULL);
2157         pthread_cond_init(&progress_cond, NULL);
2158         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2159 }
2160
2161 static void cleanup_threaded_search(void)
2162 {
2163         set_try_to_free_routine(old_try_to_free_routine);
2164         pthread_cond_destroy(&progress_cond);
2165         pthread_mutex_destroy(&read_mutex);
2166         pthread_mutex_destroy(&cache_mutex);
2167         pthread_mutex_destroy(&progress_mutex);
2168 }
2169
2170 static void *threaded_find_deltas(void *arg)
2171 {
2172         struct thread_params *me = arg;
2173
2174         while (me->remaining) {
2175                 find_deltas(me->list, &me->remaining,
2176                             me->window, me->depth, me->processed);
2177
2178                 progress_lock();
2179                 me->working = 0;
2180                 pthread_cond_signal(&progress_cond);
2181                 progress_unlock();
2182
2183                 /*
2184                  * We must not set ->data_ready before we wait on the
2185                  * condition because the main thread may have set it to 1
2186                  * before we get here. In order to be sure that new
2187                  * work is available if we see 1 in ->data_ready, it
2188                  * was initialized to 0 before this thread was spawned
2189                  * and we reset it to 0 right away.
2190                  */
2191                 pthread_mutex_lock(&me->mutex);
2192                 while (!me->data_ready)
2193                         pthread_cond_wait(&me->cond, &me->mutex);
2194                 me->data_ready = 0;
2195                 pthread_mutex_unlock(&me->mutex);
2196         }
2197         /* leave ->working 1 so that this doesn't get more work assigned */
2198         return NULL;
2199 }
2200
2201 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2202                            int window, int depth, unsigned *processed)
2203 {
2204         struct thread_params *p;
2205         int i, ret, active_threads = 0;
2206
2207         init_threaded_search();
2208
2209         if (delta_search_threads <= 1) {
2210                 find_deltas(list, &list_size, window, depth, processed);
2211                 cleanup_threaded_search();
2212                 return;
2213         }
2214         if (progress > pack_to_stdout)
2215                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2216                                 delta_search_threads);
2217         p = xcalloc(delta_search_threads, sizeof(*p));
2218
2219         /* Partition the work amongst work threads. */
2220         for (i = 0; i < delta_search_threads; i++) {
2221                 unsigned sub_size = list_size / (delta_search_threads - i);
2222
2223                 /* don't use too small segments or no deltas will be found */
2224                 if (sub_size < 2*window && i+1 < delta_search_threads)
2225                         sub_size = 0;
2226
2227                 p[i].window = window;
2228                 p[i].depth = depth;
2229                 p[i].processed = processed;
2230                 p[i].working = 1;
2231                 p[i].data_ready = 0;
2232
2233                 /* try to split chunks on "path" boundaries */
2234                 while (sub_size && sub_size < list_size &&
2235                        list[sub_size]->hash &&
2236                        list[sub_size]->hash == list[sub_size-1]->hash)
2237                         sub_size++;
2238
2239                 p[i].list = list;
2240                 p[i].list_size = sub_size;
2241                 p[i].remaining = sub_size;
2242
2243                 list += sub_size;
2244                 list_size -= sub_size;
2245         }
2246
2247         /* Start work threads. */
2248         for (i = 0; i < delta_search_threads; i++) {
2249                 if (!p[i].list_size)
2250                         continue;
2251                 pthread_mutex_init(&p[i].mutex, NULL);
2252                 pthread_cond_init(&p[i].cond, NULL);
2253                 ret = pthread_create(&p[i].thread, NULL,
2254                                      threaded_find_deltas, &p[i]);
2255                 if (ret)
2256                         die("unable to create thread: %s", strerror(ret));
2257                 active_threads++;
2258         }
2259
2260         /*
2261          * Now let's wait for work completion.  Each time a thread is done
2262          * with its work, we steal half of the remaining work from the
2263          * thread with the largest number of unprocessed objects and give
2264          * it to that newly idle thread.  This ensure good load balancing
2265          * until the remaining object list segments are simply too short
2266          * to be worth splitting anymore.
