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 int write_bitmap_index;
 100 static uint16_t write_bitmap_options = BITMAP_OPT_HASH_CACHE;
 101
 102 static int exclude_promisor_objects;
 103
 104 static int use_delta_islands;
 105
 106 static unsigned long delta_cache_size = 0;
 107 static unsigned long max_delta_cache_size = DEFAULT_DELTA_CACHE_SIZE;
 108 static unsigned long cache_max_small_delta_size = 1000;
 109
 110 static unsigned long window_memory_limit = 0;
 111
 112 static struct list_objects_filter_options filter_options;
 113
 114 enum missing_action {
 115         MA_ERROR = 0,      /* fail if any missing objects are encountered */
 116         MA_ALLOW_ANY,      /* silently allow ALL missing objects */
 117         MA_ALLOW_PROMISOR, /* silently allow all missing PROMISOR objects */
 118 };
 119 static enum missing_action arg_missing_action;
 120 static show_object_fn fn_show_object;
 121
 122 /*
 123  * stats
 124  */
 125 static uint32_t written, written_delta;
 126 static uint32_t reused, reused_delta;
 127
 128 /*
 129  * Indexed commits
 130  */
 131 static struct commit **indexed_commits;
 132 static unsigned int indexed_commits_nr;
 133 static unsigned int indexed_commits_alloc;
 134
 135 static void index_commit_for_bitmap(struct commit *commit)
 136 {
 137         if (indexed_commits_nr >= indexed_commits_alloc) {
 138                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
 139                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
 140         }
 141
 142         indexed_commits[indexed_commits_nr++] = commit;
 143 }
 144
 145 static void *get_delta(struct object_entry *entry)
 146 {
 147         unsigned long size, base_size, delta_size;
 148         void *buf, *base_buf, *delta_buf;
 149         enum object_type type;
 150
 151         buf = read_object_file(&entry->idx.oid, &type, &size);
 152         if (!buf)
 153                 die(_("unable to read %s"), oid_to_hex(&entry->idx.oid));
 154         base_buf = read_object_file(&DELTA(entry)->idx.oid, &type,
 155                                     &base_size);
 156         if (!base_buf)
 157                 die("unable to read %s",
 158                     oid_to_hex(&DELTA(entry)->idx.oid));
 159         delta_buf = diff_delta(base_buf, base_size,
 160                                buf, size, &delta_size, 0);
 161         /*
 162          * We succesfully computed this delta once but dropped it for
 163          * memory reasons. Something is very wrong if this time we
 164          * recompute and create a different delta.
 165          */
 166         if (!delta_buf || delta_size != DELTA_SIZE(entry))
 167                 BUG("delta size changed");
 168         free(buf);
 169         free(base_buf);
 170         return delta_buf;
 171 }
 172
 173 static unsigned long do_compress(void **pptr, unsigned long size)
 174 {
 175         git_zstream stream;
 176         void *in, *out;
 177         unsigned long maxsize;
 178
 179         git_deflate_init(&stream, pack_compression_level);
 180         maxsize = git_deflate_bound(&stream, size);
 181
 182         in = *pptr;
 183         out = xmalloc(maxsize);
 184         *pptr = out;
 185
 186         stream.next_in = in;
 187         stream.avail_in = size;
 188         stream.next_out = out;
 189         stream.avail_out = maxsize;
 190         while (git_deflate(&stream, Z_FINISH) == Z_OK)
 191                 ; /* nothing */
 192         git_deflate_end(&stream);
 193
 194         free(in);
 195         return stream.total_out;
 196 }
 197
 198 static unsigned long write_large_blob_data(struct git_istream *st, struct hashfile *f,
 199                                            const struct object_id *oid)
 200 {
 201         git_zstream stream;
 202         unsigned char ibuf[1024 * 16];
 203         unsigned char obuf[1024 * 16];
 204         unsigned long olen = 0;
 205
 206         git_deflate_init(&stream, pack_compression_level);
 207
 208         for (;;) {
 209                 ssize_t readlen;
 210                 int zret = Z_OK;
 211                 readlen = read_istream(st, ibuf, sizeof(ibuf));
 212                 if (readlen == -1)
 213                         die(_("unable to read %s"), oid_to_hex(oid));
 214
 215                 stream.next_in = ibuf;
 216                 stream.avail_in = readlen;
 217                 while ((stream.avail_in || readlen == 0) &&
 218                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
 219                         stream.next_out = obuf;
 220                         stream.avail_out = sizeof(obuf);
 221                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 222                         hashwrite(f, obuf, stream.next_out - obuf);
 223                         olen += stream.next_out - obuf;
 224                 }
 225                 if (stream.avail_in)
 226                         die(_("deflate error (%d)"), zret);
 227                 if (readlen == 0) {
 228                         if (zret != Z_STREAM_END)
 229                                 die(_("deflate error (%d)"), zret);
 230                         break;
 231                 }
 232         }
 233         git_deflate_end(&stream);
 234         return olen;
 235 }
 236
 237 /*
 238  * we are going to reuse the existing object data as is.  make
 239  * sure it is not corrupt.
 240  */
 241 static int check_pack_inflate(struct packed_git *p,
 242                 struct pack_window **w_curs,
 243                 off_t offset,
 244                 off_t len,
 245                 unsigned long expect)
 246 {
 247         git_zstream stream;
 248         unsigned char fakebuf[4096], *in;
 249         int st;
 250
 251         memset(&stream, 0, sizeof(stream));
 252         git_inflate_init(&stream);
 253         do {
 254                 in = use_pack(p, w_curs, offset, &stream.avail_in);
 255                 stream.next_in = in;
 256                 stream.next_out = fakebuf;
 257                 stream.avail_out = sizeof(fakebuf);
 258                 st = git_inflate(&stream, Z_FINISH);
 259                 offset += stream.next_in - in;
 260         } while (st == Z_OK || st == Z_BUF_ERROR);
 261         git_inflate_end(&stream);
 262         return (st == Z_STREAM_END &&
 263                 stream.total_out == expect &&
 264                 stream.total_in == len) ? 0 : -1;
 265 }
 266
 267 static void copy_pack_data(struct hashfile *f,
 268                 struct packed_git *p,
 269                 struct pack_window **w_curs,
 270                 off_t offset,
 271                 off_t len)
 272 {
 273         unsigned char *in;
 274         unsigned long avail;
 275
 276         while (len) {
 277                 in = use_pack(p, w_curs, offset, &avail);
 278                 if (avail > len)
 279                         avail = (unsigned long)len;
 280                 hashwrite(f, in, avail);
 281                 offset += avail;
 282                 len -= avail;
 283         }
 284 }
 285
 286 /* Return 0 if we will bust the pack-size limit */
 287 static unsigned long write_no_reuse_object(struct hashfile *f, struct object_entry *entry,
 288                                            unsigned long limit, int usable_delta)
 289 {
 290         unsigned long size, datalen;
 291         unsigned char header[MAX_PACK_OBJECT_HEADER],
 292                       dheader[MAX_PACK_OBJECT_HEADER];
 293         unsigned hdrlen;
 294         enum object_type type;
 295         void *buf;
 296         struct git_istream *st = NULL;
 297         const unsigned hashsz = the_hash_algo->rawsz;
 298
 299         if (!usable_delta) {
 300                 if (oe_type(entry) == OBJ_BLOB &&
 301                     oe_size_greater_than(&to_pack, entry, big_file_threshold) &&
 302                     (st = open_istream(&entry->idx.oid, &type, &size, NULL)) != NULL)
 303                         buf = NULL;
 304                 else {
 305                         buf = read_object_file(&entry->idx.oid, &type, &size);
 306                         if (!buf)
 307                                 die(_("unable to read %s"),
 308                                     oid_to_hex(&entry->idx.oid));
 309                 }
 310                 /*
 311                  * make sure no cached delta data remains from a
 312                  * previous attempt before a pack split occurred.
 313                  */
 314                 FREE_AND_NULL(entry->delta_data);
 315                 entry->z_delta_size = 0;
 316         } else if (entry->delta_data) {
 317                 size = DELTA_SIZE(entry);
 318                 buf = entry->delta_data;
 319                 entry->delta_data = NULL;
 320                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
 321                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 322         } else {
 323                 buf = get_delta(entry);
 324                 size = DELTA_SIZE(entry);
 325                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
 326                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 327         }
 328
 329         if (st) /* large blob case, just assume we don't compress well */
 330                 datalen = size;
 331         else if (entry->z_delta_size)
 332                 datalen = entry->z_delta_size;
 333         else
 334                 datalen = do_compress(&buf, size);
 335
 336         /*
 337          * The object header is a byte of 'type' followed by zero or
 338          * more bytes of length.
 339          */
 340         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 341                                               type, size);
 342
 343         if (type == OBJ_OFS_DELTA) {
 344                 /*
 345                  * Deltas with relative base contain an additional
 346                  * encoding of the relative offset for the delta
 347                  * base from this object's position in the pack.
 348                  */
 349                 off_t ofs = entry->idx.offset - DELTA(entry)->idx.offset;
 350                 unsigned pos = sizeof(dheader) - 1;
 351                 dheader[pos] = ofs & 127;
 352                 while (ofs >>= 7)
 353                         dheader[--pos] = 128 | (--ofs & 127);
 354                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + hashsz >= limit) {
 355                         if (st)
 356                                 close_istream(st);
 357                         free(buf);
 358                         return 0;
 359                 }
 360                 hashwrite(f, header, hdrlen);
 361                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
 362                 hdrlen += sizeof(dheader) - pos;
 363         } else if (type == OBJ_REF_DELTA) {
 364                 /*
 365                  * Deltas with a base reference contain
 366                  * additional bytes for the base object ID.
 367                  */
 368                 if (limit && hdrlen + hashsz + datalen + hashsz >= limit) {
 369                         if (st)
 370                                 close_istream(st);
 371                         free(buf);
 372                         return 0;
 373                 }
 374                 hashwrite(f, header, hdrlen);
 375                 hashwrite(f, DELTA(entry)->idx.oid.hash, hashsz);
 376                 hdrlen += hashsz;
 377         } else {
 378                 if (limit && hdrlen + datalen + hashsz >= limit) {
 379                         if (st)
 380                                 close_istream(st);
 381                         free(buf);
 382                         return 0;
 383                 }
 384                 hashwrite(f, header, hdrlen);
 385         }
 386         if (st) {
 387                 datalen = write_large_blob_data(st, f, &entry->idx.oid);
 388                 close_istream(st);
 389         } else {
 390                 hashwrite(f, buf, datalen);
 391                 free(buf);
 392         }
 393
 394         return hdrlen + datalen;
 395 }
 396
 397 /* Return 0 if we will bust the pack-size limit */
 398 static off_t write_reuse_object(struct hashfile *f, struct object_entry *entry,
 399                                 unsigned long limit, int usable_delta)
 400 {
 401         struct packed_git *p = IN_PACK(entry);
 402         struct pack_window *w_curs = NULL;
 403         struct revindex_entry *revidx;
 404         off_t offset;
 405         enum object_type type = oe_type(entry);
 406         off_t datalen;
 407         unsigned char header[MAX_PACK_OBJECT_HEADER],
 408                       dheader[MAX_PACK_OBJECT_HEADER];
 409         unsigned hdrlen;
 410         const unsigned hashsz = the_hash_algo->rawsz;
 411         unsigned long entry_size = SIZE(entry);
 412
 413         if (DELTA(entry))
 414                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
 415                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 416         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 417                                               type, entry_size);
 418
 419         offset = entry->in_pack_offset;
 420         revidx = find_pack_revindex(p, offset);
 421         datalen = revidx[1].offset - offset;
 422         if (!pack_to_stdout && p->index_version > 1 &&
 423             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 424                 error(_("bad packed object CRC for %s"),
 425                       oid_to_hex(&entry->idx.oid));
 426                 unuse_pack(&w_curs);
 427                 return write_no_reuse_object(f, entry, limit, usable_delta);
 428         }
 429
 430         offset += entry->in_pack_header_size;
 431         datalen -= entry->in_pack_header_size;
 432
 433         if (!pack_to_stdout && p->index_version == 1 &&
 434             check_pack_inflate(p, &w_curs, offset, datalen, entry_size)) {
 435                 error(_("corrupt packed object for %s"),
 436                       oid_to_hex(&entry->idx.oid));
 437                 unuse_pack(&w_curs);
 438                 return write_no_reuse_object(f, entry, limit, usable_delta);
 439         }
 440
 441         if (type == OBJ_OFS_DELTA) {
 442                 off_t ofs = entry->idx.offset - DELTA(entry)->idx.offset;
 443                 unsigned pos = sizeof(dheader) - 1;
 444                 dheader[pos] = ofs & 127;
 445                 while (ofs >>= 7)
 446                         dheader[--pos] = 128 | (--ofs & 127);
 447                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + hashsz >= limit) {
 448                         unuse_pack(&w_curs);
 449                         return 0;
 450                 }
 451                 hashwrite(f, header, hdrlen);
 452                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
 453                 hdrlen += sizeof(dheader) - pos;
 454                 reused_delta++;
 455         } else if (type == OBJ_REF_DELTA) {
 456                 if (limit && hdrlen + hashsz + datalen + hashsz >= limit) {
 457                         unuse_pack(&w_curs);
 458                         return 0;
 459                 }
 460                 hashwrite(f, header, hdrlen);
 461                 hashwrite(f, DELTA(entry)->idx.oid.hash, hashsz);
 462                 hdrlen += hashsz;
 463                 reused_delta++;
 464         } else {
 465                 if (limit && hdrlen + datalen + hashsz >= limit) {
 466                         unuse_pack(&w_curs);
 467                         return 0;
 468                 }
 469                 hashwrite(f, header, hdrlen);
 470         }
 471         copy_pack_data(f, p, &w_curs, offset, datalen);
 472         unuse_pack(&w_curs);
 473         reused++;
 474         return hdrlen + datalen;
 475 }
 476
 477 /* Return 0 if we will bust the pack-size limit */
 478 static off_t write_object(struct hashfile *f,
 479                           struct object_entry *entry,
 480                           off_t write_offset)
 481 {
 482         unsigned long limit;
 483         off_t len;
 484         int usable_delta, to_reuse;
 485
 486         if (!pack_to_stdout)
 487                 crc32_begin(f);
 488
 489         /* apply size limit if limited packsize and not first object */
 490         if (!pack_size_limit || !nr_written)
 491                 limit = 0;
 492         else if (pack_size_limit <= write_offset)
 493                 /*
 494                  * the earlier object did not fit the limit; avoid
 495                  * mistaking this with unlimited (i.e. limit = 0).
