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

   1 #include "builtin.h"
   2 #include "cache.h"
   3 #include "config.h"
   4 #include "attr.h"
   5 #include "object.h"
   6 #include "blob.h"
   7 #include "commit.h"
   8 #include "tag.h"
   9 #include "tree.h"
  10 #include "delta.h"
  11 #include "pack.h"
  12 #include "pack-revindex.h"
  13 #include "csum-file.h"
  14 #include "tree-walk.h"
  15 #include "diff.h"
  16 #include "revision.h"
  17 #include "list-objects.h"
  18 #include "pack-objects.h"
  19 #include "progress.h"
  20 #include "refs.h"
  21 #include "streaming.h"
  22 #include "thread-utils.h"
  23 #include "pack-bitmap.h"
  24 #include "reachable.h"
  25 #include "sha1-array.h"
  26 #include "argv-array.h"
  27 #include "mru.h"
  28
  29 static const char *pack_usage[] = {
  30         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
  31         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
  32         NULL
  33 };
  34
  35 /*
  36  * Objects we are going to pack are collected in the `to_pack` structure.
  37  * It contains an array (dynamically expanded) of the object data, and a map
  38  * that can resolve SHA1s to their position in the array.
  39  */
  40 static struct packing_data to_pack;
  41
  42 static struct pack_idx_entry **written_list;
  43 static uint32_t nr_result, nr_written;
  44
  45 static int non_empty;
  46 static int reuse_delta = 1, reuse_object = 1;
  47 static int keep_unreachable, unpack_unreachable, include_tag;
  48 static unsigned long 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.sha1, &type, &size);
 111         if (!buf)
 112                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
 113         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
 114         if (!base_buf)
 115                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
 116         delta_buf = diff_delta(base_buf, base_size,
 117                                buf, size, &delta_size, 0);
 118         if (!delta_buf || delta_size != entry->delta_size)
 119                 die("delta size changed");
 120         free(buf);
 121         free(base_buf);
 122         return delta_buf;
 123 }
 124
 125 static unsigned long do_compress(void **pptr, unsigned long size)
 126 {
 127         git_zstream stream;
 128         void *in, *out;
 129         unsigned long maxsize;
 130
 131         git_deflate_init(&stream, pack_compression_level);
 132         maxsize = git_deflate_bound(&stream, size);
 133
 134         in = *pptr;
 135         out = xmalloc(maxsize);
 136         *pptr = out;
 137
 138         stream.next_in = in;
 139         stream.avail_in = size;
 140         stream.next_out = out;
 141         stream.avail_out = maxsize;
 142         while (git_deflate(&stream, Z_FINISH) == Z_OK)
 143                 ; /* nothing */
 144         git_deflate_end(&stream);
 145
 146         free(in);
 147         return stream.total_out;
 148 }
 149
 150 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
 151                                            const unsigned char *sha1)
 152 {
 153         git_zstream stream;
 154         unsigned char ibuf[1024 * 16];
 155         unsigned char obuf[1024 * 16];
 156         unsigned long olen = 0;
 157
 158         git_deflate_init(&stream, pack_compression_level);
 159
 160         for (;;) {
 161                 ssize_t readlen;
 162                 int zret = Z_OK;
 163                 readlen = read_istream(st, ibuf, sizeof(ibuf));
 164                 if (readlen == -1)
 165                         die(_("unable to read %s"), sha1_to_hex(sha1));
 166
 167                 stream.next_in = ibuf;
 168                 stream.avail_in = readlen;
 169                 while ((stream.avail_in || readlen == 0) &&
 170                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
 171                         stream.next_out = obuf;
 172                         stream.avail_out = sizeof(obuf);
 173                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 174                         sha1write(f, obuf, stream.next_out - obuf);
 175                         olen += stream.next_out - obuf;
 176                 }
 177                 if (stream.avail_in)
 178                         die(_("deflate error (%d)"), zret);
 179                 if (readlen == 0) {
 180                         if (zret != Z_STREAM_END)
 181                                 die(_("deflate error (%d)"), zret);
 182                         break;
 183                 }
 184         }
 185         git_deflate_end(&stream);
 186         return olen;
 187 }
 188
 189 /*
 190  * we are going to reuse the existing object data as is.  make
 191  * sure it is not corrupt.
 192  */
 193 static int check_pack_inflate(struct packed_git *p,
 194                 struct pack_window **w_curs,
 195                 off_t offset,
 196                 off_t len,
 197                 unsigned long expect)
 198 {
 199         git_zstream stream;
 200         unsigned char fakebuf[4096], *in;
 201         int st;
 202
 203         memset(&stream, 0, sizeof(stream));
 204         git_inflate_init(&stream);
 205         do {
 206                 in = use_pack(p, w_curs, offset, &stream.avail_in);
 207                 stream.next_in = in;
 208                 stream.next_out = fakebuf;
 209                 stream.avail_out = sizeof(fakebuf);
 210                 st = git_inflate(&stream, Z_FINISH);
 211                 offset += stream.next_in - in;
 212         } while (st == Z_OK || st == Z_BUF_ERROR);
 213         git_inflate_end(&stream);
 214         return (st == Z_STREAM_END &&
 215                 stream.total_out == expect &&
 216                 stream.total_in == len) ? 0 : -1;
 217 }
 218
 219 static void copy_pack_data(struct sha1file *f,
 220                 struct packed_git *p,
 221                 struct pack_window **w_curs,
 222                 off_t offset,
 223                 off_t len)
 224 {
 225         unsigned char *in;
 226         unsigned long avail;
 227
 228         while (len) {
 229                 in = use_pack(p, w_curs, offset, &avail);
 230                 if (avail > len)
 231                         avail = (unsigned long)len;
 232                 sha1write(f, in, avail);
 233                 offset += avail;
 234                 len -= avail;
 235         }
 236 }
 237
 238 /* Return 0 if we will bust the pack-size limit */
 239 static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
 240                                            unsigned long limit, int usable_delta)
 241 {
 242         unsigned long size, datalen;
 243         unsigned char header[MAX_PACK_OBJECT_HEADER],
 244                       dheader[MAX_PACK_OBJECT_HEADER];
 245         unsigned hdrlen;
 246         enum object_type type;
 247         void *buf;
 248         struct git_istream *st = NULL;
 249
 250         if (!usable_delta) {
 251                 if (entry->type == OBJ_BLOB &&
 252                     entry->size > big_file_threshold &&
 253                     (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
 254                         buf = NULL;
 255                 else {
 256                         buf = read_sha1_file(entry->idx.sha1, &type, &size);
 257                         if (!buf)
 258                                 die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
 259                 }
 260                 /*
 261                  * make sure no cached delta data remains from a
 262                  * previous attempt before a pack split occurred.
 263                  */
 264                 free(entry->delta_data);
 265                 entry->delta_data = NULL;
 266                 entry->z_delta_size = 0;
 267         } else if (entry->delta_data) {
 268                 size = entry->delta_size;
 269                 buf = entry->delta_data;
 270                 entry->delta_data = NULL;
 271                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 272                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 273         } else {
 274                 buf = get_delta(entry);
 275                 size = entry->delta_size;
 276                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 277                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 278         }
 279
 280         if (st) /* large blob case, just assume we don't compress well */
 281                 datalen = size;
 282         else if (entry->z_delta_size)
 283                 datalen = entry->z_delta_size;
 284         else
 285                 datalen = do_compress(&buf, size);
 286
 287         /*
 288          * The object header is a byte of 'type' followed by zero or
 289          * more bytes of length.
 290          */
 291         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 292                                               type, size);
 293
 294         if (type == OBJ_OFS_DELTA) {
 295                 /*
 296                  * Deltas with relative base contain an additional
 297                  * encoding of the relative offset for the delta
 298                  * base from this object's position in the pack.
 299                  */
 300                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 301                 unsigned pos = sizeof(dheader) - 1;
 302                 dheader[pos] = ofs & 127;
 303                 while (ofs >>= 7)
 304                         dheader[--pos] = 128 | (--ofs & 127);
 305                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 306                         if (st)
 307                                 close_istream(st);
 308                         free(buf);
 309                         return 0;
 310                 }
 311                 sha1write(f, header, hdrlen);
 312                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 313                 hdrlen += sizeof(dheader) - pos;
 314         } else if (type == OBJ_REF_DELTA) {
 315                 /*
 316                  * Deltas with a base reference contain
 317                  * an additional 20 bytes for the base sha1.
 318                  */
 319                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 320                         if (st)
 321                                 close_istream(st);
 322                         free(buf);
 323                         return 0;
 324                 }
 325                 sha1write(f, header, hdrlen);
 326                 sha1write(f, entry->delta->idx.sha1, 20);
 327                 hdrlen += 20;
 328         } else {
 329                 if (limit && hdrlen + datalen + 20 >= limit) {
 330                         if (st)
 331                                 close_istream(st);
 332                         free(buf);
 333                         return 0;
 334                 }
 335                 sha1write(f, header, hdrlen);
 336         }
 337         if (st) {
 338                 datalen = write_large_blob_data(st, f, entry->idx.sha1);
 339                 close_istream(st);
 340         } else {
 341                 sha1write(f, buf, datalen);
 342                 free(buf);
 343         }
 344
 345         return hdrlen + datalen;
 346 }
 347
 348 /* Return 0 if we will bust the pack-size limit */
 349 static off_t write_reuse_object(struct sha1file *f, struct object_entry *entry,
 350                                 unsigned long limit, int usable_delta)
 351 {
 352         struct packed_git *p = entry->in_pack;
 353         struct pack_window *w_curs = NULL;
 354         struct revindex_entry *revidx;
 355         off_t offset;
 356         enum object_type type = entry->type;
 357         off_t datalen;
 358         unsigned char header[MAX_PACK_OBJECT_HEADER],
 359                       dheader[MAX_PACK_OBJECT_HEADER];
 360         unsigned hdrlen;
 361
 362         if (entry->delta)
 363                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 364                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 365         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 366                                               type, entry->size);
 367
 368         offset = entry->in_pack_offset;
 369         revidx = find_pack_revindex(p, offset);
 370         datalen = revidx[1].offset - offset;
 371         if (!pack_to_stdout && p->index_version > 1 &&
 372             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 373                 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
 374                 unuse_pack(&w_curs);
 375                 return write_no_reuse_object(f, entry, limit, usable_delta);
 376         }
 377
 378         offset += entry->in_pack_header_size;
 379         datalen -= entry->in_pack_header_size;
 380
 381         if (!pack_to_stdout && p->index_version == 1 &&
 382             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 383                 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
 384                 unuse_pack(&w_curs);
 385                 return write_no_reuse_object(f, entry, limit, usable_delta);
 386         }
 387
 388         if (type == OBJ_OFS_DELTA) {
 389                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 390                 unsigned pos = sizeof(dheader) - 1;
 391                 dheader[pos] = ofs & 127;
 392                 while (ofs >>= 7)
 393                         dheader[--pos] = 128 | (--ofs & 127);
 394                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 395                         unuse_pack(&w_curs);
 396                         return 0;
 397                 }
 398                 sha1write(f, header, hdrlen);
 399                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 400                 hdrlen += sizeof(dheader) - pos;
 401                 reused_delta++;
 402         } else if (type == OBJ_REF_DELTA) {
 403                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 404                         unuse_pack(&w_curs);
 405                         return 0;
 406                 }
 407                 sha1write(f, header, hdrlen);
 408                 sha1write(f, entry->delta->idx.sha1, 20);
 409                 hdrlen += 20;
 410                 reused_delta++;
 411         } else {
 412                 if (limit && hdrlen + datalen + 20 >= limit) {
 413                         unuse_pack(&w_curs);
 414                         return 0;
 415                 }
 416                 sha1write(f, header, hdrlen);
 417         }
 418         copy_pack_data(f, p, &w_curs, offset, datalen);
 419         unuse_pack(&w_curs);
 420         reused++;
 421         return hdrlen + datalen;
 422 }
 423
 424 /* Return 0 if we will bust the pack-size limit */
 425 static off_t write_object(struct sha1file *f,
 426                           struct object_entry *entry,
 427                           off_t write_offset)
 428 {
 429         unsigned long limit;
 430         off_t len;
 431         int usable_delta, to_reuse;
 432
 433         if (!pack_to_stdout)
 434                 crc32_begin(f);
 435
 436         /* apply size limit if limited packsize and not first object */
 437         if (!pack_size_limit || !nr_written)
 438                 limit = 0;
 439         else if (pack_size_limit <= write_offset)
 440                 /*
 441                  * the earlier object did not fit the limit; avoid
 442                  * mistaking this with unlimited (i.e. limit = 0).