2267          */
2268         while (active_threads) {
2269                 struct thread_params *target = NULL;
2270                 struct thread_params *victim = NULL;
2271                 unsigned sub_size = 0;
2272
2273                 progress_lock();
2274                 for (;;) {
2275                         for (i = 0; !target && i < delta_search_threads; i++)
2276                                 if (!p[i].working)
2277                                         target = &p[i];
2278                         if (target)
2279                                 break;
2280                         pthread_cond_wait(&progress_cond, &progress_mutex);
2281                 }
2282
2283                 for (i = 0; i < delta_search_threads; i++)
2284                         if (p[i].remaining > 2*window &&
2285                             (!victim || victim->remaining < p[i].remaining))
2286                                 victim = &p[i];
2287                 if (victim) {
2288                         sub_size = victim->remaining / 2;
2289                         list = victim->list + victim->list_size - sub_size;
2290                         while (sub_size && list[0]->hash &&
2291                                list[0]->hash == list[-1]->hash) {
2292                                 list++;
2293                                 sub_size--;
2294                         }
2295                         if (!sub_size) {
2296                                 /*
2297                                  * It is possible for some "paths" to have
2298                                  * so many objects that no hash boundary
2299                                  * might be found.  Let's just steal the
2300                                  * exact half in that case.
2301                                  */
2302                                 sub_size = victim->remaining / 2;
2303                                 list -= sub_size;
2304                         }
2305                         target->list = list;
2306                         victim->list_size -= sub_size;
2307                         victim->remaining -= sub_size;
2308                 }
2309                 target->list_size = sub_size;
2310                 target->remaining = sub_size;
2311                 target->working = 1;
2312                 progress_unlock();
2313
2314                 pthread_mutex_lock(&target->mutex);
2315                 target->data_ready = 1;
2316                 pthread_cond_signal(&target->cond);
2317                 pthread_mutex_unlock(&target->mutex);
2318
2319                 if (!sub_size) {
2320                         pthread_join(target->thread, NULL);
2321                         pthread_cond_destroy(&target->cond);
2322                         pthread_mutex_destroy(&target->mutex);
2323                         active_threads--;
2324                 }
2325         }
2326         cleanup_threaded_search();
2327         free(p);
2328 }
2329
2330 #else
2331 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2332 #endif
2333
2334 static void add_tag_chain(const struct object_id *oid)
2335 {
2336         struct tag *tag;
2337
2338         /*
2339          * We catch duplicates already in add_object_entry(), but we'd
2340          * prefer to do this extra check to avoid having to parse the
2341          * tag at all if we already know that it's being packed (e.g., if
2342          * it was included via bitmaps, we would not have parsed it
2343          * previously).
2344          */
2345         if (packlist_find(&to_pack, oid->hash, NULL))
2346                 return;
2347
2348         tag = lookup_tag(oid);
2349         while (1) {
2350                 if (!tag || parse_tag(tag) || !tag->tagged)
2351                         die("unable to pack objects reachable from tag %s",
2352                             oid_to_hex(oid));
2353
2354                 add_object_entry(tag->object.oid.hash, OBJ_TAG, NULL, 0);
2355
2356                 if (tag->tagged->type != OBJ_TAG)
2357                         return;
2358
2359                 tag = (struct tag *)tag->tagged;
2360         }
2361 }
2362
2363 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2364 {
2365         struct object_id peeled;
2366
2367         if (starts_with(path, "refs/tags/") && /* is a tag? */
2368             !peel_ref(path, peeled.hash)    && /* peelable? */
2369             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2370                 add_tag_chain(oid);
2371         return 0;
2372 }
2373
2374 static void prepare_pack(int window, int depth)
2375 {
2376         struct object_entry **delta_list;
2377         uint32_t i, nr_deltas;
2378         unsigned n;
2379
2380         get_object_details();
2381
2382         /*
2383          * If we're locally repacking then we need to be doubly careful
2384          * from now on in order to make sure no stealth corruption gets
2385          * propagated to the new pack.  Clients receiving streamed packs
2386          * should validate everything they get anyway so no need to incur
2387          * the additional cost here in that case.