 496                  */
 497                 limit = 1;
 498         else
 499                 limit = pack_size_limit - write_offset;
 500
 501         if (!DELTA(entry))
 502                 usable_delta = 0;       /* no delta */
 503         else if (!pack_size_limit)
 504                usable_delta = 1;        /* unlimited packfile */
 505         else if (DELTA(entry)->idx.offset == (off_t)-1)
 506                 usable_delta = 0;       /* base was written to another pack */
 507         else if (DELTA(entry)->idx.offset)
 508                 usable_delta = 1;       /* base already exists in this pack */
 509         else
 510                 usable_delta = 0;       /* base could end up in another pack */
 511
 512         if (!reuse_object)
 513                 to_reuse = 0;   /* explicit */
 514         else if (!IN_PACK(entry))
 515                 to_reuse = 0;   /* can't reuse what we don't have */
 516         else if (oe_type(entry) == OBJ_REF_DELTA ||
 517                  oe_type(entry) == OBJ_OFS_DELTA)
 518                                 /* check_object() decided it for us ... */
 519                 to_reuse = usable_delta;
 520                                 /* ... but pack split may override that */
 521         else if (oe_type(entry) != entry->in_pack_type)
 522                 to_reuse = 0;   /* pack has delta which is unusable */
 523         else if (DELTA(entry))
 524                 to_reuse = 0;   /* we want to pack afresh */
 525         else
 526                 to_reuse = 1;   /* we have it in-pack undeltified,
 527                                  * and we do not need to deltify it.
 528                                  */
 529
 530         if (!to_reuse)
 531                 len = write_no_reuse_object(f, entry, limit, usable_delta);
 532         else
 533                 len = write_reuse_object(f, entry, limit, usable_delta);
 534         if (!len)
 535                 return 0;
 536
 537         if (usable_delta)
 538                 written_delta++;
 539         written++;
 540         if (!pack_to_stdout)
 541                 entry->idx.crc32 = crc32_end(f);
 542         return len;
 543 }
 544
 545 enum write_one_status {
 546         WRITE_ONE_SKIP = -1, /* already written */
 547         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 548         WRITE_ONE_WRITTEN = 1, /* normal */
 549         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 550 };
 551
 552 static enum write_one_status write_one(struct hashfile *f,
 553                                        struct object_entry *e,
 554                                        off_t *offset)
 555 {
 556         off_t size;
 557         int recursing;
 558
 559         /*
 560          * we set offset to 1 (which is an impossible value) to mark
 561          * the fact that this object is involved in "write its base
 562          * first before writing a deltified object" recursion.
 563          */
 564         recursing = (e->idx.offset == 1);
 565         if (recursing) {
 566                 warning(_("recursive delta detected for object %s"),
 567                         oid_to_hex(&e->idx.oid));
 568                 return WRITE_ONE_RECURSIVE;
 569         } else if (e->idx.offset || e->preferred_base) {
 570                 /* offset is non zero if object is written already. */
 571                 return WRITE_ONE_SKIP;
 572         }
 573
 574         /* if we are deltified, write out base object first. */
 575         if (DELTA(e)) {
 576                 e->idx.offset = 1; /* now recurse */
 577                 switch (write_one(f, DELTA(e), offset)) {
 578                 case WRITE_ONE_RECURSIVE:
 579                         /* we cannot depend on this one */
 580                         SET_DELTA(e, NULL);
 581                         break;
 582                 default:
 583                         break;
 584                 case WRITE_ONE_BREAK:
 585                         e->idx.offset = recursing;
 586                         return WRITE_ONE_BREAK;
 587                 }
 588         }
 589
 590         e->idx.offset = *offset;
 591         size = write_object(f, e, *offset);
 592         if (!size) {
 593                 e->idx.offset = recursing;
 594                 return WRITE_ONE_BREAK;
 595         }
 596         written_list[nr_written++] = &e->idx;
 597
 598         /* make sure off_t is sufficiently large not to wrap */
 599         if (signed_add_overflows(*offset, size))
 600                 die(_("pack too large for current definition of off_t"));
 601         *offset += size;
 602         return WRITE_ONE_WRITTEN;
 603 }
 604
 605 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
 606                        void *cb_data)
 607 {
 608         struct object_id peeled;
 609         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
 610
 611         if (entry)
 612                 entry->tagged = 1;
 613         if (!peel_ref(path, &peeled)) {
 614                 entry = packlist_find(&to_pack, peeled.hash, NULL);
 615                 if (entry)
 616                         entry->tagged = 1;
 617         }
 618         return 0;
 619 }
 620
 621 static inline void add_to_write_order(struct object_entry **wo,
 622                                unsigned int *endp,
 623                                struct object_entry *e)
 624 {
 625         if (e->filled || oe_layer(&to_pack, e) != write_layer)
 626                 return;
 627         wo[(*endp)++] = e;
 628         e->filled = 1;
 629 }
 630
 631 static void add_descendants_to_write_order(struct object_entry **wo,
 632                                            unsigned int *endp,
 633                                            struct object_entry *e)
 634 {
 635         int add_to_order = 1;
 636         while (e) {
 637                 if (add_to_order) {
 638                         struct object_entry *s;
 639                         /* add this node... */
 640                         add_to_write_order(wo, endp, e);
 641                         /* all its siblings... */
 642                         for (s = DELTA_SIBLING(e); s; s = DELTA_SIBLING(s)) {
 643                                 add_to_write_order(wo, endp, s);
 644                         }
 645                 }
 646                 /* drop down a level to add left subtree nodes if possible */
 647                 if (DELTA_CHILD(e)) {
 648                         add_to_order = 1;
 649                         e = DELTA_CHILD(e);
 650                 } else {
 651                         add_to_order = 0;
 652                         /* our sibling might have some children, it is next */
 653                         if (DELTA_SIBLING(e)) {
 654                                 e = DELTA_SIBLING(e);
 655                                 continue;
 656                         }
 657                         /* go back to our parent node */
 658                         e = DELTA(e);
 659                         while (e && !DELTA_SIBLING(e)) {
 660                                 /* we're on the right side of a subtree, keep
 661                                  * going up until we can go right again */
 662                                 e = DELTA(e);
 663                         }
 664                         if (!e) {
 665                                 /* done- we hit our original root node */
 666                                 return;
 667                         }
 668                         /* pass it off to sibling at this level */
 669                         e = DELTA_SIBLING(e);
 670                 }
 671         };
 672 }
 673
 674 static void add_family_to_write_order(struct object_entry **wo,
 675                                       unsigned int *endp,
 676                                       struct object_entry *e)
 677 {
 678         struct object_entry *root;
 679
 680         for (root = e; DELTA(root); root = DELTA(root))
 681                 ; /* nothing */
 682         add_descendants_to_write_order(wo, endp, root);
 683 }
 684
 685 static void compute_layer_order(struct object_entry **wo, unsigned int *wo_end)
 686 {
 687         unsigned int i, last_untagged;
 688         struct object_entry *objects = to_pack.objects;
 689
 690         for (i = 0; i < to_pack.nr_objects; i++) {
 691                 if (objects[i].tagged)
 692                         break;
 693                 add_to_write_order(wo, wo_end, &objects[i]);
 694         }
 695         last_untagged = i;
 696
 697         /*
 698          * Then fill all the tagged tips.
 699          */
 700         for (; i < to_pack.nr_objects; i++) {
 701                 if (objects[i].tagged)
 702                         add_to_write_order(wo, wo_end, &objects[i]);
 703         }
 704
 705         /*
 706          * And then all remaining commits and tags.
 707          */
 708         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 709                 if (oe_type(&objects[i]) != OBJ_COMMIT &&
 710                     oe_type(&objects[i]) != OBJ_TAG)
 711                         continue;
 712                 add_to_write_order(wo, wo_end, &objects[i]);
 713         }
 714
 715         /*
 716          * And then all the trees.
 717          */
 718         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 719                 if (oe_type(&objects[i]) != OBJ_TREE)
 720                         continue;
 721                 add_to_write_order(wo, wo_end, &objects[i]);
 722         }
 723
 724         /*
 725          * Finally all the rest in really tight order
 726          */
 727         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 728                 if (!objects[i].filled && oe_layer(&to_pack, &objects[i]) == write_layer)
 729                         add_family_to_write_order(wo, wo_end, &objects[i]);
 730         }
 731 }
 732
 733 static struct object_entry **compute_write_order(void)
 734 {
 735         uint32_t max_layers = 1;
 736         unsigned int i, wo_end;
 737
 738         struct object_entry **wo;
 739         struct object_entry *objects = to_pack.objects;
 740
 741         for (i = 0; i < to_pack.nr_objects; i++) {
 742                 objects[i].tagged = 0;
 743                 objects[i].filled = 0;
 744                 SET_DELTA_CHILD(&objects[i], NULL);
 745                 SET_DELTA_SIBLING(&objects[i], NULL);
 746         }
 747
 748         /*
 749          * Fully connect delta_child/delta_sibling network.
 750          * Make sure delta_sibling is sorted in the original
 751          * recency order.
 752          */
 753         for (i = to_pack.nr_objects; i > 0;) {
 754                 struct object_entry *e = &objects[--i];
 755                 if (!DELTA(e))
 756                         continue;
 757                 /* Mark me as the first child */
 758                 e->delta_sibling_idx = DELTA(e)->delta_child_idx;
 759                 SET_DELTA_CHILD(DELTA(e), e);
 760         }
 761
 762         /*
 763          * Mark objects that are at the tip of tags.
 764          */
 765         for_each_tag_ref(mark_tagged, NULL);
 766
 767         if (use_delta_islands)
 768                 max_layers = compute_pack_layers(&to_pack);
 769
 770         ALLOC_ARRAY(wo, to_pack.nr_objects);
 771         wo_end = 0;
 772
 773         for (; write_layer < max_layers; ++write_layer)
 774                 compute_layer_order(wo, &wo_end);
 775
 776         if (wo_end != to_pack.nr_objects)
 777                 die(_("ordered %u objects, expected %"PRIu32),
 778                     wo_end, to_pack.nr_objects);
 779
 780         return wo;
 781 }
 782
 783 static off_t write_reused_pack(struct hashfile *f)
 784 {
 785         unsigned char buffer[8192];
 786         off_t to_write, total;
 787         int fd;
 788
 789         if (!is_pack_valid(reuse_packfile))
 790                 die(_("packfile is invalid: %s"), reuse_packfile->pack_name);
 791
 792         fd = git_open(reuse_packfile->pack_name);
 793         if (fd < 0)
 794                 die_errno(_("unable to open packfile for reuse: %s"),
 795                           reuse_packfile->pack_name);
 796
 797         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
 798                 die_errno(_("unable to seek in reused packfile"));
 799
 800         if (reuse_packfile_offset < 0)
 801                 reuse_packfile_offset = reuse_packfile->pack_size - the_hash_algo->rawsz;
 802
 803         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
 804
 805         while (to_write) {
 806                 int read_pack = xread(fd, buffer, sizeof(buffer));
 807
 808                 if (read_pack <= 0)
 809                         die_errno(_("unable to read from reused packfile"));
 810
 811                 if (read_pack > to_write)
 812                         read_pack = to_write;
 813
 814                 hashwrite(f, buffer, read_pack);
 815                 to_write -= read_pack;
 816
 817                 /*
 818                  * We don't know the actual number of objects written,
 819                  * only how many bytes written, how many bytes total, and
 820                  * how many objects total. So we can fake it by pretending all
 821                  * objects we are writing are the same size. This gives us a
 822                  * smooth progress meter, and at the end it matches the true
 823                  * answer.
 824                  */
 825                 written = reuse_packfile_objects *
 826                                 (((double)(total - to_write)) / total);
 827                 display_progress(progress_state, written);
 828         }
 829
 830         close(fd);
 831         written = reuse_packfile_objects;
 832         display_progress(progress_state, written);
 833         return reuse_packfile_offset - sizeof(struct pack_header);
 834 }
 835
 836 static const char no_split_warning[] = N_(
 837 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
 838 );
 839
 840 static void write_pack_file(void)
 841 {
 842         uint32_t i = 0, j;
 843         struct hashfile *f;
 844         off_t offset;
 845         uint32_t nr_remaining = nr_result;
 846         time_t last_mtime = 0;
 847         struct object_entry **write_order;
 848
 849         if (progress > pack_to_stdout)
 850                 progress_state = start_progress(_("Writing objects"), nr_result);
 851         ALLOC_ARRAY(written_list, to_pack.nr_objects);
 852         write_order = compute_write_order();
 853
 854         do {
 855                 struct object_id oid;
 856                 char *pack_tmp_name = NULL;
 857
 858                 if (pack_to_stdout)
 859                         f = hashfd_throughput(1, "<stdout>", progress_state);
 860                 else
 861                         f = create_tmp_packfile(&pack_tmp_name);
 862
 863                 offset = write_pack_header(f, nr_remaining);
 864
 865                 if (reuse_packfile) {
 866                         off_t packfile_size;
 867                         assert(pack_to_stdout);
 868
 869                         packfile_size = write_reused_pack(f);
 870                         offset += packfile_size;
 871                 }
 872
 873                 nr_written = 0;
 874                 for (; i < to_pack.nr_objects; i++) {
 875                         struct object_entry *e = write_order[i];
 876                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 877                                 break;
 878                         display_progress(progress_state, written);
 879                 }
 880
 881                 /*
 882                  * Did we write the wrong # entries in the header?
 883                  * If so, rewrite it like in fast-import
 884                  */
 885                 if (pack_to_stdout) {
 886                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_CLOSE);
 887                 } else if (nr_written == nr_remaining) {
 888                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE);
 889                 } else {
 890                         int fd = finalize_hashfile(f, oid.hash, 0);
 891                         fixup_pack_header_footer(fd, oid.hash, pack_tmp_name,
 892                                                  nr_written, oid.hash, offset);
 893                         close(fd);
 894                         if (write_bitmap_index) {
 895                                 warning(_(no_split_warning));
 896                                 write_bitmap_index = 0;
 897                         }
 898                 }
 899
 900                 if (!pack_to_stdout) {
 901                         struct stat st;
 902                         struct strbuf tmpname = STRBUF_INIT;
 903
 904                         /*
 905                          * Packs are runtime accessed in their mtime
 906                          * order since newer packs are more likely to contain
 907                          * younger objects.  So if we are creating multiple
 908                          * packs then we should modify the mtime of later ones
 909                          * to preserve this property.
