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

   1 #include "builtin.h"
   2 #include "cache.h"
   3 #include "repository.h"
   4 #include "config.h"
   5 #include "attr.h"
   6 #include "object.h"
   7 #include "blob.h"
   8 #include "commit.h"
   9 #include "tag.h"
  10 #include "tree.h"
  11 #include "delta.h"
  12 #include "pack.h"
  13 #include "pack-revindex.h"
  14 #include "csum-file.h"
  15 #include "tree-walk.h"
  16 #include "diff.h"
  17 #include "revision.h"
  18 #include "list-objects.h"
  19 #include "list-objects-filter.h"
  20 #include "list-objects-filter-options.h"
  21 #include "pack-objects.h"
  22 #include "progress.h"
  23 #include "refs.h"
  24 #include "streaming.h"
  25 #include "thread-utils.h"
  26 #include "pack-bitmap.h"
  27 #include "delta-islands.h"
  28 #include "reachable.h"
  29 #include "sha1-array.h"
  30 #include "argv-array.h"
  31 #include "list.h"
  32 #include "packfile.h"
  33 #include "object-store.h"
  34 #include "dir.h"
  35 #include "midx.h"
  36 #include "trace2.h"
  37
  38 #define IN_PACK(obj) oe_in_pack(&to_pack, obj)
  39 #define SIZE(obj) oe_size(&to_pack, obj)
  40 #define SET_SIZE(obj,size) oe_set_size(&to_pack, obj, size)
  41 #define DELTA_SIZE(obj) oe_delta_size(&to_pack, obj)
  42 #define DELTA(obj) oe_delta(&to_pack, obj)
  43 #define DELTA_CHILD(obj) oe_delta_child(&to_pack, obj)
  44 #define DELTA_SIBLING(obj) oe_delta_sibling(&to_pack, obj)
  45 #define SET_DELTA(obj, val) oe_set_delta(&to_pack, obj, val)
  46 #define SET_DELTA_EXT(obj, oid) oe_set_delta_ext(&to_pack, obj, oid)
  47 #define SET_DELTA_SIZE(obj, val) oe_set_delta_size(&to_pack, obj, val)
  48 #define SET_DELTA_CHILD(obj, val) oe_set_delta_child(&to_pack, obj, val)
  49 #define SET_DELTA_SIBLING(obj, val) oe_set_delta_sibling(&to_pack, obj, val)
  50
  51 static const char *pack_usage[] = {
  52         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
  53         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
  54         NULL
  55 };
  56
  57 /*
  58  * Objects we are going to pack are collected in the `to_pack` structure.
  59  * It contains an array (dynamically expanded) of the object data, and a map
  60  * that can resolve SHA1s to their position in the array.
  61  */
  62 static struct packing_data to_pack;
  63
  64 static struct pack_idx_entry **written_list;
  65 static uint32_t nr_result, nr_written, nr_seen;
  66 static struct bitmap_index *bitmap_git;
  67 static uint32_t write_layer;
  68
  69 static int non_empty;
  70 static int reuse_delta = 1, reuse_object = 1;
  71 static int keep_unreachable, unpack_unreachable, include_tag;
  72 static timestamp_t unpack_unreachable_expiration;
  73 static int pack_loose_unreachable;
  74 static int local;
  75 static int have_non_local_packs;
  76 static int incremental;
  77 static int ignore_packed_keep_on_disk;
  78 static int ignore_packed_keep_in_core;
  79 static int allow_ofs_delta;
  80 static struct pack_idx_option pack_idx_opts;
  81 static const char *base_name;
  82 static int progress = 1;
  83 static int window = 10;
  84 static unsigned long pack_size_limit;
  85 static int depth = 50;
  86 static int delta_search_threads;
  87 static int pack_to_stdout;
  88 static int sparse;
  89 static int thin;
  90 static int num_preferred_base;
  91 static struct progress *progress_state;
  92
  93 static struct packed_git *reuse_packfile;
  94 static uint32_t reuse_packfile_objects;
  95 static off_t reuse_packfile_offset;
  96
  97 static int use_bitmap_index_default = 1;
  98 static int use_bitmap_index = -1;
  99 static enum {
 100         WRITE_BITMAP_FALSE = 0,
 101         WRITE_BITMAP_QUIET,
 102         WRITE_BITMAP_TRUE,
 103 } write_bitmap_index;
 104 static uint16_t write_bitmap_options = BITMAP_OPT_HASH_CACHE;
 105
 106 static int exclude_promisor_objects;
 107
 108 static int use_delta_islands;
 109
 110 static unsigned long delta_cache_size = 0;
 111 static unsigned long max_delta_cache_size = DEFAULT_DELTA_CACHE_SIZE;
 112 static unsigned long cache_max_small_delta_size = 1000;
 113
 114 static unsigned long window_memory_limit = 0;
 115
 116 static struct list_objects_filter_options filter_options;
 117
 118 enum missing_action {
 119         MA_ERROR = 0,      /* fail if any missing objects are encountered */
 120         MA_ALLOW_ANY,      /* silently allow ALL missing objects */
 121         MA_ALLOW_PROMISOR, /* silently allow all missing PROMISOR objects */
 122 };
 123 static enum missing_action arg_missing_action;
 124 static show_object_fn fn_show_object;
 125
 126 /*
 127  * stats
 128  */
 129 static uint32_t written, written_delta;
 130 static uint32_t reused, reused_delta;
 131
 132 /*
 133  * Indexed commits
 134  */
 135 static struct commit **indexed_commits;
 136 static unsigned int indexed_commits_nr;
 137 static unsigned int indexed_commits_alloc;
 138
 139 static void index_commit_for_bitmap(struct commit *commit)
 140 {
 141         if (indexed_commits_nr >= indexed_commits_alloc) {
 142                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
 143                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
 144         }
 145
 146         indexed_commits[indexed_commits_nr++] = commit;
 147 }
 148
 149 static void *get_delta(struct object_entry *entry)
 150 {
 151         unsigned long size, base_size, delta_size;
 152         void *buf, *base_buf, *delta_buf;
 153         enum object_type type;
 154
 155         buf = read_object_file(&entry->idx.oid, &type, &size);
 156         if (!buf)
 157                 die(_("unable to read %s"), oid_to_hex(&entry->idx.oid));
 158         base_buf = read_object_file(&DELTA(entry)->idx.oid, &type,
 159                                     &base_size);
 160         if (!base_buf)
 161                 die("unable to read %s",
 162                     oid_to_hex(&DELTA(entry)->idx.oid));
 163         delta_buf = diff_delta(base_buf, base_size,
 164                                buf, size, &delta_size, 0);
 165         /*
 166          * We successfully computed this delta once but dropped it for
 167          * memory reasons. Something is very wrong if this time we
 168          * recompute and create a different delta.
 169          */
 170         if (!delta_buf || delta_size != DELTA_SIZE(entry))
 171                 BUG("delta size changed");
 172         free(buf);
 173         free(base_buf);
 174         return delta_buf;
 175 }
 176
 177 static unsigned long do_compress(void **pptr, unsigned long size)
 178 {
 179         git_zstream stream;
 180         void *in, *out;
 181         unsigned long maxsize;
 182
 183         git_deflate_init(&stream, pack_compression_level);
 184         maxsize = git_deflate_bound(&stream, size);
 185
 186         in = *pptr;
 187         out = xmalloc(maxsize);
 188         *pptr = out;
 189
 190         stream.next_in = in;
 191         stream.avail_in = size;
 192         stream.next_out = out;
 193         stream.avail_out = maxsize;
 194         while (git_deflate(&stream, Z_FINISH) == Z_OK)
 195                 ; /* nothing */
 196         git_deflate_end(&stream);
 197
 198         free(in);
 199         return stream.total_out;
 200 }
 201
 202 static unsigned long write_large_blob_data(struct git_istream *st, struct hashfile *f,
 203                                            const struct object_id *oid)
 204 {
 205         git_zstream stream;
 206         unsigned char ibuf[1024 * 16];
 207         unsigned char obuf[1024 * 16];
 208         unsigned long olen = 0;
 209
 210         git_deflate_init(&stream, pack_compression_level);
 211
 212         for (;;) {
 213                 ssize_t readlen;
 214                 int zret = Z_OK;
 215                 readlen = read_istream(st, ibuf, sizeof(ibuf));
 216                 if (readlen == -1)
 217                         die(_("unable to read %s"), oid_to_hex(oid));
 218
 219                 stream.next_in = ibuf;
 220                 stream.avail_in = readlen;
 221                 while ((stream.avail_in || readlen == 0) &&
 222                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
 223                         stream.next_out = obuf;
 224                         stream.avail_out = sizeof(obuf);
 225                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 226                         hashwrite(f, obuf, stream.next_out - obuf);
 227                         olen += stream.next_out - obuf;
 228                 }
 229                 if (stream.avail_in)
 230                         die(_("deflate error (%d)"), zret);
 231                 if (readlen == 0) {
 232                         if (zret != Z_STREAM_END)
 233                                 die(_("deflate error (%d)"), zret);
 234                         break;
 235                 }
 236         }
 237         git_deflate_end(&stream);
 238         return olen;
 239 }
 240
 241 /*
 242  * we are going to reuse the existing object data as is.  make
 243  * sure it is not corrupt.
 244  */
 245 static int check_pack_inflate(struct packed_git *p,
 246                 struct pack_window **w_curs,
 247                 off_t offset,
 248                 off_t len,
 249                 unsigned long expect)
 250 {
 251         git_zstream stream;
 252         unsigned char fakebuf[4096], *in;
 253         int st;
 254
 255         memset(&stream, 0, sizeof(stream));
 256         git_inflate_init(&stream);
 257         do {
 258                 in = use_pack(p, w_curs, offset, &stream.avail_in);
 259                 stream.next_in = in;
 260                 stream.next_out = fakebuf;
 261                 stream.avail_out = sizeof(fakebuf);
 262                 st = git_inflate(&stream, Z_FINISH);
 263                 offset += stream.next_in - in;
 264         } while (st == Z_OK || st == Z_BUF_ERROR);
 265         git_inflate_end(&stream);
 266         return (st == Z_STREAM_END &&
 267                 stream.total_out == expect &&
 268                 stream.total_in == len) ? 0 : -1;
 269 }
 270
 271 static void copy_pack_data(struct hashfile *f,
 272                 struct packed_git *p,
 273                 struct pack_window **w_curs,
 274                 off_t offset,
 275                 off_t len)
 276 {
 277         unsigned char *in;
 278         unsigned long avail;
 279
 280         while (len) {
 281                 in = use_pack(p, w_curs, offset, &avail);
 282                 if (avail > len)
 283                         avail = (unsigned long)len;
 284                 hashwrite(f, in, avail);
 285                 offset += avail;
 286                 len -= avail;
 287         }
 288 }
 289
 290 /* Return 0 if we will bust the pack-size limit */
 291 static unsigned long write_no_reuse_object(struct hashfile *f, struct object_entry *entry,
 292                                            unsigned long limit, int usable_delta)
 293 {
 294         unsigned long size, datalen;
 295         unsigned char header[MAX_PACK_OBJECT_HEADER],
 296                       dheader[MAX_PACK_OBJECT_HEADER];
 297         unsigned hdrlen;
 298         enum object_type type;
 299         void *buf;
 300         struct git_istream *st = NULL;
 301         const unsigned hashsz = the_hash_algo->rawsz;
 302
 303         if (!usable_delta) {
 304                 if (oe_type(entry) == OBJ_BLOB &&
 305                     oe_size_greater_than(&to_pack, entry, big_file_threshold) &&
 306                     (st = open_istream(&entry->idx.oid, &type, &size, NULL)) != NULL)
 307                         buf = NULL;
 308                 else {
 309                         buf = read_object_file(&entry->idx.oid, &type, &size);
 310                         if (!buf)
 311                                 die(_("unable to read %s"),
 312                                     oid_to_hex(&entry->idx.oid));
 313                 }
 314                 /*
 315                  * make sure no cached delta data remains from a
 316                  * previous attempt before a pack split occurred.
 317                  */
 318                 FREE_AND_NULL(entry->delta_data);
 319                 entry->z_delta_size = 0;
 320         } else if (entry->delta_data) {
 321                 size = DELTA_SIZE(entry);
 322                 buf = entry->delta_data;
 323                 entry->delta_data = NULL;
 324                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
 325                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 326         } else {
 327                 buf = get_delta(entry);
 328                 size = DELTA_SIZE(entry);
 329                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
 330                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 331         }
 332
 333         if (st) /* large blob case, just assume we don't compress well */
 334                 datalen = size;
 335         else if (entry->z_delta_size)
 336                 datalen = entry->z_delta_size;
 337         else
 338                 datalen = do_compress(&buf, size);
 339
 340         /*
 341          * The object header is a byte of 'type' followed by zero or
 342          * more bytes of length.
 343          */
 344         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 345                                               type, size);
 346
 347         if (type == OBJ_OFS_DELTA) {
 348                 /*
 349                  * Deltas with relative base contain an additional
 350                  * encoding of the relative offset for the delta
 351                  * base from this object's position in the pack.
 352                  */
 353                 off_t ofs = entry->idx.offset - DELTA(entry)->idx.offset;
 354                 unsigned pos = sizeof(dheader) - 1;
 355                 dheader[pos] = ofs & 127;
 356                 while (ofs >>= 7)
 357                         dheader[--pos] = 128 | (--ofs & 127);
 358                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + hashsz >= limit) {
 359                         if (st)
 360                                 close_istream(st);
 361                         free(buf);
 362                         return 0;
 363                 }
 364                 hashwrite(f, header, hdrlen);
 365                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
 366                 hdrlen += sizeof(dheader) - pos;
 367         } else if (type == OBJ_REF_DELTA) {
 368                 /*
 369                  * Deltas with a base reference contain
 370                  * additional bytes for the base object ID.