2388          */
2389         if (!pack_to_stdout)
2390                 do_check_packed_object_crc = 1;
2391
2392         if (!to_pack.nr_objects || !window || !depth)
2393                 return;
2394
2395         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2396         nr_deltas = n = 0;
2397
2398         for (i = 0; i < to_pack.nr_objects; i++) {
2399                 struct object_entry *entry = to_pack.objects + i;
2400
2401                 if (entry->delta)
2402                         /* This happens if we decided to reuse existing
2403                          * delta from a pack.  "reuse_delta &&" is implied.
2404                          */
2405                         continue;
2406
2407                 if (entry->size < 50)
2408                         continue;
2409
2410                 if (entry->no_try_delta)
2411                         continue;
2412
2413                 if (!entry->preferred_base) {
2414                         nr_deltas++;
2415                         if (entry->type < 0)
2416                                 die("unable to get type of object %s",
2417                                     oid_to_hex(&entry->idx.oid));
2418                 } else {
2419                         if (entry->type < 0) {
2420                                 /*
2421                                  * This object is not found, but we
2422                                  * don't have to include it anyway.
2423                                  */
2424                                 continue;
2425                         }
2426                 }
2427
2428                 delta_list[n++] = entry;
2429         }
2430
2431         if (nr_deltas && n > 1) {
2432                 unsigned nr_done = 0;
2433                 if (progress)
2434                         progress_state = start_progress(_("Compressing objects"),
2435                                                         nr_deltas);
2436                 QSORT(delta_list, n, type_size_sort);
2437                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2438                 stop_progress(&progress_state);
2439                 if (nr_done != nr_deltas)
2440                         die("inconsistency with delta count");
2441         }
2442         free(delta_list);
2443 }
2444
2445 static int git_pack_config(const char *k, const char *v, void *cb)
2446 {
2447         if (!strcmp(k, "pack.window")) {
2448                 window = git_config_int(k, v);
2449                 return 0;
2450         }
2451         if (!strcmp(k, "pack.windowmemory")) {
2452                 window_memory_limit = git_config_ulong(k, v);
2453                 return 0;
2454         }
2455         if (!strcmp(k, "pack.depth")) {
2456                 depth = git_config_int(k, v);
2457                 return 0;
2458         }
2459         if (!strcmp(k, "pack.deltacachesize")) {
2460                 max_delta_cache_size = git_config_int(k, v);
2461                 return 0;
2462         }
2463         if (!strcmp(k, "pack.deltacachelimit")) {
2464                 cache_max_small_delta_size = git_config_int(k, v);
2465                 return 0;
2466         }
2467         if (!strcmp(k, "pack.writebitmaphashcache")) {
2468                 if (git_config_bool(k, v))
2469                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2470                 else
2471                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2472         }
2473         if (!strcmp(k, "pack.usebitmaps")) {
2474                 use_bitmap_index_default = git_config_bool(k, v);
2475                 return 0;
2476         }
2477         if (!strcmp(k, "pack.threads")) {
2478                 delta_search_threads = git_config_int(k, v);
2479                 if (delta_search_threads < 0)
2480                         die("invalid number of threads specified (%d)",
2481                             delta_search_threads);
2482 #ifdef NO_PTHREADS
2483                 if (delta_search_threads != 1) {
2484                         warning("no threads support, ignoring %s", k);
2485                         delta_search_threads = 0;
2486                 }
2487 #endif
2488                 return 0;
2489         }
2490         if (!strcmp(k, "pack.indexversion")) {
2491                 pack_idx_opts.version = git_config_int(k, v);
2492                 if (pack_idx_opts.version > 2)
2493                         die("bad pack.indexversion=%"PRIu32,
2494                             pack_idx_opts.version);
2495                 return 0;
2496         }
2497         return git_default_config(k, v, cb);
2498 }
2499
2500 static void read_object_list_from_stdin(void)
2501 {
2502         char line[40 + 1 + PATH_MAX + 2];
2503         unsigned char sha1[20];
2504
2505         for (;;) {
2506                 if (!fgets(line, sizeof(line), stdin)) {
2507                         if (feof(stdin))