 910                          */
 911                         if (stat(pack_tmp_name, &st) < 0) {
 912                                 warning_errno(_("failed to stat %s"), pack_tmp_name);
 913                         } else if (!last_mtime) {
 914                                 last_mtime = st.st_mtime;
 915                         } else {
 916                                 struct utimbuf utb;
 917                                 utb.actime = st.st_atime;
 918                                 utb.modtime = --last_mtime;
 919                                 if (utime(pack_tmp_name, &utb) < 0)
 920                                         warning_errno(_("failed utime() on %s"), pack_tmp_name);
 921                         }
 922
 923                         strbuf_addf(&tmpname, "%s-", base_name);
 924
 925                         if (write_bitmap_index) {
 926                                 bitmap_writer_set_checksum(oid.hash);
 927                                 bitmap_writer_build_type_index(
 928                                         &to_pack, written_list, nr_written);
 929                         }
 930
 931                         finish_tmp_packfile(&tmpname, pack_tmp_name,
 932                                             written_list, nr_written,
 933                                             &pack_idx_opts, oid.hash);
 934
 935                         if (write_bitmap_index) {
 936                                 strbuf_addf(&tmpname, "%s.bitmap", oid_to_hex(&oid));
 937
 938                                 stop_progress(&progress_state);
 939
 940                                 bitmap_writer_show_progress(progress);
 941                                 bitmap_writer_reuse_bitmaps(&to_pack);
 942                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
 943                                 bitmap_writer_build(&to_pack);
 944                                 bitmap_writer_finish(written_list, nr_written,
 945                                                      tmpname.buf, write_bitmap_options);
 946                                 write_bitmap_index = 0;
 947                         }
 948
 949                         strbuf_release(&tmpname);
 950                         free(pack_tmp_name);
 951                         puts(oid_to_hex(&oid));
 952                 }
 953
 954                 /* mark written objects as written to previous pack */
 955                 for (j = 0; j < nr_written; j++) {
 956                         written_list[j]->offset = (off_t)-1;
 957                 }
 958                 nr_remaining -= nr_written;
 959         } while (nr_remaining && i < to_pack.nr_objects);
 960
 961         free(written_list);
 962         free(write_order);
 963         stop_progress(&progress_state);
 964         if (written != nr_result)
 965                 die(_("wrote %"PRIu32" objects while expecting %"PRIu32),
 966                     written, nr_result);
 967         trace2_data_intmax("pack-objects", the_repository,
 968                            "write_pack_file/wrote", nr_result);
 969 }
 970
 971 static int no_try_delta(const char *path)
 972 {
 973         static struct attr_check *check;
 974
 975         if (!check)
 976                 check = attr_check_initl("delta", NULL);
 977         git_check_attr(the_repository->index, path, check);
 978         if (ATTR_FALSE(check->items[0].value))
 979                 return 1;
 980         return 0;
 981 }
 982
 983 /*
 984  * When adding an object, check whether we have already added it
 985  * to our packing list. If so, we can skip. However, if we are
 986  * being asked to excludei t, but the previous mention was to include
 987  * it, make sure to adjust its flags and tweak our numbers accordingly.
 988  *
 989  * As an optimization, we pass out the index position where we would have
 990  * found the item, since that saves us from having to look it up again a
 991  * few lines later when we want to add the new entry.
 992  */
 993 static int have_duplicate_entry(const struct object_id *oid,
 994                                 int exclude,
 995                                 uint32_t *index_pos)
 996 {
 997         struct object_entry *entry;
 998
 999         entry = packlist_find(&to_pack, oid->hash, index_pos);
1000         if (!entry)
1001                 return 0;
1002
1003         if (exclude) {
1004                 if (!entry->preferred_base)
1005                         nr_result--;
1006                 entry->preferred_base = 1;
1007         }
1008
1009         return 1;
1010 }
1011
1012 static int want_found_object(int exclude, struct packed_git *p)
1013 {
1014         if (exclude)
1015                 return 1;
1016         if (incremental)
1017                 return 0;
1018
1019         /*
1020          * When asked to do --local (do not include an object that appears in a
1021          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
1022          * an object that appears in a pack marked with .keep), finding a pack
1023          * that matches the criteria is sufficient for us to decide to omit it.
1024          * However, even if this pack does not satisfy the criteria, we need to
1025          * make sure no copy of this object appears in _any_ pack that makes us
1026          * to omit the object, so we need to check all the packs.
1027          *
1028          * We can however first check whether these options can possible matter;
1029          * if they do not matter we know we want the object in generated pack.
1030          * Otherwise, we signal "-1" at the end to tell the caller that we do
1031          * not know either way, and it needs to check more packs.
1032          */
1033         if (!ignore_packed_keep_on_disk &&
1034             !ignore_packed_keep_in_core &&
1035             (!local || !have_non_local_packs))
1036                 return 1;
1037
1038         if (local && !p->pack_local)
1039                 return 0;
1040         if (p->pack_local &&
1041             ((ignore_packed_keep_on_disk && p->pack_keep) ||
1042              (ignore_packed_keep_in_core && p->pack_keep_in_core)))
1043                 return 0;
1044
1045         /* we don't know yet; keep looking for more packs */
1046         return -1;
1047 }
1048
1049 /*
1050  * Check whether we want the object in the pack (e.g., we do not want
1051  * objects found in non-local stores if the "--local" option was used).
1052  *
1053  * If the caller already knows an existing pack it wants to take the object
1054  * from, that is passed in *found_pack and *found_offset; otherwise this
1055  * function finds if there is any pack that has the object and returns the pack
1056  * and its offset in these variables.
1057  */
1058 static int want_object_in_pack(const struct object_id *oid,
1059                                int exclude,
1060                                struct packed_git **found_pack,
1061                                off_t *found_offset)
1062 {
1063         int want;
1064         struct list_head *pos;
1065         struct multi_pack_index *m;
1066
1067         if (!exclude && local && has_loose_object_nonlocal(oid))
1068                 return 0;
1069
1070         /*
1071          * If we already know the pack object lives in, start checks from that
1072          * pack - in the usual case when neither --local was given nor .keep files
1073          * are present we will determine the answer right now.
1074          */
1075         if (*found_pack) {
1076                 want = want_found_object(exclude, *found_pack);
1077                 if (want != -1)
1078                         return want;
1079         }
1080
1081         for (m = get_multi_pack_index(the_repository); m; m = m->next) {
1082                 struct pack_entry e;
1083                 if (fill_midx_entry(the_repository, oid, &e, m)) {
1084                         struct packed_git *p = e.p;
1085                         off_t offset;
1086
1087                         if (p == *found_pack)
1088                                 offset = *found_offset;
1089                         else
1090                                 offset = find_pack_entry_one(oid->hash, p);
1091
1092                         if (offset) {
1093                                 if (!*found_pack) {
1094                                         if (!is_pack_valid(p))
1095                                                 continue;
1096                                         *found_offset = offset;
1097                                         *found_pack = p;
1098                                 }
1099                                 want = want_found_object(exclude, p);
1100                                 if (want != -1)
1101                                         return want;
1102                         }
1103                 }
1104         }
1105
1106         list_for_each(pos, get_packed_git_mru(the_repository)) {
1107                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
1108                 off_t offset;
1109
1110                 if (p == *found_pack)
1111                         offset = *found_offset;
1112                 else
1113                         offset = find_pack_entry_one(oid->hash, p);
1114
1115                 if (offset) {
1116                         if (!*found_pack) {
1117                                 if (!is_pack_valid(p))
1118                                         continue;
1119                                 *found_offset = offset;
1120                                 *found_pack = p;
1121                         }
1122                         want = want_found_object(exclude, p);
1123                         if (!exclude && want > 0)
1124                                 list_move(&p->mru,
1125                                           get_packed_git_mru(the_repository));
1126                         if (want != -1)
1127                                 return want;
1128                 }
1129         }
1130
1131         return 1;
1132 }
1133
1134 static void create_object_entry(const struct object_id *oid,
1135                                 enum object_type type,
1136                                 uint32_t hash,
1137                                 int exclude,
1138                                 int no_try_delta,
1139                                 uint32_t index_pos,
1140                                 struct packed_git *found_pack,
1141                                 off_t found_offset)
1142 {
1143         struct object_entry *entry;
1144
1145         entry = packlist_alloc(&to_pack, oid->hash, index_pos);
1146         entry->hash = hash;
1147         oe_set_type(entry, type);
1148         if (exclude)
1149                 entry->preferred_base = 1;
1150         else
1151                 nr_result++;
1152         if (found_pack) {
1153                 oe_set_in_pack(&to_pack, entry, found_pack);
1154                 entry->in_pack_offset = found_offset;
1155         }
1156
1157         entry->no_try_delta = no_try_delta;
1158 }
1159
1160 static const char no_closure_warning[] = N_(
1161 "disabling bitmap writing, as some objects are not being packed"
1162 );
1163
1164 static int add_object_entry(const struct object_id *oid, enum object_type type,
1165                             const char *name, int exclude)
1166 {
1167         struct packed_git *found_pack = NULL;
1168         off_t found_offset = 0;
1169         uint32_t index_pos;
1170
1171         display_progress(progress_state, ++nr_seen);
1172
1173         if (have_duplicate_entry(oid, exclude, &index_pos))
1174                 return 0;
1175
1176         if (!want_object_in_pack(oid, exclude, &found_pack, &found_offset)) {
1177                 /* The pack is missing an object, so it will not have closure */
1178                 if (write_bitmap_index) {
1179                         warning(_(no_closure_warning));
1180                         write_bitmap_index = 0;
1181                 }
1182                 return 0;
1183         }
1184
1185         create_object_entry(oid, type, pack_name_hash(name),
1186                             exclude, name && no_try_delta(name),
1187                             index_pos, found_pack, found_offset);
1188         return 1;
1189 }
1190
1191 static int add_object_entry_from_bitmap(const struct object_id *oid,
1192                                         enum object_type type,
1193                                         int flags, uint32_t name_hash,
1194                                         struct packed_git *pack, off_t offset)
1195 {
1196         uint32_t index_pos;
1197
1198         display_progress(progress_state, ++nr_seen);
1199
1200         if (have_duplicate_entry(oid, 0, &index_pos))
1201                 return 0;
1202
1203         if (!want_object_in_pack(oid, 0, &pack, &offset))
1204                 return 0;
1205
1206         create_object_entry(oid, type, name_hash, 0, 0, index_pos, pack, offset);
1207         return 1;
1208 }
1209
1210 struct pbase_tree_cache {
1211         struct object_id oid;
1212         int ref;
1213         int temporary;
1214         void *tree_data;
1215         unsigned long tree_size;
1216 };
1217
1218 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1219 static int pbase_tree_cache_ix(const struct object_id *oid)
1220 {
1221         return oid->hash[0] % ARRAY_SIZE(pbase_tree_cache);
1222 }
1223 static int pbase_tree_cache_ix_incr(int ix)
1224 {
1225         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1226 }
1227
1228 static struct pbase_tree {
1229         struct pbase_tree *next;
1230         /* This is a phony "cache" entry; we are not
1231          * going to evict it or find it through _get()
1232          * mechanism -- this is for the toplevel node that
1233          * would almost always change with any commit.
1234          */
1235         struct pbase_tree_cache pcache;
1236 } *pbase_tree;
1237
1238 static struct pbase_tree_cache *pbase_tree_get(const struct object_id *oid)
1239 {
1240         struct pbase_tree_cache *ent, *nent;
1241         void *data;
1242         unsigned long size;
1243         enum object_type type;
1244         int neigh;
1245         int my_ix = pbase_tree_cache_ix(oid);
1246         int available_ix = -1;
1247
1248         /* pbase-tree-cache acts as a limited hashtable.
1249          * your object will be found at your index or within a few
1250          * slots after that slot if it is cached.
1251          */
1252         for (neigh = 0; neigh < 8; neigh++) {
1253                 ent = pbase_tree_cache[my_ix];
1254                 if (ent && oideq(&ent->oid, oid)) {
1255                         ent->ref++;
1256                         return ent;
1257                 }
1258                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1259                          ((0 <= available_ix) &&
1260                           (!ent && pbase_tree_cache[available_ix])))
1261                         available_ix = my_ix;
1262                 if (!ent)
1263                         break;
1264                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1265         }
1266
1267         /* Did not find one.  Either we got a bogus request or
1268          * we need to read and perhaps cache.
1269          */
1270         data = read_object_file(oid, &type, &size);
1271         if (!data)
1272                 return NULL;
1273         if (type != OBJ_TREE) {
1274                 free(data);
1275                 return NULL;
1276         }
1277
1278         /* We need to either cache or return a throwaway copy */
1279
1280         if (available_ix < 0)
1281                 ent = NULL;
1282         else {
1283                 ent = pbase_tree_cache[available_ix];
1284                 my_ix = available_ix;
1285         }
1286
1287         if (!ent) {
1288                 nent = xmalloc(sizeof(*nent));
1289                 nent->temporary = (available_ix < 0);
1290         }
1291         else {
1292                 /* evict and reuse */
1293                 free(ent->tree_data);
1294                 nent = ent;
1295         }
1296         oidcpy(&nent->oid, oid);
1297         nent->tree_data = data;
1298         nent->tree_size = size;
1299         nent->ref = 1;
1300         if (!nent->temporary)
1301                 pbase_tree_cache[my_ix] = nent;
1302         return nent;
1303 }
1304
1305 static void pbase_tree_put(struct pbase_tree_cache *cache)
1306 {
1307         if (!cache->temporary) {
1308                 cache->ref--;
1309                 return;
1310         }
1311         free(cache->tree_data);
1312         free(cache);
1313 }
1314
1315 static int name_cmp_len(const char *name)
1316 {
1317         int i;
1318         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1319                 ;
1320         return i;
1321 }
1322
1323 static void add_pbase_object(struct tree_desc *tree,
1324                              const char *name,
1325                              int cmplen,
1326                              const char *fullname)
1327 {
1328         struct name_entry entry;
1329         int cmp;
1330
1331         while (tree_entry(tree,&entry)) {
1332                 if (S_ISGITLINK(entry.mode))
1333                         continue;
1334                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1335                       memcmp(name, entry.path, cmplen);
1336                 if (cmp > 0)
1337                         continue;
1338                 if (cmp < 0)
1339                         return;
1340                 if (name[cmplen] != '/') {
1341                         add_object_entry(&entry.oid,
1342                                          object_type(entry.mode),
1343                                          fullname, 1);
1344                         return;
1345                 }
1346                 if (S_ISDIR(entry.mode)) {
1347                         struct tree_desc sub;
1348                         struct pbase_tree_cache *tree;
1349                         const char *down = name+cmplen+1;
1350                         int downlen = name_cmp_len(down);
1351
1352                         tree = pbase_tree_get(&entry.oid);
1353                         if (!tree)
1354                                 return;
1355                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1356
1357                         add_pbase_object(&sub, down, downlen, fullname);
1358                         pbase_tree_put(tree);
1359                 }
1360         }
1361 }
1362
1363 static unsigned *done_pbase_paths;
1364 static int done_pbase_paths_num;
1365 static int done_pbase_paths_alloc;
1366 static int done_pbase_path_pos(unsigned hash)
1367 {
1368         int lo = 0;
1369         int hi = done_pbase_paths_num;
1370         while (lo < hi) {
1371                 int mi = lo + (hi - lo) / 2;
1372                 if (done_pbase_paths[mi] == hash)
1373                         return mi;
1374                 if (done_pbase_paths[mi] < hash)
1375                         hi = mi;
1376                 else
1377                         lo = mi + 1;
1378         }
1379         return -lo-1;
1380 }
1381
1382 static int check_pbase_path(unsigned hash)
1383 {
1384         int pos = done_pbase_path_pos(hash);
1385         if (0 <= pos)
1386                 return 1;
1387         pos = -pos - 1;
1388         ALLOC_GROW(done_pbase_paths,
1389                    done_pbase_paths_num + 1,
1390                    done_pbase_paths_alloc);
1391         done_pbase_paths_num++;
1392         if (pos < done_pbase_paths_num)
1393                 MOVE_ARRAY(done_pbase_paths + pos + 1, done_pbase_paths + pos,
1394                            done_pbase_paths_num - pos - 1);
1395         done_pbase_paths[pos] = hash;
1396         return 0;
1397 }
1398
1399 static void add_preferred_base_object(const char *name)
1400 {
1401         struct pbase_tree *it;
1402         int cmplen;
1403         unsigned hash = pack_name_hash(name);
1404
1405         if (!num_preferred_base || check_pbase_path(hash))
1406                 return;
1407
1408         cmplen = name_cmp_len(name);
1409         for (it = pbase_tree; it; it = it->next) {
1410                 if (cmplen == 0) {
1411                         add_object_entry(&it->pcache.oid, OBJ_TREE, NULL, 1);
1412                 }
1413                 else {
1414                         struct tree_desc tree;
1415                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1416                         add_pbase_object(&tree, name, cmplen, name);
1417                 }
1418         }
1419 }
1420
1421 static void add_preferred_base(struct object_id *oid)
1422 {
1423         struct pbase_tree *it;
1424         void *data;
1425         unsigned long size;
1426         struct object_id tree_oid;
1427
1428         if (window <= num_preferred_base++)
1429                 return;
1430
1431         data = read_object_with_reference(oid, tree_type, &size, &tree_oid);
1432         if (!data)
1433                 return;
1434
1435         for (it = pbase_tree; it; it = it->next) {
1436                 if (oideq(&it->pcache.oid, &tree_oid)) {
1437                         free(data);
1438                         return;
1439                 }
1440         }
1441
1442         it = xcalloc(1, sizeof(*it));
1443         it->next = pbase_tree;
1444         pbase_tree = it;
1445
1446         oidcpy(&it->pcache.oid, &tree_oid);
1447         it->pcache.tree_data = data;
1448         it->pcache.tree_size = size;
1449 }
1450
1451 static void cleanup_preferred_base(void)
1452 {
1453         struct pbase_tree *it;
1454         unsigned i;
1455
1456         it = pbase_tree;
1457         pbase_tree = NULL;
1458         while (it) {
1459                 struct pbase_tree *tmp = it;
1460                 it = tmp->next;
1461                 free(tmp->pcache.tree_data);
1462                 free(tmp);
1463         }
1464
1465         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1466                 if (!pbase_tree_cache[i])
1467                         continue;
1468                 free(pbase_tree_cache[i]->tree_data);
1469                 FREE_AND_NULL(pbase_tree_cache[i]);
1470         }
1471
1472         FREE_AND_NULL(done_pbase_paths);
1473         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1474 }
1475
1476 /*
1477  * Return 1 iff the object specified by "delta" can be sent
1478  * literally as a delta against the base in "base_sha1". If
1479  * so, then *base_out will point to the entry in our packing
1480  * list, or NULL if we must use the external-base list.
