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 "oid-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 #include "shallow.h"
  38
  39 #define IN_PACK(obj) oe_in_pack(&to_pack, obj)
  40 #define SIZE(obj) oe_size(&to_pack, obj)
  41 #define SET_SIZE(obj,size) oe_set_size(&to_pack, obj, size)
  42 #define DELTA_SIZE(obj) oe_delta_size(&to_pack, obj)
  43 #define DELTA(obj) oe_delta(&to_pack, obj)
  44 #define DELTA_CHILD(obj) oe_delta_child(&to_pack, obj)
  45 #define DELTA_SIBLING(obj) oe_delta_sibling(&to_pack, obj)
  46 #define SET_DELTA(obj, val) oe_set_delta(&to_pack, obj, val)
  47 #define SET_DELTA_EXT(obj, oid) oe_set_delta_ext(&to_pack, obj, oid)
  48 #define SET_DELTA_SIZE(obj, val) oe_set_delta_size(&to_pack, obj, val)
  49 #define SET_DELTA_CHILD(obj, val) oe_set_delta_child(&to_pack, obj, val)
  50 #define SET_DELTA_SIBLING(obj, val) oe_set_delta_sibling(&to_pack, obj, val)
  51
  52 static const char *pack_usage[] = {
  53         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
  54         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
  55         NULL
  56 };
  57
  58 /*
  59  * Objects we are going to pack are collected in the `to_pack` structure.
  60  * It contains an array (dynamically expanded) of the object data, and a map
  61  * that can resolve SHA1s to their position in the array.
  62  */
  63 static struct packing_data to_pack;
  64
  65 static struct pack_idx_entry **written_list;
  66 static uint32_t nr_result, nr_written, nr_seen;
  67 static struct bitmap_index *bitmap_git;
  68 static uint32_t write_layer;
  69
  70 static int non_empty;
  71 static int reuse_delta = 1, reuse_object = 1;
  72 static int keep_unreachable, unpack_unreachable, include_tag;
  73 static timestamp_t unpack_unreachable_expiration;
  74 static int pack_loose_unreachable;
  75 static int local;
  76 static int have_non_local_packs;
  77 static int incremental;
  78 static int ignore_packed_keep_on_disk;
  79 static int ignore_packed_keep_in_core;
  80 static int allow_ofs_delta;
  81 static struct pack_idx_option pack_idx_opts;
  82 static const char *base_name;
  83 static int progress = 1;
  84 static int window = 10;
  85 static unsigned long pack_size_limit;
  86 static int depth = 50;
  87 static int delta_search_threads;
  88 static int pack_to_stdout;
  89 static int sparse;
  90 static int thin;
  91 static int num_preferred_base;
  92 static struct progress *progress_state;
  93
  94 static struct packed_git *reuse_packfile;
  95 static uint32_t reuse_packfile_objects;
  96 static struct bitmap *reuse_packfile_bitmap;
  97
  98 static int use_bitmap_index_default = 1;
  99 static int use_bitmap_index = -1;
 100 static int allow_pack_reuse = 1;
 101 static enum {
 102         WRITE_BITMAP_FALSE = 0,
 103         WRITE_BITMAP_QUIET,
 104         WRITE_BITMAP_TRUE,
 105 } write_bitmap_index;
 106 static uint16_t write_bitmap_options = BITMAP_OPT_HASH_CACHE;
 107
 108 static int exclude_promisor_objects;
 109
 110 static int use_delta_islands;
 111
 112 static unsigned long delta_cache_size = 0;
 113 static unsigned long max_delta_cache_size = DEFAULT_DELTA_CACHE_SIZE;
 114 static unsigned long cache_max_small_delta_size = 1000;
 115
 116 static unsigned long window_memory_limit = 0;
 117
 118 static struct list_objects_filter_options filter_options;
 119
 120 enum missing_action {
 121         MA_ERROR = 0,      /* fail if any missing objects are encountered */
 122         MA_ALLOW_ANY,      /* silently allow ALL missing objects */
 123         MA_ALLOW_PROMISOR, /* silently allow all missing PROMISOR objects */
 124 };
 125 static enum missing_action arg_missing_action;
 126 static show_object_fn fn_show_object;
 127
 128 /*
 129  * stats
 130  */
 131 static uint32_t written, written_delta;
 132 static uint32_t reused, reused_delta;
 133
 134 /*
 135  * Indexed commits
 136  */
 137 static struct commit **indexed_commits;
 138 static unsigned int indexed_commits_nr;
 139 static unsigned int indexed_commits_alloc;
 140
 141 static void index_commit_for_bitmap(struct commit *commit)
 142 {
 143         if (indexed_commits_nr >= indexed_commits_alloc) {
 144                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
 145                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
 146         }
 147
 148         indexed_commits[indexed_commits_nr++] = commit;
 149 }
 150
 151 static void *get_delta(struct object_entry *entry)
 152 {
 153         unsigned long size, base_size, delta_size;
 154         void *buf, *base_buf, *delta_buf;
 155         enum object_type type;
 156
 157         buf = read_object_file(&entry->idx.oid, &type, &size);
 158         if (!buf)
 159                 die(_("unable to read %s"), oid_to_hex(&entry->idx.oid));
 160         base_buf = read_object_file(&DELTA(entry)->idx.oid, &type,
 161                                     &base_size);
 162         if (!base_buf)
 163                 die("unable to read %s",
 164                     oid_to_hex(&DELTA(entry)->idx.oid));
 165         delta_buf = diff_delta(base_buf, base_size,
 166                                buf, size, &delta_size, 0);
 167         /*
 168          * We successfully computed this delta once but dropped it for
 169          * memory reasons. Something is very wrong if this time we
 170          * recompute and create a different delta.