2508                                 break;
2509                         if (!ferror(stdin))
2510                                 die("fgets returned NULL, not EOF, not error!");
2511                         if (errno != EINTR)
2512                                 die_errno("fgets");
2513                         clearerr(stdin);
2514                         continue;
2515                 }
2516                 if (line[0] == '-') {
2517                         if (get_sha1_hex(line+1, sha1))
2518                                 die("expected edge sha1, got garbage:\n %s",
2519                                     line);
2520                         add_preferred_base(sha1);
2521                         continue;
2522                 }
2523                 if (get_sha1_hex(line, sha1))
2524                         die("expected sha1, got garbage:\n %s", line);
2525
2526                 add_preferred_base_object(line+41);
2527                 add_object_entry(sha1, 0, line+41, 0);
2528         }
2529 }
2530
2531 #define OBJECT_ADDED (1u<<20)
2532
2533 static void show_commit(struct commit *commit, void *data)
2534 {
2535         add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2536         commit->object.flags |= OBJECT_ADDED;
2537
2538         if (write_bitmap_index)
2539                 index_commit_for_bitmap(commit);
2540 }
2541
2542 static void show_object(struct object *obj, const char *name, void *data)
2543 {
2544         add_preferred_base_object(name);
2545         add_object_entry(obj->oid.hash, obj->type, name, 0);
2546         obj->flags |= OBJECT_ADDED;
2547 }
2548
2549 static void show_edge(struct commit *commit)
2550 {
2551         add_preferred_base(commit->object.oid.hash);
2552 }
2553
2554 struct in_pack_object {
2555         off_t offset;
2556         struct object *object;
2557 };
2558
2559 struct in_pack {
2560         int alloc;
2561         int nr;
2562         struct in_pack_object *array;
2563 };
2564
2565 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2566 {
2567         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2568         in_pack->array[in_pack->nr].object = object;
2569         in_pack->nr++;
2570 }
2571
2572 /*
2573  * Compare the objects in the offset order, in order to emulate the
2574  * "git rev-list --objects" output that produced the pack originally.
2575  */
2576 static int ofscmp(const void *a_, const void *b_)
2577 {
2578         struct in_pack_object *a = (struct in_pack_object *)a_;
2579         struct in_pack_object *b = (struct in_pack_object *)b_;
2580
2581         if (a->offset < b->offset)
2582                 return -1;
2583         else if (a->offset > b->offset)
2584                 return 1;
2585         else
2586                 return oidcmp(&a->object->oid, &b->object->oid);
2587 }
2588
2589 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2590 {
2591         struct packed_git *p;
2592         struct in_pack in_pack;
2593         uint32_t i;
2594
2595         memset(&in_pack, 0, sizeof(in_pack));
2596
2597         for (p = packed_git; p; p = p->next) {
2598                 const unsigned char *sha1;
2599                 struct object *o;
2600
2601                 if (!p->pack_local || p->pack_keep)
2602                         continue;
2603                 if (open_pack_index(p))
2604                         die("cannot open pack index");
2605
2606                 ALLOC_GROW(in_pack.array,
2607                            in_pack.nr + p->num_objects,
2608                            in_pack.alloc);
2609
2610                 for (i = 0; i < p->num_objects; i++) {
2611                         sha1 = nth_packed_object_sha1(p, i);
2612                         o = lookup_unknown_object(sha1);
2613                         if (!(o->flags & OBJECT_ADDED))
2614                                 mark_in_pack_object(o, p, &in_pack);
2615                         o->flags |= OBJECT_ADDED;
2616                 }
2617         }
2618
2619         if (in_pack.nr) {
2620                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2621                 for (i = 0; i < in_pack.nr; i++) {
2622                         struct object *o = in_pack.array[i].object;
2623                         add_object_entry(o->oid.hash, o->type, "", 0);
2624                 }
2625         }
2626         free(in_pack.array);
2627 }
2628
2629 static int add_loose_object(const struct object_id *oid, const char *path,
2630                             void *data)
2631 {
2632         enum object_type type = sha1_object_info(oid->hash, NULL);
2633
2634         if (type < 0) {
2635                 warning("loose object at %s could not be examined", path);
2636                 return 0;
2637         }
2638
2639         add_object_entry(oid->hash, type, "", 0);
2640         return 0;
2641 }
2642
2643 /*
2644  * We actually don't even have to worry about reachability here.