 443                  */
 444                 limit = 1;
 445         else
 446                 limit = pack_size_limit - write_offset;
 447
 448         if (!entry->delta)
 449                 usable_delta = 0;       /* no delta */
 450         else if (!pack_size_limit)
 451                usable_delta = 1;        /* unlimited packfile */
 452         else if (entry->delta->idx.offset == (off_t)-1)
 453                 usable_delta = 0;       /* base was written to another pack */
 454         else if (entry->delta->idx.offset)
 455                 usable_delta = 1;       /* base already exists in this pack */
 456         else
 457                 usable_delta = 0;       /* base could end up in another pack */
 458
 459         if (!reuse_object)
 460                 to_reuse = 0;   /* explicit */
 461         else if (!entry->in_pack)
 462                 to_reuse = 0;   /* can't reuse what we don't have */
 463         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
 464                                 /* check_object() decided it for us ... */
 465                 to_reuse = usable_delta;
 466                                 /* ... but pack split may override that */
 467         else if (entry->type != entry->in_pack_type)
 468                 to_reuse = 0;   /* pack has delta which is unusable */
 469         else if (entry->delta)
 470                 to_reuse = 0;   /* we want to pack afresh */
 471         else
 472                 to_reuse = 1;   /* we have it in-pack undeltified,
 473                                  * and we do not need to deltify it.
 474                                  */
 475
 476         if (!to_reuse)
 477                 len = write_no_reuse_object(f, entry, limit, usable_delta);
 478         else
 479                 len = write_reuse_object(f, entry, limit, usable_delta);
 480         if (!len)
 481                 return 0;
 482
 483         if (usable_delta)
 484                 written_delta++;
 485         written++;
 486         if (!pack_to_stdout)
 487                 entry->idx.crc32 = crc32_end(f);
 488         return len;
 489 }
 490
 491 enum write_one_status {
 492         WRITE_ONE_SKIP = -1, /* already written */
 493         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 494         WRITE_ONE_WRITTEN = 1, /* normal */
 495         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 496 };
 497
 498 static enum write_one_status write_one(struct sha1file *f,
 499                                        struct object_entry *e,
 500                                        off_t *offset)
 501 {
 502         off_t size;
 503         int recursing;
 504
 505         /*
 506          * we set offset to 1 (which is an impossible value) to mark
 507          * the fact that this object is involved in "write its base
 508          * first before writing a deltified object" recursion.
 509          */
 510         recursing = (e->idx.offset == 1);
 511         if (recursing) {
 512                 warning("recursive delta detected for object %s",
 513                         sha1_to_hex(e->idx.sha1));
 514                 return WRITE_ONE_RECURSIVE;
 515         } else if (e->idx.offset || e->preferred_base) {
 516                 /* offset is non zero if object is written already. */
 517                 return WRITE_ONE_SKIP;
 518         }
 519
 520         /* if we are deltified, write out base object first. */
 521         if (e->delta) {
 522                 e->idx.offset = 1; /* now recurse */
 523                 switch (write_one(f, e->delta, offset)) {
 524                 case WRITE_ONE_RECURSIVE:
 525                         /* we cannot depend on this one */
 526                         e->delta = NULL;
 527                         break;
 528                 default:
 529                         break;
 530                 case WRITE_ONE_BREAK:
 531                         e->idx.offset = recursing;
 532                         return WRITE_ONE_BREAK;
 533                 }
 534         }
 535
 536         e->idx.offset = *offset;
 537         size = write_object(f, e, *offset);
 538         if (!size) {
 539                 e->idx.offset = recursing;
 540                 return WRITE_ONE_BREAK;
 541         }
 542         written_list[nr_written++] = &e->idx;
 543
 544         /* make sure off_t is sufficiently large not to wrap */
 545         if (signed_add_overflows(*offset, size))
 546                 die("pack too large for current definition of off_t");
 547         *offset += size;
 548         return WRITE_ONE_WRITTEN;
 549 }
 550
 551 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
 552                        void *cb_data)
 553 {
 554         unsigned char peeled[20];
 555         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
 556
 557         if (entry)
 558                 entry->tagged = 1;
 559         if (!peel_ref(path, peeled)) {
 560                 entry = packlist_find(&to_pack, peeled, NULL);
 561                 if (entry)
 562                         entry->tagged = 1;
 563         }
 564         return 0;
 565 }
 566
 567 static inline void add_to_write_order(struct object_entry **wo,
 568                                unsigned int *endp,
 569                                struct object_entry *e)
 570 {
 571         if (e->filled)
 572                 return;
 573         wo[(*endp)++] = e;
 574         e->filled = 1;
 575 }
 576
 577 static void add_descendants_to_write_order(struct object_entry **wo,
 578                                            unsigned int *endp,
 579                                            struct object_entry *e)
 580 {
 581         int add_to_order = 1;
 582         while (e) {
 583                 if (add_to_order) {
 584                         struct object_entry *s;
 585                         /* add this node... */
 586                         add_to_write_order(wo, endp, e);
 587                         /* all its siblings... */
 588                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
 589                                 add_to_write_order(wo, endp, s);
 590                         }
 591                 }
 592                 /* drop down a level to add left subtree nodes if possible */
 593                 if (e->delta_child) {
 594                         add_to_order = 1;
 595                         e = e->delta_child;
 596                 } else {
 597                         add_to_order = 0;
 598                         /* our sibling might have some children, it is next */
 599                         if (e->delta_sibling) {
 600                                 e = e->delta_sibling;
 601                                 continue;
 602                         }
 603                         /* go back to our parent node */
 604                         e = e->delta;
 605                         while (e && !e->delta_sibling) {
 606                                 /* we're on the right side of a subtree, keep
 607                                  * going up until we can go right again */
 608                                 e = e->delta;
 609                         }
 610                         if (!e) {
 611                                 /* done- we hit our original root node */
 612                                 return;
 613                         }
 614                         /* pass it off to sibling at this level */
 615                         e = e->delta_sibling;
 616                 }
 617         };
 618 }
 619
 620 static void add_family_to_write_order(struct object_entry **wo,
 621                                       unsigned int *endp,
 622                                       struct object_entry *e)
 623 {
 624         struct object_entry *root;
 625
 626         for (root = e; root->delta; root = root->delta)
 627                 ; /* nothing */
 628         add_descendants_to_write_order(wo, endp, root);
 629 }
 630
 631 static struct object_entry **compute_write_order(void)
 632 {
 633         unsigned int i, wo_end, last_untagged;
 634
 635         struct object_entry **wo;
 636         struct object_entry *objects = to_pack.objects;
 637
 638         for (i = 0; i < to_pack.nr_objects; i++) {
 639                 objects[i].tagged = 0;
 640                 objects[i].filled = 0;
 641                 objects[i].delta_child = NULL;
 642                 objects[i].delta_sibling = NULL;
 643         }
 644
 645         /*
 646          * Fully connect delta_child/delta_sibling network.
 647          * Make sure delta_sibling is sorted in the original
 648          * recency order.
 649          */
 650         for (i = to_pack.nr_objects; i > 0;) {
 651                 struct object_entry *e = &objects[--i];
 652                 if (!e->delta)
 653                         continue;
 654                 /* Mark me as the first child */
 655                 e->delta_sibling = e->delta->delta_child;
 656                 e->delta->delta_child = e;
 657         }
 658
 659         /*
 660          * Mark objects that are at the tip of tags.
 661          */
 662         for_each_tag_ref(mark_tagged, NULL);
 663
 664         /*
 665          * Give the objects in the original recency order until
 666          * we see a tagged tip.
 667          */
 668         ALLOC_ARRAY(wo, to_pack.nr_objects);
 669         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
 670                 if (objects[i].tagged)
 671                         break;
 672                 add_to_write_order(wo, &wo_end, &objects[i]);
 673         }
 674         last_untagged = i;
 675
 676         /*
 677          * Then fill all the tagged tips.
 678          */
 679         for (; i < to_pack.nr_objects; i++) {
 680                 if (objects[i].tagged)
 681                         add_to_write_order(wo, &wo_end, &objects[i]);
 682         }
 683
 684         /*
 685          * And then all remaining commits and tags.
 686          */
 687         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 688                 if (objects[i].type != OBJ_COMMIT &&
 689                     objects[i].type != OBJ_TAG)
 690                         continue;
 691                 add_to_write_order(wo, &wo_end, &objects[i]);
 692         }
 693
 694         /*
 695          * And then all the trees.
 696          */
 697         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 698                 if (objects[i].type != OBJ_TREE)
 699                         continue;
 700                 add_to_write_order(wo, &wo_end, &objects[i]);
 701         }
 702
 703         /*
 704          * Finally all the rest in really tight order
 705          */
 706         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 707                 if (!objects[i].filled)
 708                         add_family_to_write_order(wo, &wo_end, &objects[i]);
 709         }
 710
 711         if (wo_end != to_pack.nr_objects)
 712                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
 713
 714         return wo;
 715 }
 716
 717 static off_t write_reused_pack(struct sha1file *f)
 718 {
 719         unsigned char buffer[8192];
 720         off_t to_write, total;
 721         int fd;
 722
 723         if (!is_pack_valid(reuse_packfile))
 724                 die("packfile is invalid: %s", reuse_packfile->pack_name);
 725
 726         fd = git_open(reuse_packfile->pack_name);
 727         if (fd < 0)
 728                 die_errno("unable to open packfile for reuse: %s",
 729                           reuse_packfile->pack_name);
 730
 731         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
 732                 die_errno("unable to seek in reused packfile");
 733
 734         if (reuse_packfile_offset < 0)
 735                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
 736
 737         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
 738
 739         while (to_write) {
 740                 int read_pack = xread(fd, buffer, sizeof(buffer));
 741
 742                 if (read_pack <= 0)
 743                         die_errno("unable to read from reused packfile");
 744
 745                 if (read_pack > to_write)
 746                         read_pack = to_write;
 747
 748                 sha1write(f, buffer, read_pack);
 749                 to_write -= read_pack;
 750
 751                 /*
 752                  * We don't know the actual number of objects written,
 753                  * only how many bytes written, how many bytes total, and
 754                  * how many objects total. So we can fake it by pretending all
 755                  * objects we are writing are the same size. This gives us a
 756                  * smooth progress meter, and at the end it matches the true
 757                  * answer.
 758                  */
 759                 written = reuse_packfile_objects *
 760                                 (((double)(total - to_write)) / total);
 761                 display_progress(progress_state, written);
 762         }
 763
 764         close(fd);
 765         written = reuse_packfile_objects;
 766         display_progress(progress_state, written);
 767         return reuse_packfile_offset - sizeof(struct pack_header);
 768 }
 769
 770 static const char no_split_warning[] = N_(
 771 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
 772 );
 773
 774 static void write_pack_file(void)
 775 {
 776         uint32_t i = 0, j;
 777         struct sha1file *f;
 778         off_t offset;
 779         uint32_t nr_remaining = nr_result;
 780         time_t last_mtime = 0;
 781         struct object_entry **write_order;
 782
 783         if (progress > pack_to_stdout)
 784                 progress_state = start_progress(_("Writing objects"), nr_result);
 785         ALLOC_ARRAY(written_list, to_pack.nr_objects);
 786         write_order = compute_write_order();
 787
 788         do {
 789                 unsigned char sha1[20];
 790                 char *pack_tmp_name = NULL;
 791
 792                 if (pack_to_stdout)
 793                         f = sha1fd_throughput(1, "<stdout>", progress_state);
 794                 else
 795                         f = create_tmp_packfile(&pack_tmp_name);
 796
 797                 offset = write_pack_header(f, nr_remaining);
 798
 799                 if (reuse_packfile) {
 800                         off_t packfile_size;
 801                         assert(pack_to_stdout);
 802
 803                         packfile_size = write_reused_pack(f);
 804                         offset += packfile_size;
 805                 }
 806
 807                 nr_written = 0;
 808                 for (; i < to_pack.nr_objects; i++) {
 809                         struct object_entry *e = write_order[i];
 810                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 811                                 break;
 812                         display_progress(progress_state, written);
 813                 }
 814
 815                 /*
 816                  * Did we write the wrong # entries in the header?