 171          */
 172         if (!delta_buf || delta_size != DELTA_SIZE(entry))
 173                 BUG("delta size changed");
 174         free(buf);
 175         free(base_buf);
 176         return delta_buf;
 177 }
 178
 179 static unsigned long do_compress(void **pptr, unsigned long size)
 180 {
 181         git_zstream stream;
 182         void *in, *out;
 183         unsigned long maxsize;
 184
 185         git_deflate_init(&stream, pack_compression_level);
 186         maxsize = git_deflate_bound(&stream, size);
 187
 188         in = *pptr;
 189         out = xmalloc(maxsize);
 190         *pptr = out;
 191
 192         stream.next_in = in;
 193         stream.avail_in = size;
 194         stream.next_out = out;
 195         stream.avail_out = maxsize;
 196         while (git_deflate(&stream, Z_FINISH) == Z_OK)
 197                 ; /* nothing */
 198         git_deflate_end(&stream);
 199
 200         free(in);
 201         return stream.total_out;
 202 }
 203
 204 static unsigned long write_large_blob_data(struct git_istream *st, struct hashfile *f,
 205                                            const struct object_id *oid)
 206 {
 207         git_zstream stream;
 208         unsigned char ibuf[1024 * 16];
 209         unsigned char obuf[1024 * 16];
 210         unsigned long olen = 0;
 211
 212         git_deflate_init(&stream, pack_compression_level);
 213
 214         for (;;) {
 215                 ssize_t readlen;
 216                 int zret = Z_OK;
 217                 readlen = read_istream(st, ibuf, sizeof(ibuf));
 218                 if (readlen == -1)
 219                         die(_("unable to read %s"), oid_to_hex(oid));
 220
 221                 stream.next_in = ibuf;
 222                 stream.avail_in = readlen;
 223                 while ((stream.avail_in || readlen == 0) &&
 224                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
 225                         stream.next_out = obuf;
 226                         stream.avail_out = sizeof(obuf);
 227                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 228                         hashwrite(f, obuf, stream.next_out - obuf);
 229                         olen += stream.next_out - obuf;
 230                 }
 231                 if (stream.avail_in)
 232                         die(_("deflate error (%d)"), zret);
 233                 if (readlen == 0) {
 234                         if (zret != Z_STREAM_END)
 235                                 die(_("deflate error (%d)"), zret);
 236                         break;
 237                 }
 238         }
 239         git_deflate_end(&stream);
 240         return olen;
 241 }
 242
 243 /*
 244  * we are going to reuse the existing object data as is.  make
 245  * sure it is not corrupt.
 246  */
 247 static int check_pack_inflate(struct packed_git *p,
 248                 struct pack_window **w_curs,
 249                 off_t offset,
 250                 off_t len,
 251                 unsigned long expect)
 252 {
 253         git_zstream stream;
 254         unsigned char fakebuf[4096], *in;
 255         int st;
 256
 257         memset(&stream, 0, sizeof(stream));
 258         git_inflate_init(&stream);
 259         do {
 260                 in = use_pack(p, w_curs, offset, &stream.avail_in);
 261                 stream.next_in = in;
 262                 stream.next_out = fakebuf;
 263                 stream.avail_out = sizeof(fakebuf);
 264                 st = git_inflate(&stream, Z_FINISH);
 265                 offset += stream.next_in - in;
 266         } while (st == Z_OK || st == Z_BUF_ERROR);
 267         git_inflate_end(&stream);
 268         return (st == Z_STREAM_END &&
 269                 stream.total_out == expect &&
 270                 stream.total_in == len) ? 0 : -1;
 271 }
 272
 273 static void copy_pack_data(struct hashfile *f,
 274                 struct packed_git *p,
 275                 struct pack_window **w_curs,
 276                 off_t offset,
 277                 off_t len)
 278 {
 279         unsigned char *in;
 280         unsigned long avail;
 281
 282         while (len) {
 283                 in = use_pack(p, w_curs, offset, &avail);
 284                 if (avail > len)
 285                         avail = (unsigned long)len;
 286                 hashwrite(f, in, avail);
 287                 offset += avail;
 288                 len -= avail;
 289         }
 290 }
 291
 292 /* Return 0 if we will bust the pack-size limit */
 293 static unsigned long write_no_reuse_object(struct hashfile *f, struct object_entry *entry,
 294                                            unsigned long limit, int usable_delta)
 295 {
 296         unsigned long size, datalen;
 297         unsigned char header[MAX_PACK_OBJECT_HEADER],
 298                       dheader[MAX_PACK_OBJECT_HEADER];
 299         unsigned hdrlen;
 300         enum object_type type;
 301         void *buf;
 302         struct git_istream *st = NULL;
 303         const unsigned hashsz = the_hash_algo->rawsz;
 304
 305         if (!usable_delta) {
 306                 if (oe_type(entry) == OBJ_BLOB &&
 307                     oe_size_greater_than(&to_pack, entry, big_file_threshold) &&
 308                     (st = open_istream(the_repository, &entry->idx.oid, &type,
 309                                        &size, NULL)) != NULL)
 310                         buf = NULL;
 311                 else {
 312                         buf = read_object_file(&entry->idx.oid, &type, &size);
 313                         if (!buf)
 314                                 die(_("unable to read %s"),
 315                                     oid_to_hex(&entry->idx.oid));
 316                 }
 317                 /*
 318                  * make sure no cached delta data remains from a
 319                  * previous attempt before a pack split occurred.
 320                  */
 321                 FREE_AND_NULL(entry->delta_data);
 322                 entry->z_delta_size = 0;
 323         } else if (entry->delta_data) {
 324                 size = DELTA_SIZE(entry);
 325                 buf = entry->delta_data;
 326                 entry->delta_data = NULL;
 327                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
 328                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 329         } else {
 330                 buf = get_delta(entry);
 331                 size = DELTA_SIZE(entry);
 332                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
 333                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 334         }
 335
 336         if (st) /* large blob case, just assume we don't compress well */
 337                 datalen = size;
 338         else if (entry->z_delta_size)
 339                 datalen = entry->z_delta_size;
 340         else
 341                 datalen = do_compress(&buf, size);
 342
 343         /*
 344          * The object header is a byte of 'type' followed by zero or
 345          * more bytes of length.
 346          */
 347         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 348                                               type, size);
 349
 350         if (type == OBJ_OFS_DELTA) {
 351                 /*
 352                  * Deltas with relative base contain an additional
 353                  * encoding of the relative offset for the delta
 354                  * base from this object's position in the pack.
 355                  */
 356                 off_t ofs = entry->idx.offset - DELTA(entry)->idx.offset;
 357                 unsigned pos = sizeof(dheader) - 1;
 358                 dheader[pos] = ofs & 127;
 359                 while (ofs >>= 7)
 360                         dheader[--pos] = 128 | (--ofs & 127);
 361                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + hashsz >= limit) {
 362                         if (st)
 363                                 close_istream(st);
 364                         free(buf);
 365                         return 0;
 366                 }
 367                 hashwrite(f, header, hdrlen);
 368                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
 369                 hdrlen += sizeof(dheader) - pos;
 370         } else if (type == OBJ_REF_DELTA) {
 371                 /*
 372                  * Deltas with a base reference contain
 373                  * additional bytes for the base object ID.