2645  * add_object_entry will weed out duplicates, so we just add every
2646  * loose object we find.
2647  */
2648 static void add_unreachable_loose_objects(void)
2649 {
2650         for_each_loose_file_in_objdir(get_object_directory(),
2651                                       add_loose_object,
2652                                       NULL, NULL, NULL);
2653 }
2654
2655 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2656 {
2657         static struct packed_git *last_found = (void *)1;
2658         struct packed_git *p;
2659
2660         p = (last_found != (void *)1) ? last_found : packed_git;
2661
2662         while (p) {
2663                 if ((!p->pack_local || p->pack_keep) &&
2664                         find_pack_entry_one(sha1, p)) {
2665                         last_found = p;
2666                         return 1;
2667                 }
2668                 if (p == last_found)
2669                         p = packed_git;
2670                 else
2671                         p = p->next;
2672                 if (p == last_found)
2673                         p = p->next;
2674         }
2675         return 0;
2676 }
2677
2678 /*
2679  * Store a list of sha1s that are should not be discarded
2680  * because they are either written too recently, or are
2681  * reachable from another object that was.
2682  *
2683  * This is filled by get_object_list.
2684  */
2685 static struct oid_array recent_objects;
2686
2687 static int loosened_object_can_be_discarded(const struct object_id *oid,
2688                                             timestamp_t mtime)
2689 {
2690         if (!unpack_unreachable_expiration)
2691                 return 0;
2692         if (mtime > unpack_unreachable_expiration)
2693                 return 0;
2694         if (oid_array_lookup(&recent_objects, oid) >= 0)
2695                 return 0;
2696         return 1;
2697 }
2698
2699 static void loosen_unused_packed_objects(struct rev_info *revs)
2700 {
2701         struct packed_git *p;
2702         uint32_t i;
2703         struct object_id oid;
2704
2705         for (p = packed_git; p; p = p->next) {
2706                 if (!p->pack_local || p->pack_keep)
2707                         continue;
2708
2709                 if (open_pack_index(p))
2710                         die("cannot open pack index");
2711
2712                 for (i = 0; i < p->num_objects; i++) {
2713                         nth_packed_object_oid(&oid, p, i);
2714                         if (!packlist_find(&to_pack, oid.hash, NULL) &&
2715                             !has_sha1_pack_kept_or_nonlocal(oid.hash) &&
2716                             !loosened_object_can_be_discarded(&oid, p->mtime))
2717                                 if (force_object_loose(oid.hash, p->mtime))
2718                                         die("unable to force loose object");
2719                 }
2720         }
2721 }
2722
2723 /*
2724  * This tracks any options which pack-reuse code expects to be on, or which a
2725  * reader of the pack might not understand, and which would therefore prevent
2726  * blind reuse of what we have on disk.