 817                  * If so, rewrite it like in fast-import
 818                  */
 819                 if (pack_to_stdout) {
 820                         sha1close(f, sha1, CSUM_CLOSE);
 821                 } else if (nr_written == nr_remaining) {
 822                         sha1close(f, sha1, CSUM_FSYNC);
 823                 } else {
 824                         int fd = sha1close(f, sha1, 0);
 825                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
 826                                                  nr_written, sha1, offset);
 827                         close(fd);
 828                         if (write_bitmap_index) {
 829                                 warning(_(no_split_warning));
 830                                 write_bitmap_index = 0;
 831                         }
 832                 }
 833
 834                 if (!pack_to_stdout) {
 835                         struct stat st;
 836                         struct strbuf tmpname = STRBUF_INIT;
 837
 838                         /*
 839                          * Packs are runtime accessed in their mtime
 840                          * order since newer packs are more likely to contain
 841                          * younger objects.  So if we are creating multiple
 842                          * packs then we should modify the mtime of later ones
 843                          * to preserve this property.
 844                          */
 845                         if (stat(pack_tmp_name, &st) < 0) {
 846                                 warning_errno("failed to stat %s", pack_tmp_name);
 847                         } else if (!last_mtime) {
 848                                 last_mtime = st.st_mtime;
 849                         } else {
 850                                 struct utimbuf utb;
 851                                 utb.actime = st.st_atime;
 852                                 utb.modtime = --last_mtime;
 853                                 if (utime(pack_tmp_name, &utb) < 0)
 854                                         warning_errno("failed utime() on %s", pack_tmp_name);
 855                         }
 856
 857                         strbuf_addf(&tmpname, "%s-", base_name);
 858
 859                         if (write_bitmap_index) {
 860                                 bitmap_writer_set_checksum(sha1);
 861                                 bitmap_writer_build_type_index(written_list, nr_written);
 862                         }
 863
 864                         finish_tmp_packfile(&tmpname, pack_tmp_name,
 865                                             written_list, nr_written,
 866                                             &pack_idx_opts, sha1);
 867
 868                         if (write_bitmap_index) {
 869                                 strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
 870
 871                                 stop_progress(&progress_state);
 872
 873                                 bitmap_writer_show_progress(progress);
 874                                 bitmap_writer_reuse_bitmaps(&to_pack);
 875                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
 876                                 bitmap_writer_build(&to_pack);
 877                                 bitmap_writer_finish(written_list, nr_written,
 878                                                      tmpname.buf, write_bitmap_options);
 879                                 write_bitmap_index = 0;
 880                         }
 881
 882                         strbuf_release(&tmpname);
 883                         free(pack_tmp_name);
 884                         puts(sha1_to_hex(sha1));
 885                 }
 886
 887                 /* mark written objects as written to previous pack */
 888                 for (j = 0; j < nr_written; j++) {
 889                         written_list[j]->offset = (off_t)-1;
 890                 }
 891                 nr_remaining -= nr_written;
 892         } while (nr_remaining && i < to_pack.nr_objects);
 893
 894         free(written_list);
 895         free(write_order);
 896         stop_progress(&progress_state);
 897         if (written != nr_result)
 898                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
 899                         written, nr_result);
 900 }
 901
 902 static int no_try_delta(const char *path)
 903 {
 904         static struct attr_check *check;
 905
 906         if (!check)
 907                 check = attr_check_initl("delta", NULL);
 908         if (git_check_attr(path, check))
 909                 return 0;
 910         if (ATTR_FALSE(check->items[0].value))
 911                 return 1;
 912         return 0;
 913 }
 914
 915 /*
 916  * When adding an object, check whether we have already added it
 917  * to our packing list. If so, we can skip. However, if we are
 918  * being asked to excludei t, but the previous mention was to include
 919  * it, make sure to adjust its flags and tweak our numbers accordingly.
 920  *
 921  * As an optimization, we pass out the index position where we would have
 922  * found the item, since that saves us from having to look it up again a
 923  * few lines later when we want to add the new entry.
 924  */
 925 static int have_duplicate_entry(const unsigned char *sha1,
 926                                 int exclude,
 927                                 uint32_t *index_pos)
 928 {
 929         struct object_entry *entry;
 930
 931         entry = packlist_find(&to_pack, sha1, index_pos);
 932         if (!entry)
 933                 return 0;
 934
 935         if (exclude) {
 936                 if (!entry->preferred_base)
 937                         nr_result--;
 938                 entry->preferred_base = 1;
 939         }
 940
 941         return 1;
 942 }
 943
 944 static int want_found_object(int exclude, struct packed_git *p)
 945 {
 946         if (exclude)
 947                 return 1;
 948         if (incremental)
 949                 return 0;
 950
 951         /*
 952          * When asked to do --local (do not include an object that appears in a
 953          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
 954          * an object that appears in a pack marked with .keep), finding a pack
 955          * that matches the criteria is sufficient for us to decide to omit it.
 956          * However, even if this pack does not satisfy the criteria, we need to
 957          * make sure no copy of this object appears in _any_ pack that makes us
 958          * to omit the object, so we need to check all the packs.
 959          *
 960          * We can however first check whether these options can possible matter;
 961          * if they do not matter we know we want the object in generated pack.
 962          * Otherwise, we signal "-1" at the end to tell the caller that we do
 963          * not know either way, and it needs to check more packs.
 964          */
 965         if (!ignore_packed_keep &&
 966             (!local || !have_non_local_packs))
 967                 return 1;
 968
 969         if (local && !p->pack_local)
 970                 return 0;
 971         if (ignore_packed_keep && p->pack_local && p->pack_keep)
 972                 return 0;
 973
 974         /* we don't know yet; keep looking for more packs */
 975         return -1;
 976 }
 977
 978 /*
 979  * Check whether we want the object in the pack (e.g., we do not want
 980  * objects found in non-local stores if the "--local" option was used).
 981  *
 982  * If the caller already knows an existing pack it wants to take the object
 983  * from, that is passed in *found_pack and *found_offset; otherwise this
 984  * function finds if there is any pack that has the object and returns the pack
 985  * and its offset in these variables.
 986  */
 987 static int want_object_in_pack(const unsigned char *sha1,
 988                                int exclude,
 989                                struct packed_git **found_pack,
 990                                off_t *found_offset)
 991 {
 992         struct mru_entry *entry;
 993         int want;
 994
 995         if (!exclude && local && has_loose_object_nonlocal(sha1))
 996                 return 0;
 997
 998         /*
 999          * If we already know the pack object lives in, start checks from that
1000          * pack - in the usual case when neither --local was given nor .keep files
1001          * are present we will determine the answer right now.
1002          */
1003         if (*found_pack) {
1004                 want = want_found_object(exclude, *found_pack);
1005                 if (want != -1)
1006                         return want;
1007         }
1008
1009         for (entry = packed_git_mru->head; entry; entry = entry->next) {
1010                 struct packed_git *p = entry->item;
1011                 off_t offset;
1012
1013                 if (p == *found_pack)
1014                         offset = *found_offset;
1015                 else
1016                         offset = find_pack_entry_one(sha1, p);
1017
1018                 if (offset) {
1019                         if (!*found_pack) {
1020                                 if (!is_pack_valid(p))
1021                                         continue;
1022                                 *found_offset = offset;
1023                                 *found_pack = p;
1024                         }
1025                         want = want_found_object(exclude, p);
1026                         if (!exclude && want > 0)
1027                                 mru_mark(packed_git_mru, entry);
1028                         if (want != -1)
1029                                 return want;
1030                 }
1031         }
1032
1033         return 1;
1034 }
1035
1036 static void create_object_entry(const unsigned char *sha1,
1037                                 enum object_type type,
1038                                 uint32_t hash,
1039                                 int exclude,
1040                                 int no_try_delta,
1041                                 uint32_t index_pos,
1042                                 struct packed_git *found_pack,
1043                                 off_t found_offset)
1044 {
1045         struct object_entry *entry;
1046
1047         entry = packlist_alloc(&to_pack, sha1, index_pos);
1048         entry->hash = hash;
1049         if (type)
1050                 entry->type = type;
1051         if (exclude)
1052                 entry->preferred_base = 1;
1053         else
1054                 nr_result++;
1055         if (found_pack) {
1056                 entry->in_pack = found_pack;
1057                 entry->in_pack_offset = found_offset;
1058         }
1059
1060         entry->no_try_delta = no_try_delta;
1061 }
1062
1063 static const char no_closure_warning[] = N_(
1064 "disabling bitmap writing, as some objects are not being packed"
1065 );
1066
1067 static int add_object_entry(const unsigned char *sha1, enum object_type type,
1068                             const char *name, int exclude)
1069 {
1070         struct packed_git *found_pack = NULL;
1071         off_t found_offset = 0;
1072         uint32_t index_pos;
1073
1074         if (have_duplicate_entry(sha1, exclude, &index_pos))
1075                 return 0;
1076
1077         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1078                 /* The pack is missing an object, so it will not have closure */
1079                 if (write_bitmap_index) {
1080                         warning(_(no_closure_warning));
1081                         write_bitmap_index = 0;
1082                 }
1083                 return 0;
1084         }
1085
1086         create_object_entry(sha1, type, pack_name_hash(name),
1087                             exclude, name && no_try_delta(name),
1088                             index_pos, found_pack, found_offset);
1089
1090         display_progress(progress_state, nr_result);
1091         return 1;
1092 }
1093
1094 static int add_object_entry_from_bitmap(const unsigned char *sha1,
1095                                         enum object_type type,
1096                                         int flags, uint32_t name_hash,
1097                                         struct packed_git *pack, off_t offset)
1098 {
1099         uint32_t index_pos;
1100
1101         if (have_duplicate_entry(sha1, 0, &index_pos))
1102                 return 0;
1103
1104         if (!want_object_in_pack(sha1, 0, &pack, &offset))
1105                 return 0;
1106
1107         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1108
1109         display_progress(progress_state, nr_result);
1110         return 1;
1111 }
1112
1113 struct pbase_tree_cache {
1114         unsigned char sha1[20];
1115         int ref;
1116         int temporary;
1117         void *tree_data;
1118         unsigned long tree_size;
1119 };
1120
1121 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1122 static int pbase_tree_cache_ix(const unsigned char *sha1)
1123 {
1124         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1125 }
1126 static int pbase_tree_cache_ix_incr(int ix)
1127 {
1128         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1129 }
1130
1131 static struct pbase_tree {
1132         struct pbase_tree *next;
1133         /* This is a phony "cache" entry; we are not
1134          * going to evict it or find it through _get()
1135          * mechanism -- this is for the toplevel node that
1136          * would almost always change with any commit.
1137          */
1138         struct pbase_tree_cache pcache;
1139 } *pbase_tree;
1140
1141 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1142 {
1143         struct pbase_tree_cache *ent, *nent;
1144         void *data;
1145         unsigned long size;
1146         enum object_type type;
1147         int neigh;
1148         int my_ix = pbase_tree_cache_ix(sha1);
1149         int available_ix = -1;
1150
1151         /* pbase-tree-cache acts as a limited hashtable.
1152          * your object will be found at your index or within a few
1153          * slots after that slot if it is cached.
1154          */
1155         for (neigh = 0; neigh < 8; neigh++) {
1156                 ent = pbase_tree_cache[my_ix];
1157                 if (ent && !hashcmp(ent->sha1, sha1)) {
1158                         ent->ref++;
1159                         return ent;
1160                 }
1161                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1162                          ((0 <= available_ix) &&
1163                           (!ent && pbase_tree_cache[available_ix])))
1164                         available_ix = my_ix;
1165                 if (!ent)
1166                         break;
1167                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1168         }
1169
1170         /* Did not find one.  Either we got a bogus request or
1171          * we need to read and perhaps cache.