 371                  */
 372                 if (limit && hdrlen + hashsz + datalen + hashsz >= limit) {
 373                         if (st)
 374                                 close_istream(st);
 375                         free(buf);
 376                         return 0;
 377                 }
 378                 hashwrite(f, header, hdrlen);
 379                 hashwrite(f, DELTA(entry)->idx.oid.hash, hashsz);
 380                 hdrlen += hashsz;
 381         } else {
 382                 if (limit && hdrlen + datalen + hashsz >= limit) {
 383                         if (st)
 384                                 close_istream(st);
 385                         free(buf);
 386                         return 0;
 387                 }
 388                 hashwrite(f, header, hdrlen);
 389         }
 390         if (st) {
 391                 datalen = write_large_blob_data(st, f, &entry->idx.oid);
 392                 close_istream(st);
 393         } else {
 394                 hashwrite(f, buf, datalen);
 395                 free(buf);
 396         }
 397
 398         return hdrlen + datalen;
 399 }
 400
 401 /* Return 0 if we will bust the pack-size limit */
 402 static off_t write_reuse_object(struct hashfile *f, struct object_entry *entry,
 403                                 unsigned long limit, int usable_delta)
 404 {
 405         struct packed_git *p = IN_PACK(entry);
 406         struct pack_window *w_curs = NULL;
 407         struct revindex_entry *revidx;
 408         off_t offset;
 409         enum object_type type = oe_type(entry);
 410         off_t datalen;
 411         unsigned char header[MAX_PACK_OBJECT_HEADER],
 412                       dheader[MAX_PACK_OBJECT_HEADER];
 413         unsigned hdrlen;
 414         const unsigned hashsz = the_hash_algo->rawsz;
 415         unsigned long entry_size = SIZE(entry);
 416
 417         if (DELTA(entry))
 418                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
 419                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 420         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 421                                               type, entry_size);
 422
 423         offset = entry->in_pack_offset;
 424         revidx = find_pack_revindex(p, offset);
 425         datalen = revidx[1].offset - offset;
 426         if (!pack_to_stdout && p->index_version > 1 &&
 427             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 428                 error(_("bad packed object CRC for %s"),
 429                       oid_to_hex(&entry->idx.oid));
 430                 unuse_pack(&w_curs);
 431                 return write_no_reuse_object(f, entry, limit, usable_delta);
 432         }
 433
 434         offset += entry->in_pack_header_size;
 435         datalen -= entry->in_pack_header_size;
 436
 437         if (!pack_to_stdout && p->index_version == 1 &&
 438             check_pack_inflate(p, &w_curs, offset, datalen, entry_size)) {
 439                 error(_("corrupt packed object for %s"),
 440                       oid_to_hex(&entry->idx.oid));
 441                 unuse_pack(&w_curs);
 442                 return write_no_reuse_object(f, entry, limit, usable_delta);
 443         }
 444
 445         if (type == OBJ_OFS_DELTA) {
 446                 off_t ofs = entry->idx.offset - DELTA(entry)->idx.offset;
 447                 unsigned pos = sizeof(dheader) - 1;
 448                 dheader[pos] = ofs & 127;
 449                 while (ofs >>= 7)
 450                         dheader[--pos] = 128 | (--ofs & 127);
 451                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + hashsz >= limit) {
 452                         unuse_pack(&w_curs);
 453                         return 0;
 454                 }
 455                 hashwrite(f, header, hdrlen);
 456                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
 457                 hdrlen += sizeof(dheader) - pos;
 458                 reused_delta++;
 459         } else if (type == OBJ_REF_DELTA) {
 460                 if (limit && hdrlen + hashsz + datalen + hashsz >= limit) {
 461                         unuse_pack(&w_curs);
 462                         return 0;
 463                 }
 464                 hashwrite(f, header, hdrlen);
 465                 hashwrite(f, DELTA(entry)->idx.oid.hash, hashsz);
 466                 hdrlen += hashsz;
 467                 reused_delta++;
 468         } else {
 469                 if (limit && hdrlen + datalen + hashsz >= limit) {
 470                         unuse_pack(&w_curs);
 471                         return 0;
 472                 }
 473                 hashwrite(f, header, hdrlen);
 474         }
 475         copy_pack_data(f, p, &w_curs, offset, datalen);
 476         unuse_pack(&w_curs);
 477         reused++;
 478         return hdrlen + datalen;
 479 }
 480
 481 /* Return 0 if we will bust the pack-size limit */
 482 static off_t write_object(struct hashfile *f,
 483                           struct object_entry *entry,
 484                           off_t write_offset)
 485 {
 486         unsigned long limit;
 487         off_t len;
 488         int usable_delta, to_reuse;
 489
 490         if (!pack_to_stdout)
 491                 crc32_begin(f);
 492
 493         /* apply size limit if limited packsize and not first object */
 494         if (!pack_size_limit || !nr_written)
 495                 limit = 0;
 496         else if (pack_size_limit <= write_offset)
 497                 /*
 498                  * the earlier object did not fit the limit; avoid
 499                  * mistaking this with unlimited (i.e. limit = 0).
 500                  */
 501                 limit = 1;
 502         else
 503                 limit = pack_size_limit - write_offset;
 504
 505         if (!DELTA(entry))
 506                 usable_delta = 0;       /* no delta */
 507         else if (!pack_size_limit)
 508                usable_delta = 1;        /* unlimited packfile */
 509         else if (DELTA(entry)->idx.offset == (off_t)-1)
 510                 usable_delta = 0;       /* base was written to another pack */
 511         else if (DELTA(entry)->idx.offset)
 512                 usable_delta = 1;       /* base already exists in this pack */
 513         else
 514                 usable_delta = 0;       /* base could end up in another pack */
 515
 516         if (!reuse_object)
 517                 to_reuse = 0;   /* explicit */
 518         else if (!IN_PACK(entry))
 519                 to_reuse = 0;   /* can't reuse what we don't have */
 520         else if (oe_type(entry) == OBJ_REF_DELTA ||
 521                  oe_type(entry) == OBJ_OFS_DELTA)
 522                                 /* check_object() decided it for us ... */
 523                 to_reuse = usable_delta;
 524                                 /* ... but pack split may override that */
 525         else if (oe_type(entry) != entry->in_pack_type)
 526                 to_reuse = 0;   /* pack has delta which is unusable */
 527         else if (DELTA(entry))
 528                 to_reuse = 0;   /* we want to pack afresh */
 529         else
 530                 to_reuse = 1;   /* we have it in-pack undeltified,
 531                                  * and we do not need to deltify it.
 532                                  */
 533
 534         if (!to_reuse)
 535                 len = write_no_reuse_object(f, entry, limit, usable_delta);
 536         else
 537                 len = write_reuse_object(f, entry, limit, usable_delta);
 538         if (!len)
 539                 return 0;
 540
 541         if (usable_delta)
 542                 written_delta++;
 543         written++;
 544         if (!pack_to_stdout)
 545                 entry->idx.crc32 = crc32_end(f);
 546         return len;
 547 }
 548
 549 enum write_one_status {
 550         WRITE_ONE_SKIP = -1, /* already written */
 551         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 552         WRITE_ONE_WRITTEN = 1, /* normal */
 553         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 554 };
 555
 556 static enum write_one_status write_one(struct hashfile *f,
 557                                        struct object_entry *e,
 558                                        off_t *offset)
 559 {
 560         off_t size;
 561         int recursing;
 562
 563         /*
 564          * we set offset to 1 (which is an impossible value) to mark
 565          * the fact that this object is involved in "write its base
 566          * first before writing a deltified object" recursion.
 567          */
 568         recursing = (e->idx.offset == 1);
 569         if (recursing) {
 570                 warning(_("recursive delta detected for object %s"),
 571                         oid_to_hex(&e->idx.oid));
 572                 return WRITE_ONE_RECURSIVE;
 573         } else if (e->idx.offset || e->preferred_base) {
 574                 /* offset is non zero if object is written already. */
 575                 return WRITE_ONE_SKIP;
 576         }
 577
 578         /* if we are deltified, write out base object first. */
 579         if (DELTA(e)) {
 580                 e->idx.offset = 1; /* now recurse */
 581                 switch (write_one(f, DELTA(e), offset)) {
 582                 case WRITE_ONE_RECURSIVE:
 583                         /* we cannot depend on this one */
 584                         SET_DELTA(e, NULL);
 585                         break;
 586                 default:
 587                         break;
 588                 case WRITE_ONE_BREAK:
 589                         e->idx.offset = recursing;
 590                         return WRITE_ONE_BREAK;
 591                 }
 592         }
 593
 594         e->idx.offset = *offset;
 595         size = write_object(f, e, *offset);
 596         if (!size) {
 597                 e->idx.offset = recursing;
 598                 return WRITE_ONE_BREAK;
 599         }
 600         written_list[nr_written++] = &e->idx;
 601
 602         /* make sure off_t is sufficiently large not to wrap */
 603         if (signed_add_overflows(*offset, size))
 604                 die(_("pack too large for current definition of off_t"));
 605         *offset += size;
 606         return WRITE_ONE_WRITTEN;
 607 }
 608
 609 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
 610                        void *cb_data)
 611 {
 612         struct object_id peeled;
 613         struct object_entry *entry = packlist_find(&to_pack, oid);
 614
 615         if (entry)
 616                 entry->tagged = 1;
 617         if (!peel_ref(path, &peeled)) {
 618                 entry = packlist_find(&to_pack, &peeled);
 619                 if (entry)
 620                         entry->tagged = 1;
 621         }
 622         return 0;
 623 }
 624
 625 static inline void add_to_write_order(struct object_entry **wo,
 626                                unsigned int *endp,
 627                                struct object_entry *e)
 628 {
 629         if (e->filled || oe_layer(&to_pack, e) != write_layer)
 630                 return;
 631         wo[(*endp)++] = e;
 632         e->filled = 1;
 633 }
 634
 635 static void add_descendants_to_write_order(struct object_entry **wo,
 636                                            unsigned int *endp,
 637                                            struct object_entry *e)
 638 {
 639         int add_to_order = 1;
 640         while (e) {
 641                 if (add_to_order) {
 642                         struct object_entry *s;
 643                         /* add this node... */
 644                         add_to_write_order(wo, endp, e);
 645                         /* all its siblings... */
 646                         for (s = DELTA_SIBLING(e); s; s = DELTA_SIBLING(s)) {
 647                                 add_to_write_order(wo, endp, s);
 648                         }
 649                 }
 650                 /* drop down a level to add left subtree nodes if possible */
 651                 if (DELTA_CHILD(e)) {
 652                         add_to_order = 1;
 653                         e = DELTA_CHILD(e);
 654                 } else {
 655                         add_to_order = 0;
 656                         /* our sibling might have some children, it is next */
 657                         if (DELTA_SIBLING(e)) {
 658                                 e = DELTA_SIBLING(e);
 659                                 continue;
 660                         }
 661                         /* go back to our parent node */
 662                         e = DELTA(e);
 663                         while (e && !DELTA_SIBLING(e)) {
 664                                 /* we're on the right side of a subtree, keep
 665                                  * going up until we can go right again */
 666                                 e = DELTA(e);
 667                         }
 668                         if (!e) {
 669                                 /* done- we hit our original root node */
 670                                 return;
 671                         }
 672                         /* pass it off to sibling at this level */
 673                         e = DELTA_SIBLING(e);
 674                 }
 675         };
 676 }
 677
 678 static void add_family_to_write_order(struct object_entry **wo,
 679                                       unsigned int *endp,
 680                                       struct object_entry *e)
 681 {
 682         struct object_entry *root;
 683
 684         for (root = e; DELTA(root); root = DELTA(root))
 685                 ; /* nothing */
 686         add_descendants_to_write_order(wo, endp, root);
 687 }
 688
 689 static void compute_layer_order(struct object_entry **wo, unsigned int *wo_end)
 690 {
 691         unsigned int i, last_untagged;
 692         struct object_entry *objects = to_pack.objects;
 693
 694         for (i = 0; i < to_pack.nr_objects; i++) {
 695                 if (objects[i].tagged)
 696                         break;
 697                 add_to_write_order(wo, wo_end, &objects[i]);
 698         }
 699         last_untagged = i;
 700
 701         /*
 702          * Then fill all the tagged tips.
 703          */
 704         for (; i < to_pack.nr_objects; i++) {
 705                 if (objects[i].tagged)
 706                         add_to_write_order(wo, wo_end, &objects[i]);
 707         }
 708
 709         /*
 710          * And then all remaining commits and tags.
 711          */
 712         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 713                 if (oe_type(&objects[i]) != OBJ_COMMIT &&
 714                     oe_type(&objects[i]) != OBJ_TAG)
 715                         continue;
 716                 add_to_write_order(wo, wo_end, &objects[i]);
 717         }
 718
 719         /*
 720          * And then all the trees.
 721          */
 722         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 723                 if (oe_type(&objects[i]) != OBJ_TREE)
 724                         continue;
 725                 add_to_write_order(wo, wo_end, &objects[i]);
 726         }
 727
 728         /*
 729          * Finally all the rest in really tight order
 730          */
 731         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 732                 if (!objects[i].filled && oe_layer(&to_pack, &objects[i]) == write_layer)
 733                         add_family_to_write_order(wo, wo_end, &objects[i]);
 734         }
 735 }
 736
 737 static struct object_entry **compute_write_order(void)
 738 {
 739         uint32_t max_layers = 1;
 740         unsigned int i, wo_end;
 741
 742         struct object_entry **wo;
 743         struct object_entry *objects = to_pack.objects;
 744
 745         for (i = 0; i < to_pack.nr_objects; i++) {
 746                 objects[i].tagged = 0;
 747                 objects[i].filled = 0;
 748                 SET_DELTA_CHILD(&objects[i], NULL);
 749                 SET_DELTA_SIBLING(&objects[i], NULL);
 750         }
 751
 752         /*
 753          * Fully connect delta_child/delta_sibling network.
 754          * Make sure delta_sibling is sorted in the original
 755          * recency order.
 756          */
 757         for (i = to_pack.nr_objects; i > 0;) {
 758                 struct object_entry *e = &objects[--i];
 759                 if (!DELTA(e))
 760                         continue;
 761                 /* Mark me as the first child */
 762                 e->delta_sibling_idx = DELTA(e)->delta_child_idx;
 763                 SET_DELTA_CHILD(DELTA(e), e);
 764         }
 765
 766         /*
 767          * Mark objects that are at the tip of tags.
 768          */
 769         for_each_tag_ref(mark_tagged, NULL);
 770
 771         if (use_delta_islands)
 772                 max_layers = compute_pack_layers(&to_pack);
 773
 774         ALLOC_ARRAY(wo, to_pack.nr_objects);
 775         wo_end = 0;
 776
 777         for (; write_layer < max_layers; ++write_layer)
 778                 compute_layer_order(wo, &wo_end);
 779
 780         if (wo_end != to_pack.nr_objects)
 781                 die(_("ordered %u objects, expected %"PRIu32),
 782                     wo_end, to_pack.nr_objects);
 783
 784         return wo;
 785 }
 786
 787 static off_t write_reused_pack(struct hashfile *f)
 788 {
 789         unsigned char buffer[8192];
 790         off_t to_write, total;
 791         int fd;
 792
 793         if (!is_pack_valid(reuse_packfile))
 794                 die(_("packfile is invalid: %s"), reuse_packfile->pack_name);
 795
 796         fd = git_open(reuse_packfile->pack_name);
 797         if (fd < 0)
 798                 die_errno(_("unable to open packfile for reuse: %s"),
 799                           reuse_packfile->pack_name);
 800
 801         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
 802                 die_errno(_("unable to seek in reused packfile"));
 803
 804         if (reuse_packfile_offset < 0)
 805                 reuse_packfile_offset = reuse_packfile->pack_size - the_hash_algo->rawsz;
 806
 807         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
 808
 809         while (to_write) {
 810                 int read_pack = xread(fd, buffer, sizeof(buffer));
 811
 812                 if (read_pack <= 0)
 813                         die_errno(_("unable to read from reused packfile"));
 814
 815                 if (read_pack > to_write)
 816                         read_pack = to_write;
 817
 818                 hashwrite(f, buffer, read_pack);
 819                 to_write -= read_pack;
 820
 821                 /*
 822                  * We don't know the actual number of objects written,
 823                  * only how many bytes written, how many bytes total, and
 824                  * how many objects total. So we can fake it by pretending all
 825                  * objects we are writing are the same size. This gives us a
 826                  * smooth progress meter, and at the end it matches the true
 827                  * answer.
 828                  */
 829                 written = reuse_packfile_objects *
 830                                 (((double)(total - to_write)) / total);
 831                 display_progress(progress_state, written);
 832         }
 833
 834         close(fd);
 835         written = reuse_packfile_objects;
 836         display_progress(progress_state, written);
 837         return reuse_packfile_offset - sizeof(struct pack_header);
 838 }
 839
 840 static const char no_split_warning[] = N_(
 841 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
 842 );
 843
 844 static void write_pack_file(void)
 845 {
 846         uint32_t i = 0, j;
 847         struct hashfile *f;
 848         off_t offset;
 849         uint32_t nr_remaining = nr_result;
 850         time_t last_mtime = 0;
 851         struct object_entry **write_order;
 852
 853         if (progress > pack_to_stdout)
 854                 progress_state = start_progress(_("Writing objects"), nr_result);
 855         ALLOC_ARRAY(written_list, to_pack.nr_objects);
 856         write_order = compute_write_order();
 857
 858         do {
 859                 struct object_id oid;
 860                 char *pack_tmp_name = NULL;
 861
 862                 if (pack_to_stdout)
 863                         f = hashfd_throughput(1, "<stdout>", progress_state);
 864                 else
 865                         f = create_tmp_packfile(&pack_tmp_name);
 866
 867                 offset = write_pack_header(f, nr_remaining);
 868
 869                 if (reuse_packfile) {
 870                         off_t packfile_size;
 871                         assert(pack_to_stdout);
 872
 873                         packfile_size = write_reused_pack(f);
 874                         offset += packfile_size;
 875                 }
 876
 877                 nr_written = 0;
 878                 for (; i < to_pack.nr_objects; i++) {
 879                         struct object_entry *e = write_order[i];
 880                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 881                                 break;
 882                         display_progress(progress_state, written);
 883                 }
 884
 885                 /*
 886                  * Did we write the wrong # entries in the header?