 374                  */
 375                 if (limit && hdrlen + hashsz + datalen + hashsz >= limit) {
 376                         if (st)
 377                                 close_istream(st);
 378                         free(buf);
 379                         return 0;
 380                 }
 381                 hashwrite(f, header, hdrlen);
 382                 hashwrite(f, DELTA(entry)->idx.oid.hash, hashsz);
 383                 hdrlen += hashsz;
 384         } else {
 385                 if (limit && hdrlen + datalen + hashsz >= limit) {
 386                         if (st)
 387                                 close_istream(st);
 388                         free(buf);
 389                         return 0;
 390                 }
 391                 hashwrite(f, header, hdrlen);
 392         }
 393         if (st) {
 394                 datalen = write_large_blob_data(st, f, &entry->idx.oid);
 395                 close_istream(st);
 396         } else {
 397                 hashwrite(f, buf, datalen);
 398                 free(buf);
 399         }
 400
 401         return hdrlen + datalen;
 402 }
 403
 404 /* Return 0 if we will bust the pack-size limit */
 405 static off_t write_reuse_object(struct hashfile *f, struct object_entry *entry,
 406                                 unsigned long limit, int usable_delta)
 407 {
 408         struct packed_git *p = IN_PACK(entry);
 409         struct pack_window *w_curs = NULL;
 410         struct revindex_entry *revidx;
 411         off_t offset;
 412         enum object_type type = oe_type(entry);
 413         off_t datalen;
 414         unsigned char header[MAX_PACK_OBJECT_HEADER],
 415                       dheader[MAX_PACK_OBJECT_HEADER];
 416         unsigned hdrlen;
 417         const unsigned hashsz = the_hash_algo->rawsz;
 418         unsigned long entry_size = SIZE(entry);
 419
 420         if (DELTA(entry))
 421                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
 422                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 423         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 424                                               type, entry_size);
 425
 426         offset = entry->in_pack_offset;
 427         revidx = find_pack_revindex(p, offset);
 428         datalen = revidx[1].offset - offset;
 429         if (!pack_to_stdout && p->index_version > 1 &&
 430             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 431                 error(_("bad packed object CRC for %s"),
 432                       oid_to_hex(&entry->idx.oid));
 433                 unuse_pack(&w_curs);
 434                 return write_no_reuse_object(f, entry, limit, usable_delta);
 435         }
 436
 437         offset += entry->in_pack_header_size;
 438         datalen -= entry->in_pack_header_size;
 439
 440         if (!pack_to_stdout && p->index_version == 1 &&
 441             check_pack_inflate(p, &w_curs, offset, datalen, entry_size)) {
 442                 error(_("corrupt packed object for %s"),
 443                       oid_to_hex(&entry->idx.oid));
 444                 unuse_pack(&w_curs);
 445                 return write_no_reuse_object(f, entry, limit, usable_delta);
 446         }
 447
 448         if (type == OBJ_OFS_DELTA) {
 449                 off_t ofs = entry->idx.offset - DELTA(entry)->idx.offset;
 450                 unsigned pos = sizeof(dheader) - 1;
 451                 dheader[pos] = ofs & 127;
 452                 while (ofs >>= 7)
 453                         dheader[--pos] = 128 | (--ofs & 127);
 454                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + hashsz >= limit) {
 455                         unuse_pack(&w_curs);
 456                         return 0;
 457                 }
 458                 hashwrite(f, header, hdrlen);
 459                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
 460                 hdrlen += sizeof(dheader) - pos;
 461                 reused_delta++;
 462         } else if (type == OBJ_REF_DELTA) {
 463                 if (limit && hdrlen + hashsz + datalen + hashsz >= limit) {
 464                         unuse_pack(&w_curs);
 465                         return 0;
 466                 }
 467                 hashwrite(f, header, hdrlen);
 468                 hashwrite(f, DELTA(entry)->idx.oid.hash, hashsz);
 469                 hdrlen += hashsz;
 470                 reused_delta++;
 471         } else {
 472                 if (limit && hdrlen + datalen + hashsz >= limit) {
 473                         unuse_pack(&w_curs);
 474                         return 0;
 475                 }
 476                 hashwrite(f, header, hdrlen);
 477         }
 478         copy_pack_data(f, p, &w_curs, offset, datalen);
 479         unuse_pack(&w_curs);
 480         reused++;
 481         return hdrlen + datalen;
 482 }
 483
 484 /* Return 0 if we will bust the pack-size limit */
 485 static off_t write_object(struct hashfile *f,
 486                           struct object_entry *entry,
 487                           off_t write_offset)
 488 {
 489         unsigned long limit;
 490         off_t len;
 491         int usable_delta, to_reuse;
 492
 493         if (!pack_to_stdout)
 494                 crc32_begin(f);
 495
 496         /* apply size limit if limited packsize and not first object */
 497         if (!pack_size_limit || !nr_written)
 498                 limit = 0;
 499         else if (pack_size_limit <= write_offset)
 500                 /*
 501                  * the earlier object did not fit the limit; avoid
 502                  * mistaking this with unlimited (i.e. limit = 0).
 503                  */
 504                 limit = 1;
 505         else
 506                 limit = pack_size_limit - write_offset;
 507
 508         if (!DELTA(entry))
 509                 usable_delta = 0;       /* no delta */
 510         else if (!pack_size_limit)
 511                usable_delta = 1;        /* unlimited packfile */
 512         else if (DELTA(entry)->idx.offset == (off_t)-1)
 513                 usable_delta = 0;       /* base was written to another pack */
 514         else if (DELTA(entry)->idx.offset)
 515                 usable_delta = 1;       /* base already exists in this pack */
 516         else
 517                 usable_delta = 0;       /* base could end up in another pack */
 518
 519         if (!reuse_object)
 520                 to_reuse = 0;   /* explicit */
 521         else if (!IN_PACK(entry))
 522                 to_reuse = 0;   /* can't reuse what we don't have */
 523         else if (oe_type(entry) == OBJ_REF_DELTA ||
 524                  oe_type(entry) == OBJ_OFS_DELTA)
 525                                 /* check_object() decided it for us ... */
 526                 to_reuse = usable_delta;
 527                                 /* ... but pack split may override that */
 528         else if (oe_type(entry) != entry->in_pack_type)
 529                 to_reuse = 0;   /* pack has delta which is unusable */
 530         else if (DELTA(entry))
 531                 to_reuse = 0;   /* we want to pack afresh */
 532         else
 533                 to_reuse = 1;   /* we have it in-pack undeltified,
 534                                  * and we do not need to deltify it.