1172          */
1173         data = read_sha1_file(sha1, &type, &size);
1174         if (!data)
1175                 return NULL;
1176         if (type != OBJ_TREE) {
1177                 free(data);
1178                 return NULL;
1179         }
1180
1181         /* We need to either cache or return a throwaway copy */
1182
1183         if (available_ix < 0)
1184                 ent = NULL;
1185         else {
1186                 ent = pbase_tree_cache[available_ix];
1187                 my_ix = available_ix;
1188         }
1189
1190         if (!ent) {
1191                 nent = xmalloc(sizeof(*nent));
1192                 nent->temporary = (available_ix < 0);
1193         }
1194         else {
1195                 /* evict and reuse */
1196                 free(ent->tree_data);
1197                 nent = ent;
1198         }
1199         hashcpy(nent->sha1, sha1);
1200         nent->tree_data = data;
1201         nent->tree_size = size;
1202         nent->ref = 1;
1203         if (!nent->temporary)
1204                 pbase_tree_cache[my_ix] = nent;
1205         return nent;
1206 }
1207
1208 static void pbase_tree_put(struct pbase_tree_cache *cache)
1209 {
1210         if (!cache->temporary) {
1211                 cache->ref--;
1212                 return;
1213         }
1214         free(cache->tree_data);
1215         free(cache);
1216 }
1217
1218 static int name_cmp_len(const char *name)
1219 {
1220         int i;
1221         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1222                 ;
1223         return i;
1224 }
1225
1226 static void add_pbase_object(struct tree_desc *tree,
1227                              const char *name,
1228                              int cmplen,
1229                              const char *fullname)
1230 {
1231         struct name_entry entry;
1232         int cmp;
1233
1234         while (tree_entry(tree,&entry)) {
1235                 if (S_ISGITLINK(entry.mode))
1236                         continue;
1237                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1238                       memcmp(name, entry.path, cmplen);
1239                 if (cmp > 0)
1240                         continue;
1241                 if (cmp < 0)
1242                         return;
1243                 if (name[cmplen] != '/') {
1244                         add_object_entry(entry.oid->hash,
1245                                          object_type(entry.mode),
1246                                          fullname, 1);
1247                         return;
1248                 }
1249                 if (S_ISDIR(entry.mode)) {
1250                         struct tree_desc sub;
1251                         struct pbase_tree_cache *tree;
1252                         const char *down = name+cmplen+1;
1253                         int downlen = name_cmp_len(down);
1254
1255                         tree = pbase_tree_get(entry.oid->hash);
1256                         if (!tree)
1257                                 return;
1258                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1259
1260                         add_pbase_object(&sub, down, downlen, fullname);
1261                         pbase_tree_put(tree);
1262                 }
1263         }
1264 }
1265
1266 static unsigned *done_pbase_paths;
1267 static int done_pbase_paths_num;
1268 static int done_pbase_paths_alloc;
1269 static int done_pbase_path_pos(unsigned hash)
1270 {
1271         int lo = 0;
1272         int hi = done_pbase_paths_num;
1273         while (lo < hi) {
1274                 int mi = (hi + lo) / 2;
1275                 if (done_pbase_paths[mi] == hash)
1276                         return mi;
1277                 if (done_pbase_paths[mi] < hash)
1278                         hi = mi;
1279                 else
1280                         lo = mi + 1;
1281         }
1282         return -lo-1;
1283 }
1284
1285 static int check_pbase_path(unsigned hash)
1286 {
1287         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1288         if (0 <= pos)
1289                 return 1;
1290         pos = -pos - 1;
1291         ALLOC_GROW(done_pbase_paths,
1292                    done_pbase_paths_num + 1,
1293                    done_pbase_paths_alloc);
1294         done_pbase_paths_num++;
1295         if (pos < done_pbase_paths_num)
1296                 memmove(done_pbase_paths + pos + 1,
1297                         done_pbase_paths + pos,
1298                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1299         done_pbase_paths[pos] = hash;
1300         return 0;
1301 }
1302
1303 static void add_preferred_base_object(const char *name)
1304 {
1305         struct pbase_tree *it;
1306         int cmplen;
1307         unsigned hash = pack_name_hash(name);
1308
1309         if (!num_preferred_base || check_pbase_path(hash))
1310                 return;
1311
1312         cmplen = name_cmp_len(name);
1313         for (it = pbase_tree; it; it = it->next) {
1314                 if (cmplen == 0) {
1315                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1316                 }
1317                 else {
1318                         struct tree_desc tree;
1319                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1320                         add_pbase_object(&tree, name, cmplen, name);
1321                 }
1322         }
1323 }
1324
1325 static void add_preferred_base(unsigned char *sha1)
1326 {
1327         struct pbase_tree *it;
1328         void *data;
1329         unsigned long size;
1330         unsigned char tree_sha1[20];
1331
1332         if (window <= num_preferred_base++)
1333                 return;
1334
1335         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1336         if (!data)
1337                 return;
1338
1339         for (it = pbase_tree; it; it = it->next) {
1340                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1341                         free(data);
1342                         return;
1343                 }
1344         }
1345
1346         it = xcalloc(1, sizeof(*it));
1347         it->next = pbase_tree;
1348         pbase_tree = it;
1349
1350         hashcpy(it->pcache.sha1, tree_sha1);
1351         it->pcache.tree_data = data;
1352         it->pcache.tree_size = size;
1353 }
1354
1355 static void cleanup_preferred_base(void)
1356 {
1357         struct pbase_tree *it;
1358         unsigned i;
1359
1360         it = pbase_tree;
1361         pbase_tree = NULL;
1362         while (it) {
1363                 struct pbase_tree *this = it;
1364                 it = this->next;
1365                 free(this->pcache.tree_data);
1366                 free(this);
1367         }
1368
1369         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1370                 if (!pbase_tree_cache[i])
1371                         continue;
1372                 free(pbase_tree_cache[i]->tree_data);
1373                 free(pbase_tree_cache[i]);
1374                 pbase_tree_cache[i] = NULL;
1375         }
1376
1377         free(done_pbase_paths);
1378         done_pbase_paths = NULL;
1379         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1380 }
1381
1382 static void check_object(struct object_entry *entry)
1383 {
1384         if (entry->in_pack) {
1385                 struct packed_git *p = entry->in_pack;
1386                 struct pack_window *w_curs = NULL;
1387                 const unsigned char *base_ref = NULL;
1388                 struct object_entry *base_entry;
1389                 unsigned long used, used_0;
1390                 unsigned long avail;
1391                 off_t ofs;
1392                 unsigned char *buf, c;
1393
1394                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1395
1396                 /*
1397                  * We want in_pack_type even if we do not reuse delta
1398                  * since non-delta representations could still be reused.
1399                  */
1400                 used = unpack_object_header_buffer(buf, avail,
1401                                                    &entry->in_pack_type,
1402                                                    &entry->size);
1403                 if (used == 0)
1404                         goto give_up;
1405
1406                 /*
1407                  * Determine if this is a delta and if so whether we can
1408                  * reuse it or not.  Otherwise let's find out as cheaply as
1409                  * possible what the actual type and size for this object is.
1410                  */
1411                 switch (entry->in_pack_type) {
1412                 default:
1413                         /* Not a delta hence we've already got all we need. */
1414                         entry->type = entry->in_pack_type;
1415                         entry->in_pack_header_size = used;
1416                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1417                                 goto give_up;
1418                         unuse_pack(&w_curs);
1419                         return;
1420                 case OBJ_REF_DELTA:
1421                         if (reuse_delta && !entry->preferred_base)
1422                                 base_ref = use_pack(p, &w_curs,
1423                                                 entry->in_pack_offset + used, NULL);
1424                         entry->in_pack_header_size = used + 20;
1425                         break;
1426                 case OBJ_OFS_DELTA:
1427                         buf = use_pack(p, &w_curs,
1428                                        entry->in_pack_offset + used, NULL);
1429                         used_0 = 0;
1430                         c = buf[used_0++];
1431                         ofs = c & 127;
1432                         while (c & 128) {
1433                                 ofs += 1;
1434                                 if (!ofs || MSB(ofs, 7)) {
1435                                         error("delta base offset overflow in pack for %s",
1436                                               sha1_to_hex(entry->idx.sha1));
1437                                         goto give_up;
1438                                 }
1439                                 c = buf[used_0++];
1440                                 ofs = (ofs << 7) + (c & 127);
1441                         }
1442                         ofs = entry->in_pack_offset - ofs;
1443                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1444                                 error("delta base offset out of bound for %s",
1445                                       sha1_to_hex(entry->idx.sha1));
1446                                 goto give_up;
1447                         }
1448                         if (reuse_delta && !entry->preferred_base) {
1449                                 struct revindex_entry *revidx;
1450                                 revidx = find_pack_revindex(p, ofs);
1451                                 if (!revidx)
1452                                         goto give_up;
1453                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1454                         }
1455                         entry->in_pack_header_size = used + used_0;
1456                         break;
1457                 }
1458
1459                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1460                         /*
1461                          * If base_ref was set above that means we wish to
1462                          * reuse delta data, and we even found that base
1463                          * in the list of objects we want to pack. Goodie!
1464                          *
1465                          * Depth value does not matter - find_deltas() will
1466                          * never consider reused delta as the base object to
1467                          * deltify other objects against, in order to avoid
1468                          * circular deltas.
1469                          */
1470                         entry->type = entry->in_pack_type;
1471                         entry->delta = base_entry;
1472                         entry->delta_size = entry->size;
1473                         entry->delta_sibling = base_entry->delta_child;
1474                         base_entry->delta_child = entry;
1475                         unuse_pack(&w_curs);
1476                         return;
1477                 }
1478
1479                 if (entry->type) {
1480                         /*
1481                          * This must be a delta and we already know what the
1482                          * final object type is.  Let's extract the actual
1483                          * object size from the delta header.
1484                          */
1485                         entry->size = get_size_from_delta(p, &w_curs,
1486                                         entry->in_pack_offset + entry->in_pack_header_size);
1487                         if (entry->size == 0)
1488                                 goto give_up;
1489                         unuse_pack(&w_curs);
1490                         return;
1491                 }
1492
1493                 /*
1494                  * No choice but to fall back to the recursive delta walk
1495                  * with sha1_object_info() to find about the object type
1496                  * at this point...
1497                  */
1498                 give_up:
1499                 unuse_pack(&w_curs);
1500         }
1501
1502         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1503         /*
1504          * The error condition is checked in prepare_pack().  This is
1505          * to permit a missing preferred base object to be ignored
1506          * as a preferred base.  Doing so can result in a larger
1507          * pack file, but the transfer will still take place.
1508          */
1509 }
1510
1511 static int pack_offset_sort(const void *_a, const void *_b)
1512 {
1513         const struct object_entry *a = *(struct object_entry **)_a;
1514         const struct object_entry *b = *(struct object_entry **)_b;
1515
1516         /* avoid filesystem trashing with loose objects */
1517         if (!a->in_pack && !b->in_pack)
1518                 return hashcmp(a->idx.sha1, b->idx.sha1);
1519
1520         if (a->in_pack < b->in_pack)
1521                 return -1;
1522         if (a->in_pack > b->in_pack)
1523                 return 1;
1524         return a->in_pack_offset < b->in_pack_offset ? -1 :
1525                         (a->in_pack_offset > b->in_pack_offset);
1526 }
1527
1528 /*
1529  * Drop an on-disk delta we were planning to reuse. Naively, this would
1530  * just involve blanking out the "delta" field, but we have to deal
1531  * with some extra book-keeping:
1532  *
1533  *   1. Removing ourselves from the delta_sibling linked list.
1534  *
1535  *   2. Updating our size/type to the non-delta representation. These were
1536  *      either not recorded initially (size) or overwritten with the delta type
1537  *      (type) when check_object() decided to reuse the delta.
1538  *
1539  *   3. Resetting our delta depth, as we are now a base object.
1540  */
1541 static void drop_reused_delta(struct object_entry *entry)
1542 {
1543         struct object_entry **p = &entry->delta->delta_child;
1544         struct object_info oi = OBJECT_INFO_INIT;
1545
1546         while (*p) {
1547                 if (*p == entry)
1548                         *p = (*p)->delta_sibling;
1549                 else
1550                         p = &(*p)->delta_sibling;
1551         }
1552         entry->delta = NULL;
1553         entry->depth = 0;
1554
1555         oi.sizep = &entry->size;
1556         oi.typep = &entry->type;
1557         if (packed_object_info(entry->in_pack, entry->in_pack_offset, &oi) < 0) {
1558                 /*
1559                  * We failed to get the info from this pack for some reason;
1560                  * fall back to sha1_object_info, which may find another copy.
1561                  * And if that fails, the error will be recorded in entry->type
1562                  * and dealt with in prepare_pack().
1563                  */
1564                 entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1565         }
1566 }
1567
1568 /*
1569  * Follow the chain of deltas from this entry onward, throwing away any links
1570  * that cause us to hit a cycle (as determined by the DFS state flags in
1571  * the entries).
1572  *
1573  * We also detect too-long reused chains that would violate our --depth
1574  * limit.
1575  */
1576 static void break_delta_chains(struct object_entry *entry)
1577 {
1578         /*
1579          * The actual depth of each object we will write is stored as an int,
1580          * as it cannot exceed our int "depth" limit. But before we break
1581          * changes based no that limit, we may potentially go as deep as the
1582          * number of objects, which is elsewhere bounded to a uint32_t.
1583          */
1584         uint32_t total_depth;
1585         struct object_entry *cur, *next;
1586
1587         for (cur = entry, total_depth = 0;
1588              cur;
1589              cur = cur->delta, total_depth++) {
1590                 if (cur->dfs_state == DFS_DONE) {
1591                         /*
1592                          * We've already seen this object and know it isn't
1593                          * part of a cycle. We do need to append its depth
1594                          * to our count.