1481  *
1482  * Depth value does not matter - find_deltas() will
1483  * never consider reused delta as the base object to
1484  * deltify other objects against, in order to avoid
1485  * circular deltas.
1486  */
1487 static int can_reuse_delta(const unsigned char *base_sha1,
1488                            struct object_entry *delta,
1489                            struct object_entry **base_out)
1490 {
1491         struct object_entry *base;
1492         struct object_id base_oid;
1493
1494         if (!base_sha1)
1495                 return 0;
1496
1497         /*
1498          * First see if we're already sending the base (or it's explicitly in
1499          * our "excluded" list).
1500          */
1501         base = packlist_find(&to_pack, base_sha1, NULL);
1502         if (base) {
1503                 if (!in_same_island(&delta->idx.oid, &base->idx.oid))
1504                         return 0;
1505                 *base_out = base;
1506                 return 1;
1507         }
1508
1509         /*
1510          * Otherwise, reachability bitmaps may tell us if the receiver has it,
1511          * even if it was buried too deep in history to make it into the
1512          * packing list.
1513          */
1514         oidread(&base_oid, base_sha1);
1515         if (thin && bitmap_has_oid_in_uninteresting(bitmap_git, &base_oid)) {
1516                 if (use_delta_islands) {
1517                         if (!in_same_island(&delta->idx.oid, &base_oid))
1518                                 return 0;
1519                 }
1520                 *base_out = NULL;
1521                 return 1;
1522         }
1523
1524         return 0;
1525 }
1526
1527 static void check_object(struct object_entry *entry)
1528 {
1529         unsigned long canonical_size;
1530
1531         if (IN_PACK(entry)) {
1532                 struct packed_git *p = IN_PACK(entry);
1533                 struct pack_window *w_curs = NULL;
1534                 const unsigned char *base_ref = NULL;
1535                 struct object_entry *base_entry;
1536                 unsigned long used, used_0;
1537                 unsigned long avail;
1538                 off_t ofs;
1539                 unsigned char *buf, c;
1540                 enum object_type type;
1541                 unsigned long in_pack_size;
1542
1543                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1544
1545                 /*
1546                  * We want in_pack_type even if we do not reuse delta
1547                  * since non-delta representations could still be reused.
1548                  */
1549                 used = unpack_object_header_buffer(buf, avail,
1550                                                    &type,
1551                                                    &in_pack_size);
1552                 if (used == 0)
1553                         goto give_up;
1554
1555                 if (type < 0)
1556                         BUG("invalid type %d", type);
1557                 entry->in_pack_type = type;
1558
1559                 /*
1560                  * Determine if this is a delta and if so whether we can
1561                  * reuse it or not.  Otherwise let's find out as cheaply as
1562                  * possible what the actual type and size for this object is.
1563                  */
1564                 switch (entry->in_pack_type) {
1565                 default:
1566                         /* Not a delta hence we've already got all we need. */
1567                         oe_set_type(entry, entry->in_pack_type);
1568                         SET_SIZE(entry, in_pack_size);
1569                         entry->in_pack_header_size = used;
1570                         if (oe_type(entry) < OBJ_COMMIT || oe_type(entry) > OBJ_BLOB)
1571                                 goto give_up;
1572                         unuse_pack(&w_curs);
1573                         return;
1574                 case OBJ_REF_DELTA:
1575                         if (reuse_delta && !entry->preferred_base)
1576                                 base_ref = use_pack(p, &w_curs,
1577                                                 entry->in_pack_offset + used, NULL);
1578                         entry->in_pack_header_size = used + the_hash_algo->rawsz;
1579                         break;
1580                 case OBJ_OFS_DELTA:
1581                         buf = use_pack(p, &w_curs,
1582                                        entry->in_pack_offset + used, NULL);
1583                         used_0 = 0;
1584                         c = buf[used_0++];
1585                         ofs = c & 127;
1586                         while (c & 128) {
1587                                 ofs += 1;
1588                                 if (!ofs || MSB(ofs, 7)) {
1589                                         error(_("delta base offset overflow in pack for %s"),
1590                                               oid_to_hex(&entry->idx.oid));
1591                                         goto give_up;
1592                                 }
1593                                 c = buf[used_0++];
1594                                 ofs = (ofs << 7) + (c & 127);
1595                         }
1596                         ofs = entry->in_pack_offset - ofs;
1597                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1598                                 error(_("delta base offset out of bound for %s"),
1599                                       oid_to_hex(&entry->idx.oid));
1600                                 goto give_up;
1601                         }
1602                         if (reuse_delta && !entry->preferred_base) {
1603                                 struct revindex_entry *revidx;
1604                                 revidx = find_pack_revindex(p, ofs);
1605                                 if (!revidx)
1606                                         goto give_up;
1607                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1608                         }
1609                         entry->in_pack_header_size = used + used_0;
1610                         break;
1611                 }
1612
1613                 if (can_reuse_delta(base_ref, entry, &base_entry)) {
1614                         oe_set_type(entry, entry->in_pack_type);
1615                         SET_SIZE(entry, in_pack_size); /* delta size */
1616                         SET_DELTA_SIZE(entry, in_pack_size);
1617
1618                         if (base_entry) {
1619                                 SET_DELTA(entry, base_entry);
1620                                 entry->delta_sibling_idx = base_entry->delta_child_idx;
1621                                 SET_DELTA_CHILD(base_entry, entry);
1622                         } else {
1623                                 SET_DELTA_EXT(entry, base_ref);
1624                         }
1625
1626                         unuse_pack(&w_curs);
1627                         return;
1628                 }
1629
1630                 if (oe_type(entry)) {
1631                         off_t delta_pos;
1632
1633                         /*
1634                          * This must be a delta and we already know what the
1635                          * final object type is.  Let's extract the actual
1636                          * object size from the delta header.
1637                          */
1638                         delta_pos = entry->in_pack_offset + entry->in_pack_header_size;
1639                         canonical_size = get_size_from_delta(p, &w_curs, delta_pos);
1640                         if (canonical_size == 0)
1641                                 goto give_up;
1642                         SET_SIZE(entry, canonical_size);
1643                         unuse_pack(&w_curs);
1644                         return;
1645                 }
1646
1647                 /*
1648                  * No choice but to fall back to the recursive delta walk
1649                  * with oid_object_info() to find about the object type
1650                  * at this point...
1651                  */
1652                 give_up:
1653                 unuse_pack(&w_curs);
1654         }
1655
1656         oe_set_type(entry,
1657                     oid_object_info(the_repository, &entry->idx.oid, &canonical_size));
1658         if (entry->type_valid) {
1659                 SET_SIZE(entry, canonical_size);
1660         } else {
1661                 /*
1662                  * Bad object type is checked in prepare_pack().  This is
1663                  * to permit a missing preferred base object to be ignored
1664                  * as a preferred base.  Doing so can result in a larger
1665                  * pack file, but the transfer will still take place.
1666                  */
1667         }
1668 }
1669
1670 static int pack_offset_sort(const void *_a, const void *_b)
1671 {
1672         const struct object_entry *a = *(struct object_entry **)_a;
1673         const struct object_entry *b = *(struct object_entry **)_b;
1674         const struct packed_git *a_in_pack = IN_PACK(a);
1675         const struct packed_git *b_in_pack = IN_PACK(b);
1676
1677         /* avoid filesystem trashing with loose objects */
1678         if (!a_in_pack && !b_in_pack)
1679                 return oidcmp(&a->idx.oid, &b->idx.oid);
1680
1681         if (a_in_pack < b_in_pack)
1682                 return -1;
1683         if (a_in_pack > b_in_pack)
1684                 return 1;
1685         return a->in_pack_offset < b->in_pack_offset ? -1 :
1686                         (a->in_pack_offset > b->in_pack_offset);
1687 }
1688
1689 /*
1690  * Drop an on-disk delta we were planning to reuse. Naively, this would
1691  * just involve blanking out the "delta" field, but we have to deal
1692  * with some extra book-keeping:
1693  *
1694  *   1. Removing ourselves from the delta_sibling linked list.
1695  *
1696  *   2. Updating our size/type to the non-delta representation. These were
1697  *      either not recorded initially (size) or overwritten with the delta type
1698  *      (type) when check_object() decided to reuse the delta.
1699  *
1700  *   3. Resetting our delta depth, as we are now a base object.
1701  */
1702 static void drop_reused_delta(struct object_entry *entry)
1703 {
1704         unsigned *idx = &to_pack.objects[entry->delta_idx - 1].delta_child_idx;
1705         struct object_info oi = OBJECT_INFO_INIT;
1706         enum object_type type;
1707         unsigned long size;
1708
1709         while (*idx) {
1710                 struct object_entry *oe = &to_pack.objects[*idx - 1];
1711
1712                 if (oe == entry)
1713                         *idx = oe->delta_sibling_idx;
1714                 else
1715                         idx = &oe->delta_sibling_idx;
1716         }
1717         SET_DELTA(entry, NULL);
1718         entry->depth = 0;
1719
1720         oi.sizep = &size;
1721         oi.typep = &type;
1722         if (packed_object_info(the_repository, IN_PACK(entry), entry->in_pack_offset, &oi) < 0) {
1723                 /*
1724                  * We failed to get the info from this pack for some reason;
1725                  * fall back to oid_object_info, which may find another copy.
1726                  * And if that fails, the error will be recorded in oe_type(entry)
1727                  * and dealt with in prepare_pack().
1728                  */
1729                 oe_set_type(entry,
1730                             oid_object_info(the_repository, &entry->idx.oid, &size));
1731         } else {
1732                 oe_set_type(entry, type);
1733         }
1734         SET_SIZE(entry, size);
1735 }
1736
1737 /*
1738  * Follow the chain of deltas from this entry onward, throwing away any links
1739  * that cause us to hit a cycle (as determined by the DFS state flags in
1740  * the entries).
1741  *
1742  * We also detect too-long reused chains that would violate our --depth
1743  * limit.
1744  */
1745 static void break_delta_chains(struct object_entry *entry)
1746 {
1747         /*
1748          * The actual depth of each object we will write is stored as an int,
1749          * as it cannot exceed our int "depth" limit. But before we break
1750          * changes based no that limit, we may potentially go as deep as the
1751          * number of objects, which is elsewhere bounded to a uint32_t.
1752          */
1753         uint32_t total_depth;
1754         struct object_entry *cur, *next;
1755
1756         for (cur = entry, total_depth = 0;
1757              cur;
1758              cur = DELTA(cur), total_depth++) {
1759                 if (cur->dfs_state == DFS_DONE) {
1760                         /*
1761                          * We've already seen this object and know it isn't
1762                          * part of a cycle. We do need to append its depth
1763                          * to our count.
1764                          */
1765                         total_depth += cur->depth;
1766                         break;
1767                 }
1768
1769                 /*
1770                  * We break cycles before looping, so an ACTIVE state (or any
1771                  * other cruft which made its way into the state variable)
1772                  * is a bug.
1773                  */
1774                 if (cur->dfs_state != DFS_NONE)
1775                         BUG("confusing delta dfs state in first pass: %d",
1776                             cur->dfs_state);
1777
1778                 /*
1779                  * Now we know this is the first time we've seen the object. If
1780                  * it's not a delta, we're done traversing, but we'll mark it
1781                  * done to save time on future traversals.
1782                  */
1783                 if (!DELTA(cur)) {
1784                         cur->dfs_state = DFS_DONE;
1785                         break;
1786                 }
1787
1788                 /*
1789                  * Mark ourselves as active and see if the next step causes
1790                  * us to cycle to another active object. It's important to do
1791                  * this _before_ we loop, because it impacts where we make the
1792                  * cut, and thus how our total_depth counter works.
1793                  * E.g., We may see a partial loop like:
1794                  *
1795                  *   A -> B -> C -> D -> B
1796                  *
1797                  * Cutting B->C breaks the cycle. But now the depth of A is
1798                  * only 1, and our total_depth counter is at 3. The size of the
1799                  * error is always one less than the size of the cycle we
1800                  * broke. Commits C and D were "lost" from A's chain.
1801                  *
1802                  * If we instead cut D->B, then the depth of A is correct at 3.
1803                  * We keep all commits in the chain that we examined.
1804                  */
1805                 cur->dfs_state = DFS_ACTIVE;
1806                 if (DELTA(cur)->dfs_state == DFS_ACTIVE) {
1807                         drop_reused_delta(cur);
1808                         cur->dfs_state = DFS_DONE;
1809                         break;
1810                 }
1811         }
1812
1813         /*
1814          * And now that we've gone all the way to the bottom of the chain, we
1815          * need to clear the active flags and set the depth fields as
1816          * appropriate. Unlike the loop above, which can quit when it drops a
1817          * delta, we need to keep going to look for more depth cuts. So we need
1818          * an extra "next" pointer to keep going after we reset cur->delta.
1819          */
1820         for (cur = entry; cur; cur = next) {
1821                 next = DELTA(cur);
1822
1823                 /*
1824                  * We should have a chain of zero or more ACTIVE states down to
1825                  * a final DONE. We can quit after the DONE, because either it
1826                  * has no bases, or we've already handled them in a previous
1827                  * call.