 887                  * If so, rewrite it like in fast-import
 888                  */
 889                 if (pack_to_stdout) {
 890                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_CLOSE);
 891                 } else if (nr_written == nr_remaining) {
 892                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE);
 893                 } else {
 894                         int fd = finalize_hashfile(f, oid.hash, 0);
 895                         fixup_pack_header_footer(fd, oid.hash, pack_tmp_name,
 896                                                  nr_written, oid.hash, offset);
 897                         close(fd);
 898                         if (write_bitmap_index) {
 899                                 if (write_bitmap_index != WRITE_BITMAP_QUIET)
 900                                         warning(_(no_split_warning));
 901                                 write_bitmap_index = 0;
 902                         }
 903                 }
 904
 905                 if (!pack_to_stdout) {
 906                         struct stat st;
 907                         struct strbuf tmpname = STRBUF_INIT;
 908
 909                         /*
 910                          * Packs are runtime accessed in their mtime
 911                          * order since newer packs are more likely to contain
 912                          * younger objects.  So if we are creating multiple
 913                          * packs then we should modify the mtime of later ones
 914                          * to preserve this property.
 915                          */
 916                         if (stat(pack_tmp_name, &st) < 0) {
 917                                 warning_errno(_("failed to stat %s"), pack_tmp_name);
 918                         } else if (!last_mtime) {
 919                                 last_mtime = st.st_mtime;
 920                         } else {
 921                                 struct utimbuf utb;
 922                                 utb.actime = st.st_atime;
 923                                 utb.modtime = --last_mtime;
 924                                 if (utime(pack_tmp_name, &utb) < 0)
 925                                         warning_errno(_("failed utime() on %s"), pack_tmp_name);
 926                         }
 927
 928                         strbuf_addf(&tmpname, "%s-", base_name);
 929
 930                         if (write_bitmap_index) {
 931                                 bitmap_writer_set_checksum(oid.hash);
 932                                 bitmap_writer_build_type_index(
 933                                         &to_pack, written_list, nr_written);
 934                         }
 935
 936                         finish_tmp_packfile(&tmpname, pack_tmp_name,
 937                                             written_list, nr_written,
 938                                             &pack_idx_opts, oid.hash);
 939
 940                         if (write_bitmap_index) {
 941                                 strbuf_addf(&tmpname, "%s.bitmap", oid_to_hex(&oid));
 942
 943                                 stop_progress(&progress_state);
 944
 945                                 bitmap_writer_show_progress(progress);
 946                                 bitmap_writer_reuse_bitmaps(&to_pack);
 947                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
 948                                 bitmap_writer_build(&to_pack);
 949                                 bitmap_writer_finish(written_list, nr_written,
 950                                                      tmpname.buf, write_bitmap_options);
 951                                 write_bitmap_index = 0;
 952                         }
 953
 954                         strbuf_release(&tmpname);
 955                         free(pack_tmp_name);
 956                         puts(oid_to_hex(&oid));
 957                 }
 958
 959                 /* mark written objects as written to previous pack */
 960                 for (j = 0; j < nr_written; j++) {
 961                         written_list[j]->offset = (off_t)-1;
 962                 }
 963                 nr_remaining -= nr_written;
 964         } while (nr_remaining && i < to_pack.nr_objects);
 965
 966         free(written_list);
 967         free(write_order);
 968         stop_progress(&progress_state);
 969         if (written != nr_result)
 970                 die(_("wrote %"PRIu32" objects while expecting %"PRIu32),
 971                     written, nr_result);
 972         trace2_data_intmax("pack-objects", the_repository,
 973                            "write_pack_file/wrote", nr_result);
 974 }
 975
 976 static int no_try_delta(const char *path)
 977 {
 978         static struct attr_check *check;
 979
 980         if (!check)
 981                 check = attr_check_initl("delta", NULL);
 982         git_check_attr(the_repository->index, path, check);
 983         if (ATTR_FALSE(check->items[0].value))
 984                 return 1;
 985         return 0;
 986 }
 987
 988 /*
 989  * When adding an object, check whether we have already added it
 990  * to our packing list. If so, we can skip. However, if we are
 991  * being asked to excludei t, but the previous mention was to include
 992  * it, make sure to adjust its flags and tweak our numbers accordingly.
 993  *
 994  * As an optimization, we pass out the index position where we would have
 995  * found the item, since that saves us from having to look it up again a
 996  * few lines later when we want to add the new entry.
 997  */
 998 static int have_duplicate_entry(const struct object_id *oid,
 999                                 int exclude)
1000 {
1001         struct object_entry *entry;
1002
1003         entry = packlist_find(&to_pack, oid);
1004         if (!entry)
1005                 return 0;
1006
1007         if (exclude) {
1008                 if (!entry->preferred_base)
1009                         nr_result--;
1010                 entry->preferred_base = 1;
1011         }
1012
1013         return 1;
1014 }
1015
1016 static int want_found_object(int exclude, struct packed_git *p)
1017 {
1018         if (exclude)
1019                 return 1;
1020         if (incremental)
1021                 return 0;
1022
1023         /*
1024          * When asked to do --local (do not include an object that appears in a
1025          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
1026          * an object that appears in a pack marked with .keep), finding a pack
1027          * that matches the criteria is sufficient for us to decide to omit it.
1028          * However, even if this pack does not satisfy the criteria, we need to
1029          * make sure no copy of this object appears in _any_ pack that makes us
1030          * to omit the object, so we need to check all the packs.
1031          *
1032          * We can however first check whether these options can possible matter;
1033          * if they do not matter we know we want the object in generated pack.
1034          * Otherwise, we signal "-1" at the end to tell the caller that we do
1035          * not know either way, and it needs to check more packs.
1036          */
1037         if (!ignore_packed_keep_on_disk &&
1038             !ignore_packed_keep_in_core &&
1039             (!local || !have_non_local_packs))
1040                 return 1;
1041
1042         if (local && !p->pack_local)
1043                 return 0;
1044         if (p->pack_local &&
1045             ((ignore_packed_keep_on_disk && p->pack_keep) ||
1046              (ignore_packed_keep_in_core && p->pack_keep_in_core)))
1047                 return 0;
1048
1049         /* we don't know yet; keep looking for more packs */
1050         return -1;
1051 }
1052
1053 /*
1054  * Check whether we want the object in the pack (e.g., we do not want
1055  * objects found in non-local stores if the "--local" option was used).
1056  *
1057  * If the caller already knows an existing pack it wants to take the object
1058  * from, that is passed in *found_pack and *found_offset; otherwise this
1059  * function finds if there is any pack that has the object and returns the pack
1060  * and its offset in these variables.
1061  */
1062 static int want_object_in_pack(const struct object_id *oid,
1063                                int exclude,
1064                                struct packed_git **found_pack,
1065                                off_t *found_offset)
1066 {
1067         int want;
1068         struct list_head *pos;
1069         struct multi_pack_index *m;
1070
1071         if (!exclude && local && has_loose_object_nonlocal(oid))
1072                 return 0;
1073
1074         /*
1075          * If we already know the pack object lives in, start checks from that
1076          * pack - in the usual case when neither --local was given nor .keep files
1077          * are present we will determine the answer right now.
1078          */
1079         if (*found_pack) {
1080                 want = want_found_object(exclude, *found_pack);
1081                 if (want != -1)
1082                         return want;
1083         }
1084
1085         for (m = get_multi_pack_index(the_repository); m; m = m->next) {
1086                 struct pack_entry e;
1087                 if (fill_midx_entry(the_repository, oid, &e, m)) {
1088                         struct packed_git *p = e.p;
1089                         off_t offset;
1090
1091                         if (p == *found_pack)
1092                                 offset = *found_offset;
1093                         else
1094                                 offset = find_pack_entry_one(oid->hash, p);
1095
1096                         if (offset) {
1097                                 if (!*found_pack) {
1098                                         if (!is_pack_valid(p))
1099                                                 continue;
1100                                         *found_offset = offset;
1101                                         *found_pack = p;
1102                                 }
1103                                 want = want_found_object(exclude, p);
1104                                 if (want != -1)
1105                                         return want;
1106                         }
1107                 }
1108         }
1109
1110         list_for_each(pos, get_packed_git_mru(the_repository)) {
1111                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
1112                 off_t offset;
1113
1114                 if (p == *found_pack)
1115                         offset = *found_offset;
1116                 else
1117                         offset = find_pack_entry_one(oid->hash, p);
1118
1119                 if (offset) {
1120                         if (!*found_pack) {
1121                                 if (!is_pack_valid(p))
1122                                         continue;
1123                                 *found_offset = offset;
1124                                 *found_pack = p;
1125                         }
1126                         want = want_found_object(exclude, p);
1127                         if (!exclude && want > 0)
1128                                 list_move(&p->mru,
1129                                           get_packed_git_mru(the_repository));
1130                         if (want != -1)
1131                                 return want;
1132                 }
1133         }
1134
1135         return 1;
1136 }
1137
1138 static void create_object_entry(const struct object_id *oid,
1139                                 enum object_type type,
1140                                 uint32_t hash,
1141                                 int exclude,
1142                                 int no_try_delta,
1143                                 struct packed_git *found_pack,
1144                                 off_t found_offset)
1145 {
1146         struct object_entry *entry;
1147
1148         entry = packlist_alloc(&to_pack, oid);
1149         entry->hash = hash;
1150         oe_set_type(entry, type);
1151         if (exclude)
1152                 entry->preferred_base = 1;
1153         else
1154                 nr_result++;
1155         if (found_pack) {
1156                 oe_set_in_pack(&to_pack, entry, found_pack);
1157                 entry->in_pack_offset = found_offset;
1158         }
1159
1160         entry->no_try_delta = no_try_delta;
1161 }
1162
1163 static const char no_closure_warning[] = N_(
1164 "disabling bitmap writing, as some objects are not being packed"
1165 );
1166
1167 static int add_object_entry(const struct object_id *oid, enum object_type type,
1168                             const char *name, int exclude)
1169 {
1170         struct packed_git *found_pack = NULL;
1171         off_t found_offset = 0;
1172
1173         display_progress(progress_state, ++nr_seen);
1174
1175         if (have_duplicate_entry(oid, exclude))
1176                 return 0;
1177
1178         if (!want_object_in_pack(oid, exclude, &found_pack, &found_offset)) {
1179                 /* The pack is missing an object, so it will not have closure */
1180                 if (write_bitmap_index) {
1181                         if (write_bitmap_index != WRITE_BITMAP_QUIET)
1182                                 warning(_(no_closure_warning));
1183                         write_bitmap_index = 0;
1184                 }
1185                 return 0;
1186         }
1187
1188         create_object_entry(oid, type, pack_name_hash(name),
1189                             exclude, name && no_try_delta(name),
1190                             found_pack, found_offset);
1191         return 1;
1192 }
1193
1194 static int add_object_entry_from_bitmap(const struct object_id *oid,
1195                                         enum object_type type,
1196                                         int flags, uint32_t name_hash,
1197                                         struct packed_git *pack, off_t offset)
1198 {
1199         display_progress(progress_state, ++nr_seen);
1200
1201         if (have_duplicate_entry(oid, 0))
1202                 return 0;
1203
1204         if (!want_object_in_pack(oid, 0, &pack, &offset))
1205                 return 0;
1206
1207         create_object_entry(oid, type, name_hash, 0, 0, pack, offset);
1208         return 1;
1209 }
1210
1211 struct pbase_tree_cache {
1212         struct object_id oid;
1213         int ref;
1214         int temporary;
1215         void *tree_data;
1216         unsigned long tree_size;
1217 };
1218
1219 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1220 static int pbase_tree_cache_ix(const struct object_id *oid)
1221 {
1222         return oid->hash[0] % ARRAY_SIZE(pbase_tree_cache);
1223 }
1224 static int pbase_tree_cache_ix_incr(int ix)
1225 {
1226         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1227 }
1228
1229 static struct pbase_tree {
1230         struct pbase_tree *next;
1231         /* This is a phony "cache" entry; we are not
1232          * going to evict it or find it through _get()
1233          * mechanism -- this is for the toplevel node that
1234          * would almost always change with any commit.
1235          */
1236         struct pbase_tree_cache pcache;
1237 } *pbase_tree;
1238
1239 static struct pbase_tree_cache *pbase_tree_get(const struct object_id *oid)
1240 {
1241         struct pbase_tree_cache *ent, *nent;
1242         void *data;
1243         unsigned long size;
1244         enum object_type type;
1245         int neigh;
1246         int my_ix = pbase_tree_cache_ix(oid);
1247         int available_ix = -1;
1248
1249         /* pbase-tree-cache acts as a limited hashtable.
1250          * your object will be found at your index or within a few
1251          * slots after that slot if it is cached.
1252          */
1253         for (neigh = 0; neigh < 8; neigh++) {
1254                 ent = pbase_tree_cache[my_ix];
1255                 if (ent && oideq(&ent->oid, oid)) {
1256                         ent->ref++;
1257                         return ent;
1258                 }
1259                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1260                          ((0 <= available_ix) &&
1261                           (!ent && pbase_tree_cache[available_ix])))
1262                         available_ix = my_ix;
1263                 if (!ent)
1264                         break;
1265                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1266         }
1267
1268         /* Did not find one.  Either we got a bogus request or
1269          * we need to read and perhaps cache.