 535                                  */
 536
 537         if (!to_reuse)
 538                 len = write_no_reuse_object(f, entry, limit, usable_delta);
 539         else
 540                 len = write_reuse_object(f, entry, limit, usable_delta);
 541         if (!len)
 542                 return 0;
 543
 544         if (usable_delta)
 545                 written_delta++;
 546         written++;
 547         if (!pack_to_stdout)
 548                 entry->idx.crc32 = crc32_end(f);
 549         return len;
 550 }
 551
 552 enum write_one_status {
 553         WRITE_ONE_SKIP = -1, /* already written */
 554         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 555         WRITE_ONE_WRITTEN = 1, /* normal */
 556         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 557 };
 558
 559 static enum write_one_status write_one(struct hashfile *f,
 560                                        struct object_entry *e,
 561                                        off_t *offset)
 562 {
 563         off_t size;
 564         int recursing;
 565
 566         /*
 567          * we set offset to 1 (which is an impossible value) to mark
 568          * the fact that this object is involved in "write its base
 569          * first before writing a deltified object" recursion.
 570          */
 571         recursing = (e->idx.offset == 1);
 572         if (recursing) {
 573                 warning(_("recursive delta detected for object %s"),
 574                         oid_to_hex(&e->idx.oid));
 575                 return WRITE_ONE_RECURSIVE;
 576         } else if (e->idx.offset || e->preferred_base) {
 577                 /* offset is non zero if object is written already. */
 578                 return WRITE_ONE_SKIP;
 579         }
 580
 581         /* if we are deltified, write out base object first. */
 582         if (DELTA(e)) {
 583                 e->idx.offset = 1; /* now recurse */
 584                 switch (write_one(f, DELTA(e), offset)) {
 585                 case WRITE_ONE_RECURSIVE:
 586                         /* we cannot depend on this one */
 587                         SET_DELTA(e, NULL);
 588                         break;
 589                 default:
 590                         break;
 591                 case WRITE_ONE_BREAK:
 592                         e->idx.offset = recursing;
 593                         return WRITE_ONE_BREAK;
 594                 }
 595         }
 596
 597         e->idx.offset = *offset;
 598         size = write_object(f, e, *offset);
 599         if (!size) {
 600                 e->idx.offset = recursing;
 601                 return WRITE_ONE_BREAK;
 602         }
 603         written_list[nr_written++] = &e->idx;
 604
 605         /* make sure off_t is sufficiently large not to wrap */
 606         if (signed_add_overflows(*offset, size))
 607                 die(_("pack too large for current definition of off_t"));
 608         *offset += size;
 609         return WRITE_ONE_WRITTEN;
 610 }
 611
 612 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
 613                        void *cb_data)
 614 {
 615         struct object_id peeled;
 616         struct object_entry *entry = packlist_find(&to_pack, oid);
 617
 618         if (entry)
 619                 entry->tagged = 1;
 620         if (!peel_ref(path, &peeled)) {
 621                 entry = packlist_find(&to_pack, &peeled);
 622                 if (entry)
 623                         entry->tagged = 1;
 624         }
 625         return 0;
 626 }
 627
 628 static inline void add_to_write_order(struct object_entry **wo,
 629                                unsigned int *endp,
 630                                struct object_entry *e)
 631 {
 632         if (e->filled || oe_layer(&to_pack, e) != write_layer)
 633                 return;
 634         wo[(*endp)++] = e;
 635         e->filled = 1;
 636 }
 637
 638 static void add_descendants_to_write_order(struct object_entry **wo,
 639                                            unsigned int *endp,
 640                                            struct object_entry *e)
 641 {
 642         int add_to_order = 1;
 643         while (e) {
 644                 if (add_to_order) {
 645                         struct object_entry *s;
 646                         /* add this node... */
 647                         add_to_write_order(wo, endp, e);
 648                         /* all its siblings... */
 649                         for (s = DELTA_SIBLING(e); s; s = DELTA_SIBLING(s)) {
 650                                 add_to_write_order(wo, endp, s);
 651                         }
 652                 }
 653                 /* drop down a level to add left subtree nodes if possible */
 654                 if (DELTA_CHILD(e)) {
 655                         add_to_order = 1;
 656                         e = DELTA_CHILD(e);
 657                 } else {
 658                         add_to_order = 0;
 659                         /* our sibling might have some children, it is next */
 660                         if (DELTA_SIBLING(e)) {
 661                                 e = DELTA_SIBLING(e);
 662                                 continue;
 663                         }
 664                         /* go back to our parent node */
 665                         e = DELTA(e);
 666                         while (e && !DELTA_SIBLING(e)) {
 667                                 /* we're on the right side of a subtree, keep
 668                                  * going up until we can go right again */
 669                                 e = DELTA(e);
 670                         }
 671                         if (!e) {
 672                                 /* done- we hit our original root node */
 673                                 return;
 674                         }
 675                         /* pass it off to sibling at this level */
 676                         e = DELTA_SIBLING(e);
 677                 }
 678         };
 679 }
 680
 681 static void add_family_to_write_order(struct object_entry **wo,
 682                                       unsigned int *endp,
 683                                       struct object_entry *e)
 684 {
 685         struct object_entry *root;
 686
 687         for (root = e; DELTA(root); root = DELTA(root))
 688                 ; /* nothing */
 689         add_descendants_to_write_order(wo, endp, root);
 690 }
 691
 692 static void compute_layer_order(struct object_entry **wo, unsigned int *wo_end)
 693 {
 694         unsigned int i, last_untagged;
 695         struct object_entry *objects = to_pack.objects;
 696
 697         for (i = 0; i < to_pack.nr_objects; i++) {
 698                 if (objects[i].tagged)
 699                         break;
 700                 add_to_write_order(wo, wo_end, &objects[i]);
 701         }
 702         last_untagged = i;
 703
 704         /*
 705          * Then fill all the tagged tips.
 706          */
 707         for (; i < to_pack.nr_objects; i++) {
 708                 if (objects[i].tagged)
 709                         add_to_write_order(wo, wo_end, &objects[i]);
 710         }
 711
 712         /*
 713          * And then all remaining commits and tags.