2727  */
2728 static int pack_options_allow_reuse(void)
2729 {
2730         return pack_to_stdout &&
2731                allow_ofs_delta &&
2732                !ignore_packed_keep &&
2733                (!local || !have_non_local_packs) &&
2734                !incremental;
2735 }
2736
2737 static int get_object_list_from_bitmap(struct rev_info *revs)
2738 {
2739         if (prepare_bitmap_walk(revs) < 0)
2740                 return -1;
2741
2742         if (pack_options_allow_reuse() &&
2743             !reuse_partial_packfile_from_bitmap(
2744                         &reuse_packfile,
2745                         &reuse_packfile_objects,
2746                         &reuse_packfile_offset)) {
2747                 assert(reuse_packfile_objects);
2748                 nr_result += reuse_packfile_objects;
2749                 display_progress(progress_state, nr_result);
2750         }
2751
2752         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2753         return 0;
2754 }
2755
2756 static void record_recent_object(struct object *obj,
2757                                  const char *name,
2758                                  void *data)
2759 {
2760         oid_array_append(&recent_objects, &obj->oid);
2761 }
2762
2763 static void record_recent_commit(struct commit *commit, void *data)
2764 {
2765         oid_array_append(&recent_objects, &commit->object.oid);
2766 }
2767
2768 static void get_object_list(int ac, const char **av)
2769 {
2770         struct rev_info revs;
2771         char line[1000];
2772         int flags = 0;
2773
2774         init_revisions(&revs, NULL);
2775         save_commit_buffer = 0;
2776         setup_revisions(ac, av, &revs, NULL);
2777
2778         /* make sure shallows are read */
2779         is_repository_shallow();
2780
2781         while (fgets(line, sizeof(line), stdin) != NULL) {
2782                 int len = strlen(line);
2783                 if (len && line[len - 1] == '\n')
2784                         line[--len] = 0;
2785                 if (!len)
2786                         break;
2787                 if (*line == '-') {
2788                         if (!strcmp(line, "--not")) {
2789                                 flags ^= UNINTERESTING;
2790                                 write_bitmap_index = 0;
2791                                 continue;
2792                         }
2793                         if (starts_with(line, "--shallow ")) {
2794                                 struct object_id oid;
2795                                 if (get_oid_hex(line + 10, &oid))
2796                                         die("not an SHA-1 '%s'", line + 10);
2797                                 register_shallow(&oid);
2798                                 use_bitmap_index = 0;
2799                                 continue;
2800                         }
2801                         die("not a rev '%s'", line);
2802                 }
2803                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2804                         die("bad revision '%s'", line);
2805         }
2806
2807         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2808                 return;
2809
2810         if (prepare_revision_walk(&revs))
2811                 die("revision walk setup failed");
2812         mark_edges_uninteresting(&revs, show_edge);
2813         traverse_commit_list(&revs, show_commit, show_object, NULL);
2814
2815         if (unpack_unreachable_expiration) {
2816                 revs.ignore_missing_links = 1;
2817                 if (add_unseen_recent_objects_to_traversal(&revs,
2818                                 unpack_unreachable_expiration))
2819                         die("unable to add recent objects");
2820                 if (prepare_revision_walk(&revs))
2821                         die("revision walk setup failed");
2822                 traverse_commit_list(&revs, record_recent_commit,
2823                                      record_recent_object, NULL);
2824         }
2825
2826         if (keep_unreachable)
2827                 add_objects_in_unpacked_packs(&revs);
2828         if (pack_loose_unreachable)
2829                 add_unreachable_loose_objects();
2830         if (unpack_unreachable)
2831                 loosen_unused_packed_objects(&revs);
2832
2833         oid_array_clear(&recent_objects);
2834 }
2835
2836 static int option_parse_index_version(const struct option *opt,
2837                                       const char *arg, int unset)
2838 {
2839         char *c;
2840         const char *val = arg;
2841         pack_idx_opts.version = strtoul(val, &c, 10);
2842         if (pack_idx_opts.version > 2)
2843                 die(_("unsupported index version %s"), val);
2844         if (*c == ',' && c[1])
2845                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2846         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2847                 die(_("bad index version '%s'"), val);
2848         return 0;
2849 }
2850
2851 static int option_parse_unpack_unreachable(const struct option *opt,
2852                                            const char *arg, int unset)
2853 {
2854         if (unset) {
2855                 unpack_unreachable = 0;
2856                 unpack_unreachable_expiration = 0;
2857         }
2858         else {
2859                 unpack_unreachable = 1;
2860                 if (arg)
2861                         unpack_unreachable_expiration = approxidate(arg);
2862         }
2863         return 0;
2864 }
2865
2866 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2867 {
2868         int use_internal_rev_list = 0;
2869         int thin = 0;
2870         int shallow = 0;
2871         int all_progress_implied = 0;