1595                          */
1596                         total_depth += cur->depth;
1597                         break;
1598                 }
1599
1600                 /*
1601                  * We break cycles before looping, so an ACTIVE state (or any
1602                  * other cruft which made its way into the state variable)
1603                  * is a bug.
1604                  */
1605                 if (cur->dfs_state != DFS_NONE)
1606                         die("BUG: confusing delta dfs state in first pass: %d",
1607                             cur->dfs_state);
1608
1609                 /*
1610                  * Now we know this is the first time we've seen the object. If
1611                  * it's not a delta, we're done traversing, but we'll mark it
1612                  * done to save time on future traversals.
1613                  */
1614                 if (!cur->delta) {
1615                         cur->dfs_state = DFS_DONE;
1616                         break;
1617                 }
1618
1619                 /*
1620                  * Mark ourselves as active and see if the next step causes
1621                  * us to cycle to another active object. It's important to do
1622                  * this _before_ we loop, because it impacts where we make the
1623                  * cut, and thus how our total_depth counter works.
1624                  * E.g., We may see a partial loop like:
1625                  *
1626                  *   A -> B -> C -> D -> B
1627                  *
1628                  * Cutting B->C breaks the cycle. But now the depth of A is
1629                  * only 1, and our total_depth counter is at 3. The size of the
1630                  * error is always one less than the size of the cycle we
1631                  * broke. Commits C and D were "lost" from A's chain.
1632                  *
1633                  * If we instead cut D->B, then the depth of A is correct at 3.
1634                  * We keep all commits in the chain that we examined.
1635                  */
1636                 cur->dfs_state = DFS_ACTIVE;
1637                 if (cur->delta->dfs_state == DFS_ACTIVE) {
1638                         drop_reused_delta(cur);
1639                         cur->dfs_state = DFS_DONE;
1640                         break;
1641                 }
1642         }
1643
1644         /*
1645          * And now that we've gone all the way to the bottom of the chain, we
1646          * need to clear the active flags and set the depth fields as
1647          * appropriate. Unlike the loop above, which can quit when it drops a
1648          * delta, we need to keep going to look for more depth cuts. So we need
1649          * an extra "next" pointer to keep going after we reset cur->delta.
1650          */
1651         for (cur = entry; cur; cur = next) {
1652                 next = cur->delta;
1653
1654                 /*
1655                  * We should have a chain of zero or more ACTIVE states down to
1656                  * a final DONE. We can quit after the DONE, because either it
1657                  * has no bases, or we've already handled them in a previous
1658                  * call.
1659                  */
1660                 if (cur->dfs_state == DFS_DONE)
1661                         break;
1662                 else if (cur->dfs_state != DFS_ACTIVE)
1663                         die("BUG: confusing delta dfs state in second pass: %d",
1664                             cur->dfs_state);
1665
1666                 /*
1667                  * If the total_depth is more than depth, then we need to snip
1668                  * the chain into two or more smaller chains that don't exceed
1669                  * the maximum depth. Most of the resulting chains will contain
1670                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1671                  * the last chain (i.e., the one containing entry) will contain
1672                  * whatever entries are left over, namely
1673                  * (total_depth % (depth + 1)) of them.
1674                  *
1675                  * Since we are iterating towards decreasing depth, we need to
1676                  * decrement total_depth as we go, and we need to write to the
1677                  * entry what its final depth will be after all of the
1678                  * snipping. Since we're snipping into chains of length (depth
1679                  * + 1) entries, the final depth of an entry will be its
1680                  * original depth modulo (depth + 1). Any time we encounter an
1681                  * entry whose final depth is supposed to be zero, we snip it
1682                  * from its delta base, thereby making it so.
1683                  */
1684                 cur->depth = (total_depth--) % (depth + 1);
1685                 if (!cur->depth)
1686                         drop_reused_delta(cur);
1687
1688                 cur->dfs_state = DFS_DONE;
1689         }
1690 }
1691
1692 static void get_object_details(void)
1693 {
1694         uint32_t i;
1695         struct object_entry **sorted_by_offset;
1696
1697         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1698         for (i = 0; i < to_pack.nr_objects; i++)
1699                 sorted_by_offset[i] = to_pack.objects + i;
1700         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1701
1702         for (i = 0; i < to_pack.nr_objects; i++) {
1703                 struct object_entry *entry = sorted_by_offset[i];
1704                 check_object(entry);
1705                 if (big_file_threshold < entry->size)
1706                         entry->no_try_delta = 1;
1707         }
1708
1709         /*
1710          * This must happen in a second pass, since we rely on the delta
1711          * information for the whole list being completed.
1712          */
1713         for (i = 0; i < to_pack.nr_objects; i++)
1714                 break_delta_chains(&to_pack.objects[i]);
1715
1716         free(sorted_by_offset);
1717 }
1718
1719 /*
1720  * We search for deltas in a list sorted by type, by filename hash, and then
1721  * by size, so that we see progressively smaller and smaller files.
1722  * That's because we prefer deltas to be from the bigger file
1723  * to the smaller -- deletes are potentially cheaper, but perhaps
1724  * more importantly, the bigger file is likely the more recent
1725  * one.  The deepest deltas are therefore the oldest objects which are
1726  * less susceptible to be accessed often.
1727  */
1728 static int type_size_sort(const void *_a, const void *_b)
1729 {
1730         const struct object_entry *a = *(struct object_entry **)_a;
1731         const struct object_entry *b = *(struct object_entry **)_b;
1732
1733         if (a->type > b->type)
1734                 return -1;
1735         if (a->type < b->type)
1736                 return 1;
1737         if (a->hash > b->hash)
1738                 return -1;
1739         if (a->hash < b->hash)
1740                 return 1;
1741         if (a->preferred_base > b->preferred_base)
1742                 return -1;
1743         if (a->preferred_base < b->preferred_base)
1744                 return 1;
1745         if (a->size > b->size)
1746                 return -1;
1747         if (a->size < b->size)
1748                 return 1;
1749         return a < b ? -1 : (a > b);  /* newest first */
1750 }
1751
1752 struct unpacked {
1753         struct object_entry *entry;
1754         void *data;
1755         struct delta_index *index;
1756         unsigned depth;
1757 };
1758
1759 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1760                            unsigned long delta_size)
1761 {
1762         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1763                 return 0;
1764
1765         if (delta_size < cache_max_small_delta_size)
1766                 return 1;
1767
1768         /* cache delta, if objects are large enough compared to delta size */
1769         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1770                 return 1;
1771
1772         return 0;
1773 }
1774
1775 #ifndef NO_PTHREADS
1776
1777 static pthread_mutex_t read_mutex;
1778 #define read_lock()             pthread_mutex_lock(&read_mutex)
1779 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1780
1781 static pthread_mutex_t cache_mutex;
1782 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1783 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1784
1785 static pthread_mutex_t progress_mutex;
1786 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1787 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1788
1789 #else
1790
1791 #define read_lock()             (void)0
1792 #define read_unlock()           (void)0
1793 #define cache_lock()            (void)0
1794 #define cache_unlock()          (void)0
1795 #define progress_lock()         (void)0
1796 #define progress_unlock()       (void)0
1797
1798 #endif
1799
1800 static int try_delta(struct unpacked *trg, struct unpacked *src,
1801                      unsigned max_depth, unsigned long *mem_usage)
1802 {
1803         struct object_entry *trg_entry = trg->entry;
1804         struct object_entry *src_entry = src->entry;
1805         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1806         unsigned ref_depth;
1807         enum object_type type;
1808         void *delta_buf;
1809
1810         /* Don't bother doing diffs between different types */
1811         if (trg_entry->type != src_entry->type)
1812                 return -1;
1813
1814         /*
1815          * We do not bother to try a delta that we discarded on an
1816          * earlier try, but only when reusing delta data.  Note that
1817          * src_entry that is marked as the preferred_base should always
1818          * be considered, as even if we produce a suboptimal delta against
1819          * it, we will still save the transfer cost, as we already know
1820          * the other side has it and we won't send src_entry at all.
1821          */
1822         if (reuse_delta && trg_entry->in_pack &&
1823             trg_entry->in_pack == src_entry->in_pack &&
1824             !src_entry->preferred_base &&
1825             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1826             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1827                 return 0;
1828
1829         /* Let's not bust the allowed depth. */
1830         if (src->depth >= max_depth)
1831                 return 0;
1832
1833         /* Now some size filtering heuristics. */
1834         trg_size = trg_entry->size;
1835         if (!trg_entry->delta) {
1836                 max_size = trg_size/2 - 20;
1837                 ref_depth = 1;
1838         } else {
1839                 max_size = trg_entry->delta_size;
1840                 ref_depth = trg->depth;
1841         }
1842         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1843                                                 (max_depth - ref_depth + 1);
1844         if (max_size == 0)
1845                 return 0;
1846         src_size = src_entry->size;
1847         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1848         if (sizediff >= max_size)
1849                 return 0;
1850         if (trg_size < src_size / 32)
1851                 return 0;
1852
1853         /* Load data if not already done */
1854         if (!trg->data) {
1855                 read_lock();
1856                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1857                 read_unlock();
1858                 if (!trg->data)
1859                         die("object %s cannot be read",
1860                             sha1_to_hex(trg_entry->idx.sha1));
1861                 if (sz != trg_size)
1862                         die("object %s inconsistent object length (%lu vs %lu)",
1863                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1864                 *mem_usage += sz;
1865         }
1866         if (!src->data) {
1867                 read_lock();
1868                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1869                 read_unlock();
1870                 if (!src->data) {
1871                         if (src_entry->preferred_base) {
1872                                 static int warned = 0;
1873                                 if (!warned++)
1874                                         warning("object %s cannot be read",
1875                                                 sha1_to_hex(src_entry->idx.sha1));
1876                                 /*
1877                                  * Those objects are not included in the
1878                                  * resulting pack.  Be resilient and ignore
1879                                  * them if they can't be read, in case the
1880                                  * pack could be created nevertheless.
1881                                  */
1882                                 return 0;
1883                         }
1884                         die("object %s cannot be read",
1885                             sha1_to_hex(src_entry->idx.sha1));
1886                 }
1887                 if (sz != src_size)
1888                         die("object %s inconsistent object length (%lu vs %lu)",
1889                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1890                 *mem_usage += sz;
1891         }
1892         if (!src->index) {
1893                 src->index = create_delta_index(src->data, src_size);
1894                 if (!src->index) {
1895                         static int warned = 0;
1896                         if (!warned++)
1897                                 warning("suboptimal pack - out of memory");
1898                         return 0;
1899                 }
1900                 *mem_usage += sizeof_delta_index(src->index);
1901         }
1902
1903         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1904         if (!delta_buf)
1905                 return 0;
1906
1907         if (trg_entry->delta) {
1908                 /* Prefer only shallower same-sized deltas. */
1909                 if (delta_size == trg_entry->delta_size &&
1910                     src->depth + 1 >= trg->depth) {
1911                         free(delta_buf);
1912                         return 0;
1913                 }
1914         }
1915
1916         /*
1917          * Handle memory allocation outside of the cache
1918          * accounting lock.  Compiler will optimize the strangeness
1919          * away when NO_PTHREADS is defined.
1920          */
1921         free(trg_entry->delta_data);
1922         cache_lock();
1923         if (trg_entry->delta_data) {
1924                 delta_cache_size -= trg_entry->delta_size;
1925                 trg_entry->delta_data = NULL;
1926         }
1927         if (delta_cacheable(src_size, trg_size, delta_size)) {
1928                 delta_cache_size += delta_size;
1929                 cache_unlock();
1930                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1931         } else {
1932                 cache_unlock();
1933                 free(delta_buf);
1934         }
1935
1936         trg_entry->delta = src_entry;
1937         trg_entry->delta_size = delta_size;
1938         trg->depth = src->depth + 1;
1939
1940         return 1;
1941 }
1942
1943 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1944 {
1945         struct object_entry *child = me->delta_child;
1946         unsigned int m = n;
1947         while (child) {
1948                 unsigned int c = check_delta_limit(child, n + 1);
1949                 if (m < c)
1950                         m = c;
1951                 child = child->delta_sibling;
1952         }
1953         return m;
1954 }
1955
1956 static unsigned long free_unpacked(struct unpacked *n)
1957 {
1958         unsigned long freed_mem = sizeof_delta_index(n->index);
1959         free_delta_index(n->index);
1960         n->index = NULL;
1961         if (n->data) {
1962                 freed_mem += n->entry->size;
1963                 free(n->data);
1964                 n->data = NULL;
1965         }
1966         n->entry = NULL;
1967         n->depth = 0;
1968         return freed_mem;
1969 }
1970
1971 static void find_deltas(struct object_entry **list, unsigned *list_size,
1972                         int window, int depth, unsigned *processed)
1973 {
1974         uint32_t i, idx = 0, count = 0;
1975         struct unpacked *array;
1976         unsigned long mem_usage = 0;
1977
1978         array = xcalloc(window, sizeof(struct unpacked));
1979
1980         for (;;) {
1981                 struct object_entry *entry;
1982                 struct unpacked *n = array + idx;
1983                 int j, max_depth, best_base = -1;
1984
1985                 progress_lock();
1986                 if (!*list_size) {
1987                         progress_unlock();
1988                         break;
1989                 }
1990                 entry = *list++;
1991                 (*list_size)--;
1992                 if (!entry->preferred_base) {
1993                         (*processed)++;
1994                         display_progress(progress_state, *processed);
1995                 }
1996                 progress_unlock();
1997
1998                 mem_usage -= free_unpacked(n);
1999                 n->entry = entry;
2000
2001                 while (window_memory_limit &&
2002                        mem_usage > window_memory_limit &&
2003                        count > 1) {
2004                         uint32_t tail = (idx + window - count) % window;
2005                         mem_usage -= free_unpacked(array + tail);
2006                         count--;
2007                 }
2008
2009                 /* We do not compute delta to *create* objects we are not
2010                  * going to pack.