 714          */
 715         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 716                 if (oe_type(&objects[i]) != OBJ_COMMIT &&
 717                     oe_type(&objects[i]) != OBJ_TAG)
 718                         continue;
 719                 add_to_write_order(wo, wo_end, &objects[i]);
 720         }
 721
 722         /*
 723          * And then all the trees.
 724          */
 725         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 726                 if (oe_type(&objects[i]) != OBJ_TREE)
 727                         continue;
 728                 add_to_write_order(wo, wo_end, &objects[i]);
 729         }
 730
 731         /*
 732          * Finally all the rest in really tight order
 733          */
 734         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 735                 if (!objects[i].filled && oe_layer(&to_pack, &objects[i]) == write_layer)
 736                         add_family_to_write_order(wo, wo_end, &objects[i]);
 737         }
 738 }
 739
 740 static struct object_entry **compute_write_order(void)
 741 {
 742         uint32_t max_layers = 1;
 743         unsigned int i, wo_end;
 744
 745         struct object_entry **wo;
 746         struct object_entry *objects = to_pack.objects;
 747
 748         for (i = 0; i < to_pack.nr_objects; i++) {
 749                 objects[i].tagged = 0;
 750                 objects[i].filled = 0;
 751                 SET_DELTA_CHILD(&objects[i], NULL);
 752                 SET_DELTA_SIBLING(&objects[i], NULL);
 753         }
 754
 755         /*
 756          * Fully connect delta_child/delta_sibling network.
 757          * Make sure delta_sibling is sorted in the original
 758          * recency order.
 759          */
 760         for (i = to_pack.nr_objects; i > 0;) {
 761                 struct object_entry *e = &objects[--i];
 762                 if (!DELTA(e))
 763                         continue;
 764                 /* Mark me as the first child */
 765                 e->delta_sibling_idx = DELTA(e)->delta_child_idx;
 766                 SET_DELTA_CHILD(DELTA(e), e);
 767         }
 768
 769         /*
 770          * Mark objects that are at the tip of tags.
 771          */
 772         for_each_tag_ref(mark_tagged, NULL);
 773
 774         if (use_delta_islands)
 775                 max_layers = compute_pack_layers(&to_pack);
 776
 777         ALLOC_ARRAY(wo, to_pack.nr_objects);
 778         wo_end = 0;
 779
 780         for (; write_layer < max_layers; ++write_layer)
 781                 compute_layer_order(wo, &wo_end);
 782
 783         if (wo_end != to_pack.nr_objects)
 784                 die(_("ordered %u objects, expected %"PRIu32),
 785                     wo_end, to_pack.nr_objects);
 786
 787         return wo;
 788 }
 789
 790
 791 /*
 792  * A reused set of objects. All objects in a chunk have the same
 793  * relative position in the original packfile and the generated
 794  * packfile.
 795  */
 796
 797 static struct reused_chunk {
 798         /* The offset of the first object of this chunk in the original
 799          * packfile. */
 800         off_t original;
 801         /* The offset of the first object of this chunk in the generated
 802          * packfile minus "original". */
 803         off_t difference;
 804 } *reused_chunks;
 805 static int reused_chunks_nr;
 806 static int reused_chunks_alloc;
 807
 808 static void record_reused_object(off_t where, off_t offset)
 809 {
 810         if (reused_chunks_nr && reused_chunks[reused_chunks_nr-1].difference == offset)
 811                 return;
 812
 813         ALLOC_GROW(reused_chunks, reused_chunks_nr + 1,
 814                    reused_chunks_alloc);
 815         reused_chunks[reused_chunks_nr].original = where;
 816         reused_chunks[reused_chunks_nr].difference = offset;
 817         reused_chunks_nr++;
 818 }
 819
 820 /*
 821  * Binary search to find the chunk that "where" is in. Note
 822  * that we're not looking for an exact match, just the first
 823  * chunk that contains it (which implicitly ends at the start
 824  * of the next chunk.
 825  */
 826 static off_t find_reused_offset(off_t where)
 827 {
 828         int lo = 0, hi = reused_chunks_nr;
 829         while (lo < hi) {
 830                 int mi = lo + ((hi - lo) / 2);
 831                 if (where == reused_chunks[mi].original)
 832                         return reused_chunks[mi].difference;
 833                 if (where < reused_chunks[mi].original)
 834                         hi = mi;
 835                 else
 836                         lo = mi + 1;
 837         }
 838
 839         /*
 840          * The first chunk starts at zero, so we can't have gone below
 841          * there.
 842          */
 843         assert(lo);
 844         return reused_chunks[lo-1].difference;
 845 }
 846
 847 static void write_reused_pack_one(size_t pos, struct hashfile *out,
 848                                   struct pack_window **w_curs)
 849 {
 850         off_t offset, next, cur;
 851         enum object_type type;
 852         unsigned long size;
 853
 854         offset = reuse_packfile->revindex[pos].offset;
 855         next = reuse_packfile->revindex[pos + 1].offset;
 856
 857         record_reused_object(offset, offset - hashfile_total(out));
 858
 859         cur = offset;
 860         type = unpack_object_header(reuse_packfile, w_curs, &cur, &size);
 861         assert(type >= 0);
 862
 863         if (type == OBJ_OFS_DELTA) {
 864                 off_t base_offset;
 865                 off_t fixup;
 866
 867                 unsigned char header[MAX_PACK_OBJECT_HEADER];
 868                 unsigned len;
 869
 870                 base_offset = get_delta_base(reuse_packfile, w_curs, &cur, type, offset);
 871                 assert(base_offset != 0);
 872
 873                 /* Convert to REF_DELTA if we must... */
 874                 if (!allow_ofs_delta) {
 875                         int base_pos = find_revindex_position(reuse_packfile, base_offset);
 876                         struct object_id base_oid;
 877
 878                         nth_packed_object_id(&base_oid, reuse_packfile,
 879                                              reuse_packfile->revindex[base_pos].nr);
 880
 881                         len = encode_in_pack_object_header(header, sizeof(header),
 882                                                            OBJ_REF_DELTA, size);
 883                         hashwrite(out, header, len);
 884                         hashwrite(out, base_oid.hash, the_hash_algo->rawsz);
 885                         copy_pack_data(out, reuse_packfile, w_curs, cur, next - cur);
 886                         return;
 887                 }
 888