1270          */
1271         data = read_object_file(oid, &type, &size);
1272         if (!data)
1273                 return NULL;
1274         if (type != OBJ_TREE) {
1275                 free(data);
1276                 return NULL;
1277         }
1278
1279         /* We need to either cache or return a throwaway copy */
1280
1281         if (available_ix < 0)
1282                 ent = NULL;
1283         else {
1284                 ent = pbase_tree_cache[available_ix];
1285                 my_ix = available_ix;
1286         }
1287
1288         if (!ent) {
1289                 nent = xmalloc(sizeof(*nent));
1290                 nent->temporary = (available_ix < 0);
1291         }
1292         else {
1293                 /* evict and reuse */
1294                 free(ent->tree_data);
1295                 nent = ent;
1296         }
1297         oidcpy(&nent->oid, oid);
1298         nent->tree_data = data;
1299         nent->tree_size = size;
1300         nent->ref = 1;
1301         if (!nent->temporary)
1302                 pbase_tree_cache[my_ix] = nent;
1303         return nent;
1304 }
1305
1306 static void pbase_tree_put(struct pbase_tree_cache *cache)
1307 {
1308         if (!cache->temporary) {
1309                 cache->ref--;
1310                 return;
1311         }
1312         free(cache->tree_data);
1313         free(cache);
1314 }
1315
1316 static int name_cmp_len(const char *name)
1317 {
1318         int i;
1319         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1320                 ;
1321         return i;
1322 }
1323
1324 static void add_pbase_object(struct tree_desc *tree,
1325                              const char *name,
1326                              int cmplen,
1327                              const char *fullname)
1328 {
1329         struct name_entry entry;
1330         int cmp;
1331
1332         while (tree_entry(tree,&entry)) {
1333                 if (S_ISGITLINK(entry.mode))
1334                         continue;
1335                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1336                       memcmp(name, entry.path, cmplen);
1337                 if (cmp > 0)
1338                         continue;
1339                 if (cmp < 0)
1340                         return;
1341                 if (name[cmplen] != '/') {
1342                         add_object_entry(&entry.oid,
1343                                          object_type(entry.mode),
1344                                          fullname, 1);
1345                         return;
1346                 }
1347                 if (S_ISDIR(entry.mode)) {
1348                         struct tree_desc sub;
1349                         struct pbase_tree_cache *tree;
1350                         const char *down = name+cmplen+1;
1351                         int downlen = name_cmp_len(down);
1352
1353                         tree = pbase_tree_get(&entry.oid);
1354                         if (!tree)
1355                                 return;
1356                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1357
1358                         add_pbase_object(&sub, down, downlen, fullname);
1359                         pbase_tree_put(tree);
1360                 }
1361         }
1362 }
1363
1364 static unsigned *done_pbase_paths;
1365 static int done_pbase_paths_num;
1366 static int done_pbase_paths_alloc;
1367 static int done_pbase_path_pos(unsigned hash)
1368 {
1369         int lo = 0;
1370         int hi = done_pbase_paths_num;
1371         while (lo < hi) {
1372                 int mi = lo + (hi - lo) / 2;
1373                 if (done_pbase_paths[mi] == hash)
1374                         return mi;
1375                 if (done_pbase_paths[mi] < hash)
1376                         hi = mi;
1377                 else
1378                         lo = mi + 1;
1379         }
1380         return -lo-1;
1381 }
1382
1383 static int check_pbase_path(unsigned hash)
1384 {
1385         int pos = done_pbase_path_pos(hash);
1386         if (0 <= pos)
1387                 return 1;
1388         pos = -pos - 1;
1389         ALLOC_GROW(done_pbase_paths,
1390                    done_pbase_paths_num + 1,
1391                    done_pbase_paths_alloc);
1392         done_pbase_paths_num++;
1393         if (pos < done_pbase_paths_num)
1394                 MOVE_ARRAY(done_pbase_paths + pos + 1, done_pbase_paths + pos,
1395                            done_pbase_paths_num - pos - 1);
1396         done_pbase_paths[pos] = hash;
1397         return 0;
1398 }
1399
1400 static void add_preferred_base_object(const char *name)
1401 {
1402         struct pbase_tree *it;
1403         int cmplen;
1404         unsigned hash = pack_name_hash(name);
1405
1406         if (!num_preferred_base || check_pbase_path(hash))
1407                 return;
1408
1409         cmplen = name_cmp_len(name);
1410         for (it = pbase_tree; it; it = it->next) {
1411                 if (cmplen == 0) {
1412                         add_object_entry(&it->pcache.oid, OBJ_TREE, NULL, 1);
1413                 }
1414                 else {
1415                         struct tree_desc tree;
1416                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1417                         add_pbase_object(&tree, name, cmplen, name);
1418                 }
1419         }
1420 }
1421
1422 static void add_preferred_base(struct object_id *oid)
1423 {
1424         struct pbase_tree *it;
1425         void *data;
1426         unsigned long size;
1427         struct object_id tree_oid;
1428
1429         if (window <= num_preferred_base++)
1430                 return;
1431
1432         data = read_object_with_reference(the_repository, oid,
1433                                           tree_type, &size, &tree_oid);
1434         if (!data)
1435                 return;
1436
1437         for (it = pbase_tree; it; it = it->next) {
1438                 if (oideq(&it->pcache.oid, &tree_oid)) {
1439                         free(data);
1440                         return;
1441                 }
1442         }
1443
1444         it = xcalloc(1, sizeof(*it));
1445         it->next = pbase_tree;
1446         pbase_tree = it;
1447
1448         oidcpy(&it->pcache.oid, &tree_oid);
1449         it->pcache.tree_data = data;
1450         it->pcache.tree_size = size;
1451 }
1452
1453 static void cleanup_preferred_base(void)
1454 {
1455         struct pbase_tree *it;
1456         unsigned i;
1457
1458         it = pbase_tree;
1459         pbase_tree = NULL;
1460         while (it) {
1461                 struct pbase_tree *tmp = it;
1462                 it = tmp->next;
1463                 free(tmp->pcache.tree_data);
1464                 free(tmp);
1465         }
1466
1467         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1468                 if (!pbase_tree_cache[i])
1469                         continue;
1470                 free(pbase_tree_cache[i]->tree_data);
1471                 FREE_AND_NULL(pbase_tree_cache[i]);
1472         }
1473
1474         FREE_AND_NULL(done_pbase_paths);
1475         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1476 }
1477
1478 /*
1479  * Return 1 iff the object specified by "delta" can be sent
1480  * literally as a delta against the base in "base_sha1". If
1481  * so, then *base_out will point to the entry in our packing
1482  * list, or NULL if we must use the external-base list.
1483  *
1484  * Depth value does not matter - find_deltas() will
1485  * never consider reused delta as the base object to
1486  * deltify other objects against, in order to avoid
1487  * circular deltas.
1488  */
1489 static int can_reuse_delta(const unsigned char *base_sha1,
1490                            struct object_entry *delta,
1491                            struct object_entry **base_out)
1492 {
1493         struct object_entry *base;
1494         struct object_id base_oid;
1495
1496         if (!base_sha1)
1497                 return 0;
1498
1499         oidread(&base_oid, base_sha1);
1500
1501         /*
1502          * First see if we're already sending the base (or it's explicitly in
1503          * our "excluded" list).
1504          */
1505         base = packlist_find(&to_pack, &base_oid);
1506         if (base) {
1507                 if (!in_same_island(&delta->idx.oid, &base->idx.oid))
1508                         return 0;
1509                 *base_out = base;
1510                 return 1;
1511         }
1512
1513         /*
1514          * Otherwise, reachability bitmaps may tell us if the receiver has it,
1515          * even if it was buried too deep in history to make it into the
1516          * packing list.
1517          */
1518         if (thin && bitmap_has_oid_in_uninteresting(bitmap_git, &base_oid)) {
1519                 if (use_delta_islands) {
1520                         if (!in_same_island(&delta->idx.oid, &base_oid))
1521                                 return 0;
1522                 }
1523                 *base_out = NULL;
1524                 return 1;
1525         }
1526
1527         return 0;
1528 }
1529
1530 static void check_object(struct object_entry *entry)
1531 {
1532         unsigned long canonical_size;
1533
1534         if (IN_PACK(entry)) {
1535                 struct packed_git *p = IN_PACK(entry);
1536                 struct pack_window *w_curs = NULL;
1537                 const unsigned char *base_ref = NULL;
1538                 struct object_entry *base_entry;
1539                 unsigned long used, used_0;
1540                 unsigned long avail;
1541                 off_t ofs;
1542                 unsigned char *buf, c;
1543                 enum object_type type;
1544                 unsigned long in_pack_size;
1545
1546                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1547
1548                 /*
1549                  * We want in_pack_type even if we do not reuse delta
1550                  * since non-delta representations could still be reused.
1551                  */
1552                 used = unpack_object_header_buffer(buf, avail,
1553                                                    &type,
1554                                                    &in_pack_size);
1555                 if (used == 0)
1556                         goto give_up;
1557
1558                 if (type < 0)
1559                         BUG("invalid type %d", type);
1560                 entry->in_pack_type = type;
1561
1562                 /*
1563                  * Determine if this is a delta and if so whether we can
1564                  * reuse it or not.  Otherwise let's find out as cheaply as
1565                  * possible what the actual type and size for this object is.
1566                  */
1567                 switch (entry->in_pack_type) {
1568                 default:
1569                         /* Not a delta hence we've already got all we need. */
1570                         oe_set_type(entry, entry->in_pack_type);
1571                         SET_SIZE(entry, in_pack_size);
1572                         entry->in_pack_header_size = used;
1573                         if (oe_type(entry) < OBJ_COMMIT || oe_type(entry) > OBJ_BLOB)
1574                                 goto give_up;
1575                         unuse_pack(&w_curs);
1576                         return;
1577                 case OBJ_REF_DELTA:
1578                         if (reuse_delta && !entry->preferred_base)
1579                                 base_ref = use_pack(p, &w_curs,
1580                                                 entry->in_pack_offset + used, NULL);
1581                         entry->in_pack_header_size = used + the_hash_algo->rawsz;
1582                         break;
1583                 case OBJ_OFS_DELTA:
1584                         buf = use_pack(p, &w_curs,
1585                                        entry->in_pack_offset + used, NULL);
1586                         used_0 = 0;
1587                         c = buf[used_0++];
1588                         ofs = c & 127;
1589                         while (c & 128) {
1590                                 ofs += 1;
1591                                 if (!ofs || MSB(ofs, 7)) {
1592                                         error(_("delta base offset overflow in pack for %s"),
1593                                               oid_to_hex(&entry->idx.oid));
1594                                         goto give_up;
1595                                 }
1596                                 c = buf[used_0++];
1597                                 ofs = (ofs << 7) + (c & 127);
1598                         }
1599                         ofs = entry->in_pack_offset - ofs;
1600                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1601                                 error(_("delta base offset out of bound for %s"),
1602                                       oid_to_hex(&entry->idx.oid));
1603                                 goto give_up;
1604                         }
1605                         if (reuse_delta && !entry->preferred_base) {
1606                                 struct revindex_entry *revidx;
1607                                 revidx = find_pack_revindex(p, ofs);
1608                                 if (!revidx)
1609                                         goto give_up;
1610                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1611                         }
1612                         entry->in_pack_header_size = used + used_0;
1613                         break;
1614                 }
1615
1616                 if (can_reuse_delta(base_ref, entry, &base_entry)) {
1617                         oe_set_type(entry, entry->in_pack_type);
1618                         SET_SIZE(entry, in_pack_size); /* delta size */
1619                         SET_DELTA_SIZE(entry, in_pack_size);
1620
1621                         if (base_entry) {
1622                                 SET_DELTA(entry, base_entry);
1623                                 entry->delta_sibling_idx = base_entry->delta_child_idx;
1624                                 SET_DELTA_CHILD(base_entry, entry);
1625                         } else {
1626                                 SET_DELTA_EXT(entry, base_ref);
1627                         }
1628
1629                         unuse_pack(&w_curs);
1630                         return;
1631                 }
1632
1633                 if (oe_type(entry)) {
1634                         off_t delta_pos;
1635
1636                         /*
1637                          * This must be a delta and we already know what the
1638                          * final object type is.  Let's extract the actual
1639                          * object size from the delta header.
1640                          */
1641                         delta_pos = entry->in_pack_offset + entry->in_pack_header_size;
1642                         canonical_size = get_size_from_delta(p, &w_curs, delta_pos);
1643                         if (canonical_size == 0)
1644                                 goto give_up;
1645                         SET_SIZE(entry, canonical_size);
1646                         unuse_pack(&w_curs);
1647                         return;
1648                 }
1649
1650                 /*
1651                  * No choice but to fall back to the recursive delta walk
1652                  * with oid_object_info() to find about the object type
1653                  * at this point...
1654                  */
1655                 give_up:
1656                 unuse_pack(&w_curs);
1657         }
1658
1659         oe_set_type(entry,
1660                     oid_object_info(the_repository, &entry->idx.oid, &canonical_size));
1661         if (entry->type_valid) {
1662                 SET_SIZE(entry, canonical_size);
1663         } else {
1664                 /*
1665                  * Bad object type is checked in prepare_pack().  This is
1666                  * to permit a missing preferred base object to be ignored
1667                  * as a preferred base.  Doing so can result in a larger
1668                  * pack file, but the transfer will still take place.
1669                  */
1670         }
1671 }
1672
1673 static int pack_offset_sort(const void *_a, const void *_b)
1674 {
1675         const struct object_entry *a = *(struct object_entry **)_a;
1676         const struct object_entry *b = *(struct object_entry **)_b;
1677         const struct packed_git *a_in_pack = IN_PACK(a);
1678         const struct packed_git *b_in_pack = IN_PACK(b);
1679
1680         /* avoid filesystem trashing with loose objects */
1681         if (!a_in_pack && !b_in_pack)
1682                 return oidcmp(&a->idx.oid, &b->idx.oid);
1683
1684         if (a_in_pack < b_in_pack)
1685                 return -1;
1686         if (a_in_pack > b_in_pack)
1687                 return 1;
1688         return a->in_pack_offset < b->in_pack_offset ? -1 :
1689                         (a->in_pack_offset > b->in_pack_offset);
1690 }
1691
1692 /*
1693  * Drop an on-disk delta we were planning to reuse. Naively, this would
1694  * just involve blanking out the "delta" field, but we have to deal
1695  * with some extra book-keeping:
1696  *
1697  *   1. Removing ourselves from the delta_sibling linked list.
1698  *
1699  *   2. Updating our size/type to the non-delta representation. These were
1700  *      either not recorded initially (size) or overwritten with the delta type
1701  *      (type) when check_object() decided to reuse the delta.
1702  *
1703  *   3. Resetting our delta depth, as we are now a base object.
1704  */
1705 static void drop_reused_delta(struct object_entry *entry)
1706 {
1707         unsigned *idx = &to_pack.objects[entry->delta_idx - 1].delta_child_idx;
1708         struct object_info oi = OBJECT_INFO_INIT;
1709         enum object_type type;
1710         unsigned long size;
1711
1712         while (*idx) {
1713                 struct object_entry *oe = &to_pack.objects[*idx - 1];
1714
1715                 if (oe == entry)
1716                         *idx = oe->delta_sibling_idx;
1717                 else
1718                         idx = &oe->delta_sibling_idx;
1719         }
1720         SET_DELTA(entry, NULL);
1721         entry->depth = 0;
1722
1723         oi.sizep = &size;
1724         oi.typep = &type;
1725         if (packed_object_info(the_repository, IN_PACK(entry), entry->in_pack_offset, &oi) < 0) {
1726                 /*
1727                  * We failed to get the info from this pack for some reason;
1728                  * fall back to oid_object_info, which may find another copy.
1729                  * And if that fails, the error will be recorded in oe_type(entry)
1730                  * and dealt with in prepare_pack().
1731                  */
1732                 oe_set_type(entry,
1733                             oid_object_info(the_repository, &entry->idx.oid, &size));
1734         } else {
1735                 oe_set_type(entry, type);
1736         }
1737         SET_SIZE(entry, size);
1738 }
1739
1740 /*
1741  * Follow the chain of deltas from this entry onward, throwing away any links
1742  * that cause us to hit a cycle (as determined by the DFS state flags in
1743  * the entries).
1744  *
1745  * We also detect too-long reused chains that would violate our --depth
1746  * limit.
1747  */
1748 static void break_delta_chains(struct object_entry *entry)
1749 {
1750         /*
1751          * The actual depth of each object we will write is stored as an int,
1752          * as it cannot exceed our int "depth" limit. But before we break
1753          * changes based no that limit, we may potentially go as deep as the
1754          * number of objects, which is elsewhere bounded to a uint32_t.
1755          */
1756         uint32_t total_depth;
1757         struct object_entry *cur, *next;
1758
1759         for (cur = entry, total_depth = 0;
1760              cur;
1761              cur = DELTA(cur), total_depth++) {
1762                 if (cur->dfs_state == DFS_DONE) {
1763                         /*
1764                          * We've already seen this object and know it isn't
1765                          * part of a cycle. We do need to append its depth
1766                          * to our count.
1767                          */
1768                         total_depth += cur->depth;
1769                         break;
1770                 }
1771
1772                 /*
1773                  * We break cycles before looping, so an ACTIVE state (or any
1774                  * other cruft which made its way into the state variable)
1775                  * is a bug.
1776                  */
1777                 if (cur->dfs_state != DFS_NONE)
1778                         BUG("confusing delta dfs state in first pass: %d",
1779                             cur->dfs_state);
1780
1781                 /*
1782                  * Now we know this is the first time we've seen the object. If
1783                  * it's not a delta, we're done traversing, but we'll mark it
1784                  * done to save time on future traversals.