2011                  */
2012                 if (entry->preferred_base)
2013                         goto next;
2014
2015                 /*
2016                  * If the current object is at pack edge, take the depth the
2017                  * objects that depend on the current object into account
2018                  * otherwise they would become too deep.
2019                  */
2020                 max_depth = depth;
2021                 if (entry->delta_child) {
2022                         max_depth -= check_delta_limit(entry, 0);
2023                         if (max_depth <= 0)
2024                                 goto next;
2025                 }
2026
2027                 j = window;
2028                 while (--j > 0) {
2029                         int ret;
2030                         uint32_t other_idx = idx + j;
2031                         struct unpacked *m;
2032                         if (other_idx >= window)
2033                                 other_idx -= window;
2034                         m = array + other_idx;
2035                         if (!m->entry)
2036                                 break;
2037                         ret = try_delta(n, m, max_depth, &mem_usage);
2038                         if (ret < 0)
2039                                 break;
2040                         else if (ret > 0)
2041                                 best_base = other_idx;
2042                 }
2043
2044                 /*
2045                  * If we decided to cache the delta data, then it is best
2046                  * to compress it right away.  First because we have to do
2047                  * it anyway, and doing it here while we're threaded will
2048                  * save a lot of time in the non threaded write phase,
2049                  * as well as allow for caching more deltas within
2050                  * the same cache size limit.
2051                  * ...
2052                  * But only if not writing to stdout, since in that case
2053                  * the network is most likely throttling writes anyway,
2054                  * and therefore it is best to go to the write phase ASAP
2055                  * instead, as we can afford spending more time compressing
2056                  * between writes at that moment.
2057                  */
2058                 if (entry->delta_data && !pack_to_stdout) {
2059                         entry->z_delta_size = do_compress(&entry->delta_data,
2060                                                           entry->delta_size);
2061                         cache_lock();
2062                         delta_cache_size -= entry->delta_size;
2063                         delta_cache_size += entry->z_delta_size;
2064                         cache_unlock();
2065                 }
2066
2067                 /* if we made n a delta, and if n is already at max
2068                  * depth, leaving it in the window is pointless.  we
2069                  * should evict it first.
2070                  */
2071                 if (entry->delta && max_depth <= n->depth)
2072                         continue;
2073
2074                 /*
2075                  * Move the best delta base up in the window, after the
2076                  * currently deltified object, to keep it longer.  It will
2077                  * be the first base object to be attempted next.
2078                  */
2079                 if (entry->delta) {
2080                         struct unpacked swap = array[best_base];
2081                         int dist = (window + idx - best_base) % window;
2082                         int dst = best_base;
2083                         while (dist--) {
2084                                 int src = (dst + 1) % window;
2085                                 array[dst] = array[src];
2086                                 dst = src;
2087                         }
2088                         array[dst] = swap;
2089                 }
2090
2091                 next:
2092                 idx++;
2093                 if (count + 1 < window)
2094                         count++;
2095                 if (idx >= window)
2096                         idx = 0;
2097         }
2098
2099         for (i = 0; i < window; ++i) {
2100                 free_delta_index(array[i].index);
2101                 free(array[i].data);
2102         }
2103         free(array);
2104 }
2105
2106 #ifndef NO_PTHREADS
2107
2108 static void try_to_free_from_threads(size_t size)
2109 {
2110         read_lock();
2111         release_pack_memory(size);
2112         read_unlock();
2113 }
2114
2115 static try_to_free_t old_try_to_free_routine;
2116
2117 /*
2118  * The main thread waits on the condition that (at least) one of the workers
2119  * has stopped working (which is indicated in the .working member of
2120  * struct thread_params).
2121  * When a work thread has completed its work, it sets .working to 0 and
2122  * signals the main thread and waits on the condition that .data_ready
2123  * becomes 1.
2124  */
2125
2126 struct thread_params {
2127         pthread_t thread;
2128         struct object_entry **list;
2129         unsigned list_size;
2130         unsigned remaining;
2131         int window;
2132         int depth;
2133         int working;
2134         int data_ready;
2135         pthread_mutex_t mutex;
2136         pthread_cond_t cond;
2137         unsigned *processed;
2138 };
2139
2140 static pthread_cond_t progress_cond;
2141
2142 /*
2143  * Mutex and conditional variable can't be statically-initialized on Windows.
2144  */
2145 static void init_threaded_search(void)
2146 {
2147         init_recursive_mutex(&read_mutex);
2148         pthread_mutex_init(&cache_mutex, NULL);
2149         pthread_mutex_init(&progress_mutex, NULL);
2150         pthread_cond_init(&progress_cond, NULL);
2151         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2152 }
2153
2154 static void cleanup_threaded_search(void)
2155 {
2156         set_try_to_free_routine(old_try_to_free_routine);
2157         pthread_cond_destroy(&progress_cond);
2158         pthread_mutex_destroy(&read_mutex);
2159         pthread_mutex_destroy(&cache_mutex);
2160         pthread_mutex_destroy(&progress_mutex);
2161 }
2162
2163 static void *threaded_find_deltas(void *arg)
2164 {
2165         struct thread_params *me = arg;
2166
2167         while (me->remaining) {
2168                 find_deltas(me->list, &me->remaining,
2169                             me->window, me->depth, me->processed);
2170
2171                 progress_lock();
2172                 me->working = 0;
2173                 pthread_cond_signal(&progress_cond);
2174                 progress_unlock();
2175
2176                 /*
2177                  * We must not set ->data_ready before we wait on the
2178                  * condition because the main thread may have set it to 1
2179                  * before we get here. In order to be sure that new
2180                  * work is available if we see 1 in ->data_ready, it
2181                  * was initialized to 0 before this thread was spawned
2182                  * and we reset it to 0 right away.
2183                  */
2184                 pthread_mutex_lock(&me->mutex);
2185                 while (!me->data_ready)
2186                         pthread_cond_wait(&me->cond, &me->mutex);
2187                 me->data_ready = 0;
2188                 pthread_mutex_unlock(&me->mutex);
2189         }
2190         /* leave ->working 1 so that this doesn't get more work assigned */
2191         return NULL;
2192 }
2193
2194 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2195                            int window, int depth, unsigned *processed)
2196 {
2197         struct thread_params *p;
2198         int i, ret, active_threads = 0;
2199
2200         init_threaded_search();
2201
2202         if (delta_search_threads <= 1) {
2203                 find_deltas(list, &list_size, window, depth, processed);
2204                 cleanup_threaded_search();
2205                 return;
2206         }
2207         if (progress > pack_to_stdout)
2208                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2209                                 delta_search_threads);
2210         p = xcalloc(delta_search_threads, sizeof(*p));
2211
2212         /* Partition the work amongst work threads. */
2213         for (i = 0; i < delta_search_threads; i++) {
2214                 unsigned sub_size = list_size / (delta_search_threads - i);
2215
2216                 /* don't use too small segments or no deltas will be found */
2217                 if (sub_size < 2*window && i+1 < delta_search_threads)
2218                         sub_size = 0;
2219
2220                 p[i].window = window;
2221                 p[i].depth = depth;
2222                 p[i].processed = processed;
2223                 p[i].working = 1;
2224                 p[i].data_ready = 0;
2225
2226                 /* try to split chunks on "path" boundaries */
2227                 while (sub_size && sub_size < list_size &&
2228                        list[sub_size]->hash &&
2229                        list[sub_size]->hash == list[sub_size-1]->hash)
2230                         sub_size++;
2231
2232                 p[i].list = list;
2233                 p[i].list_size = sub_size;
2234                 p[i].remaining = sub_size;
2235
2236                 list += sub_size;
2237                 list_size -= sub_size;
2238         }
2239
2240         /* Start work threads. */
2241         for (i = 0; i < delta_search_threads; i++) {
2242                 if (!p[i].list_size)
2243                         continue;
2244                 pthread_mutex_init(&p[i].mutex, NULL);
2245                 pthread_cond_init(&p[i].cond, NULL);
2246                 ret = pthread_create(&p[i].thread, NULL,
2247                                      threaded_find_deltas, &p[i]);
2248                 if (ret)
2249                         die("unable to create thread: %s", strerror(ret));
2250                 active_threads++;
2251         }
2252
2253         /*
2254          * Now let's wait for work completion.  Each time a thread is done
2255          * with its work, we steal half of the remaining work from the
2256          * thread with the largest number of unprocessed objects and give
2257          * it to that newly idle thread.  This ensure good load balancing
2258          * until the remaining object list segments are simply too short
2259          * to be worth splitting anymore.
2260          */
2261         while (active_threads) {
2262                 struct thread_params *target = NULL;
2263                 struct thread_params *victim = NULL;
2264                 unsigned sub_size = 0;
2265
2266                 progress_lock();
2267                 for (;;) {
2268                         for (i = 0; !target && i < delta_search_threads; i++)
2269                                 if (!p[i].working)
2270                                         target = &p[i];
2271                         if (target)
2272                                 break;
2273                         pthread_cond_wait(&progress_cond, &progress_mutex);
2274                 }
2275
2276                 for (i = 0; i < delta_search_threads; i++)
2277                         if (p[i].remaining > 2*window &&
2278                             (!victim || victim->remaining < p[i].remaining))
2279                                 victim = &p[i];
2280                 if (victim) {
2281                         sub_size = victim->remaining / 2;
2282                         list = victim->list + victim->list_size - sub_size;
2283                         while (sub_size && list[0]->hash &&
2284                                list[0]->hash == list[-1]->hash) {
2285                                 list++;
2286                                 sub_size--;
2287                         }
2288                         if (!sub_size) {
2289                                 /*
2290                                  * It is possible for some "paths" to have
2291                                  * so many objects that no hash boundary
2292                                  * might be found.  Let's just steal the
2293                                  * exact half in that case.
2294                                  */
2295                                 sub_size = victim->remaining / 2;
2296                                 list -= sub_size;
2297                         }
2298                         target->list = list;
2299                         victim->list_size -= sub_size;
2300                         victim->remaining -= sub_size;
2301                 }
2302                 target->list_size = sub_size;
2303                 target->remaining = sub_size;
2304                 target->working = 1;
2305                 progress_unlock();
2306
2307                 pthread_mutex_lock(&target->mutex);
2308                 target->data_ready = 1;
2309                 pthread_cond_signal(&target->cond);
2310                 pthread_mutex_unlock(&target->mutex);
2311
2312                 if (!sub_size) {
2313                         pthread_join(target->thread, NULL);
2314                         pthread_cond_destroy(&target->cond);
2315                         pthread_mutex_destroy(&target->mutex);
2316                         active_threads--;
2317                 }
2318         }
2319         cleanup_threaded_search();
2320         free(p);
2321 }
2322
2323 #else
2324 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2325 #endif
2326
2327 static void add_tag_chain(const struct object_id *oid)
2328 {
2329         struct tag *tag;
2330
2331         /*
2332          * We catch duplicates already in add_object_entry(), but we'd
2333          * prefer to do this extra check to avoid having to parse the
2334          * tag at all if we already know that it's being packed (e.g., if
2335          * it was included via bitmaps, we would not have parsed it
2336          * previously).