1828                  */
1829                 if (cur->dfs_state == DFS_DONE)
1830                         break;
1831                 else if (cur->dfs_state != DFS_ACTIVE)
1832                         BUG("confusing delta dfs state in second pass: %d",
1833                             cur->dfs_state);
1834
1835                 /*
1836                  * If the total_depth is more than depth, then we need to snip
1837                  * the chain into two or more smaller chains that don't exceed
1838                  * the maximum depth. Most of the resulting chains will contain
1839                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1840                  * the last chain (i.e., the one containing entry) will contain
1841                  * whatever entries are left over, namely
1842                  * (total_depth % (depth + 1)) of them.
1843                  *
1844                  * Since we are iterating towards decreasing depth, we need to
1845                  * decrement total_depth as we go, and we need to write to the
1846                  * entry what its final depth will be after all of the
1847                  * snipping. Since we're snipping into chains of length (depth
1848                  * + 1) entries, the final depth of an entry will be its
1849                  * original depth modulo (depth + 1). Any time we encounter an
1850                  * entry whose final depth is supposed to be zero, we snip it
1851                  * from its delta base, thereby making it so.
1852                  */
1853                 cur->depth = (total_depth--) % (depth + 1);
1854                 if (!cur->depth)
1855                         drop_reused_delta(cur);
1856
1857                 cur->dfs_state = DFS_DONE;
1858         }
1859 }
1860
1861 static void get_object_details(void)
1862 {
1863         uint32_t i;
1864         struct object_entry **sorted_by_offset;
1865
1866         if (progress)
1867                 progress_state = start_progress(_("Counting objects"),
1868                                                 to_pack.nr_objects);
1869
1870         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1871         for (i = 0; i < to_pack.nr_objects; i++)
1872                 sorted_by_offset[i] = to_pack.objects + i;
1873         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1874
1875         for (i = 0; i < to_pack.nr_objects; i++) {
1876                 struct object_entry *entry = sorted_by_offset[i];
1877                 check_object(entry);
1878                 if (entry->type_valid &&
1879                     oe_size_greater_than(&to_pack, entry, big_file_threshold))
1880                         entry->no_try_delta = 1;
1881                 display_progress(progress_state, i + 1);
1882         }
1883         stop_progress(&progress_state);
1884
1885         /*
1886          * This must happen in a second pass, since we rely on the delta
1887          * information for the whole list being completed.
1888          */
1889         for (i = 0; i < to_pack.nr_objects; i++)
1890                 break_delta_chains(&to_pack.objects[i]);
1891
1892         free(sorted_by_offset);
1893 }
1894
1895 /*
1896  * We search for deltas in a list sorted by type, by filename hash, and then
1897  * by size, so that we see progressively smaller and smaller files.
1898  * That's because we prefer deltas to be from the bigger file
1899  * to the smaller -- deletes are potentially cheaper, but perhaps
1900  * more importantly, the bigger file is likely the more recent
1901  * one.  The deepest deltas are therefore the oldest objects which are
1902  * less susceptible to be accessed often.
1903  */
1904 static int type_size_sort(const void *_a, const void *_b)
1905 {
1906         const struct object_entry *a = *(struct object_entry **)_a;
1907         const struct object_entry *b = *(struct object_entry **)_b;
1908         const enum object_type a_type = oe_type(a);
1909         const enum object_type b_type = oe_type(b);
1910         const unsigned long a_size = SIZE(a);
1911         const unsigned long b_size = SIZE(b);
1912
1913         if (a_type > b_type)
1914                 return -1;
1915         if (a_type < b_type)
1916                 return 1;
1917         if (a->hash > b->hash)
1918                 return -1;
1919         if (a->hash < b->hash)
1920                 return 1;
1921         if (a->preferred_base > b->preferred_base)
1922                 return -1;
1923         if (a->preferred_base < b->preferred_base)
1924                 return 1;
1925         if (use_delta_islands) {
1926                 const int island_cmp = island_delta_cmp(&a->idx.oid, &b->idx.oid);
1927                 if (island_cmp)
1928                         return island_cmp;
1929         }
1930         if (a_size > b_size)
1931                 return -1;
1932         if (a_size < b_size)
1933                 return 1;
1934         return a < b ? -1 : (a > b);  /* newest first */
1935 }
1936
1937 struct unpacked {
1938         struct object_entry *entry;
1939         void *data;
1940         struct delta_index *index;
1941         unsigned depth;
1942 };
1943
1944 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1945                            unsigned long delta_size)
1946 {
1947         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1948                 return 0;
1949
1950         if (delta_size < cache_max_small_delta_size)
1951                 return 1;
1952
1953         /* cache delta, if objects are large enough compared to delta size */
1954         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1955                 return 1;
1956
1957         return 0;
1958 }
1959
1960 /* Protect delta_cache_size */
1961 static pthread_mutex_t cache_mutex;
1962 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1963 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1964
1965 /*
1966  * Protect object list partitioning (e.g. struct thread_param) and
1967  * progress_state
1968  */
1969 static pthread_mutex_t progress_mutex;
1970 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1971 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1972
1973 /*
1974  * Access to struct object_entry is unprotected since each thread owns
1975  * a portion of the main object list. Just don't access object entries
1976  * ahead in the list because they can be stolen and would need
1977  * progress_mutex for protection.
1978  */
1979
1980 /*
1981  * Return the size of the object without doing any delta
1982  * reconstruction (so non-deltas are true object sizes, but deltas
1983  * return the size of the delta data).
1984  */
1985 unsigned long oe_get_size_slow(struct packing_data *pack,
1986                                const struct object_entry *e)
1987 {
1988         struct packed_git *p;
1989         struct pack_window *w_curs;
1990         unsigned char *buf;
1991         enum object_type type;
1992         unsigned long used, avail, size;
1993
1994         if (e->type_ != OBJ_OFS_DELTA && e->type_ != OBJ_REF_DELTA) {
1995                 packing_data_lock(&to_pack);
1996                 if (oid_object_info(the_repository, &e->idx.oid, &size) < 0)
1997                         die(_("unable to get size of %s"),
1998                             oid_to_hex(&e->idx.oid));
1999                 packing_data_unlock(&to_pack);
2000                 return size;
2001         }
2002
2003         p = oe_in_pack(pack, e);
2004         if (!p)
2005                 BUG("when e->type is a delta, it must belong to a pack");
2006
2007         packing_data_lock(&to_pack);
2008         w_curs = NULL;
2009         buf = use_pack(p, &w_curs, e->in_pack_offset, &avail);
2010         used = unpack_object_header_buffer(buf, avail, &type, &size);
2011         if (used == 0)
2012                 die(_("unable to parse object header of %s"),
2013                     oid_to_hex(&e->idx.oid));
2014
2015         unuse_pack(&w_curs);
2016         packing_data_unlock(&to_pack);
2017         return size;
2018 }
2019
2020 static int try_delta(struct unpacked *trg, struct unpacked *src,
2021                      unsigned max_depth, unsigned long *mem_usage)
2022 {
2023         struct object_entry *trg_entry = trg->entry;
2024         struct object_entry *src_entry = src->entry;
2025         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
2026         unsigned ref_depth;
2027         enum object_type type;
2028         void *delta_buf;
2029
2030         /* Don't bother doing diffs between different types */
2031         if (oe_type(trg_entry) != oe_type(src_entry))
2032                 return -1;
2033
2034         /*
2035          * We do not bother to try a delta that we discarded on an
2036          * earlier try, but only when reusing delta data.  Note that
2037          * src_entry that is marked as the preferred_base should always
2038          * be considered, as even if we produce a suboptimal delta against
2039          * it, we will still save the transfer cost, as we already know
2040          * the other side has it and we won't send src_entry at all.
2041          */
2042         if (reuse_delta && IN_PACK(trg_entry) &&
2043             IN_PACK(trg_entry) == IN_PACK(src_entry) &&
2044             !src_entry->preferred_base &&
2045             trg_entry->in_pack_type != OBJ_REF_DELTA &&
2046             trg_entry->in_pack_type != OBJ_OFS_DELTA)
2047                 return 0;
2048
2049         /* Let's not bust the allowed depth. */
2050         if (src->depth >= max_depth)
2051                 return 0;
2052
2053         /* Now some size filtering heuristics. */
2054         trg_size = SIZE(trg_entry);
2055         if (!DELTA(trg_entry)) {
2056                 max_size = trg_size/2 - the_hash_algo->rawsz;
2057                 ref_depth = 1;
2058         } else {
2059                 max_size = DELTA_SIZE(trg_entry);
2060                 ref_depth = trg->depth;
2061         }
2062         max_size = (uint64_t)max_size * (max_depth - src->depth) /
2063                                                 (max_depth - ref_depth + 1);
2064         if (max_size == 0)
2065                 return 0;
2066         src_size = SIZE(src_entry);
2067         sizediff = src_size < trg_size ? trg_size - src_size : 0;
2068         if (sizediff >= max_size)
2069                 return 0;
2070         if (trg_size < src_size / 32)
2071                 return 0;
2072
2073         if (!in_same_island(&trg->entry->idx.oid, &src->entry->idx.oid))
2074                 return 0;
2075
2076         /* Load data if not already done */
2077         if (!trg->data) {
2078                 packing_data_lock(&to_pack);
2079                 trg->data = read_object_file(&trg_entry->idx.oid, &type, &sz);
2080                 packing_data_unlock(&to_pack);
2081                 if (!trg->data)
2082                         die(_("object %s cannot be read"),
2083                             oid_to_hex(&trg_entry->idx.oid));
2084                 if (sz != trg_size)
2085                         die(_("object %s inconsistent object length (%"PRIuMAX" vs %"PRIuMAX")"),
2086                             oid_to_hex(&trg_entry->idx.oid), (uintmax_t)sz,
2087                             (uintmax_t)trg_size);
2088                 *mem_usage += sz;
2089         }
2090         if (!src->data) {
2091                 packing_data_lock(&to_pack);
2092                 src->data = read_object_file(&src_entry->idx.oid, &type, &sz);
2093                 packing_data_unlock(&to_pack);
2094                 if (!src->data) {
2095                         if (src_entry->preferred_base) {
2096                                 static int warned = 0;
2097                                 if (!warned++)
2098                                         warning(_("object %s cannot be read"),
2099                                                 oid_to_hex(&src_entry->idx.oid));
2100                                 /*
2101                                  * Those objects are not included in the
2102                                  * resulting pack.  Be resilient and ignore
2103                                  * them if they can't be read, in case the
2104                                  * pack could be created nevertheless.
2105                                  */
2106                                 return 0;
2107                         }
2108                         die(_("object %s cannot be read"),
2109                             oid_to_hex(&src_entry->idx.oid));
2110                 }
2111                 if (sz != src_size)
2112                         die(_("object %s inconsistent object length (%"PRIuMAX" vs %"PRIuMAX")"),
2113                             oid_to_hex(&src_entry->idx.oid), (uintmax_t)sz,
2114                             (uintmax_t)src_size);
2115                 *mem_usage += sz;
2116         }
2117         if (!src->index) {
2118                 src->index = create_delta_index(src->data, src_size);
2119                 if (!src->index) {
2120                         static int warned = 0;
2121                         if (!warned++)
2122                                 warning(_("suboptimal pack - out of memory"));
2123                         return 0;
2124                 }
2125                 *mem_usage += sizeof_delta_index(src->index);
2126         }
2127
2128         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
2129         if (!delta_buf)
2130                 return 0;
2131
2132         if (DELTA(trg_entry)) {
2133                 /* Prefer only shallower same-sized deltas. */
2134                 if (delta_size == DELTA_SIZE(trg_entry) &&
2135                     src->depth + 1 >= trg->depth) {
2136                         free(delta_buf);
2137                         return 0;
2138                 }
2139         }
2140
2141         /*
2142          * Handle memory allocation outside of the cache
2143          * accounting lock.  Compiler will optimize the strangeness
2144          * away when NO_PTHREADS is defined.
2145          */
2146         free(trg_entry->delta_data);
2147         cache_lock();
2148         if (trg_entry->delta_data) {
2149                 delta_cache_size -= DELTA_SIZE(trg_entry);
2150                 trg_entry->delta_data = NULL;
2151         }
2152         if (delta_cacheable(src_size, trg_size, delta_size)) {
2153                 delta_cache_size += delta_size;
2154                 cache_unlock();
2155                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
2156         } else {
2157                 cache_unlock();
2158                 free(delta_buf);
2159         }
2160
2161         SET_DELTA(trg_entry, src_entry);
2162         SET_DELTA_SIZE(trg_entry, delta_size);
2163         trg->depth = src->depth + 1;
2164
2165         return 1;
2166 }
2167
2168 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
2169 {
2170         struct object_entry *child = DELTA_CHILD(me);
2171         unsigned int m = n;
2172         while (child) {
2173                 const unsigned int c = check_delta_limit(child, n + 1);
2174                 if (m < c)
2175                         m = c;
2176                 child = DELTA_SIBLING(child);
2177         }
2178         return m;
2179 }
2180
2181 static unsigned long free_unpacked(struct unpacked *n)
2182 {
2183         unsigned long freed_mem = sizeof_delta_index(n->index);
2184         free_delta_index(n->index);
2185         n->index = NULL;
2186         if (n->data) {
2187                 freed_mem += SIZE(n->entry);
2188                 FREE_AND_NULL(n->data);
2189         }
2190         n->entry = NULL;
2191         n->depth = 0;
2192         return freed_mem;
2193 }
2194
2195 static void find_deltas(struct object_entry **list, unsigned *list_size,
2196                         int window, int depth, unsigned *processed)
2197 {
2198         uint32_t i, idx = 0, count = 0;
2199         struct unpacked *array;
2200         unsigned long mem_usage = 0;
2201
2202         array = xcalloc(window, sizeof(struct unpacked));
2203
2204         for (;;) {
2205                 struct object_entry *entry;
2206                 struct unpacked *n = array + idx;
2207                 int j, max_depth, best_base = -1;
2208
2209                 progress_lock();
2210                 if (!*list_size) {
2211                         progress_unlock();
2212                         break;
2213                 }
2214                 entry = *list++;
2215                 (*list_size)--;
2216                 if (!entry->preferred_base) {
2217                         (*processed)++;
2218                         display_progress(progress_state, *processed);
2219                 }
2220                 progress_unlock();
2221
2222                 mem_usage -= free_unpacked(n);
2223                 n->entry = entry;
2224
2225                 while (window_memory_limit &&
2226                        mem_usage > window_memory_limit &&
2227                        count > 1) {
2228                         const uint32_t tail = (idx + window - count) % window;
2229                         mem_usage -= free_unpacked(array + tail);
2230                         count--;
2231                 }
2232
2233                 /* We do not compute delta to *create* objects we are not
2234                  * going to pack.
2235                  */
2236                 if (entry->preferred_base)
2237                         goto next;
2238
2239                 /*
2240                  * If the current object is at pack edge, take the depth the
2241                  * objects that depend on the current object into account
2242                  * otherwise they would become too deep.