1785                  */
1786                 if (!DELTA(cur)) {
1787                         cur->dfs_state = DFS_DONE;
1788                         break;
1789                 }
1790
1791                 /*
1792                  * Mark ourselves as active and see if the next step causes
1793                  * us to cycle to another active object. It's important to do
1794                  * this _before_ we loop, because it impacts where we make the
1795                  * cut, and thus how our total_depth counter works.
1796                  * E.g., We may see a partial loop like:
1797                  *
1798                  *   A -> B -> C -> D -> B
1799                  *
1800                  * Cutting B->C breaks the cycle. But now the depth of A is
1801                  * only 1, and our total_depth counter is at 3. The size of the
1802                  * error is always one less than the size of the cycle we
1803                  * broke. Commits C and D were "lost" from A's chain.
1804                  *
1805                  * If we instead cut D->B, then the depth of A is correct at 3.
1806                  * We keep all commits in the chain that we examined.
1807                  */
1808                 cur->dfs_state = DFS_ACTIVE;
1809                 if (DELTA(cur)->dfs_state == DFS_ACTIVE) {
1810                         drop_reused_delta(cur);
1811                         cur->dfs_state = DFS_DONE;
1812                         break;
1813                 }
1814         }
1815
1816         /*
1817          * And now that we've gone all the way to the bottom of the chain, we
1818          * need to clear the active flags and set the depth fields as
1819          * appropriate. Unlike the loop above, which can quit when it drops a
1820          * delta, we need to keep going to look for more depth cuts. So we need
1821          * an extra "next" pointer to keep going after we reset cur->delta.
1822          */
1823         for (cur = entry; cur; cur = next) {
1824                 next = DELTA(cur);
1825
1826                 /*
1827                  * We should have a chain of zero or more ACTIVE states down to
1828                  * a final DONE. We can quit after the DONE, because either it
1829                  * has no bases, or we've already handled them in a previous
1830                  * call.
1831                  */
1832                 if (cur->dfs_state == DFS_DONE)
1833                         break;
1834                 else if (cur->dfs_state != DFS_ACTIVE)
1835                         BUG("confusing delta dfs state in second pass: %d",
1836                             cur->dfs_state);
1837
1838                 /*
1839                  * If the total_depth is more than depth, then we need to snip
1840                  * the chain into two or more smaller chains that don't exceed
1841                  * the maximum depth. Most of the resulting chains will contain
1842                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1843                  * the last chain (i.e., the one containing entry) will contain
1844                  * whatever entries are left over, namely
1845                  * (total_depth % (depth + 1)) of them.
1846                  *
1847                  * Since we are iterating towards decreasing depth, we need to
1848                  * decrement total_depth as we go, and we need to write to the
1849                  * entry what its final depth will be after all of the
1850                  * snipping. Since we're snipping into chains of length (depth
1851                  * + 1) entries, the final depth of an entry will be its
1852                  * original depth modulo (depth + 1). Any time we encounter an
1853                  * entry whose final depth is supposed to be zero, we snip it
1854                  * from its delta base, thereby making it so.
1855                  */
1856                 cur->depth = (total_depth--) % (depth + 1);
1857                 if (!cur->depth)
1858                         drop_reused_delta(cur);
1859
1860                 cur->dfs_state = DFS_DONE;
1861         }
1862 }
1863
1864 static void get_object_details(void)
1865 {
1866         uint32_t i;
1867         struct object_entry **sorted_by_offset;
1868
1869         if (progress)
1870                 progress_state = start_progress(_("Counting objects"),
1871                                                 to_pack.nr_objects);
1872
1873         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1874         for (i = 0; i < to_pack.nr_objects; i++)
1875                 sorted_by_offset[i] = to_pack.objects + i;
1876         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1877
1878         for (i = 0; i < to_pack.nr_objects; i++) {
1879                 struct object_entry *entry = sorted_by_offset[i];
1880                 check_object(entry);
1881                 if (entry->type_valid &&
1882                     oe_size_greater_than(&to_pack, entry, big_file_threshold))
1883                         entry->no_try_delta = 1;
1884                 display_progress(progress_state, i + 1);
1885         }
1886         stop_progress(&progress_state);
1887
1888         /*
1889          * This must happen in a second pass, since we rely on the delta
1890          * information for the whole list being completed.
1891          */
1892         for (i = 0; i < to_pack.nr_objects; i++)
1893                 break_delta_chains(&to_pack.objects[i]);
1894
1895         free(sorted_by_offset);
1896 }
1897
1898 /*
1899  * We search for deltas in a list sorted by type, by filename hash, and then
1900  * by size, so that we see progressively smaller and smaller files.
1901  * That's because we prefer deltas to be from the bigger file
1902  * to the smaller -- deletes are potentially cheaper, but perhaps
1903  * more importantly, the bigger file is likely the more recent
1904  * one.  The deepest deltas are therefore the oldest objects which are
1905  * less susceptible to be accessed often.
1906  */
1907 static int type_size_sort(const void *_a, const void *_b)
1908 {
1909         const struct object_entry *a = *(struct object_entry **)_a;
1910         const struct object_entry *b = *(struct object_entry **)_b;
1911         const enum object_type a_type = oe_type(a);
1912         const enum object_type b_type = oe_type(b);
1913         const unsigned long a_size = SIZE(a);
1914         const unsigned long b_size = SIZE(b);
1915
1916         if (a_type > b_type)
1917                 return -1;
1918         if (a_type < b_type)
1919                 return 1;
1920         if (a->hash > b->hash)
1921                 return -1;
1922         if (a->hash < b->hash)
1923                 return 1;
1924         if (a->preferred_base > b->preferred_base)
1925                 return -1;
1926         if (a->preferred_base < b->preferred_base)
1927                 return 1;
1928         if (use_delta_islands) {
1929                 const int island_cmp = island_delta_cmp(&a->idx.oid, &b->idx.oid);
1930                 if (island_cmp)
1931                         return island_cmp;
1932         }
1933         if (a_size > b_size)
1934                 return -1;
1935         if (a_size < b_size)
1936                 return 1;
1937         return a < b ? -1 : (a > b);  /* newest first */
1938 }
1939
1940 struct unpacked {
1941         struct object_entry *entry;
1942         void *data;
1943         struct delta_index *index;
1944         unsigned depth;
1945 };
1946
1947 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1948                            unsigned long delta_size)
1949 {
1950         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1951                 return 0;
1952
1953         if (delta_size < cache_max_small_delta_size)
1954                 return 1;
1955
1956         /* cache delta, if objects are large enough compared to delta size */
1957         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1958                 return 1;
1959
1960         return 0;
1961 }
1962
1963 /* Protect delta_cache_size */
1964 static pthread_mutex_t cache_mutex;
1965 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1966 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1967
1968 /*
1969  * Protect object list partitioning (e.g. struct thread_param) and
1970  * progress_state
1971  */
1972 static pthread_mutex_t progress_mutex;
1973 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1974 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1975
1976 /*
1977  * Access to struct object_entry is unprotected since each thread owns
1978  * a portion of the main object list. Just don't access object entries
1979  * ahead in the list because they can be stolen and would need
1980  * progress_mutex for protection.
1981  */
1982
1983 /*
1984  * Return the size of the object without doing any delta
1985  * reconstruction (so non-deltas are true object sizes, but deltas
1986  * return the size of the delta data).
1987  */
1988 unsigned long oe_get_size_slow(struct packing_data *pack,
1989                                const struct object_entry *e)
1990 {
1991         struct packed_git *p;
1992         struct pack_window *w_curs;
1993         unsigned char *buf;
1994         enum object_type type;
1995         unsigned long used, avail, size;
1996
1997         if (e->type_ != OBJ_OFS_DELTA && e->type_ != OBJ_REF_DELTA) {
1998                 packing_data_lock(&to_pack);
1999                 if (oid_object_info(the_repository, &e->idx.oid, &size) < 0)
2000                         die(_("unable to get size of %s"),
2001                             oid_to_hex(&e->idx.oid));
2002                 packing_data_unlock(&to_pack);
2003                 return size;
2004         }
2005
2006         p = oe_in_pack(pack, e);
2007         if (!p)
2008                 BUG("when e->type is a delta, it must belong to a pack");
2009
2010         packing_data_lock(&to_pack);
2011         w_curs = NULL;
2012         buf = use_pack(p, &w_curs, e->in_pack_offset, &avail);
2013         used = unpack_object_header_buffer(buf, avail, &type, &size);
2014         if (used == 0)
2015                 die(_("unable to parse object header of %s"),
2016                     oid_to_hex(&e->idx.oid));
2017
2018         unuse_pack(&w_curs);
2019         packing_data_unlock(&to_pack);
2020         return size;
2021 }
2022
2023 static int try_delta(struct unpacked *trg, struct unpacked *src,
2024                      unsigned max_depth, unsigned long *mem_usage)
2025 {
2026         struct object_entry *trg_entry = trg->entry;
2027         struct object_entry *src_entry = src->entry;
2028         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
2029         unsigned ref_depth;
2030         enum object_type type;
2031         void *delta_buf;
2032
2033         /* Don't bother doing diffs between different types */
2034         if (oe_type(trg_entry) != oe_type(src_entry))
2035                 return -1;
2036
2037         /*
2038          * We do not bother to try a delta that we discarded on an
2039          * earlier try, but only when reusing delta data.  Note that
2040          * src_entry that is marked as the preferred_base should always
2041          * be considered, as even if we produce a suboptimal delta against
2042          * it, we will still save the transfer cost, as we already know
2043          * the other side has it and we won't send src_entry at all.
2044          */
2045         if (reuse_delta && IN_PACK(trg_entry) &&
2046             IN_PACK(trg_entry) == IN_PACK(src_entry) &&
2047             !src_entry->preferred_base &&
2048             trg_entry->in_pack_type != OBJ_REF_DELTA &&
2049             trg_entry->in_pack_type != OBJ_OFS_DELTA)
2050                 return 0;
2051
2052         /* Let's not bust the allowed depth. */
2053         if (src->depth >= max_depth)
2054                 return 0;
2055
2056         /* Now some size filtering heuristics. */
2057         trg_size = SIZE(trg_entry);
2058         if (!DELTA(trg_entry)) {
2059                 max_size = trg_size/2 - the_hash_algo->rawsz;
2060                 ref_depth = 1;
2061         } else {
2062                 max_size = DELTA_SIZE(trg_entry);
2063                 ref_depth = trg->depth;
2064         }
2065         max_size = (uint64_t)max_size * (max_depth - src->depth) /
2066                                                 (max_depth - ref_depth + 1);
2067         if (max_size == 0)
2068                 return 0;
2069         src_size = SIZE(src_entry);
2070         sizediff = src_size < trg_size ? trg_size - src_size : 0;
2071         if (sizediff >= max_size)
2072                 return 0;
2073         if (trg_size < src_size / 32)
2074                 return 0;
2075
2076         if (!in_same_island(&trg->entry->idx.oid, &src->entry->idx.oid))
2077                 return 0;
2078
2079         /* Load data if not already done */
2080         if (!trg->data) {
2081                 packing_data_lock(&to_pack);
2082                 trg->data = read_object_file(&trg_entry->idx.oid, &type, &sz);
2083                 packing_data_unlock(&to_pack);
2084                 if (!trg->data)
2085                         die(_("object %s cannot be read"),
2086                             oid_to_hex(&trg_entry->idx.oid));
2087                 if (sz != trg_size)
2088                         die(_("object %s inconsistent object length (%"PRIuMAX" vs %"PRIuMAX")"),
2089                             oid_to_hex(&trg_entry->idx.oid), (uintmax_t)sz,
2090                             (uintmax_t)trg_size);
2091                 *mem_usage += sz;
2092         }
2093         if (!src->data) {
2094                 packing_data_lock(&to_pack);
2095                 src->data = read_object_file(&src_entry->idx.oid, &type, &sz);
2096                 packing_data_unlock(&to_pack);
2097                 if (!src->data) {
2098                         if (src_entry->preferred_base) {
2099                                 static int warned = 0;
2100                                 if (!warned++)
2101                                         warning(_("object %s cannot be read"),
2102                                                 oid_to_hex(&src_entry->idx.oid));
2103                                 /*
2104                                  * Those objects are not included in the
2105                                  * resulting pack.  Be resilient and ignore
2106                                  * them if they can't be read, in case the
2107                                  * pack could be created nevertheless.
2108                                  */
2109                                 return 0;
2110                         }
2111                         die(_("object %s cannot be read"),
2112                             oid_to_hex(&src_entry->idx.oid));
2113                 }
2114                 if (sz != src_size)
2115                         die(_("object %s inconsistent object length (%"PRIuMAX" vs %"PRIuMAX")"),
2116                             oid_to_hex(&src_entry->idx.oid), (uintmax_t)sz,
2117                             (uintmax_t)src_size);
2118                 *mem_usage += sz;
2119         }
2120         if (!src->index) {
2121                 src->index = create_delta_index(src->data, src_size);
2122                 if (!src->index) {
2123                         static int warned = 0;
2124                         if (!warned++)
2125                                 warning(_("suboptimal pack - out of memory"));
2126                         return 0;
2127                 }
2128                 *mem_usage += sizeof_delta_index(src->index);
2129         }
2130
2131         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
2132         if (!delta_buf)
2133                 return 0;
2134
2135         if (DELTA(trg_entry)) {
2136                 /* Prefer only shallower same-sized deltas. */
2137                 if (delta_size == DELTA_SIZE(trg_entry) &&
2138                     src->depth + 1 >= trg->depth) {
2139                         free(delta_buf);
2140                         return 0;
2141                 }
2142         }
2143
2144         /*
2145          * Handle memory allocation outside of the cache
2146          * accounting lock.  Compiler will optimize the strangeness
2147          * away when NO_PTHREADS is defined.
2148          */
2149         free(trg_entry->delta_data);
2150         cache_lock();
2151         if (trg_entry->delta_data) {
2152                 delta_cache_size -= DELTA_SIZE(trg_entry);
2153                 trg_entry->delta_data = NULL;
2154         }
2155         if (delta_cacheable(src_size, trg_size, delta_size)) {
2156                 delta_cache_size += delta_size;
2157                 cache_unlock();
2158                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
2159         } else {
2160                 cache_unlock();
2161                 free(delta_buf);
2162         }
2163
2164         SET_DELTA(trg_entry, src_entry);
2165         SET_DELTA_SIZE(trg_entry, delta_size);
2166         trg->depth = src->depth + 1;
2167
2168         return 1;
2169 }
2170
2171 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
2172 {
2173         struct object_entry *child = DELTA_CHILD(me);
2174         unsigned int m = n;
2175         while (child) {
2176                 const unsigned int c = check_delta_limit(child, n + 1);
2177                 if (m < c)
2178                         m = c;
2179                 child = DELTA_SIBLING(child);
2180         }
2181         return m;
2182 }
2183
2184 static unsigned long free_unpacked(struct unpacked *n)
2185 {
2186         unsigned long freed_mem = sizeof_delta_index(n->index);
2187         free_delta_index(n->index);
2188         n->index = NULL;
2189         if (n->data) {
2190                 freed_mem += SIZE(n->entry);
2191                 FREE_AND_NULL(n->data);
2192         }
2193         n->entry = NULL;
2194         n->depth = 0;
2195         return freed_mem;
2196 }
2197
2198 static void find_deltas(struct object_entry **list, unsigned *list_size,
2199                         int window, int depth, unsigned *processed)
2200 {
2201         uint32_t i, idx = 0, count = 0;
2202         struct unpacked *array;
2203         unsigned long mem_usage = 0;
2204
2205         array = xcalloc(window, sizeof(struct unpacked));
2206
2207         for (;;) {
2208                 struct object_entry *entry;
2209                 struct unpacked *n = array + idx;
2210                 int j, max_depth, best_base = -1;
2211
2212                 progress_lock();
2213                 if (!*list_size) {
2214                         progress_unlock();
2215                         break;
2216                 }
2217                 entry = *list++;
2218                 (*list_size)--;
2219                 if (!entry->preferred_base) {
2220                         (*processed)++;
2221                         display_progress(progress_state, *processed);
2222                 }
2223                 progress_unlock();
2224
2225                 mem_usage -= free_unpacked(n);
2226                 n->entry = entry;
2227
2228                 while (window_memory_limit &&
2229                        mem_usage > window_memory_limit &&
2230                        count > 1) {
2231                         const uint32_t tail = (idx + window - count) % window;
2232                         mem_usage -= free_unpacked(array + tail);
2233                         count--;
2234                 }
2235
2236                 /* We do not compute delta to *create* objects we are not
2237                  * going to pack.