2337          */
2338         if (packlist_find(&to_pack, oid->hash, NULL))
2339                 return;
2340
2341         tag = lookup_tag(oid->hash);
2342         while (1) {
2343                 if (!tag || parse_tag(tag) || !tag->tagged)
2344                         die("unable to pack objects reachable from tag %s",
2345                             oid_to_hex(oid));
2346
2347                 add_object_entry(tag->object.oid.hash, OBJ_TAG, NULL, 0);
2348
2349                 if (tag->tagged->type != OBJ_TAG)
2350                         return;
2351
2352                 tag = (struct tag *)tag->tagged;
2353         }
2354 }
2355
2356 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2357 {
2358         struct object_id peeled;
2359
2360         if (starts_with(path, "refs/tags/") && /* is a tag? */
2361             !peel_ref(path, peeled.hash)    && /* peelable? */
2362             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2363                 add_tag_chain(oid);
2364         return 0;
2365 }
2366
2367 static void prepare_pack(int window, int depth)
2368 {
2369         struct object_entry **delta_list;
2370         uint32_t i, nr_deltas;
2371         unsigned n;
2372
2373         get_object_details();
2374
2375         /*
2376          * If we're locally repacking then we need to be doubly careful
2377          * from now on in order to make sure no stealth corruption gets
2378          * propagated to the new pack.  Clients receiving streamed packs
2379          * should validate everything they get anyway so no need to incur
2380          * the additional cost here in that case.
2381          */
2382         if (!pack_to_stdout)
2383                 do_check_packed_object_crc = 1;
2384
2385         if (!to_pack.nr_objects || !window || !depth)
2386                 return;
2387
2388         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2389         nr_deltas = n = 0;
2390
2391         for (i = 0; i < to_pack.nr_objects; i++) {
2392                 struct object_entry *entry = to_pack.objects + i;
2393
2394                 if (entry->delta)
2395                         /* This happens if we decided to reuse existing
2396                          * delta from a pack.  "reuse_delta &&" is implied.
2397                          */
2398                         continue;
2399
2400                 if (entry->size < 50)
2401                         continue;
2402
2403                 if (entry->no_try_delta)
2404                         continue;
2405
2406                 if (!entry->preferred_base) {
2407                         nr_deltas++;
2408                         if (entry->type < 0)
2409                                 die("unable to get type of object %s",
2410                                     sha1_to_hex(entry->idx.sha1));
2411                 } else {
2412                         if (entry->type < 0) {
2413                                 /*
2414                                  * This object is not found, but we
2415                                  * don't have to include it anyway.
2416                                  */
2417                                 continue;
2418                         }
2419                 }
2420
2421                 delta_list[n++] = entry;
2422         }
2423
2424         if (nr_deltas && n > 1) {
2425                 unsigned nr_done = 0;
2426                 if (progress)
2427                         progress_state = start_progress(_("Compressing objects"),
2428                                                         nr_deltas);
2429                 QSORT(delta_list, n, type_size_sort);
2430                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2431                 stop_progress(&progress_state);
2432                 if (nr_done != nr_deltas)
2433                         die("inconsistency with delta count");
2434         }
2435         free(delta_list);
2436 }
2437
2438 static int git_pack_config(const char *k, const char *v, void *cb)
2439 {
2440         if (!strcmp(k, "pack.window")) {
2441                 window = git_config_int(k, v);
2442                 return 0;
2443         }
2444         if (!strcmp(k, "pack.windowmemory")) {
2445                 window_memory_limit = git_config_ulong(k, v);
2446                 return 0;
2447         }
2448         if (!strcmp(k, "pack.depth")) {
2449                 depth = git_config_int(k, v);
2450                 return 0;
2451         }
2452         if (!strcmp(k, "pack.deltacachesize")) {
2453                 max_delta_cache_size = git_config_int(k, v);
2454                 return 0;
2455         }
2456         if (!strcmp(k, "pack.deltacachelimit")) {
2457                 cache_max_small_delta_size = git_config_int(k, v);
2458                 return 0;
2459         }
2460         if (!strcmp(k, "pack.writebitmaphashcache")) {
2461                 if (git_config_bool(k, v))
2462                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2463                 else
2464                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2465         }
2466         if (!strcmp(k, "pack.usebitmaps")) {
2467                 use_bitmap_index_default = git_config_bool(k, v);
2468                 return 0;
2469         }
2470         if (!strcmp(k, "pack.threads")) {
2471                 delta_search_threads = git_config_int(k, v);
2472                 if (delta_search_threads < 0)
2473                         die("invalid number of threads specified (%d)",
2474                             delta_search_threads);
2475 #ifdef NO_PTHREADS
2476                 if (delta_search_threads != 1)
2477                         warning("no threads support, ignoring %s", k);
2478 #endif
2479                 return 0;
2480         }
2481         if (!strcmp(k, "pack.indexversion")) {
2482                 pack_idx_opts.version = git_config_int(k, v);
2483                 if (pack_idx_opts.version > 2)
2484                         die("bad pack.indexversion=%"PRIu32,
2485                             pack_idx_opts.version);
2486                 return 0;
2487         }
2488         return git_default_config(k, v, cb);
2489 }
2490
2491 static void read_object_list_from_stdin(void)
2492 {
2493         char line[40 + 1 + PATH_MAX + 2];
2494         unsigned char sha1[20];
2495
2496         for (;;) {
2497                 if (!fgets(line, sizeof(line), stdin)) {
2498                         if (feof(stdin))
2499                                 break;
2500                         if (!ferror(stdin))
2501                                 die("fgets returned NULL, not EOF, not error!");
2502                         if (errno != EINTR)
2503                                 die_errno("fgets");
2504                         clearerr(stdin);
2505                         continue;
2506                 }
2507                 if (line[0] == '-') {
2508                         if (get_sha1_hex(line+1, sha1))
2509                                 die("expected edge sha1, got garbage:\n %s",
2510                                     line);
2511                         add_preferred_base(sha1);
2512                         continue;
2513                 }
2514                 if (get_sha1_hex(line, sha1))
2515                         die("expected sha1, got garbage:\n %s", line);
2516
2517                 add_preferred_base_object(line+41);
2518                 add_object_entry(sha1, 0, line+41, 0);
2519         }
2520 }
2521
2522 #define OBJECT_ADDED (1u<<20)
2523
2524 static void show_commit(struct commit *commit, void *data)
2525 {
2526         add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2527         commit->object.flags |= OBJECT_ADDED;
2528
2529         if (write_bitmap_index)
2530                 index_commit_for_bitmap(commit);
2531 }
2532
2533 static void show_object(struct object *obj, const char *name, void *data)
2534 {
2535         add_preferred_base_object(name);
2536         add_object_entry(obj->oid.hash, obj->type, name, 0);
2537         obj->flags |= OBJECT_ADDED;
2538 }
2539
2540 static void show_edge(struct commit *commit)
2541 {
2542         add_preferred_base(commit->object.oid.hash);
2543 }
2544
2545 struct in_pack_object {
2546         off_t offset;
2547         struct object *object;
2548 };
2549
2550 struct in_pack {
2551         int alloc;
2552         int nr;
2553         struct in_pack_object *array;
2554 };
2555
2556 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2557 {
2558         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2559         in_pack->array[in_pack->nr].object = object;
2560         in_pack->nr++;
2561 }
2562
2563 /*
2564  * Compare the objects in the offset order, in order to emulate the
2565  * "git rev-list --objects" output that produced the pack originally.
2566  */
2567 static int ofscmp(const void *a_, const void *b_)
2568 {
2569         struct in_pack_object *a = (struct in_pack_object *)a_;
2570         struct in_pack_object *b = (struct in_pack_object *)b_;
2571
2572         if (a->offset < b->offset)
2573                 return -1;
2574         else if (a->offset > b->offset)
2575                 return 1;
2576         else
2577                 return oidcmp(&a->object->oid, &b->object->oid);
2578 }
2579
2580 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2581 {
2582         struct packed_git *p;
2583         struct in_pack in_pack;
2584         uint32_t i;
2585
2586         memset(&in_pack, 0, sizeof(in_pack));
2587
2588         for (p = packed_git; p; p = p->next) {
2589                 const unsigned char *sha1;
2590                 struct object *o;
2591
2592                 if (!p->pack_local || p->pack_keep)
2593                         continue;
2594                 if (open_pack_index(p))
2595                         die("cannot open pack index");
2596
2597                 ALLOC_GROW(in_pack.array,
2598                            in_pack.nr + p->num_objects,
2599                            in_pack.alloc);
2600
2601                 for (i = 0; i < p->num_objects; i++) {
2602                         sha1 = nth_packed_object_sha1(p, i);
2603                         o = lookup_unknown_object(sha1);
2604                         if (!(o->flags & OBJECT_ADDED))
2605                                 mark_in_pack_object(o, p, &in_pack);
2606                         o->flags |= OBJECT_ADDED;
2607                 }
2608         }
2609
2610         if (in_pack.nr) {
2611                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2612                 for (i = 0; i < in_pack.nr; i++) {
2613                         struct object *o = in_pack.array[i].object;
2614                         add_object_entry(o->oid.hash, o->type, "", 0);
2615                 }
2616         }
2617         free(in_pack.array);
2618 }
2619
2620 static int add_loose_object(const struct object_id *oid, const char *path,
2621                             void *data)
2622 {
2623         enum object_type type = sha1_object_info(oid->hash, NULL);
2624
2625         if (type < 0) {
2626                 warning("loose object at %s could not be examined", path);
2627                 return 0;
2628         }
2629
2630         add_object_entry(oid->hash, type, "", 0);
2631         return 0;
2632 }
2633
2634 /*
2635  * We actually don't even have to worry about reachability here.
2636  * add_object_entry will weed out duplicates, so we just add every
2637  * loose object we find.
2638  */
2639 static void add_unreachable_loose_objects(void)
2640 {
2641         for_each_loose_file_in_objdir(get_object_directory(),
2642                                       add_loose_object,
2643                                       NULL, NULL, NULL);
2644 }
2645
2646 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2647 {
2648         static struct packed_git *last_found = (void *)1;
2649         struct packed_git *p;
2650
2651         p = (last_found != (void *)1) ? last_found : packed_git;
2652
2653         while (p) {
2654                 if ((!p->pack_local || p->pack_keep) &&
2655                         find_pack_entry_one(sha1, p)) {
2656                         last_found = p;
2657                         return 1;
2658                 }
2659                 if (p == last_found)
2660                         p = packed_git;
2661                 else
2662                         p = p->next;
2663                 if (p == last_found)
2664                         p = p->next;
2665         }
2666         return 0;
2667 }
2668
2669 /*
2670  * Store a list of sha1s that are should not be discarded
2671  * because they are either written too recently, or are
2672  * reachable from another object that was.
2673  *
2674  * This is filled by get_object_list.
2675  */
2676 static struct oid_array recent_objects;
2677
2678 static int loosened_object_can_be_discarded(const struct object_id *oid,
2679                                             unsigned long mtime)
2680 {
2681         if (!unpack_unreachable_expiration)
2682                 return 0;
2683         if (mtime > unpack_unreachable_expiration)
2684                 return 0;
2685         if (oid_array_lookup(&recent_objects, oid) >= 0)
2686                 return 0;
2687         return 1;
2688 }
2689
2690 static void loosen_unused_packed_objects(struct rev_info *revs)
2691 {
2692         struct packed_git *p;
2693         uint32_t i;
2694         struct object_id oid;
2695
2696         for (p = packed_git; p; p = p->next) {
2697                 if (!p->pack_local || p->pack_keep)
2698                         continue;
2699
2700                 if (open_pack_index(p))
2701                         die("cannot open pack index");
2702
2703                 for (i = 0; i < p->num_objects; i++) {
2704                         nth_packed_object_oid(&oid, p, i);
2705                         if (!packlist_find(&to_pack, oid.hash, NULL) &&
2706                             !has_sha1_pack_kept_or_nonlocal(oid.hash) &&
2707                             !loosened_object_can_be_discarded(&oid, p->mtime))
2708                                 if (force_object_loose(oid.hash, p->mtime))
2709                                         die("unable to force loose object");
2710                 }
2711         }
2712 }
2713
2714 /*
2715  * This tracks any options which pack-reuse code expects to be on, or which a
2716  * reader of the pack might not understand, and which would therefore prevent
2717  * blind reuse of what we have on disk.