2872         struct argv_array rp = ARGV_ARRAY_INIT;
2873         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2874         int rev_list_index = 0;
2875         struct option pack_objects_options[] = {
2876                 OPT_SET_INT('q', "quiet", &progress,
2877                             N_("do not show progress meter"), 0),
2878                 OPT_SET_INT(0, "progress", &progress,
2879                             N_("show progress meter"), 1),
2880                 OPT_SET_INT(0, "all-progress", &progress,
2881                             N_("show progress meter during object writing phase"), 2),
2882                 OPT_BOOL(0, "all-progress-implied",
2883                          &all_progress_implied,
2884                          N_("similar to --all-progress when progress meter is shown")),
2885                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2886                   N_("write the pack index file in the specified idx format version"),
2887                   0, option_parse_index_version },
2888                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2889                               N_("maximum size of each output pack file")),
2890                 OPT_BOOL(0, "local", &local,
2891                          N_("ignore borrowed objects from alternate object store")),
2892                 OPT_BOOL(0, "incremental", &incremental,
2893                          N_("ignore packed objects")),
2894                 OPT_INTEGER(0, "window", &window,
2895                             N_("limit pack window by objects")),
2896                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2897                               N_("limit pack window by memory in addition to object limit")),
2898                 OPT_INTEGER(0, "depth", &depth,
2899                             N_("maximum length of delta chain allowed in the resulting pack")),
2900                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2901                          N_("reuse existing deltas")),
2902                 OPT_BOOL(0, "reuse-object", &reuse_object,
2903                          N_("reuse existing objects")),
2904                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2905                          N_("use OFS_DELTA objects")),
2906                 OPT_INTEGER(0, "threads", &delta_search_threads,
2907                             N_("use threads when searching for best delta matches")),
2908                 OPT_BOOL(0, "non-empty", &non_empty,
2909                          N_("do not create an empty pack output")),
2910                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2911                          N_("read revision arguments from standard input")),
2912                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2913                   N_("limit the objects to those that are not yet packed"),
2914                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2915                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2916                   N_("include objects reachable from any reference"),
2917                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2918                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2919                   N_("include objects referred by reflog entries"),
2920                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2921                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2922                   N_("include objects referred to by the index"),
2923                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2924                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2925                          N_("output pack to stdout")),
2926                 OPT_BOOL(0, "include-tag", &include_tag,
2927                          N_("include tag objects that refer to objects to be packed")),
2928                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2929                          N_("keep unreachable objects")),
2930                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
2931                          N_("pack loose unreachable objects")),
2932                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2933                   N_("unpack unreachable objects newer than <time>"),
2934                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2935                 OPT_BOOL(0, "thin", &thin,
2936                          N_("create thin packs")),
2937                 OPT_BOOL(0, "shallow", &shallow,
2938                          N_("create packs suitable for shallow fetches")),
2939                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2940                          N_("ignore packs that have companion .keep file")),
2941                 OPT_INTEGER(0, "compression", &pack_compression_level,
2942                             N_("pack compression level")),
2943                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2944                             N_("do not hide commits by grafts"), 0),
2945                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2946                          N_("use a bitmap index if available to speed up counting objects")),
2947                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2948                          N_("write a bitmap index together with the pack index")),
2949                 OPT_END(),
2950         };
2951
2952         check_replace_refs = 0;
2953
2954         reset_pack_idx_option(&pack_idx_opts);
2955         git_config(git_pack_config, NULL);
2956
2957         progress = isatty(2);
2958         argc = parse_options(argc, argv, prefix, pack_objects_options,
2959                              pack_usage, 0);