2238                  */
2239                 if (entry->preferred_base)
2240                         goto next;
2241
2242                 /*
2243                  * If the current object is at pack edge, take the depth the
2244                  * objects that depend on the current object into account
2245                  * otherwise they would become too deep.
2246                  */
2247                 max_depth = depth;
2248                 if (DELTA_CHILD(entry)) {
2249                         max_depth -= check_delta_limit(entry, 0);
2250                         if (max_depth <= 0)
2251                                 goto next;
2252                 }
2253
2254                 j = window;
2255                 while (--j > 0) {
2256                         int ret;
2257                         uint32_t other_idx = idx + j;
2258                         struct unpacked *m;
2259                         if (other_idx >= window)
2260                                 other_idx -= window;
2261                         m = array + other_idx;
2262                         if (!m->entry)
2263                                 break;
2264                         ret = try_delta(n, m, max_depth, &mem_usage);
2265                         if (ret < 0)
2266                                 break;
2267                         else if (ret > 0)
2268                                 best_base = other_idx;
2269                 }
2270
2271                 /*
2272                  * If we decided to cache the delta data, then it is best
2273                  * to compress it right away.  First because we have to do
2274                  * it anyway, and doing it here while we're threaded will
2275                  * save a lot of time in the non threaded write phase,
2276                  * as well as allow for caching more deltas within
2277                  * the same cache size limit.
2278                  * ...
2279                  * But only if not writing to stdout, since in that case
2280                  * the network is most likely throttling writes anyway,
2281                  * and therefore it is best to go to the write phase ASAP
2282                  * instead, as we can afford spending more time compressing
2283                  * between writes at that moment.
2284                  */
2285                 if (entry->delta_data && !pack_to_stdout) {
2286                         unsigned long size;
2287
2288                         size = do_compress(&entry->delta_data, DELTA_SIZE(entry));
2289                         if (size < (1U << OE_Z_DELTA_BITS)) {
2290                                 entry->z_delta_size = size;
2291                                 cache_lock();
2292                                 delta_cache_size -= DELTA_SIZE(entry);
2293                                 delta_cache_size += entry->z_delta_size;
2294                                 cache_unlock();
2295                         } else {
2296                                 FREE_AND_NULL(entry->delta_data);
2297                                 entry->z_delta_size = 0;
2298                         }
2299                 }
2300
2301                 /* if we made n a delta, and if n is already at max
2302                  * depth, leaving it in the window is pointless.  we
2303                  * should evict it first.
2304                  */
2305                 if (DELTA(entry) && max_depth <= n->depth)
2306                         continue;
2307
2308                 /*
2309                  * Move the best delta base up in the window, after the
2310                  * currently deltified object, to keep it longer.  It will
2311                  * be the first base object to be attempted next.
2312                  */
2313                 if (DELTA(entry)) {
2314                         struct unpacked swap = array[best_base];
2315                         int dist = (window + idx - best_base) % window;
2316                         int dst = best_base;
2317                         while (dist--) {
2318                                 int src = (dst + 1) % window;
2319                                 array[dst] = array[src];
2320                                 dst = src;
2321                         }
2322                         array[dst] = swap;
2323                 }
2324
2325                 next:
2326                 idx++;
2327                 if (count + 1 < window)
2328                         count++;
2329                 if (idx >= window)
2330                         idx = 0;
2331         }
2332
2333         for (i = 0; i < window; ++i) {
2334                 free_delta_index(array[i].index);
2335                 free(array[i].data);
2336         }
2337         free(array);
2338 }
2339
2340 /*
2341  * The main object list is split into smaller lists, each is handed to
2342  * one worker.
2343  *
2344  * The main thread waits on the condition that (at least) one of the workers
2345  * has stopped working (which is indicated in the .working member of
2346  * struct thread_params).
2347  *
2348  * When a work thread has completed its work, it sets .working to 0 and
2349  * signals the main thread and waits on the condition that .data_ready
2350  * becomes 1.
2351  *
2352  * The main thread steals half of the work from the worker that has
2353  * most work left to hand it to the idle worker.
2354  */
2355
2356 struct thread_params {
2357         pthread_t thread;
2358         struct object_entry **list;
2359         unsigned list_size;
2360         unsigned remaining;
2361         int window;
2362         int depth;
2363         int working;
2364         int data_ready;
2365         pthread_mutex_t mutex;
2366         pthread_cond_t cond;
2367         unsigned *processed;
2368 };
2369
2370 static pthread_cond_t progress_cond;
2371
2372 /*
2373  * Mutex and conditional variable can't be statically-initialized on Windows.
2374  */
2375 static void init_threaded_search(void)
2376 {
2377         pthread_mutex_init(&cache_mutex, NULL);
2378         pthread_mutex_init(&progress_mutex, NULL);
2379         pthread_cond_init(&progress_cond, NULL);
2380 }
2381
2382 static void cleanup_threaded_search(void)
2383 {
2384         pthread_cond_destroy(&progress_cond);
2385         pthread_mutex_destroy(&cache_mutex);
2386         pthread_mutex_destroy(&progress_mutex);
2387 }
2388
2389 static void *threaded_find_deltas(void *arg)
2390 {
2391         struct thread_params *me = arg;
2392
2393         progress_lock();
2394         while (me->remaining) {
2395                 progress_unlock();
2396
2397                 find_deltas(me->list, &me->remaining,
2398                             me->window, me->depth, me->processed);
2399
2400                 progress_lock();
2401                 me->working = 0;
2402                 pthread_cond_signal(&progress_cond);
2403                 progress_unlock();
2404
2405                 /*
2406                  * We must not set ->data_ready before we wait on the
2407                  * condition because the main thread may have set it to 1
2408                  * before we get here. In order to be sure that new
2409                  * work is available if we see 1 in ->data_ready, it
2410                  * was initialized to 0 before this thread was spawned
2411                  * and we reset it to 0 right away.
2412                  */
2413                 pthread_mutex_lock(&me->mutex);
2414                 while (!me->data_ready)
2415                         pthread_cond_wait(&me->cond, &me->mutex);
2416                 me->data_ready = 0;
2417                 pthread_mutex_unlock(&me->mutex);
2418
2419                 progress_lock();
2420         }
2421         progress_unlock();
2422         /* leave ->working 1 so that this doesn't get more work assigned */
2423         return NULL;
2424 }
2425
2426 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2427                            int window, int depth, unsigned *processed)
2428 {
2429         struct thread_params *p;
2430         int i, ret, active_threads = 0;
2431
2432         init_threaded_search();
2433
2434         if (delta_search_threads <= 1) {
2435                 find_deltas(list, &list_size, window, depth, processed);
2436                 cleanup_threaded_search();
2437                 return;
2438         }
2439         if (progress > pack_to_stdout)
2440                 fprintf_ln(stderr, _("Delta compression using up to %d threads"),
2441                            delta_search_threads);
2442         p = xcalloc(delta_search_threads, sizeof(*p));
2443
2444         /* Partition the work amongst work threads. */
2445         for (i = 0; i < delta_search_threads; i++) {
2446                 unsigned sub_size = list_size / (delta_search_threads - i);
2447
2448                 /* don't use too small segments or no deltas will be found */
2449                 if (sub_size < 2*window && i+1 < delta_search_threads)
2450                         sub_size = 0;
2451
2452                 p[i].window = window;
2453                 p[i].depth = depth;
2454                 p[i].processed = processed;
2455                 p[i].working = 1;
2456                 p[i].data_ready = 0;
2457
2458                 /* try to split chunks on "path" boundaries */
2459                 while (sub_size && sub_size < list_size &&
2460                        list[sub_size]->hash &&
2461                        list[sub_size]->hash == list[sub_size-1]->hash)
2462                         sub_size++;
2463
2464                 p[i].list = list;
2465                 p[i].list_size = sub_size;
2466                 p[i].remaining = sub_size;
2467
2468                 list += sub_size;
2469                 list_size -= sub_size;
2470         }
2471
2472         /* Start work threads. */
2473         for (i = 0; i < delta_search_threads; i++) {
2474                 if (!p[i].list_size)
2475                         continue;
2476                 pthread_mutex_init(&p[i].mutex, NULL);
2477                 pthread_cond_init(&p[i].cond, NULL);
2478                 ret = pthread_create(&p[i].thread, NULL,
2479                                      threaded_find_deltas, &p[i]);
2480                 if (ret)
2481                         die(_("unable to create thread: %s"), strerror(ret));
2482                 active_threads++;
2483         }
2484
2485         /*
2486          * Now let's wait for work completion.  Each time a thread is done
2487          * with its work, we steal half of the remaining work from the
2488          * thread with the largest number of unprocessed objects and give
2489          * it to that newly idle thread.  This ensure good load balancing
2490          * until the remaining object list segments are simply too short
2491          * to be worth splitting anymore.
2492          */
2493         while (active_threads) {
2494                 struct thread_params *target = NULL;
2495                 struct thread_params *victim = NULL;
2496                 unsigned sub_size = 0;
2497
2498                 progress_lock();
2499                 for (;;) {
2500                         for (i = 0; !target && i < delta_search_threads; i++)
2501                                 if (!p[i].working)
2502                                         target = &p[i];
2503                         if (target)
2504                                 break;
2505                         pthread_cond_wait(&progress_cond, &progress_mutex);
2506                 }
2507
2508                 for (i = 0; i < delta_search_threads; i++)
2509                         if (p[i].remaining > 2*window &&
2510                             (!victim || victim->remaining < p[i].remaining))
2511                                 victim = &p[i];
2512                 if (victim) {
2513                         sub_size = victim->remaining / 2;
2514                         list = victim->list + victim->list_size - sub_size;
2515                         while (sub_size && list[0]->hash &&
2516                                list[0]->hash == list[-1]->hash) {
2517                                 list++;
2518                                 sub_size--;
2519                         }
2520                         if (!sub_size) {
2521                                 /*
2522                                  * It is possible for some "paths" to have
2523                                  * so many objects that no hash boundary
2524                                  * might be found.  Let's just steal the
2525                                  * exact half in that case.
2526                                  */
2527                                 sub_size = victim->remaining / 2;
2528                                 list -= sub_size;
2529                         }
2530                         target->list = list;
2531                         victim->list_size -= sub_size;
2532                         victim->remaining -= sub_size;
2533                 }
2534                 target->list_size = sub_size;
2535                 target->remaining = sub_size;
2536                 target->working = 1;
2537                 progress_unlock();
2538
2539                 pthread_mutex_lock(&target->mutex);
2540                 target->data_ready = 1;
2541                 pthread_cond_signal(&target->cond);
2542                 pthread_mutex_unlock(&target->mutex);
2543
2544                 if (!sub_size) {
2545                         pthread_join(target->thread, NULL);
2546                         pthread_cond_destroy(&target->cond);
2547                         pthread_mutex_destroy(&target->mutex);
2548                         active_threads--;
2549                 }
2550         }
2551         cleanup_threaded_search();
2552         free(p);
2553 }
2554
2555 static void add_tag_chain(const struct object_id *oid)
2556 {
2557         struct tag *tag;
2558
2559         /*
2560          * We catch duplicates already in add_object_entry(), but we'd
2561          * prefer to do this extra check to avoid having to parse the
2562          * tag at all if we already know that it's being packed (e.g., if
2563          * it was included via bitmaps, we would not have parsed it
2564          * previously).
2565          */
2566         if (packlist_find(&to_pack, oid))
2567                 return;
2568
2569         tag = lookup_tag(the_repository, oid);
2570         while (1) {
2571                 if (!tag || parse_tag(tag) || !tag->tagged)
2572                         die(_("unable to pack objects reachable from tag %s"),
2573                             oid_to_hex(oid));
2574
2575                 add_object_entry(&tag->object.oid, OBJ_TAG, NULL, 0);
2576
2577                 if (tag->tagged->type != OBJ_TAG)
2578                         return;
2579
2580                 tag = (struct tag *)tag->tagged;
2581         }
2582 }
2583
2584 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2585 {
2586         struct object_id peeled;
2587
2588         if (starts_with(path, "refs/tags/") && /* is a tag? */
2589             !peel_ref(path, &peeled)    && /* peelable? */
2590             packlist_find(&to_pack, &peeled))      /* object packed? */
2591                 add_tag_chain(oid);
2592         return 0;
2593 }
2594
2595 static void prepare_pack(int window, int depth)
2596 {
2597         struct object_entry **delta_list;
2598         uint32_t i, nr_deltas;
2599         unsigned n;
2600
2601         if (use_delta_islands)
2602                 resolve_tree_islands(the_repository, progress, &to_pack);
2603
2604         get_object_details();
2605
2606         /*
2607          * If we're locally repacking then we need to be doubly careful
2608          * from now on in order to make sure no stealth corruption gets
2609          * propagated to the new pack.  Clients receiving streamed packs
2610          * should validate everything they get anyway so no need to incur
2611          * the additional cost here in that case.
2612          */
2613         if (!pack_to_stdout)
2614                 do_check_packed_object_crc = 1;
2615
2616         if (!to_pack.nr_objects || !window || !depth)
2617                 return;
2618
2619         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2620         nr_deltas = n = 0;
2621
2622         for (i = 0; i < to_pack.nr_objects; i++) {
2623                 struct object_entry *entry = to_pack.objects + i;
2624
2625                 if (DELTA(entry))
2626                         /* This happens if we decided to reuse existing
2627                          * delta from a pack.  "reuse_delta &&" is implied.
2628                          */
2629                         continue;
2630
2631                 if (!entry->type_valid ||
2632                     oe_size_less_than(&to_pack, entry, 50))
2633                         continue;
2634
2635                 if (entry->no_try_delta)
2636                         continue;
2637
2638                 if (!entry->preferred_base) {
2639                         nr_deltas++;
2640                         if (oe_type(entry) < 0)
2641                                 die(_("unable to get type of object %s"),
2642                                     oid_to_hex(&entry->idx.oid));
2643                 } else {
2644                         if (oe_type(entry) < 0) {
2645                                 /*
2646                                  * This object is not found, but we
2647                                  * don't have to include it anyway.