 889                 /* Otherwise see if we need to rewrite the offset... */
 890                 fixup = find_reused_offset(offset) -
 891                         find_reused_offset(base_offset);
 892                 if (fixup) {
 893                         unsigned char ofs_header[10];
 894                         unsigned i, ofs_len;
 895                         off_t ofs = offset - base_offset - fixup;
 896
 897                         len = encode_in_pack_object_header(header, sizeof(header),
 898                                                            OBJ_OFS_DELTA, size);
 899
 900                         i = sizeof(ofs_header) - 1;
 901                         ofs_header[i] = ofs & 127;
 902                         while (ofs >>= 7)
 903                                 ofs_header[--i] = 128 | (--ofs & 127);
 904
 905                         ofs_len = sizeof(ofs_header) - i;
 906
 907                         hashwrite(out, header, len);
 908                         hashwrite(out, ofs_header + sizeof(ofs_header) - ofs_len, ofs_len);
 909                         copy_pack_data(out, reuse_packfile, w_curs, cur, next - cur);
 910                         return;
 911                 }
 912
 913                 /* ...otherwise we have no fixup, and can write it verbatim */
 914         }
 915
 916         copy_pack_data(out, reuse_packfile, w_curs, offset, next - offset);
 917 }
 918
 919 static size_t write_reused_pack_verbatim(struct hashfile *out,
 920                                          struct pack_window **w_curs)
 921 {
 922         size_t pos = 0;
 923
 924         while (pos < reuse_packfile_bitmap->word_alloc &&
 925                         reuse_packfile_bitmap->words[pos] == (eword_t)~0)
 926                 pos++;
 927
 928         if (pos) {
 929                 off_t to_write;
 930
 931                 written = (pos * BITS_IN_EWORD);
 932                 to_write = reuse_packfile->revindex[written].offset
 933                         - sizeof(struct pack_header);
 934
 935                 /* We're recording one chunk, not one object. */
 936                 record_reused_object(sizeof(struct pack_header), 0);
 937                 hashflush(out);
 938                 copy_pack_data(out, reuse_packfile, w_curs,
 939                         sizeof(struct pack_header), to_write);
 940
 941                 display_progress(progress_state, written);
 942         }
 943         return pos;
 944 }
 945
 946 static void write_reused_pack(struct hashfile *f)
 947 {
 948         size_t i = 0;
 949         uint32_t offset;
 950         struct pack_window *w_curs = NULL;
 951
 952         if (allow_ofs_delta)
 953                 i = write_reused_pack_verbatim(f, &w_curs);
 954
 955         for (; i < reuse_packfile_bitmap->word_alloc; ++i) {
 956                 eword_t word = reuse_packfile_bitmap->words[i];
 957                 size_t pos = (i * BITS_IN_EWORD);
 958
 959                 for (offset = 0; offset < BITS_IN_EWORD; ++offset) {
 960                         if ((word >> offset) == 0)
 961                                 break;
 962
 963                         offset += ewah_bit_ctz64(word >> offset);
 964                         write_reused_pack_one(pos + offset, f, &w_curs);
 965                         display_progress(progress_state, ++written);
 966                 }
 967         }
 968
 969         unuse_pack(&w_curs);
 970 }
 971
 972 static const char no_split_warning[] = N_(
 973 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
 974 );
 975
 976 static void write_pack_file(void)
 977 {
 978         uint32_t i = 0, j;
 979         struct hashfile *f;
 980         off_t offset;
 981         uint32_t nr_remaining = nr_result;
 982         time_t last_mtime = 0;
 983         struct object_entry **write_order;
 984
 985         if (progress > pack_to_stdout)
 986                 progress_state = start_progress(_("Writing objects"), nr_result);
 987         ALLOC_ARRAY(written_list, to_pack.nr_objects);
 988         write_order = compute_write_order();
 989
 990         do {
 991                 struct object_id oid;
 992                 char *pack_tmp_name = NULL;
 993
 994                 if (pack_to_stdout)
 995                         f = hashfd_throughput(1, "<stdout>", progress_state);
 996                 else
 997                         f = create_tmp_packfile(&pack_tmp_name);
 998
 999                 offset = write_pack_header(f, nr_remaining);
1000
1001                 if (reuse_packfile) {
1002                         assert(pack_to_stdout);
1003                         write_reused_pack(f);
1004                         offset = hashfile_total(f);
1005                 }
1006
1007                 nr_written = 0;
1008                 for (; i < to_pack.nr_objects; i++) {
1009                         struct object_entry *e = write_order[i];
1010                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
1011                                 break;
1012                         display_progress(progress_state, written);
1013                 }
1014
1015                 /*
1016                  * Did we write the wrong # entries in the header?
1017                  * If so, rewrite it like in fast-import
1018                  */
1019                 if (pack_to_stdout) {
1020                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_CLOSE);
1021                 } else if (nr_written == nr_remaining) {
1022                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE);
1023                 } else {
1024                         int fd = finalize_hashfile(f, oid.hash, 0);
1025                         fixup_pack_header_footer(fd, oid.hash, pack_tmp_name,
1026                                                  nr_written, oid.hash, offset);
1027                         close(fd);
1028                         if (write_bitmap_index) {
1029                                 if (write_bitmap_index != WRITE_BITMAP_QUIET)
1030                                         warning(_(no_split_warning));
1031                                 write_bitmap_index = 0;
1032                         }
1033                 }
1034
1035                 if (!pack_to_stdout) {
1036                         struct stat st;
1037                         struct strbuf tmpname = STRBUF_INIT;
1038
1039                         /*
1040                          * Packs are runtime accessed in their mtime
1041                          * order since newer packs are more likely to contain
1042                          * younger objects.  So if we are creating multiple
1043                          * packs then we should modify the mtime of later ones
1044                          * to preserve this property.