2718  */
2719 static int pack_options_allow_reuse(void)
2720 {
2721         return pack_to_stdout && allow_ofs_delta;
2722 }
2723
2724 static int get_object_list_from_bitmap(struct rev_info *revs)
2725 {
2726         if (prepare_bitmap_walk(revs) < 0)
2727                 return -1;
2728
2729         if (pack_options_allow_reuse() &&
2730             !reuse_partial_packfile_from_bitmap(
2731                         &reuse_packfile,
2732                         &reuse_packfile_objects,
2733                         &reuse_packfile_offset)) {
2734                 assert(reuse_packfile_objects);
2735                 nr_result += reuse_packfile_objects;
2736                 display_progress(progress_state, nr_result);
2737         }
2738
2739         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2740         return 0;
2741 }
2742
2743 static void record_recent_object(struct object *obj,
2744                                  const char *name,
2745                                  void *data)
2746 {
2747         oid_array_append(&recent_objects, &obj->oid);
2748 }
2749
2750 static void record_recent_commit(struct commit *commit, void *data)
2751 {
2752         oid_array_append(&recent_objects, &commit->object.oid);
2753 }
2754
2755 static void get_object_list(int ac, const char **av)
2756 {
2757         struct rev_info revs;
2758         char line[1000];
2759         int flags = 0;
2760
2761         init_revisions(&revs, NULL);
2762         save_commit_buffer = 0;
2763         setup_revisions(ac, av, &revs, NULL);
2764
2765         /* make sure shallows are read */
2766         is_repository_shallow();
2767
2768         while (fgets(line, sizeof(line), stdin) != NULL) {
2769                 int len = strlen(line);
2770                 if (len && line[len - 1] == '\n')
2771                         line[--len] = 0;
2772                 if (!len)
2773                         break;
2774                 if (*line == '-') {
2775                         if (!strcmp(line, "--not")) {
2776                                 flags ^= UNINTERESTING;
2777                                 write_bitmap_index = 0;
2778                                 continue;
2779                         }
2780                         if (starts_with(line, "--shallow ")) {
2781                                 unsigned char sha1[20];
2782                                 if (get_sha1_hex(line + 10, sha1))
2783                                         die("not an SHA-1 '%s'", line + 10);
2784                                 register_shallow(sha1);
2785                                 use_bitmap_index = 0;
2786                                 continue;
2787                         }
2788                         die("not a rev '%s'", line);
2789                 }
2790                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2791                         die("bad revision '%s'", line);
2792         }
2793
2794         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2795                 return;
2796
2797         if (prepare_revision_walk(&revs))
2798                 die("revision walk setup failed");
2799         mark_edges_uninteresting(&revs, show_edge);
2800         traverse_commit_list(&revs, show_commit, show_object, NULL);
2801
2802         if (unpack_unreachable_expiration) {
2803                 revs.ignore_missing_links = 1;
2804                 if (add_unseen_recent_objects_to_traversal(&revs,
2805                                 unpack_unreachable_expiration))
2806                         die("unable to add recent objects");
2807                 if (prepare_revision_walk(&revs))
2808                         die("revision walk setup failed");
2809                 traverse_commit_list(&revs, record_recent_commit,
2810                                      record_recent_object, NULL);
2811         }
2812
2813         if (keep_unreachable)
2814                 add_objects_in_unpacked_packs(&revs);
2815         if (pack_loose_unreachable)
2816                 add_unreachable_loose_objects();
2817         if (unpack_unreachable)
2818                 loosen_unused_packed_objects(&revs);
2819
2820         oid_array_clear(&recent_objects);
2821 }
2822
2823 static int option_parse_index_version(const struct option *opt,
2824                                       const char *arg, int unset)
2825 {
2826         char *c;
2827         const char *val = arg;
2828         pack_idx_opts.version = strtoul(val, &c, 10);
2829         if (pack_idx_opts.version > 2)
2830                 die(_("unsupported index version %s"), val);
2831         if (*c == ',' && c[1])
2832                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2833         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2834                 die(_("bad index version '%s'"), val);
2835         return 0;
2836 }
2837
2838 static int option_parse_unpack_unreachable(const struct option *opt,
2839                                            const char *arg, int unset)
2840 {
2841         if (unset) {
2842                 unpack_unreachable = 0;
2843                 unpack_unreachable_expiration = 0;
2844         }
2845         else {
2846                 unpack_unreachable = 1;
2847                 if (arg)
2848                         unpack_unreachable_expiration = approxidate(arg);
2849         }
2850         return 0;
2851 }
2852
2853 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2854 {
2855         int use_internal_rev_list = 0;
2856         int thin = 0;
2857         int shallow = 0;
2858         int all_progress_implied = 0;
2859         struct argv_array rp = ARGV_ARRAY_INIT;
2860         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2861         int rev_list_index = 0;
2862         struct option pack_objects_options[] = {
2863                 OPT_SET_INT('q', "quiet", &progress,
2864                             N_("do not show progress meter"), 0),
2865                 OPT_SET_INT(0, "progress", &progress,
2866                             N_("show progress meter"), 1),
2867                 OPT_SET_INT(0, "all-progress", &progress,
2868                             N_("show progress meter during object writing phase"), 2),
2869                 OPT_BOOL(0, "all-progress-implied",
2870                          &all_progress_implied,
2871                          N_("similar to --all-progress when progress meter is shown")),
2872                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2873                   N_("write the pack index file in the specified idx format version"),
2874                   0, option_parse_index_version },
2875                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2876                               N_("maximum size of each output pack file")),
2877                 OPT_BOOL(0, "local", &local,
2878                          N_("ignore borrowed objects from alternate object store")),
2879                 OPT_BOOL(0, "incremental", &incremental,
2880                          N_("ignore packed objects")),
2881                 OPT_INTEGER(0, "window", &window,
2882                             N_("limit pack window by objects")),
2883                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2884                               N_("limit pack window by memory in addition to object limit")),
2885                 OPT_INTEGER(0, "depth", &depth,
2886                             N_("maximum length of delta chain allowed in the resulting pack")),
2887                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2888                          N_("reuse existing deltas")),
2889                 OPT_BOOL(0, "reuse-object", &reuse_object,
2890                          N_("reuse existing objects")),
2891                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2892                          N_("use OFS_DELTA objects")),
2893                 OPT_INTEGER(0, "threads", &delta_search_threads,
2894                             N_("use threads when searching for best delta matches")),
2895                 OPT_BOOL(0, "non-empty", &non_empty,
2896                          N_("do not create an empty pack output")),
2897                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2898                          N_("read revision arguments from standard input")),
2899                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2900                   N_("limit the objects to those that are not yet packed"),
2901                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2902                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2903                   N_("include objects reachable from any reference"),
2904                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2905                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2906                   N_("include objects referred by reflog entries"),
2907                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2908                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2909                   N_("include objects referred to by the index"),
2910                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2911                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2912                          N_("output pack to stdout")),
2913                 OPT_BOOL(0, "include-tag", &include_tag,
2914                          N_("include tag objects that refer to objects to be packed")),
2915                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2916                          N_("keep unreachable objects")),
2917                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
2918                          N_("pack loose unreachable objects")),
2919                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2920                   N_("unpack unreachable objects newer than <time>"),
2921                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2922                 OPT_BOOL(0, "thin", &thin,
2923                          N_("create thin packs")),
2924                 OPT_BOOL(0, "shallow", &shallow,
2925                          N_("create packs suitable for shallow fetches")),
2926                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2927                          N_("ignore packs that have companion .keep file")),
2928                 OPT_INTEGER(0, "compression", &pack_compression_level,
2929                             N_("pack compression level")),
2930                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2931                             N_("do not hide commits by grafts"), 0),
2932                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2933                          N_("use a bitmap index if available to speed up counting objects")),
2934                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2935                          N_("write a bitmap index together with the pack index")),
2936                 OPT_END(),
2937         };
2938
2939         check_replace_refs = 0;
2940
2941         reset_pack_idx_option(&pack_idx_opts);
2942         git_config(git_pack_config, NULL);
2943
2944         progress = isatty(2);
2945         argc = parse_options(argc, argv, prefix, pack_objects_options,
2946                              pack_usage, 0);
2947
2948         if (argc) {
2949                 base_name = argv[0];
2950                 argc--;
2951         }
2952         if (pack_to_stdout != !base_name || argc)
2953                 usage_with_options(pack_usage, pack_objects_options);
2954
2955         argv_array_push(&rp, "pack-objects");
2956         if (thin) {
2957                 use_internal_rev_list = 1;
2958                 argv_array_push(&rp, shallow
2959                                 ? "--objects-edge-aggressive"
2960                                 : "--objects-edge");
2961         } else
2962                 argv_array_push(&rp, "--objects");
2963
2964         if (rev_list_all) {
2965                 use_internal_rev_list = 1;
2966                 argv_array_push(&rp, "--all");
2967         }
2968         if (rev_list_reflog) {
2969                 use_internal_rev_list = 1;
2970                 argv_array_push(&rp, "--reflog");
2971         }
2972         if (rev_list_index) {
2973                 use_internal_rev_list = 1;
2974                 argv_array_push(&rp, "--indexed-objects");
2975         }
2976         if (rev_list_unpacked) {
2977                 use_internal_rev_list = 1;
2978                 argv_array_push(&rp, "--unpacked");
2979         }
2980
2981         if (!reuse_object)
2982                 reuse_delta = 0;
2983         if (pack_compression_level == -1)
2984                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2985         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2986                 die("bad pack compression level %d", pack_compression_level);
2987
2988         if (!delta_search_threads)      /* --threads=0 means autodetect */
2989                 delta_search_threads = online_cpus();
2990
2991 #ifdef NO_PTHREADS
2992         if (delta_search_threads != 1)
2993                 warning("no threads support, ignoring --threads");
2994 #endif
2995         if (!pack_to_stdout && !pack_size_limit)
2996                 pack_size_limit = pack_size_limit_cfg;
2997         if (pack_to_stdout && pack_size_limit)
2998                 die("--max-pack-size cannot be used to build a pack for transfer.");
2999         if (pack_size_limit && pack_size_limit < 1024*1024) {
3000                 warning("minimum pack size limit is 1 MiB");
3001                 pack_size_limit = 1024*1024;
3002         }
3003
3004         if (!pack_to_stdout && thin)
3005                 die("--thin cannot be used to build an indexable pack.");
3006
3007         if (keep_unreachable && unpack_unreachable)
3008                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
3009         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3010                 unpack_unreachable_expiration = 0;
3011
3012         /*
3013          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3014          *
3015          * - to produce good pack (with bitmap index not-yet-packed objects are
3016          *   packed in suboptimal order).
3017          *
3018          * - to use more robust pack-generation codepath (avoiding possible
3019          *   bugs in bitmap code and possible bitmap index corruption).
3020          */
3021         if (!pack_to_stdout)
3022                 use_bitmap_index_default = 0;
3023
3024         if (use_bitmap_index < 0)
3025                 use_bitmap_index = use_bitmap_index_default;
3026
3027         /* "hard" reasons not to use bitmaps; these just won't work at all */
3028         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
3029                 use_bitmap_index = 0;
3030
3031         if (pack_to_stdout || !rev_list_all)
3032                 write_bitmap_index = 0;
3033
3034         if (progress && all_progress_implied)
3035                 progress = 2;
3036
3037         prepare_packed_git();
3038         if (ignore_packed_keep) {
3039                 struct packed_git *p;
3040                 for (p = packed_git; p; p = p->next)
3041                         if (p->pack_local && p->pack_keep)
3042                                 break;
3043                 if (!p) /* no keep-able packs found */
3044                         ignore_packed_keep = 0;
3045         }
3046         if (local) {
3047                 /*
3048                  * unlike ignore_packed_keep above, we do not want to
3049                  * unset "local" based on looking at packs, as it
3050                  * also covers non-local objects
3051                  */
3052                 struct packed_git *p;
3053                 for (p = packed_git; p; p = p->next) {
3054                         if (!p->pack_local) {
3055                                 have_non_local_packs = 1;
3056                                 break;
3057                         }
3058                 }
3059         }
3060
3061         if (progress)
3062                 progress_state = start_progress(_("Counting objects"), 0);
3063         if (!use_internal_rev_list)
3064                 read_object_list_from_stdin();
3065         else {
3066                 get_object_list(rp.argc, rp.argv);
3067                 argv_array_clear(&rp);
3068         }
3069         cleanup_preferred_base();
3070         if (include_tag && nr_result)
3071                 for_each_ref(add_ref_tag, NULL);
3072         stop_progress(&progress_state);
3073
3074         if (non_empty && !nr_result)
3075                 return 0;
3076         if (nr_result)
3077                 prepare_pack(window, depth);
3078         write_pack_file();
3079         if (progress)
3080                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
3081                         " reused %"PRIu32" (delta %"PRIu32")\n",
3082                         written, written_delta, reused, reused_delta);
3083         return 0;
3084 }