2648                                  */
2649                                 continue;
2650                         }
2651                 }
2652
2653                 delta_list[n++] = entry;
2654         }
2655
2656         if (nr_deltas && n > 1) {
2657                 unsigned nr_done = 0;
2658                 if (progress)
2659                         progress_state = start_progress(_("Compressing objects"),
2660                                                         nr_deltas);
2661                 QSORT(delta_list, n, type_size_sort);
2662                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2663                 stop_progress(&progress_state);
2664                 if (nr_done != nr_deltas)
2665                         die(_("inconsistency with delta count"));
2666         }
2667         free(delta_list);
2668 }
2669
2670 static int git_pack_config(const char *k, const char *v, void *cb)
2671 {
2672         if (!strcmp(k, "pack.window")) {
2673                 window = git_config_int(k, v);
2674                 return 0;
2675         }
2676         if (!strcmp(k, "pack.windowmemory")) {
2677                 window_memory_limit = git_config_ulong(k, v);
2678                 return 0;
2679         }
2680         if (!strcmp(k, "pack.depth")) {
2681                 depth = git_config_int(k, v);
2682                 return 0;
2683         }
2684         if (!strcmp(k, "pack.deltacachesize")) {
2685                 max_delta_cache_size = git_config_int(k, v);
2686                 return 0;
2687         }
2688         if (!strcmp(k, "pack.deltacachelimit")) {
2689                 cache_max_small_delta_size = git_config_int(k, v);
2690                 return 0;
2691         }
2692         if (!strcmp(k, "pack.writebitmaphashcache")) {
2693                 if (git_config_bool(k, v))
2694                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2695                 else
2696                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2697         }
2698         if (!strcmp(k, "pack.usebitmaps")) {
2699                 use_bitmap_index_default = git_config_bool(k, v);
2700                 return 0;
2701         }
2702         if (!strcmp(k, "pack.threads")) {
2703                 delta_search_threads = git_config_int(k, v);
2704                 if (delta_search_threads < 0)
2705                         die(_("invalid number of threads specified (%d)"),
2706                             delta_search_threads);
2707                 if (!HAVE_THREADS && delta_search_threads != 1) {
2708                         warning(_("no threads support, ignoring %s"), k);
2709                         delta_search_threads = 0;
2710                 }
2711                 return 0;
2712         }
2713         if (!strcmp(k, "pack.indexversion")) {
2714                 pack_idx_opts.version = git_config_int(k, v);
2715                 if (pack_idx_opts.version > 2)
2716                         die(_("bad pack.indexversion=%"PRIu32),
2717                             pack_idx_opts.version);
2718                 return 0;
2719         }
2720         return git_default_config(k, v, cb);
2721 }
2722
2723 static void read_object_list_from_stdin(void)
2724 {
2725         char line[GIT_MAX_HEXSZ + 1 + PATH_MAX + 2];
2726         struct object_id oid;
2727         const char *p;
2728
2729         for (;;) {
2730                 if (!fgets(line, sizeof(line), stdin)) {
2731                         if (feof(stdin))
2732                                 break;
2733                         if (!ferror(stdin))
2734                                 die("BUG: fgets returned NULL, not EOF, not error!");
2735                         if (errno != EINTR)
2736                                 die_errno("fgets");
2737                         clearerr(stdin);
2738                         continue;
2739                 }
2740                 if (line[0] == '-') {
2741                         if (get_oid_hex(line+1, &oid))
2742                                 die(_("expected edge object ID, got garbage:\n %s"),
2743                                     line);
2744                         add_preferred_base(&oid);
2745                         continue;
2746                 }
2747                 if (parse_oid_hex(line, &oid, &p))
2748                         die(_("expected object ID, got garbage:\n %s"), line);
2749
2750                 add_preferred_base_object(p + 1);
2751                 add_object_entry(&oid, OBJ_NONE, p + 1, 0);
2752         }
2753 }
2754
2755 /* Remember to update object flag allocation in object.h */
2756 #define OBJECT_ADDED (1u<<20)
2757
2758 static void show_commit(struct commit *commit, void *data)
2759 {
2760         add_object_entry(&commit->object.oid, OBJ_COMMIT, NULL, 0);
2761         commit->object.flags |= OBJECT_ADDED;
2762
2763         if (write_bitmap_index)
2764                 index_commit_for_bitmap(commit);
2765
2766         if (use_delta_islands)
2767                 propagate_island_marks(commit);
2768 }
2769
2770 static void show_object(struct object *obj, const char *name, void *data)
2771 {
2772         add_preferred_base_object(name);
2773         add_object_entry(&obj->oid, obj->type, name, 0);
2774         obj->flags |= OBJECT_ADDED;
2775
2776         if (use_delta_islands) {
2777                 const char *p;
2778                 unsigned depth;
2779                 struct object_entry *ent;
2780
2781                 /* the empty string is a root tree, which is depth 0 */
2782                 depth = *name ? 1 : 0;
2783                 for (p = strchr(name, '/'); p; p = strchr(p + 1, '/'))
2784                         depth++;
2785
2786                 ent = packlist_find(&to_pack, &obj->oid);
2787                 if (ent && depth > oe_tree_depth(&to_pack, ent))
2788                         oe_set_tree_depth(&to_pack, ent, depth);
2789         }
2790 }
2791
2792 static void show_object__ma_allow_any(struct object *obj, const char *name, void *data)
2793 {
2794         assert(arg_missing_action == MA_ALLOW_ANY);
2795
2796         /*
2797          * Quietly ignore ALL missing objects.  This avoids problems with
2798          * staging them now and getting an odd error later.
2799          */
2800         if (!has_object_file(&obj->oid))
2801                 return;
2802
2803         show_object(obj, name, data);
2804 }
2805
2806 static void show_object__ma_allow_promisor(struct object *obj, const char *name, void *data)
2807 {
2808         assert(arg_missing_action == MA_ALLOW_PROMISOR);
2809
2810         /*
2811          * Quietly ignore EXPECTED missing objects.  This avoids problems with
2812          * staging them now and getting an odd error later.
2813          */
2814         if (!has_object_file(&obj->oid) && is_promisor_object(&obj->oid))
2815                 return;
2816
2817         show_object(obj, name, data);
2818 }
2819
2820 static int option_parse_missing_action(const struct option *opt,
2821                                        const char *arg, int unset)
2822 {
2823         assert(arg);
2824         assert(!unset);
2825
2826         if (!strcmp(arg, "error")) {
2827                 arg_missing_action = MA_ERROR;
2828                 fn_show_object = show_object;
2829                 return 0;
2830         }
2831
2832         if (!strcmp(arg, "allow-any")) {
2833                 arg_missing_action = MA_ALLOW_ANY;
2834                 fetch_if_missing = 0;
2835                 fn_show_object = show_object__ma_allow_any;
2836                 return 0;
2837         }
2838
2839         if (!strcmp(arg, "allow-promisor")) {
2840                 arg_missing_action = MA_ALLOW_PROMISOR;
2841                 fetch_if_missing = 0;
2842                 fn_show_object = show_object__ma_allow_promisor;
2843                 return 0;
2844         }
2845
2846         die(_("invalid value for --missing"));
2847         return 0;
2848 }
2849
2850 static void show_edge(struct commit *commit)
2851 {
2852         add_preferred_base(&commit->object.oid);
2853 }
2854
2855 struct in_pack_object {
2856         off_t offset;
2857         struct object *object;
2858 };
2859
2860 struct in_pack {
2861         unsigned int alloc;
2862         unsigned int nr;
2863         struct in_pack_object *array;
2864 };
2865
2866 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2867 {
2868         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2869         in_pack->array[in_pack->nr].object = object;
2870         in_pack->nr++;
2871 }
2872
2873 /*
2874  * Compare the objects in the offset order, in order to emulate the
2875  * "git rev-list --objects" output that produced the pack originally.
2876  */
2877 static int ofscmp(const void *a_, const void *b_)
2878 {
2879         struct in_pack_object *a = (struct in_pack_object *)a_;
2880         struct in_pack_object *b = (struct in_pack_object *)b_;
2881
2882         if (a->offset < b->offset)
2883                 return -1;
2884         else if (a->offset > b->offset)
2885                 return 1;
2886         else
2887                 return oidcmp(&a->object->oid, &b->object->oid);
2888 }
2889
2890 static void add_objects_in_unpacked_packs(void)
2891 {
2892         struct packed_git *p;
2893         struct in_pack in_pack;
2894         uint32_t i;
2895
2896         memset(&in_pack, 0, sizeof(in_pack));
2897
2898         for (p = get_all_packs(the_repository); p; p = p->next) {
2899                 struct object_id oid;
2900                 struct object *o;
2901
2902                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
2903                         continue;
2904                 if (open_pack_index(p))
2905                         die(_("cannot open pack index"));
2906
2907                 ALLOC_GROW(in_pack.array,
2908                            in_pack.nr + p->num_objects,
2909                            in_pack.alloc);
2910
2911                 for (i = 0; i < p->num_objects; i++) {
2912                         nth_packed_object_oid(&oid, p, i);
2913                         o = lookup_unknown_object(&oid);
2914                         if (!(o->flags & OBJECT_ADDED))
2915                                 mark_in_pack_object(o, p, &in_pack);
2916                         o->flags |= OBJECT_ADDED;
2917                 }
2918         }
2919
2920         if (in_pack.nr) {
2921                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2922                 for (i = 0; i < in_pack.nr; i++) {
2923                         struct object *o = in_pack.array[i].object;
2924                         add_object_entry(&o->oid, o->type, "", 0);
2925                 }
2926         }
2927         free(in_pack.array);
2928 }
2929
2930 static int add_loose_object(const struct object_id *oid, const char *path,
2931                             void *data)
2932 {
2933         enum object_type type = oid_object_info(the_repository, oid, NULL);
2934
2935         if (type < 0) {
2936                 warning(_("loose object at %s could not be examined"), path);
2937                 return 0;
2938         }
2939
2940         add_object_entry(oid, type, "", 0);
2941         return 0;
2942 }
2943
2944 /*
2945  * We actually don't even have to worry about reachability here.
2946  * add_object_entry will weed out duplicates, so we just add every
2947  * loose object we find.
2948  */
2949 static void add_unreachable_loose_objects(void)
2950 {
2951         for_each_loose_file_in_objdir(get_object_directory(),
2952                                       add_loose_object,
2953                                       NULL, NULL, NULL);
2954 }
2955
2956 static int has_sha1_pack_kept_or_nonlocal(const struct object_id *oid)
2957 {
2958         static struct packed_git *last_found = (void *)1;
2959         struct packed_git *p;
2960
2961         p = (last_found != (void *)1) ? last_found :
2962                                         get_all_packs(the_repository);
2963
2964         while (p) {
2965                 if ((!p->pack_local || p->pack_keep ||
2966                                 p->pack_keep_in_core) &&
2967                         find_pack_entry_one(oid->hash, p)) {
2968                         last_found = p;
2969                         return 1;
2970                 }
2971                 if (p == last_found)
2972                         p = get_all_packs(the_repository);
2973                 else
2974                         p = p->next;
2975                 if (p == last_found)
2976                         p = p->next;
2977         }
2978         return 0;
2979 }
2980
2981 /*
2982  * Store a list of sha1s that are should not be discarded
2983  * because they are either written too recently, or are
2984  * reachable from another object that was.
2985  *
2986  * This is filled by get_object_list.
2987  */
2988 static struct oid_array recent_objects;
2989
2990 static int loosened_object_can_be_discarded(const struct object_id *oid,
2991                                             timestamp_t mtime)
2992 {
2993         if (!unpack_unreachable_expiration)
2994                 return 0;
2995         if (mtime > unpack_unreachable_expiration)
2996                 return 0;
2997         if (oid_array_lookup(&recent_objects, oid) >= 0)
2998                 return 0;
2999         return 1;
3000 }
3001
3002 static void loosen_unused_packed_objects(void)
3003 {
3004         struct packed_git *p;
3005         uint32_t i;
3006         struct object_id oid;
3007
3008         for (p = get_all_packs(the_repository); p; p = p->next) {
3009                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
3010                         continue;
3011
3012                 if (open_pack_index(p))
3013                         die(_("cannot open pack index"));
3014
3015                 for (i = 0; i < p->num_objects; i++) {
3016                         nth_packed_object_oid(&oid, p, i);
3017                         if (!packlist_find(&to_pack, &oid) &&
3018                             !has_sha1_pack_kept_or_nonlocal(&oid) &&
3019                             !loosened_object_can_be_discarded(&oid, p->mtime))
3020                                 if (force_object_loose(&oid, p->mtime))
3021                                         die(_("unable to force loose object"));
3022                 }
3023         }
3024 }
3025
3026 /*
3027  * This tracks any options which pack-reuse code expects to be on, or which a
3028  * reader of the pack might not understand, and which would therefore prevent
3029  * blind reuse of what we have on disk.