1045                          */
1046                         if (stat(pack_tmp_name, &st) < 0) {
1047                                 warning_errno(_("failed to stat %s"), pack_tmp_name);
1048                         } else if (!last_mtime) {
1049                                 last_mtime = st.st_mtime;
1050                         } else {
1051                                 struct utimbuf utb;
1052                                 utb.actime = st.st_atime;
1053                                 utb.modtime = --last_mtime;
1054                                 if (utime(pack_tmp_name, &utb) < 0)
1055                                         warning_errno(_("failed utime() on %s"), pack_tmp_name);
1056                         }
1057
1058                         strbuf_addf(&tmpname, "%s-", base_name);
1059
1060                         if (write_bitmap_index) {
1061                                 bitmap_writer_set_checksum(oid.hash);
1062                                 bitmap_writer_build_type_index(
1063                                         &to_pack, written_list, nr_written);
1064                         }
1065
1066                         finish_tmp_packfile(&tmpname, pack_tmp_name,
1067                                             written_list, nr_written,
1068                                             &pack_idx_opts, oid.hash);
1069
1070                         if (write_bitmap_index) {
1071                                 strbuf_addf(&tmpname, "%s.bitmap", oid_to_hex(&oid));
1072
1073                                 stop_progress(&progress_state);
1074
1075                                 bitmap_writer_show_progress(progress);
1076                                 bitmap_writer_reuse_bitmaps(&to_pack);
1077                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
1078                                 bitmap_writer_build(&to_pack);
1079                                 bitmap_writer_finish(written_list, nr_written,
1080                                                      tmpname.buf, write_bitmap_options);
1081                                 write_bitmap_index = 0;
1082                         }
1083
1084                         strbuf_release(&tmpname);
1085                         free(pack_tmp_name);
1086                         puts(oid_to_hex(&oid));
1087                 }
1088
1089                 /* mark written objects as written to previous pack */
1090                 for (j = 0; j < nr_written; j++) {
1091                         written_list[j]->offset = (off_t)-1;
1092                 }
1093                 nr_remaining -= nr_written;
1094         } while (nr_remaining && i < to_pack.nr_objects);
1095
1096         free(written_list);
1097         free(write_order);
1098         stop_progress(&progress_state);
1099         if (written != nr_result)
1100                 die(_("wrote %"PRIu32" objects while expecting %"PRIu32),
1101                     written, nr_result);
1102         trace2_data_intmax("pack-objects", the_repository,
1103                            "write_pack_file/wrote", nr_result);
1104 }
1105
1106 static int no_try_delta(const char *path)
1107 {
1108         static struct attr_check *check;
1109
1110         if (!check)
1111                 check = attr_check_initl("delta", NULL);
1112         git_check_attr(the_repository->index, path, check);
1113         if (ATTR_FALSE(check->items[0].value))
1114                 return 1;
1115         return 0;
1116 }
1117
1118 /*
1119  * When adding an object, check whether we have already added it
1120  * to our packing list. If so, we can skip. However, if we are
1121  * being asked to excludei t, but the previous mention was to include
1122  * it, make sure to adjust its flags and tweak our numbers accordingly.
1123  *
1124  * As an optimization, we pass out the index position where we would have
1125  * found the item, since that saves us from having to look it up again a
1126  * few lines later when we want to add the new entry.
1127  */
1128 static int have_duplicate_entry(const struct object_id *oid,
1129                                 int exclude)
1130 {
1131         struct object_entry *entry;
1132
1133         if (reuse_packfile_bitmap &&
1134             bitmap_walk_contains(bitmap_git, reuse_packfile_bitmap, oid))
1135                 return 1;
1136
1137         entry = packlist_find(&to_pack, oid);
1138         if (!entry)
1139                 return 0;
1140
1141         if (exclude) {
1142                 if (!entry->preferred_base)
1143                         nr_result--;
1144                 entry->preferred_base = 1;
1145         }
1146
1147         return 1;
1148 }
1149
1150 static int want_found_object(int exclude, struct packed_git *p)
1151 {
1152         if (exclude)
1153                 return 1;
1154         if (incremental)
1155                 return 0;
1156
1157         /*
1158          * When asked to do --local (do not include an object that appears in a
1159          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
1160          * an object that appears in a pack marked with .keep), finding a pack
1161          * that matches the criteria is sufficient for us to decide to omit it.
1162          * However, even if this pack does not satisfy the criteria, we need to
1163          * make sure no copy of this object appears in _any_ pack that makes us
1164          * to omit the object, so we need to check all the packs.
1165          *
1166          * We can however first check whether these options can possible matter;
1167          * if they do not matter we know we want the object in generated pack.
1168          * Otherwise, we signal "-1" at the end to tell the caller that we do
1169          * not know either way, and it needs to check more packs.
1170          */
1171         if (!ignore_packed_keep_on_disk &&
1172             !ignore_packed_keep_in_core &&
1173             (!local || !have_non_local_packs))
1174                 return 1;
1175
1176         if (local && !p->pack_local)
1177                 return 0;
1178         if (p->pack_local &&
1179             ((ignore_packed_keep_on_disk && p->pack_keep) ||
1180              (ignore_packed_keep_in_core && p->pack_keep_in_core)))
1181                 return 0;
1182
1183         /* we don't know yet; keep looking for more packs */
1184         return -1;
1185 }
1186
1187 /*
1188  * Check whether we want the object in the pack (e.g., we do not want
1189  * objects found in non-local stores if the "--local" option was used).
1190  *
1191  * If the caller already knows an existing pack it wants to take the object
1192  * from, that is passed in *found_pack and *found_offset; otherwise this
1193  * function finds if there is any pack that has the object and returns the pack
1194  * and its offset in these variables.
1195  */
1196 static int want_object_in_pack(const struct object_id *oid,
1197                                int exclude,
1198                                struct packed_git **found_pack,
1199                                off_t *found_offset)
1200 {
1201         int want;
1202         struct list_head *pos;
1203         struct multi_pack_index *m;
1204
1205         if (!exclude && local && has_loose_object_nonlocal(oid))
1206                 return 0;
1207
1208         /*
1209          * If we already know the pack object lives in, start checks from that
1210          * pack - in the usual case when neither --local was given nor .keep files
1211          * are present we will determine the answer right now.