2960
2961         if (argc) {
2962                 base_name = argv[0];
2963                 argc--;
2964         }
2965         if (pack_to_stdout != !base_name || argc)
2966                 usage_with_options(pack_usage, pack_objects_options);
2967
2968         argv_array_push(&rp, "pack-objects");
2969         if (thin) {
2970                 use_internal_rev_list = 1;
2971                 argv_array_push(&rp, shallow
2972                                 ? "--objects-edge-aggressive"
2973                                 : "--objects-edge");
2974         } else
2975                 argv_array_push(&rp, "--objects");
2976
2977         if (rev_list_all) {
2978                 use_internal_rev_list = 1;
2979                 argv_array_push(&rp, "--all");
2980         }
2981         if (rev_list_reflog) {
2982                 use_internal_rev_list = 1;
2983                 argv_array_push(&rp, "--reflog");
2984         }
2985         if (rev_list_index) {
2986                 use_internal_rev_list = 1;
2987                 argv_array_push(&rp, "--indexed-objects");
2988         }
2989         if (rev_list_unpacked) {
2990                 use_internal_rev_list = 1;
2991                 argv_array_push(&rp, "--unpacked");
2992         }
2993
2994         if (!reuse_object)
2995                 reuse_delta = 0;
2996         if (pack_compression_level == -1)
2997                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2998         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2999                 die("bad pack compression level %d", pack_compression_level);
3000
3001         if (!delta_search_threads)      /* --threads=0 means autodetect */
3002                 delta_search_threads = online_cpus();
3003
3004 #ifdef NO_PTHREADS
3005         if (delta_search_threads != 1)
3006                 warning("no threads support, ignoring --threads");
3007 #endif
3008         if (!pack_to_stdout && !pack_size_limit)
3009                 pack_size_limit = pack_size_limit_cfg;
3010         if (pack_to_stdout && pack_size_limit)
3011                 die("--max-pack-size cannot be used to build a pack for transfer.");
3012         if (pack_size_limit && pack_size_limit < 1024*1024) {
3013                 warning("minimum pack size limit is 1 MiB");
3014                 pack_size_limit = 1024*1024;
3015         }
3016
3017         if (!pack_to_stdout && thin)
3018                 die("--thin cannot be used to build an indexable pack.");
3019
3020         if (keep_unreachable && unpack_unreachable)
3021                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
3022         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3023                 unpack_unreachable_expiration = 0;
3024
3025         /*
3026          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3027          *
3028          * - to produce good pack (with bitmap index not-yet-packed objects are
3029          *   packed in suboptimal order).
3030          *
3031          * - to use more robust pack-generation codepath (avoiding possible
3032          *   bugs in bitmap code and possible bitmap index corruption).
3033          */
3034         if (!pack_to_stdout)
3035                 use_bitmap_index_default = 0;
3036
3037         if (use_bitmap_index < 0)
3038                 use_bitmap_index = use_bitmap_index_default;
3039
3040         /* "hard" reasons not to use bitmaps; these just won't work at all */
3041         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
3042                 use_bitmap_index = 0;
3043
3044         if (pack_to_stdout || !rev_list_all)
3045                 write_bitmap_index = 0;
3046
3047         if (progress && all_progress_implied)
3048                 progress = 2;
3049
3050         prepare_packed_git();
3051         if (ignore_packed_keep) {
3052                 struct packed_git *p;
3053                 for (p = packed_git; p; p = p->next)
3054                         if (p->pack_local && p->pack_keep)
3055                                 break;
3056                 if (!p) /* no keep-able packs found */
3057                         ignore_packed_keep = 0;
3058         }
3059         if (local) {
3060                 /*
3061                  * unlike ignore_packed_keep above, we do not want to
3062                  * unset "local" based on looking at packs, as it
3063                  * also covers non-local objects
3064                  */
3065                 struct packed_git *p;
3066                 for (p = packed_git; p; p = p->next) {
3067                         if (!p->pack_local) {
3068                                 have_non_local_packs = 1;
3069                                 break;
3070                         }
3071                 }
3072         }
3073
3074         if (progress)
3075                 progress_state = start_progress(_("Counting objects"), 0);
3076         if (!use_internal_rev_list)
3077                 read_object_list_from_stdin();
3078         else {
3079                 get_object_list(rp.argc, rp.argv);
3080                 argv_array_clear(&rp);
3081         }
3082         cleanup_preferred_base();
3083         if (include_tag && nr_result)
3084                 for_each_ref(add_ref_tag, NULL);
3085         stop_progress(&progress_state);
3086
3087         if (non_empty && !nr_result)
3088                 return 0;
3089         if (nr_result)
3090                 prepare_pack(window, depth);
3091         write_pack_file();
3092         if (progress)
3093                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
3094                         " reused %"PRIu32" (delta %"PRIu32")\n",
3095                         written, written_delta, reused, reused_delta);
3096         return 0;
3097 }