2243                  */
2244                 max_depth = depth;
2245                 if (DELTA_CHILD(entry)) {
2246                         max_depth -= check_delta_limit(entry, 0);
2247                         if (max_depth <= 0)
2248                                 goto next;
2249                 }
2250
2251                 j = window;
2252                 while (--j > 0) {
2253                         int ret;
2254                         uint32_t other_idx = idx + j;
2255                         struct unpacked *m;
2256                         if (other_idx >= window)
2257                                 other_idx -= window;
2258                         m = array + other_idx;
2259                         if (!m->entry)
2260                                 break;
2261                         ret = try_delta(n, m, max_depth, &mem_usage);
2262                         if (ret < 0)
2263                                 break;
2264                         else if (ret > 0)
2265                                 best_base = other_idx;
2266                 }
2267
2268                 /*
2269                  * If we decided to cache the delta data, then it is best
2270                  * to compress it right away.  First because we have to do
2271                  * it anyway, and doing it here while we're threaded will
2272                  * save a lot of time in the non threaded write phase,
2273                  * as well as allow for caching more deltas within
2274                  * the same cache size limit.
2275                  * ...
2276                  * But only if not writing to stdout, since in that case
2277                  * the network is most likely throttling writes anyway,
2278                  * and therefore it is best to go to the write phase ASAP
2279                  * instead, as we can afford spending more time compressing
2280                  * between writes at that moment.
2281                  */
2282                 if (entry->delta_data && !pack_to_stdout) {
2283                         unsigned long size;
2284
2285                         size = do_compress(&entry->delta_data, DELTA_SIZE(entry));
2286                         if (size < (1U << OE_Z_DELTA_BITS)) {
2287                                 entry->z_delta_size = size;
2288                                 cache_lock();
2289                                 delta_cache_size -= DELTA_SIZE(entry);
2290                                 delta_cache_size += entry->z_delta_size;
2291                                 cache_unlock();
2292                         } else {
2293                                 FREE_AND_NULL(entry->delta_data);
2294                                 entry->z_delta_size = 0;
2295                         }
2296                 }
2297
2298                 /* if we made n a delta, and if n is already at max
2299                  * depth, leaving it in the window is pointless.  we
2300                  * should evict it first.
2301                  */
2302                 if (DELTA(entry) && max_depth <= n->depth)
2303                         continue;
2304
2305                 /*
2306                  * Move the best delta base up in the window, after the
2307                  * currently deltified object, to keep it longer.  It will
2308                  * be the first base object to be attempted next.
2309                  */
2310                 if (DELTA(entry)) {
2311                         struct unpacked swap = array[best_base];
2312                         int dist = (window + idx - best_base) % window;
2313                         int dst = best_base;
2314                         while (dist--) {
2315                                 int src = (dst + 1) % window;
2316                                 array[dst] = array[src];
2317                                 dst = src;
2318                         }
2319                         array[dst] = swap;
2320                 }
2321
2322                 next:
2323                 idx++;
2324                 if (count + 1 < window)
2325                         count++;
2326                 if (idx >= window)
2327                         idx = 0;
2328         }
2329
2330         for (i = 0; i < window; ++i) {
2331                 free_delta_index(array[i].index);
2332                 free(array[i].data);
2333         }
2334         free(array);
2335 }
2336
2337 static void try_to_free_from_threads(size_t size)
2338 {
2339         packing_data_lock(&to_pack);
2340         release_pack_memory(size);
2341         packing_data_unlock(&to_pack);
2342 }
2343
2344 static try_to_free_t old_try_to_free_routine;
2345
2346 /*
2347  * The main object list is split into smaller lists, each is handed to
2348  * one worker.
2349  *
2350  * The main thread waits on the condition that (at least) one of the workers
2351  * has stopped working (which is indicated in the .working member of
2352  * struct thread_params).
2353  *
2354  * When a work thread has completed its work, it sets .working to 0 and
2355  * signals the main thread and waits on the condition that .data_ready
2356  * becomes 1.
2357  *
2358  * The main thread steals half of the work from the worker that has
2359  * most work left to hand it to the idle worker.
2360  */
2361
2362 struct thread_params {
2363         pthread_t thread;
2364         struct object_entry **list;
2365         unsigned list_size;
2366         unsigned remaining;
2367         int window;
2368         int depth;
2369         int working;
2370         int data_ready;
2371         pthread_mutex_t mutex;
2372         pthread_cond_t cond;
2373         unsigned *processed;
2374 };
2375
2376 static pthread_cond_t progress_cond;
2377
2378 /*
2379  * Mutex and conditional variable can't be statically-initialized on Windows.
2380  */
2381 static void init_threaded_search(void)
2382 {
2383         pthread_mutex_init(&cache_mutex, NULL);
2384         pthread_mutex_init(&progress_mutex, NULL);
2385         pthread_cond_init(&progress_cond, NULL);
2386         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2387 }
2388
2389 static void cleanup_threaded_search(void)
2390 {
2391         set_try_to_free_routine(old_try_to_free_routine);
2392         pthread_cond_destroy(&progress_cond);
2393         pthread_mutex_destroy(&cache_mutex);
2394         pthread_mutex_destroy(&progress_mutex);
2395 }
2396
2397 static void *threaded_find_deltas(void *arg)
2398 {
2399         struct thread_params *me = arg;
2400
2401         progress_lock();
2402         while (me->remaining) {
2403                 progress_unlock();
2404
2405                 find_deltas(me->list, &me->remaining,
2406                             me->window, me->depth, me->processed);
2407
2408                 progress_lock();
2409                 me->working = 0;
2410                 pthread_cond_signal(&progress_cond);
2411                 progress_unlock();
2412
2413                 /*
2414                  * We must not set ->data_ready before we wait on the
2415                  * condition because the main thread may have set it to 1
2416                  * before we get here. In order to be sure that new
2417                  * work is available if we see 1 in ->data_ready, it
2418                  * was initialized to 0 before this thread was spawned
2419                  * and we reset it to 0 right away.
2420                  */
2421                 pthread_mutex_lock(&me->mutex);
2422                 while (!me->data_ready)
2423                         pthread_cond_wait(&me->cond, &me->mutex);
2424                 me->data_ready = 0;
2425                 pthread_mutex_unlock(&me->mutex);
2426
2427                 progress_lock();
2428         }
2429         progress_unlock();
2430         /* leave ->working 1 so that this doesn't get more work assigned */
2431         return NULL;
2432 }
2433
2434 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2435                            int window, int depth, unsigned *processed)
2436 {
2437         struct thread_params *p;
2438         int i, ret, active_threads = 0;
2439
2440         init_threaded_search();
2441
2442         if (delta_search_threads <= 1) {
2443                 find_deltas(list, &list_size, window, depth, processed);
2444                 cleanup_threaded_search();
2445                 return;
2446         }
2447         if (progress > pack_to_stdout)
2448                 fprintf_ln(stderr, _("Delta compression using up to %d threads"),
2449                            delta_search_threads);
2450         p = xcalloc(delta_search_threads, sizeof(*p));
2451
2452         /* Partition the work amongst work threads. */
2453         for (i = 0; i < delta_search_threads; i++) {
2454                 unsigned sub_size = list_size / (delta_search_threads - i);
2455
2456                 /* don't use too small segments or no deltas will be found */
2457                 if (sub_size < 2*window && i+1 < delta_search_threads)
2458                         sub_size = 0;
2459
2460                 p[i].window = window;
2461                 p[i].depth = depth;
2462                 p[i].processed = processed;
2463                 p[i].working = 1;
2464                 p[i].data_ready = 0;
2465
2466                 /* try to split chunks on "path" boundaries */
2467                 while (sub_size && sub_size < list_size &&
2468                        list[sub_size]->hash &&
2469                        list[sub_size]->hash == list[sub_size-1]->hash)
2470                         sub_size++;
2471
2472                 p[i].list = list;
2473                 p[i].list_size = sub_size;
2474                 p[i].remaining = sub_size;
2475
2476                 list += sub_size;
2477                 list_size -= sub_size;
2478         }
2479
2480         /* Start work threads. */
2481         for (i = 0; i < delta_search_threads; i++) {
2482                 if (!p[i].list_size)
2483                         continue;
2484                 pthread_mutex_init(&p[i].mutex, NULL);
2485                 pthread_cond_init(&p[i].cond, NULL);
2486                 ret = pthread_create(&p[i].thread, NULL,
2487                                      threaded_find_deltas, &p[i]);
2488                 if (ret)
2489                         die(_("unable to create thread: %s"), strerror(ret));
2490                 active_threads++;
2491         }
2492
2493         /*
2494          * Now let's wait for work completion.  Each time a thread is done
2495          * with its work, we steal half of the remaining work from the
2496          * thread with the largest number of unprocessed objects and give
2497          * it to that newly idle thread.  This ensure good load balancing
2498          * until the remaining object list segments are simply too short
2499          * to be worth splitting anymore.
2500          */
2501         while (active_threads) {
2502                 struct thread_params *target = NULL;
2503                 struct thread_params *victim = NULL;
2504                 unsigned sub_size = 0;
2505
2506                 progress_lock();
2507                 for (;;) {
2508                         for (i = 0; !target && i < delta_search_threads; i++)
2509                                 if (!p[i].working)
2510                                         target = &p[i];
2511                         if (target)
2512                                 break;
2513                         pthread_cond_wait(&progress_cond, &progress_mutex);
2514                 }
2515
2516                 for (i = 0; i < delta_search_threads; i++)
2517                         if (p[i].remaining > 2*window &&
2518                             (!victim || victim->remaining < p[i].remaining))
2519                                 victim = &p[i];
2520                 if (victim) {
2521                         sub_size = victim->remaining / 2;
2522                         list = victim->list + victim->list_size - sub_size;
2523                         while (sub_size && list[0]->hash &&
2524                                list[0]->hash == list[-1]->hash) {
2525                                 list++;
2526                                 sub_size--;
2527                         }
2528                         if (!sub_size) {
2529                                 /*
2530                                  * It is possible for some "paths" to have
2531                                  * so many objects that no hash boundary
2532                                  * might be found.  Let's just steal the
2533                                  * exact half in that case.
2534                                  */
2535                                 sub_size = victim->remaining / 2;
2536                                 list -= sub_size;
2537                         }
2538                         target->list = list;
2539                         victim->list_size -= sub_size;
2540                         victim->remaining -= sub_size;
2541                 }
2542                 target->list_size = sub_size;
2543                 target->remaining = sub_size;
2544                 target->working = 1;
2545                 progress_unlock();
2546
2547                 pthread_mutex_lock(&target->mutex);
2548                 target->data_ready = 1;
2549                 pthread_cond_signal(&target->cond);
2550                 pthread_mutex_unlock(&target->mutex);
2551
2552                 if (!sub_size) {
2553                         pthread_join(target->thread, NULL);
2554                         pthread_cond_destroy(&target->cond);
2555                         pthread_mutex_destroy(&target->mutex);
2556                         active_threads--;
2557                 }
2558         }
2559         cleanup_threaded_search();
2560         free(p);
2561 }
2562
2563 static void add_tag_chain(const struct object_id *oid)
2564 {
2565         struct tag *tag;
2566
2567         /*
2568          * We catch duplicates already in add_object_entry(), but we'd
2569          * prefer to do this extra check to avoid having to parse the
2570          * tag at all if we already know that it's being packed (e.g., if
2571          * it was included via bitmaps, we would not have parsed it
2572          * previously).
2573          */
2574         if (packlist_find(&to_pack, oid->hash, NULL))
2575                 return;
2576
2577         tag = lookup_tag(the_repository, oid);
2578         while (1) {
2579                 if (!tag || parse_tag(tag) || !tag->tagged)
2580                         die(_("unable to pack objects reachable from tag %s"),
2581                             oid_to_hex(oid));
2582
2583                 add_object_entry(&tag->object.oid, OBJ_TAG, NULL, 0);
2584
2585                 if (tag->tagged->type != OBJ_TAG)
2586                         return;
2587
2588                 tag = (struct tag *)tag->tagged;
2589         }
2590 }
2591
2592 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2593 {
2594         struct object_id peeled;
2595
2596         if (starts_with(path, "refs/tags/") && /* is a tag? */
2597             !peel_ref(path, &peeled)    && /* peelable? */
2598             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2599                 add_tag_chain(oid);
2600         return 0;
2601 }
2602
2603 static void prepare_pack(int window, int depth)
2604 {
2605         struct object_entry **delta_list;
2606         uint32_t i, nr_deltas;
2607         unsigned n;
2608
2609         if (use_delta_islands)
2610                 resolve_tree_islands(the_repository, progress, &to_pack);
2611
2612         get_object_details();
2613
2614         /*
2615          * If we're locally repacking then we need to be doubly careful
2616          * from now on in order to make sure no stealth corruption gets
2617          * propagated to the new pack.  Clients receiving streamed packs
2618          * should validate everything they get anyway so no need to incur
2619          * the additional cost here in that case.
2620          */
2621         if (!pack_to_stdout)
2622                 do_check_packed_object_crc = 1;
2623
2624         if (!to_pack.nr_objects || !window || !depth)
2625                 return;
2626
2627         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2628         nr_deltas = n = 0;
2629
2630         for (i = 0; i < to_pack.nr_objects; i++) {
2631                 struct object_entry *entry = to_pack.objects + i;
2632
2633                 if (DELTA(entry))
2634                         /* This happens if we decided to reuse existing
2635                          * delta from a pack.  "reuse_delta &&" is implied.
2636                          */
2637                         continue;
2638
2639                 if (!entry->type_valid ||
2640                     oe_size_less_than(&to_pack, entry, 50))
2641                         continue;
2642
2643                 if (entry->no_try_delta)
2644                         continue;
2645
2646                 if (!entry->preferred_base) {
2647                         nr_deltas++;
2648                         if (oe_type(entry) < 0)
2649                                 die(_("unable to get type of object %s"),
2650                                     oid_to_hex(&entry->idx.oid));
2651                 } else {
2652                         if (oe_type(entry) < 0) {
2653                                 /*
2654                                  * This object is not found, but we
2655                                  * don't have to include it anyway.