3030  */
3031 static int pack_options_allow_reuse(void)
3032 {
3033         return pack_to_stdout &&
3034                allow_ofs_delta &&
3035                !ignore_packed_keep_on_disk &&
3036                !ignore_packed_keep_in_core &&
3037                (!local || !have_non_local_packs) &&
3038                !incremental;
3039 }
3040
3041 static int get_object_list_from_bitmap(struct rev_info *revs)
3042 {
3043         if (!(bitmap_git = prepare_bitmap_walk(revs)))
3044                 return -1;
3045
3046         if (pack_options_allow_reuse() &&
3047             !reuse_partial_packfile_from_bitmap(
3048                         bitmap_git,
3049                         &reuse_packfile,
3050                         &reuse_packfile_objects,
3051                         &reuse_packfile_offset)) {
3052                 assert(reuse_packfile_objects);
3053                 nr_result += reuse_packfile_objects;
3054                 display_progress(progress_state, nr_result);
3055         }
3056
3057         traverse_bitmap_commit_list(bitmap_git, &add_object_entry_from_bitmap);
3058         return 0;
3059 }
3060
3061 static void record_recent_object(struct object *obj,
3062                                  const char *name,
3063                                  void *data)
3064 {
3065         oid_array_append(&recent_objects, &obj->oid);
3066 }
3067
3068 static void record_recent_commit(struct commit *commit, void *data)
3069 {
3070         oid_array_append(&recent_objects, &commit->object.oid);
3071 }
3072
3073 static void get_object_list(int ac, const char **av)
3074 {
3075         struct rev_info revs;
3076         struct setup_revision_opt s_r_opt = {
3077                 .allow_exclude_promisor_objects = 1,
3078         };
3079         char line[1000];
3080         int flags = 0;
3081         int save_warning;
3082
3083         repo_init_revisions(the_repository, &revs, NULL);
3084         save_commit_buffer = 0;
3085         setup_revisions(ac, av, &revs, &s_r_opt);
3086
3087         /* make sure shallows are read */
3088         is_repository_shallow(the_repository);
3089
3090         save_warning = warn_on_object_refname_ambiguity;
3091         warn_on_object_refname_ambiguity = 0;
3092
3093         while (fgets(line, sizeof(line), stdin) != NULL) {
3094                 int len = strlen(line);
3095                 if (len && line[len - 1] == '\n')
3096                         line[--len] = 0;
3097                 if (!len)
3098                         break;
3099                 if (*line == '-') {
3100                         if (!strcmp(line, "--not")) {
3101                                 flags ^= UNINTERESTING;
3102                                 write_bitmap_index = 0;
3103                                 continue;
3104                         }
3105                         if (starts_with(line, "--shallow ")) {
3106                                 struct object_id oid;
3107                                 if (get_oid_hex(line + 10, &oid))
3108                                         die("not an SHA-1 '%s'", line + 10);
3109                                 register_shallow(the_repository, &oid);
3110                                 use_bitmap_index = 0;
3111                                 continue;
3112                         }
3113                         die(_("not a rev '%s'"), line);
3114                 }
3115                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
3116                         die(_("bad revision '%s'"), line);
3117         }
3118
3119         warn_on_object_refname_ambiguity = save_warning;
3120
3121         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
3122                 return;
3123
3124         if (use_delta_islands)
3125                 load_delta_islands(the_repository, progress);
3126
3127         if (prepare_revision_walk(&revs))
3128                 die(_("revision walk setup failed"));
3129         mark_edges_uninteresting(&revs, show_edge, sparse);
3130
3131         if (!fn_show_object)
3132                 fn_show_object = show_object;
3133         traverse_commit_list_filtered(&filter_options, &revs,
3134                                       show_commit, fn_show_object, NULL,
3135                                       NULL);
3136
3137         if (unpack_unreachable_expiration) {
3138                 revs.ignore_missing_links = 1;
3139                 if (add_unseen_recent_objects_to_traversal(&revs,
3140                                 unpack_unreachable_expiration))
3141                         die(_("unable to add recent objects"));
3142                 if (prepare_revision_walk(&revs))
3143                         die(_("revision walk setup failed"));
3144                 traverse_commit_list(&revs, record_recent_commit,
3145                                      record_recent_object, NULL);
3146         }
3147
3148         if (keep_unreachable)
3149                 add_objects_in_unpacked_packs();
3150         if (pack_loose_unreachable)
3151                 add_unreachable_loose_objects();
3152         if (unpack_unreachable)
3153                 loosen_unused_packed_objects();
3154
3155         oid_array_clear(&recent_objects);
3156 }
3157
3158 static void add_extra_kept_packs(const struct string_list *names)
3159 {
3160         struct packed_git *p;
3161
3162         if (!names->nr)
3163                 return;
3164
3165         for (p = get_all_packs(the_repository); p; p = p->next) {
3166                 const char *name = basename(p->pack_name);
3167                 int i;
3168
3169                 if (!p->pack_local)
3170                         continue;
3171
3172                 for (i = 0; i < names->nr; i++)
3173                         if (!fspathcmp(name, names->items[i].string))
3174                                 break;
3175
3176                 if (i < names->nr) {
3177                         p->pack_keep_in_core = 1;
3178                         ignore_packed_keep_in_core = 1;
3179                         continue;
3180                 }
3181         }
3182 }
3183
3184 static int option_parse_index_version(const struct option *opt,
3185                                       const char *arg, int unset)
3186 {
3187         char *c;
3188         const char *val = arg;
3189
3190         BUG_ON_OPT_NEG(unset);
3191
3192         pack_idx_opts.version = strtoul(val, &c, 10);
3193         if (pack_idx_opts.version > 2)
3194                 die(_("unsupported index version %s"), val);
3195         if (*c == ',' && c[1])
3196                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
3197         if (*c || pack_idx_opts.off32_limit & 0x80000000)
3198                 die(_("bad index version '%s'"), val);
3199         return 0;
3200 }
3201
3202 static int option_parse_unpack_unreachable(const struct option *opt,
3203                                            const char *arg, int unset)
3204 {
3205         if (unset) {
3206                 unpack_unreachable = 0;
3207                 unpack_unreachable_expiration = 0;
3208         }
3209         else {
3210                 unpack_unreachable = 1;
3211                 if (arg)
3212                         unpack_unreachable_expiration = approxidate(arg);
3213         }
3214         return 0;
3215 }
3216
3217 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
3218 {
3219         int use_internal_rev_list = 0;
3220         int shallow = 0;
3221         int all_progress_implied = 0;
3222         struct argv_array rp = ARGV_ARRAY_INIT;
3223         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
3224         int rev_list_index = 0;
3225         struct string_list keep_pack_list = STRING_LIST_INIT_NODUP;
3226         struct option pack_objects_options[] = {
3227                 OPT_SET_INT('q', "quiet", &progress,
3228                             N_("do not show progress meter"), 0),
3229                 OPT_SET_INT(0, "progress", &progress,
3230                             N_("show progress meter"), 1),
3231                 OPT_SET_INT(0, "all-progress", &progress,
3232                             N_("show progress meter during object writing phase"), 2),
3233                 OPT_BOOL(0, "all-progress-implied",
3234                          &all_progress_implied,
3235                          N_("similar to --all-progress when progress meter is shown")),
3236                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("<version>[,<offset>]"),
3237                   N_("write the pack index file in the specified idx format version"),
3238                   PARSE_OPT_NONEG, option_parse_index_version },
3239                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
3240                               N_("maximum size of each output pack file")),
3241                 OPT_BOOL(0, "local", &local,
3242                          N_("ignore borrowed objects from alternate object store")),
3243                 OPT_BOOL(0, "incremental", &incremental,
3244                          N_("ignore packed objects")),
3245                 OPT_INTEGER(0, "window", &window,
3246                             N_("limit pack window by objects")),
3247                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
3248                               N_("limit pack window by memory in addition to object limit")),
3249                 OPT_INTEGER(0, "depth", &depth,
3250                             N_("maximum length of delta chain allowed in the resulting pack")),
3251                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
3252                          N_("reuse existing deltas")),
3253                 OPT_BOOL(0, "reuse-object", &reuse_object,
3254                          N_("reuse existing objects")),
3255                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
3256                          N_("use OFS_DELTA objects")),
3257                 OPT_INTEGER(0, "threads", &delta_search_threads,
3258                             N_("use threads when searching for best delta matches")),
3259                 OPT_BOOL(0, "non-empty", &non_empty,
3260                          N_("do not create an empty pack output")),
3261                 OPT_BOOL(0, "revs", &use_internal_rev_list,
3262                          N_("read revision arguments from standard input")),
3263                 OPT_SET_INT_F(0, "unpacked", &rev_list_unpacked,
3264                               N_("limit the objects to those that are not yet packed"),
3265                               1, PARSE_OPT_NONEG),
3266                 OPT_SET_INT_F(0, "all", &rev_list_all,
3267                               N_("include objects reachable from any reference"),
3268                               1, PARSE_OPT_NONEG),
3269                 OPT_SET_INT_F(0, "reflog", &rev_list_reflog,
3270                               N_("include objects referred by reflog entries"),
3271                               1, PARSE_OPT_NONEG),
3272                 OPT_SET_INT_F(0, "indexed-objects", &rev_list_index,
3273                               N_("include objects referred to by the index"),
3274                               1, PARSE_OPT_NONEG),
3275                 OPT_BOOL(0, "stdout", &pack_to_stdout,
3276                          N_("output pack to stdout")),
3277                 OPT_BOOL(0, "include-tag", &include_tag,
3278                          N_("include tag objects that refer to objects to be packed")),
3279                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
3280                          N_("keep unreachable objects")),
3281                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
3282                          N_("pack loose unreachable objects")),
3283                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
3284                   N_("unpack unreachable objects newer than <time>"),
3285                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
3286                 OPT_BOOL(0, "sparse", &sparse,
3287                          N_("use the sparse reachability algorithm")),
3288                 OPT_BOOL(0, "thin", &thin,
3289                          N_("create thin packs")),
3290                 OPT_BOOL(0, "shallow", &shallow,
3291                          N_("create packs suitable for shallow fetches")),
3292                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep_on_disk,
3293                          N_("ignore packs that have companion .keep file")),
3294                 OPT_STRING_LIST(0, "keep-pack", &keep_pack_list, N_("name"),
3295                                 N_("ignore this pack")),
3296                 OPT_INTEGER(0, "compression", &pack_compression_level,
3297                             N_("pack compression level")),
3298                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
3299                             N_("do not hide commits by grafts"), 0),
3300                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
3301                          N_("use a bitmap index if available to speed up counting objects")),
3302                 OPT_SET_INT(0, "write-bitmap-index", &write_bitmap_index,
3303                             N_("write a bitmap index together with the pack index"),
3304                             WRITE_BITMAP_TRUE),
3305                 OPT_SET_INT_F(0, "write-bitmap-index-quiet",
3306                               &write_bitmap_index,
3307                               N_("write a bitmap index if possible"),
3308                               WRITE_BITMAP_QUIET, PARSE_OPT_HIDDEN),
3309                 OPT_PARSE_LIST_OBJECTS_FILTER(&filter_options),
3310                 { OPTION_CALLBACK, 0, "missing", NULL, N_("action"),
3311                   N_("handling for missing objects"), PARSE_OPT_NONEG,
3312                   option_parse_missing_action },
3313                 OPT_BOOL(0, "exclude-promisor-objects", &exclude_promisor_objects,
3314                          N_("do not pack objects in promisor packfiles")),
3315                 OPT_BOOL(0, "delta-islands", &use_delta_islands,
3316                          N_("respect islands during delta compression")),
3317                 OPT_END(),
3318         };
3319
3320         if (DFS_NUM_STATES > (1 << OE_DFS_STATE_BITS))
3321                 BUG("too many dfs states, increase OE_DFS_STATE_BITS");
3322
3323         read_replace_refs = 0;
3324
3325         sparse = git_env_bool("GIT_TEST_PACK_SPARSE", 0);
3326         prepare_repo_settings(the_repository);
3327         if (!sparse && the_repository->settings.pack_use_sparse != -1)
3328                 sparse = the_repository->settings.pack_use_sparse;
3329
3330         reset_pack_idx_option(&pack_idx_opts);
3331         git_config(git_pack_config, NULL);
3332
3333         progress = isatty(2);
3334         argc = parse_options(argc, argv, prefix, pack_objects_options,
3335                              pack_usage, 0);
3336
3337         if (argc) {
3338                 base_name = argv[0];
3339                 argc--;
3340         }
3341         if (pack_to_stdout != !base_name || argc)
3342                 usage_with_options(pack_usage, pack_objects_options);
3343
3344         if (depth >= (1 << OE_DEPTH_BITS)) {
3345                 warning(_("delta chain depth %d is too deep, forcing %d"),
3346                         depth, (1 << OE_DEPTH_BITS) - 1);
3347                 depth = (1 << OE_DEPTH_BITS) - 1;
3348         }
3349         if (cache_max_small_delta_size >= (1U << OE_Z_DELTA_BITS)) {
3350                 warning(_("pack.deltaCacheLimit is too high, forcing %d"),
3351                         (1U << OE_Z_DELTA_BITS) - 1);
3352                 cache_max_small_delta_size = (1U << OE_Z_DELTA_BITS) - 1;
3353         }
3354
3355         argv_array_push(&rp, "pack-objects");
3356         if (thin) {
3357                 use_internal_rev_list = 1;
3358                 argv_array_push(&rp, shallow
3359                                 ? "--objects-edge-aggressive"
3360                                 : "--objects-edge");
3361         } else
3362                 argv_array_push(&rp, "--objects");
3363
3364         if (rev_list_all) {
3365                 use_internal_rev_list = 1;
3366                 argv_array_push(&rp, "--all");
3367         }
3368         if (rev_list_reflog) {
3369                 use_internal_rev_list = 1;
3370                 argv_array_push(&rp, "--reflog");
3371         }
3372         if (rev_list_index) {
3373                 use_internal_rev_list = 1;
3374                 argv_array_push(&rp, "--indexed-objects");
3375         }
3376         if (rev_list_unpacked) {
3377                 use_internal_rev_list = 1;
3378                 argv_array_push(&rp, "--unpacked");
3379         }
3380
3381         if (exclude_promisor_objects) {
3382                 use_internal_rev_list = 1;
3383                 fetch_if_missing = 0;
3384                 argv_array_push(&rp, "--exclude-promisor-objects");
3385         }
3386         if (unpack_unreachable || keep_unreachable || pack_loose_unreachable)
3387                 use_internal_rev_list = 1;
3388
3389         if (!reuse_object)
3390                 reuse_delta = 0;
3391         if (pack_compression_level == -1)
3392                 pack_compression_level = Z_DEFAULT_COMPRESSION;
3393         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
3394                 die(_("bad pack compression level %d"), pack_compression_level);
3395
3396         if (!delta_search_threads)      /* --threads=0 means autodetect */
3397                 delta_search_threads = online_cpus();
3398
3399         if (!HAVE_THREADS && delta_search_threads != 1)
3400                 warning(_("no threads support, ignoring --threads"));
3401         if (!pack_to_stdout && !pack_size_limit)
3402                 pack_size_limit = pack_size_limit_cfg;
3403         if (pack_to_stdout && pack_size_limit)
3404                 die(_("--max-pack-size cannot be used to build a pack for transfer"));
3405         if (pack_size_limit && pack_size_limit < 1024*1024) {
3406                 warning(_("minimum pack size limit is 1 MiB"));
3407                 pack_size_limit = 1024*1024;
3408         }
3409
3410         if (!pack_to_stdout && thin)
3411                 die(_("--thin cannot be used to build an indexable pack"));
3412
3413         if (keep_unreachable && unpack_unreachable)
3414                 die(_("--keep-unreachable and --unpack-unreachable are incompatible"));
3415         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3416                 unpack_unreachable_expiration = 0;
3417
3418         if (filter_options.choice) {
3419                 if (!pack_to_stdout)
3420                         die(_("cannot use --filter without --stdout"));
3421                 use_bitmap_index = 0;
3422         }
3423
3424         /*
3425          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3426          *
3427          * - to produce good pack (with bitmap index not-yet-packed objects are
3428          *   packed in suboptimal order).
3429          *
3430          * - to use more robust pack-generation codepath (avoiding possible
3431          *   bugs in bitmap code and possible bitmap index corruption).
3432          */
3433         if (!pack_to_stdout)
3434                 use_bitmap_index_default = 0;
3435
3436         if (use_bitmap_index < 0)
3437                 use_bitmap_index = use_bitmap_index_default;
3438
3439         /* "hard" reasons not to use bitmaps; these just won't work at all */
3440         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow(the_repository))
3441                 use_bitmap_index = 0;
3442
3443         if (pack_to_stdout || !rev_list_all)
3444                 write_bitmap_index = 0;
3445
3446         if (use_delta_islands)
3447                 argv_array_push(&rp, "--topo-order");
3448
3449         if (progress && all_progress_implied)
3450                 progress = 2;
3451
3452         add_extra_kept_packs(&keep_pack_list);
3453         if (ignore_packed_keep_on_disk) {
3454                 struct packed_git *p;
3455                 for (p = get_all_packs(the_repository); p; p = p->next)
3456                         if (p->pack_local && p->pack_keep)
3457                                 break;
3458                 if (!p) /* no keep-able packs found */
3459                         ignore_packed_keep_on_disk = 0;
3460         }
3461         if (local) {
3462                 /*
3463                  * unlike ignore_packed_keep_on_disk above, we do not
3464                  * want to unset "local" based on looking at packs, as
3465                  * it also covers non-local objects
3466                  */
3467                 struct packed_git *p;
3468                 for (p = get_all_packs(the_repository); p; p = p->next) {
3469                         if (!p->pack_local) {
3470                                 have_non_local_packs = 1;
3471                                 break;
3472                         }
3473                 }
3474         }
3475
3476         trace2_region_enter("pack-objects", "enumerate-objects",
3477                             the_repository);
3478         prepare_packing_data(the_repository, &to_pack);
3479
3480         if (progress)
3481                 progress_state = start_progress(_("Enumerating objects"), 0);
3482         if (!use_internal_rev_list)
3483                 read_object_list_from_stdin();
3484         else {
3485                 get_object_list(rp.argc, rp.argv);
3486                 argv_array_clear(&rp);
3487         }
3488         cleanup_preferred_base();
3489         if (include_tag && nr_result)
3490                 for_each_ref(add_ref_tag, NULL);
3491         stop_progress(&progress_state);
3492         trace2_region_leave("pack-objects", "enumerate-objects",
3493                             the_repository);
3494
3495         if (non_empty && !nr_result)
3496                 return 0;
3497         if (nr_result) {
3498                 trace2_region_enter("pack-objects", "prepare-pack",
3499                                     the_repository);
3500                 prepare_pack(window, depth);
3501                 trace2_region_leave("pack-objects", "prepare-pack",
3502                                     the_repository);
3503         }
3504
3505         trace2_region_enter("pack-objects", "write-pack-file", the_repository);
3506         write_pack_file();
3507         trace2_region_leave("pack-objects", "write-pack-file", the_repository);
3508
3509         if (progress)
3510                 fprintf_ln(stderr,
3511                            _("Total %"PRIu32" (delta %"PRIu32"),"
3512                              " reused %"PRIu32" (delta %"PRIu32")"),
3513                            written, written_delta, reused, reused_delta);
3514         return 0;
3515 }