2656                                  */
2657                                 continue;
2658                         }
2659                 }
2660
2661                 delta_list[n++] = entry;
2662         }
2663
2664         if (nr_deltas && n > 1) {
2665                 unsigned nr_done = 0;
2666                 if (progress)
2667                         progress_state = start_progress(_("Compressing objects"),
2668                                                         nr_deltas);
2669                 QSORT(delta_list, n, type_size_sort);
2670                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2671                 stop_progress(&progress_state);
2672                 if (nr_done != nr_deltas)
2673                         die(_("inconsistency with delta count"));
2674         }
2675         free(delta_list);
2676 }
2677
2678 static int git_pack_config(const char *k, const char *v, void *cb)
2679 {
2680         if (!strcmp(k, "pack.window")) {
2681                 window = git_config_int(k, v);
2682                 return 0;
2683         }
2684         if (!strcmp(k, "pack.windowmemory")) {
2685                 window_memory_limit = git_config_ulong(k, v);
2686                 return 0;
2687         }
2688         if (!strcmp(k, "pack.depth")) {
2689                 depth = git_config_int(k, v);
2690                 return 0;
2691         }
2692         if (!strcmp(k, "pack.deltacachesize")) {
2693                 max_delta_cache_size = git_config_int(k, v);
2694                 return 0;
2695         }
2696         if (!strcmp(k, "pack.deltacachelimit")) {
2697                 cache_max_small_delta_size = git_config_int(k, v);
2698                 return 0;
2699         }
2700         if (!strcmp(k, "pack.writebitmaphashcache")) {
2701                 if (git_config_bool(k, v))
2702                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2703                 else
2704                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2705         }
2706         if (!strcmp(k, "pack.usebitmaps")) {
2707                 use_bitmap_index_default = git_config_bool(k, v);
2708                 return 0;
2709         }
2710         if (!strcmp(k, "pack.usesparse")) {
2711                 sparse = git_config_bool(k, v);
2712                 return 0;
2713         }
2714         if (!strcmp(k, "pack.threads")) {
2715                 delta_search_threads = git_config_int(k, v);
2716                 if (delta_search_threads < 0)
2717                         die(_("invalid number of threads specified (%d)"),
2718                             delta_search_threads);
2719                 if (!HAVE_THREADS && delta_search_threads != 1) {
2720                         warning(_("no threads support, ignoring %s"), k);
2721                         delta_search_threads = 0;
2722                 }
2723                 return 0;
2724         }
2725         if (!strcmp(k, "pack.indexversion")) {
2726                 pack_idx_opts.version = git_config_int(k, v);
2727                 if (pack_idx_opts.version > 2)
2728                         die(_("bad pack.indexversion=%"PRIu32),
2729                             pack_idx_opts.version);
2730                 return 0;
2731         }
2732         return git_default_config(k, v, cb);
2733 }
2734
2735 static void read_object_list_from_stdin(void)
2736 {
2737         char line[GIT_MAX_HEXSZ + 1 + PATH_MAX + 2];
2738         struct object_id oid;
2739         const char *p;
2740
2741         for (;;) {
2742                 if (!fgets(line, sizeof(line), stdin)) {
2743                         if (feof(stdin))
2744                                 break;
2745                         if (!ferror(stdin))
2746                                 die("BUG: fgets returned NULL, not EOF, not error!");
2747                         if (errno != EINTR)
2748                                 die_errno("fgets");
2749                         clearerr(stdin);
2750                         continue;
2751                 }
2752                 if (line[0] == '-') {
2753                         if (get_oid_hex(line+1, &oid))
2754                                 die(_("expected edge object ID, got garbage:\n %s"),
2755                                     line);
2756                         add_preferred_base(&oid);
2757                         continue;
2758                 }
2759                 if (parse_oid_hex(line, &oid, &p))
2760                         die(_("expected object ID, got garbage:\n %s"), line);
2761
2762                 add_preferred_base_object(p + 1);
2763                 add_object_entry(&oid, OBJ_NONE, p + 1, 0);
2764         }
2765 }
2766
2767 /* Remember to update object flag allocation in object.h */
2768 #define OBJECT_ADDED (1u<<20)
2769
2770 static void show_commit(struct commit *commit, void *data)
2771 {
2772         add_object_entry(&commit->object.oid, OBJ_COMMIT, NULL, 0);
2773         commit->object.flags |= OBJECT_ADDED;
2774
2775         if (write_bitmap_index)
2776                 index_commit_for_bitmap(commit);
2777
2778         if (use_delta_islands)
2779                 propagate_island_marks(commit);
2780 }
2781
2782 static void show_object(struct object *obj, const char *name, void *data)
2783 {
2784         add_preferred_base_object(name);
2785         add_object_entry(&obj->oid, obj->type, name, 0);
2786         obj->flags |= OBJECT_ADDED;
2787
2788         if (use_delta_islands) {
2789                 const char *p;
2790                 unsigned depth;
2791                 struct object_entry *ent;
2792
2793                 /* the empty string is a root tree, which is depth 0 */
2794                 depth = *name ? 1 : 0;
2795                 for (p = strchr(name, '/'); p; p = strchr(p + 1, '/'))
2796                         depth++;
2797
2798                 ent = packlist_find(&to_pack, obj->oid.hash, NULL);
2799                 if (ent && depth > oe_tree_depth(&to_pack, ent))
2800                         oe_set_tree_depth(&to_pack, ent, depth);
2801         }
2802 }
2803
2804 static void show_object__ma_allow_any(struct object *obj, const char *name, void *data)
2805 {
2806         assert(arg_missing_action == MA_ALLOW_ANY);
2807
2808         /*
2809          * Quietly ignore ALL missing objects.  This avoids problems with
2810          * staging them now and getting an odd error later.
2811          */
2812         if (!has_object_file(&obj->oid))
2813                 return;
2814
2815         show_object(obj, name, data);
2816 }
2817
2818 static void show_object__ma_allow_promisor(struct object *obj, const char *name, void *data)
2819 {
2820         assert(arg_missing_action == MA_ALLOW_PROMISOR);
2821
2822         /*
2823          * Quietly ignore EXPECTED missing objects.  This avoids problems with
2824          * staging them now and getting an odd error later.
2825          */
2826         if (!has_object_file(&obj->oid) && is_promisor_object(&obj->oid))
2827                 return;
2828
2829         show_object(obj, name, data);
2830 }
2831
2832 static int option_parse_missing_action(const struct option *opt,
2833                                        const char *arg, int unset)
2834 {
2835         assert(arg);
2836         assert(!unset);
2837
2838         if (!strcmp(arg, "error")) {
2839                 arg_missing_action = MA_ERROR;
2840                 fn_show_object = show_object;
2841                 return 0;
2842         }
2843
2844         if (!strcmp(arg, "allow-any")) {
2845                 arg_missing_action = MA_ALLOW_ANY;
2846                 fetch_if_missing = 0;
2847                 fn_show_object = show_object__ma_allow_any;
2848                 return 0;
2849         }
2850
2851         if (!strcmp(arg, "allow-promisor")) {
2852                 arg_missing_action = MA_ALLOW_PROMISOR;
2853                 fetch_if_missing = 0;
2854                 fn_show_object = show_object__ma_allow_promisor;
2855                 return 0;
2856         }
2857
2858         die(_("invalid value for --missing"));
2859         return 0;
2860 }
2861
2862 static void show_edge(struct commit *commit)
2863 {
2864         add_preferred_base(&commit->object.oid);
2865 }
2866
2867 struct in_pack_object {
2868         off_t offset;
2869         struct object *object;
2870 };
2871
2872 struct in_pack {
2873         unsigned int alloc;
2874         unsigned int nr;
2875         struct in_pack_object *array;
2876 };
2877
2878 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2879 {
2880         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2881         in_pack->array[in_pack->nr].object = object;
2882         in_pack->nr++;
2883 }
2884
2885 /*
2886  * Compare the objects in the offset order, in order to emulate the
2887  * "git rev-list --objects" output that produced the pack originally.
2888  */
2889 static int ofscmp(const void *a_, const void *b_)
2890 {
2891         struct in_pack_object *a = (struct in_pack_object *)a_;
2892         struct in_pack_object *b = (struct in_pack_object *)b_;
2893
2894         if (a->offset < b->offset)
2895                 return -1;
2896         else if (a->offset > b->offset)
2897                 return 1;
2898         else
2899                 return oidcmp(&a->object->oid, &b->object->oid);
2900 }
2901
2902 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2903 {
2904         struct packed_git *p;
2905         struct in_pack in_pack;
2906         uint32_t i;
2907
2908         memset(&in_pack, 0, sizeof(in_pack));
2909
2910         for (p = get_all_packs(the_repository); p; p = p->next) {
2911                 struct object_id oid;
2912                 struct object *o;
2913
2914                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
2915                         continue;
2916                 if (open_pack_index(p))
2917                         die(_("cannot open pack index"));
2918
2919                 ALLOC_GROW(in_pack.array,
2920                            in_pack.nr + p->num_objects,
2921                            in_pack.alloc);
2922
2923                 for (i = 0; i < p->num_objects; i++) {
2924                         nth_packed_object_oid(&oid, p, i);
2925                         o = lookup_unknown_object(oid.hash);
2926                         if (!(o->flags & OBJECT_ADDED))
2927                                 mark_in_pack_object(o, p, &in_pack);
2928                         o->flags |= OBJECT_ADDED;
2929                 }
2930         }
2931
2932         if (in_pack.nr) {
2933                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2934                 for (i = 0; i < in_pack.nr; i++) {
2935                         struct object *o = in_pack.array[i].object;
2936                         add_object_entry(&o->oid, o->type, "", 0);
2937                 }
2938         }
2939         free(in_pack.array);
2940 }
2941
2942 static int add_loose_object(const struct object_id *oid, const char *path,
2943                             void *data)
2944 {
2945         enum object_type type = oid_object_info(the_repository, oid, NULL);
2946
2947         if (type < 0) {
2948                 warning(_("loose object at %s could not be examined"), path);
2949                 return 0;
2950         }
2951
2952         add_object_entry(oid, type, "", 0);
2953         return 0;
2954 }
2955
2956 /*
2957  * We actually don't even have to worry about reachability here.
2958  * add_object_entry will weed out duplicates, so we just add every
2959  * loose object we find.
2960  */
2961 static void add_unreachable_loose_objects(void)
2962 {
2963         for_each_loose_file_in_objdir(get_object_directory(),
2964                                       add_loose_object,
2965                                       NULL, NULL, NULL);
2966 }
2967
2968 static int has_sha1_pack_kept_or_nonlocal(const struct object_id *oid)
2969 {
2970         static struct packed_git *last_found = (void *)1;
2971         struct packed_git *p;
2972
2973         p = (last_found != (void *)1) ? last_found :
2974                                         get_all_packs(the_repository);
2975
2976         while (p) {
2977                 if ((!p->pack_local || p->pack_keep ||
2978                                 p->pack_keep_in_core) &&
2979                         find_pack_entry_one(oid->hash, p)) {
2980                         last_found = p;
2981                         return 1;
2982                 }
2983                 if (p == last_found)
2984                         p = get_all_packs(the_repository);
2985                 else
2986                         p = p->next;
2987                 if (p == last_found)
2988                         p = p->next;
2989         }
2990         return 0;
2991 }
2992
2993 /*
2994  * Store a list of sha1s that are should not be discarded
2995  * because they are either written too recently, or are
2996  * reachable from another object that was.
2997  *
2998  * This is filled by get_object_list.
2999  */
3000 static struct oid_array recent_objects;
3001
3002 static int loosened_object_can_be_discarded(const struct object_id *oid,
3003                                             timestamp_t mtime)
3004 {
3005         if (!unpack_unreachable_expiration)
3006                 return 0;
3007         if (mtime > unpack_unreachable_expiration)
3008                 return 0;
3009         if (oid_array_lookup(&recent_objects, oid) >= 0)
3010                 return 0;
3011         return 1;
3012 }
3013
3014 static void loosen_unused_packed_objects(struct rev_info *revs)
3015 {
3016         struct packed_git *p;
3017         uint32_t i;
3018         struct object_id oid;
3019
3020         for (p = get_all_packs(the_repository); p; p = p->next) {
3021                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
3022                         continue;
3023
3024                 if (open_pack_index(p))
3025                         die(_("cannot open pack index"));
3026
3027                 for (i = 0; i < p->num_objects; i++) {
3028                         nth_packed_object_oid(&oid, p, i);
3029                         if (!packlist_find(&to_pack, oid.hash, NULL) &&
3030                             !has_sha1_pack_kept_or_nonlocal(&oid) &&
3031                             !loosened_object_can_be_discarded(&oid, p->mtime))
3032                                 if (force_object_loose(&oid, p->mtime))
3033                                         die(_("unable to force loose object"));
3034                 }
3035         }
3036 }
3037
3038 /*
3039  * This tracks any options which pack-reuse code expects to be on, or which a
3040  * reader of the pack might not understand, and which would therefore prevent
3041  * blind reuse of what we have on disk.
3042  */
3043 static int pack_options_allow_reuse(void)
3044 {
3045         return pack_to_stdout &&
3046                allow_ofs_delta &&
3047                !ignore_packed_keep_on_disk &&
3048                !ignore_packed_keep_in_core &&
3049                (!local || !have_non_local_packs) &&
3050                !incremental;
3051 }
3052
3053 static int get_object_list_from_bitmap(struct rev_info *revs)
3054 {
3055         if (!(bitmap_git = prepare_bitmap_walk(revs)))
3056                 return -1;
3057
3058         if (pack_options_allow_reuse() &&
3059             !reuse_partial_packfile_from_bitmap(
3060                         bitmap_git,
3061                         &reuse_packfile,
3062                         &reuse_packfile_objects,
3063                         &reuse_packfile_offset)) {
3064                 assert(reuse_packfile_objects);
3065                 nr_result += reuse_packfile_objects;
3066                 display_progress(progress_state, nr_result);
3067         }
3068
3069         traverse_bitmap_commit_list(bitmap_git, &add_object_entry_from_bitmap);
3070         return 0;
3071 }
3072
3073 static void record_recent_object(struct object *obj,
3074                                  const char *name,
3075                                  void *data)
3076 {
3077         oid_array_append(&recent_objects, &obj->oid);
3078 }
3079
3080 static void record_recent_commit(struct commit *commit, void *data)
3081 {
3082         oid_array_append(&recent_objects, &commit->object.oid);
3083 }
3084
3085 static void get_object_list(int ac, const char **av)
3086 {
3087         struct rev_info revs;
3088         struct setup_revision_opt s_r_opt = {
3089                 .allow_exclude_promisor_objects = 1,
3090         };
3091         char line[1000];
3092         int flags = 0;
3093         int save_warning;
3094
3095         repo_init_revisions(the_repository, &revs, NULL);
3096         save_commit_buffer = 0;
3097         setup_revisions(ac, av, &revs, &s_r_opt);
3098
3099         /* make sure shallows are read */
3100         is_repository_shallow(the_repository);
3101
3102         save_warning = warn_on_object_refname_ambiguity;
3103         warn_on_object_refname_ambiguity = 0;
3104
3105         while (fgets(line, sizeof(line), stdin) != NULL) {
3106                 int len = strlen(line);
3107                 if (len && line[len - 1] == '\n')
3108                         line[--len] = 0;
3109                 if (!len)
3110                         break;
3111                 if (*line == '-') {
3112                         if (!strcmp(line, "--not")) {
3113                                 flags ^= UNINTERESTING;
3114                                 write_bitmap_index = 0;
3115                                 continue;
3116                         }
3117                         if (starts_with(line, "--shallow ")) {
3118                                 struct object_id oid;
3119                                 if (get_oid_hex(line + 10, &oid))
3120                                         die("not an SHA-1 '%s'", line + 10);
3121                                 register_shallow(the_repository, &oid);
3122                                 use_bitmap_index = 0;
3123                                 continue;
3124                         }
3125                         die(_("not a rev '%s'"), line);
3126                 }
3127                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
3128                         die(_("bad revision '%s'"), line);
3129         }
3130
3131         warn_on_object_refname_ambiguity = save_warning;
3132
3133         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
3134                 return;
3135
3136         if (use_delta_islands)
3137                 load_delta_islands(the_repository);
3138
3139         if (prepare_revision_walk(&revs))
3140                 die(_("revision walk setup failed"));
3141         mark_edges_uninteresting(&revs, show_edge, sparse);
3142
3143         if (!fn_show_object)
3144                 fn_show_object = show_object;
3145         traverse_commit_list_filtered(&filter_options, &revs,
3146                                       show_commit, fn_show_object, NULL,
3147                                       NULL);
3148
3149         if (unpack_unreachable_expiration) {
3150                 revs.ignore_missing_links = 1;
3151                 if (add_unseen_recent_objects_to_traversal(&revs,
3152                                 unpack_unreachable_expiration))
3153                         die(_("unable to add recent objects"));
3154                 if (prepare_revision_walk(&revs))
3155                         die(_("revision walk setup failed"));
3156                 traverse_commit_list(&revs, record_recent_commit,
3157                                      record_recent_object, NULL);
3158         }
3159
3160         if (keep_unreachable)
3161                 add_objects_in_unpacked_packs(&revs);
3162         if (pack_loose_unreachable)
3163                 add_unreachable_loose_objects();
3164         if (unpack_unreachable)
3165                 loosen_unused_packed_objects(&revs);
3166
3167         oid_array_clear(&recent_objects);
3168 }
3169
3170 static void add_extra_kept_packs(const struct string_list *names)
3171 {
3172         struct packed_git *p;
3173
3174         if (!names->nr)
3175                 return;
3176
3177         for (p = get_all_packs(the_repository); p; p = p->next) {
3178                 const char *name = basename(p->pack_name);
3179                 int i;
3180
3181                 if (!p->pack_local)
3182                         continue;
3183
3184                 for (i = 0; i < names->nr; i++)
3185                         if (!fspathcmp(name, names->items[i].string))
3186                                 break;
3187
3188                 if (i < names->nr) {
3189                         p->pack_keep_in_core = 1;
3190                         ignore_packed_keep_in_core = 1;
3191                         continue;
3192                 }
3193         }
3194 }
3195
3196 static int option_parse_index_version(const struct option *opt,
3197                                       const char *arg, int unset)
3198 {
3199         char *c;
3200         const char *val = arg;
3201
3202         BUG_ON_OPT_NEG(unset);
3203
3204         pack_idx_opts.version = strtoul(val, &c, 10);
3205         if (pack_idx_opts.version > 2)
3206                 die(_("unsupported index version %s"), val);
3207         if (*c == ',' && c[1])
3208                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
3209         if (*c || pack_idx_opts.off32_limit & 0x80000000)
3210                 die(_("bad index version '%s'"), val);
3211         return 0;
3212 }
3213
3214 static int option_parse_unpack_unreachable(const struct option *opt,
3215                                            const char *arg, int unset)
3216 {
3217         if (unset) {
3218                 unpack_unreachable = 0;
3219                 unpack_unreachable_expiration = 0;
3220         }
3221         else {
3222                 unpack_unreachable = 1;
3223                 if (arg)
3224                         unpack_unreachable_expiration = approxidate(arg);
3225         }
3226         return 0;
3227 }
3228
3229 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
3230 {
3231         int use_internal_rev_list = 0;
3232         int shallow = 0;
3233         int all_progress_implied = 0;
3234         struct argv_array rp = ARGV_ARRAY_INIT;
3235         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
3236         int rev_list_index = 0;
3237         struct string_list keep_pack_list = STRING_LIST_INIT_NODUP;
3238         struct option pack_objects_options[] = {
3239                 OPT_SET_INT('q', "quiet", &progress,
3240                             N_("do not show progress meter"), 0),
3241                 OPT_SET_INT(0, "progress", &progress,
3242                             N_("show progress meter"), 1),
3243                 OPT_SET_INT(0, "all-progress", &progress,
3244                             N_("show progress meter during object writing phase"), 2),
3245                 OPT_BOOL(0, "all-progress-implied",
3246                          &all_progress_implied,
3247                          N_("similar to --all-progress when progress meter is shown")),
3248                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("<version>[,<offset>]"),
3249                   N_("write the pack index file in the specified idx format version"),
3250                   PARSE_OPT_NONEG, option_parse_index_version },
3251                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
3252                               N_("maximum size of each output pack file")),
3253                 OPT_BOOL(0, "local", &local,
3254                          N_("ignore borrowed objects from alternate object store")),
3255                 OPT_BOOL(0, "incremental", &incremental,
3256                          N_("ignore packed objects")),
3257                 OPT_INTEGER(0, "window", &window,
3258                             N_("limit pack window by objects")),
3259                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
3260                               N_("limit pack window by memory in addition to object limit")),
3261                 OPT_INTEGER(0, "depth", &depth,
3262                             N_("maximum length of delta chain allowed in the resulting pack")),
3263                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
3264                          N_("reuse existing deltas")),
3265                 OPT_BOOL(0, "reuse-object", &reuse_object,
3266                          N_("reuse existing objects")),
3267                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
3268                          N_("use OFS_DELTA objects")),
3269                 OPT_INTEGER(0, "threads", &delta_search_threads,
3270                             N_("use threads when searching for best delta matches")),
3271                 OPT_BOOL(0, "non-empty", &non_empty,
3272                          N_("do not create an empty pack output")),
3273                 OPT_BOOL(0, "revs", &use_internal_rev_list,
3274                          N_("read revision arguments from standard input")),
3275                 OPT_SET_INT_F(0, "unpacked", &rev_list_unpacked,
3276                               N_("limit the objects to those that are not yet packed"),
3277                               1, PARSE_OPT_NONEG),
3278                 OPT_SET_INT_F(0, "all", &rev_list_all,
3279                               N_("include objects reachable from any reference"),
3280                               1, PARSE_OPT_NONEG),
3281                 OPT_SET_INT_F(0, "reflog", &rev_list_reflog,
3282                               N_("include objects referred by reflog entries"),
3283                               1, PARSE_OPT_NONEG),
3284                 OPT_SET_INT_F(0, "indexed-objects", &rev_list_index,
3285                               N_("include objects referred to by the index"),
3286                               1, PARSE_OPT_NONEG),
3287                 OPT_BOOL(0, "stdout", &pack_to_stdout,
3288                          N_("output pack to stdout")),
3289                 OPT_BOOL(0, "include-tag", &include_tag,
3290                          N_("include tag objects that refer to objects to be packed")),
3291                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
3292                          N_("keep unreachable objects")),
3293                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
3294                          N_("pack loose unreachable objects")),
3295                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
3296                   N_("unpack unreachable objects newer than <time>"),
3297                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
3298                 OPT_BOOL(0, "sparse", &sparse,
3299                          N_("use the sparse reachability algorithm")),
3300                 OPT_BOOL(0, "thin", &thin,
3301                          N_("create thin packs")),
3302                 OPT_BOOL(0, "shallow", &shallow,
3303                          N_("create packs suitable for shallow fetches")),
3304                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep_on_disk,
3305                          N_("ignore packs that have companion .keep file")),
3306                 OPT_STRING_LIST(0, "keep-pack", &keep_pack_list, N_("name"),
3307                                 N_("ignore this pack")),
3308                 OPT_INTEGER(0, "compression", &pack_compression_level,
3309                             N_("pack compression level")),
3310                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
3311                             N_("do not hide commits by grafts"), 0),
3312                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
3313                          N_("use a bitmap index if available to speed up counting objects")),
3314                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
3315                          N_("write a bitmap index together with the pack index")),
3316                 OPT_PARSE_LIST_OBJECTS_FILTER(&filter_options),
3317                 { OPTION_CALLBACK, 0, "missing", NULL, N_("action"),
3318                   N_("handling for missing objects"), PARSE_OPT_NONEG,
3319                   option_parse_missing_action },
3320                 OPT_BOOL(0, "exclude-promisor-objects", &exclude_promisor_objects,
3321                          N_("do not pack objects in promisor packfiles")),
3322                 OPT_BOOL(0, "delta-islands", &use_delta_islands,
3323                          N_("respect islands during delta compression")),
3324                 OPT_END(),
3325         };
3326
3327         if (DFS_NUM_STATES > (1 << OE_DFS_STATE_BITS))
3328                 BUG("too many dfs states, increase OE_DFS_STATE_BITS");
3329
3330         read_replace_refs = 0;
3331
3332         sparse = git_env_bool("GIT_TEST_PACK_SPARSE", 0);
3333         reset_pack_idx_option(&pack_idx_opts);
3334         git_config(git_pack_config, NULL);
3335
3336         progress = isatty(2);
3337         argc = parse_options(argc, argv, prefix, pack_objects_options,
3338                              pack_usage, 0);
3339
3340         if (argc) {
3341                 base_name = argv[0];
3342                 argc--;
3343         }
3344         if (pack_to_stdout != !base_name || argc)
3345                 usage_with_options(pack_usage, pack_objects_options);
3346
3347         if (depth >= (1 << OE_DEPTH_BITS)) {
3348                 warning(_("delta chain depth %d is too deep, forcing %d"),
3349                         depth, (1 << OE_DEPTH_BITS) - 1);
3350                 depth = (1 << OE_DEPTH_BITS) - 1;
3351         }
3352         if (cache_max_small_delta_size >= (1U << OE_Z_DELTA_BITS)) {
3353                 warning(_("pack.deltaCacheLimit is too high, forcing %d"),
3354                         (1U << OE_Z_DELTA_BITS) - 1);
3355                 cache_max_small_delta_size = (1U << OE_Z_DELTA_BITS) - 1;
3356         }
3357
3358         argv_array_push(&rp, "pack-objects");
3359         if (thin) {
3360                 use_internal_rev_list = 1;
3361                 argv_array_push(&rp, shallow
3362                                 ? "--objects-edge-aggressive"
3363                                 : "--objects-edge");
3364         } else
3365                 argv_array_push(&rp, "--objects");
3366
3367         if (rev_list_all) {
3368                 use_internal_rev_list = 1;
3369                 argv_array_push(&rp, "--all");
3370         }
3371         if (rev_list_reflog) {
3372                 use_internal_rev_list = 1;
3373                 argv_array_push(&rp, "--reflog");
3374         }
3375         if (rev_list_index) {
3376                 use_internal_rev_list = 1;
3377                 argv_array_push(&rp, "--indexed-objects");
3378         }
3379         if (rev_list_unpacked) {
3380                 use_internal_rev_list = 1;
3381                 argv_array_push(&rp, "--unpacked");
3382         }
3383
3384         if (exclude_promisor_objects) {
3385                 use_internal_rev_list = 1;
3386                 fetch_if_missing = 0;
3387                 argv_array_push(&rp, "--exclude-promisor-objects");
3388         }
3389         if (unpack_unreachable || keep_unreachable || pack_loose_unreachable)
3390                 use_internal_rev_list = 1;
3391
3392         if (!reuse_object)
3393                 reuse_delta = 0;
3394         if (pack_compression_level == -1)
3395                 pack_compression_level = Z_DEFAULT_COMPRESSION;
3396         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
3397                 die(_("bad pack compression level %d"), pack_compression_level);
3398
3399         if (!delta_search_threads)      /* --threads=0 means autodetect */
3400                 delta_search_threads = online_cpus();
3401
3402         if (!HAVE_THREADS && delta_search_threads != 1)
3403                 warning(_("no threads support, ignoring --threads"));
3404         if (!pack_to_stdout && !pack_size_limit)
3405                 pack_size_limit = pack_size_limit_cfg;
3406         if (pack_to_stdout && pack_size_limit)
3407                 die(_("--max-pack-size cannot be used to build a pack for transfer"));
3408         if (pack_size_limit && pack_size_limit < 1024*1024) {
3409                 warning(_("minimum pack size limit is 1 MiB"));
3410                 pack_size_limit = 1024*1024;
3411         }
3412
3413         if (!pack_to_stdout && thin)
3414                 die(_("--thin cannot be used to build an indexable pack"));
3415
3416         if (keep_unreachable && unpack_unreachable)
3417                 die(_("--keep-unreachable and --unpack-unreachable are incompatible"));
3418         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3419                 unpack_unreachable_expiration = 0;
3420
3421         if (filter_options.choice) {
3422                 if (!pack_to_stdout)
3423                         die(_("cannot use --filter without --stdout"));
3424                 use_bitmap_index = 0;
3425         }
3426
3427         /*
3428          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3429          *
3430          * - to produce good pack (with bitmap index not-yet-packed objects are
3431          *   packed in suboptimal order).
3432          *
3433          * - to use more robust pack-generation codepath (avoiding possible
3434          *   bugs in bitmap code and possible bitmap index corruption).
3435          */
3436         if (!pack_to_stdout)
3437                 use_bitmap_index_default = 0;
3438
3439         if (use_bitmap_index < 0)
3440                 use_bitmap_index = use_bitmap_index_default;
3441
3442         /* "hard" reasons not to use bitmaps; these just won't work at all */
3443         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow(the_repository))
3444                 use_bitmap_index = 0;
3445
3446         if (pack_to_stdout || !rev_list_all)
3447                 write_bitmap_index = 0;
3448
3449         if (use_delta_islands)
3450                 argv_array_push(&rp, "--topo-order");
3451
3452         if (progress && all_progress_implied)
3453                 progress = 2;
3454
3455         add_extra_kept_packs(&keep_pack_list);
3456         if (ignore_packed_keep_on_disk) {
3457                 struct packed_git *p;
3458                 for (p = get_all_packs(the_repository); p; p = p->next)
3459                         if (p->pack_local && p->pack_keep)
3460                                 break;
3461                 if (!p) /* no keep-able packs found */
3462                         ignore_packed_keep_on_disk = 0;
3463         }
3464         if (local) {
3465                 /*
3466                  * unlike ignore_packed_keep_on_disk above, we do not
3467                  * want to unset "local" based on looking at packs, as
3468                  * it also covers non-local objects
3469                  */
3470                 struct packed_git *p;
3471                 for (p = get_all_packs(the_repository); p; p = p->next) {
3472                         if (!p->pack_local) {
3473                                 have_non_local_packs = 1;
3474                                 break;
3475                         }
3476                 }
3477         }
3478
3479         trace2_region_enter("pack-objects", "enumerate-objects",
3480                             the_repository);
3481         prepare_packing_data(the_repository, &to_pack);
3482
3483         if (progress)
3484                 progress_state = start_progress(_("Enumerating objects"), 0);
3485         if (!use_internal_rev_list)
3486                 read_object_list_from_stdin();
3487         else {
3488                 get_object_list(rp.argc, rp.argv);
3489                 argv_array_clear(&rp);
3490         }
3491         cleanup_preferred_base();
3492         if (include_tag && nr_result)
3493                 for_each_ref(add_ref_tag, NULL);
3494         stop_progress(&progress_state);
3495         trace2_region_leave("pack-objects", "enumerate-objects",
3496                             the_repository);
3497
3498         if (non_empty && !nr_result)
3499                 return 0;
3500         if (nr_result) {
3501                 trace2_region_enter("pack-objects", "prepare-pack",
3502                                     the_repository);
3503                 prepare_pack(window, depth);
3504                 trace2_region_leave("pack-objects", "prepare-pack",
3505                                     the_repository);
3506         }
3507
3508         trace2_region_enter("pack-objects", "write-pack-file", the_repository);
3509         write_pack_file();
3510         trace2_region_leave("pack-objects", "write-pack-file", the_repository);
3511
3512         if (progress)
3513                 fprintf_ln(stderr,
3514                            _("Total %"PRIu32" (delta %"PRIu32"),"
3515                              " reused %"PRIu32" (delta %"PRIu32")"),
3516                            written, written_delta, reused, reused_delta);
3517         return 0;
3518 }