1212          */
1213         if (*found_pack) {
1214                 want = want_found_object(exclude, *found_pack);
1215                 if (want != -1)
1216                         return want;
1217         }
1218
1219         for (m = get_multi_pack_index(the_repository); m; m = m->next) {
1220                 struct pack_entry e;
1221                 if (fill_midx_entry(the_repository, oid, &e, m)) {
1222                         struct packed_git *p = e.p;
1223                         off_t offset;
1224
1225                         if (p == *found_pack)
1226                                 offset = *found_offset;
1227                         else
1228                                 offset = find_pack_entry_one(oid->hash, p);
1229
1230                         if (offset) {
1231                                 if (!*found_pack) {
1232                                         if (!is_pack_valid(p))
1233                                                 continue;
1234                                         *found_offset = offset;
1235                                         *found_pack = p;
1236                                 }
1237                                 want = want_found_object(exclude, p);
1238                                 if (want != -1)
1239                                         return want;
1240                         }
1241                 }
1242         }
1243
1244         list_for_each(pos, get_packed_git_mru(the_repository)) {
1245                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
1246                 off_t offset;
1247
1248                 if (p == *found_pack)
1249                         offset = *found_offset;
1250                 else
1251                         offset = find_pack_entry_one(oid->hash, p);
1252
1253                 if (offset) {
1254                         if (!*found_pack) {
1255                                 if (!is_pack_valid(p))
1256                                         continue;
1257                                 *found_offset = offset;
1258                                 *found_pack = p;
1259                         }
1260                         want = want_found_object(exclude, p);
1261                         if (!exclude && want > 0)
1262                                 list_move(&p->mru,
1263                                           get_packed_git_mru(the_repository));
1264                         if (want != -1)
1265                                 return want;
1266                 }
1267         }
1268
1269         return 1;
1270 }
1271
1272 static void create_object_entry(const struct object_id *oid,
1273                                 enum object_type type,
1274                                 uint32_t hash,
1275                                 int exclude,
1276                                 int no_try_delta,
1277                                 struct packed_git *found_pack,
1278                                 off_t found_offset)
1279 {
1280         struct object_entry *entry;
1281
1282         entry = packlist_alloc(&to_pack, oid);
1283         entry->hash = hash;
1284         oe_set_type(entry, type);
1285         if (exclude)
1286                 entry->preferred_base = 1;
1287         else
1288                 nr_result++;
1289         if (found_pack) {
1290                 oe_set_in_pack(&to_pack, entry, found_pack);
1291                 entry->in_pack_offset = found_offset;
1292         }
1293
1294         entry->no_try_delta = no_try_delta;
1295 }
1296
1297 static const char no_closure_warning[] = N_(
1298 "disabling bitmap writing, as some objects are not being packed"
1299 );
1300
1301 static int add_object_entry(const struct object_id *oid, enum object_type type,
1302                             const char *name, int exclude)
1303 {
1304         struct packed_git *found_pack = NULL;
1305         off_t found_offset = 0;
1306
1307         display_progress(progress_state, ++nr_seen);
1308
1309         if (have_duplicate_entry(oid, exclude))
1310                 return 0;
1311
1312         if (!want_object_in_pack(oid, exclude, &found_pack, &found_offset)) {
1313                 /* The pack is missing an object, so it will not have closure */
1314                 if (write_bitmap_index) {
1315                         if (write_bitmap_index != WRITE_BITMAP_QUIET)
1316                                 warning(_(no_closure_warning));
1317                         write_bitmap_index = 0;
1318                 }
1319                 return 0;
1320         }
1321
1322         create_object_entry(oid, type, pack_name_hash(name),
1323                             exclude, name && no_try_delta(name),
1324                             found_pack, found_offset);
1325         return 1;
1326 }
1327
1328 static int add_object_entry_from_bitmap(const struct object_id *oid,
1329                                         enum object_type type,
1330                                         int flags, uint32_t name_hash,
1331                                         struct packed_git *pack, off_t offset)
1332 {
1333         display_progress(progress_state, ++nr_seen);
1334
1335         if (have_duplicate_entry(oid, 0))
1336                 return 0;
1337
1338         if (!want_object_in_pack(oid, 0, &pack, &offset))
1339                 return 0;
1340
1341         create_object_entry(oid, type, name_hash, 0, 0, pack, offset);
1342         return 1;
1343 }
1344
1345 struct pbase_tree_cache {
1346         struct object_id oid;
1347         int ref;
1348         int temporary;
1349         void *tree_data;
1350         unsigned long tree_size;
1351 };
1352
1353 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1354 static int pbase_tree_cache_ix(const struct object_id *oid)
1355 {
1356         return oid->hash[0] % ARRAY_SIZE(pbase_tree_cache);
1357 }
1358 static int pbase_tree_cache_ix_incr(int ix)
1359 {
1360         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1361 }
1362
1363 static struct pbase_tree {
1364         struct pbase_tree *next;
1365         /* This is a phony "cache" entry; we are not
1366          * going to evict it or find it through _get()
1367          * mechanism -- this is for the toplevel node that
1368          * would almost always change with any commit.
1369          */
1370         struct pbase_tree_cache pcache;
1371 } *pbase_tree;
1372
1373 static struct pbase_tree_cache *pbase_tree_get(const struct object_id *oid)
1374 {
1375         struct pbase_tree_cache *ent, *nent;
1376         void *data;
1377         unsigned long size;
1378         enum object_type type;
1379         int neigh;
1380         int my_ix = pbase_tree_cache_ix(oid);
1381         int available_ix = -1;
1382
1383         /* pbase-tree-cache acts as a limited hashtable.
1384          * your object will be found at your index or within a few
1385          * slots after that slot if it is cached.
1386          */
1387         for (neigh = 0; neigh < 8; neigh++) {
1388                 ent = pbase_tree_cache[my_ix];
1389                 if (ent && oideq(&ent->oid, oid)) {
1390                         ent->ref++;
1391                         return ent;
1392                 }
1393                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1394                          ((0 <= available_ix) &&
1395                           (!ent && pbase_tree_cache[available_ix])))
1396                         available_ix = my_ix;
1397                 if (!ent)
1398                         break;
1399                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1400         }
1401
1402         /* Did not find one.  Either we got a bogus request or
1403          * we need to read and perhaps cache.
1404          */
1405         data = read_object_file(oid, &type, &size);
1406         if (!data)
1407                 return NULL;
1408         if (type != OBJ_TREE) {
1409                 free(data);
1410                 return NULL;
1411         }
1412
1413         /* We need to either cache or return a throwaway copy */
1414
1415         if (available_ix < 0)
1416                 ent = NULL;
1417         else {
1418                 ent = pbase_tree_cache[available_ix];
1419                 my_ix = available_ix;
1420         }
1421
1422         if (!ent) {
1423                 nent = xmalloc(sizeof(*nent));
1424                 nent->temporary = (available_ix < 0);
1425         }
1426         else {
1427                 /* evict and reuse */
1428                 free(ent->tree_data);
1429                 nent = ent;
1430         }
1431         oidcpy(&nent->oid, oid);
1432         nent->tree_data = data;
1433         nent->tree_size = size;
1434         nent->ref = 1;
1435         if (!nent->temporary)
1436                 pbase_tree_cache[my_ix] = nent;
1437         return nent;
1438 }
1439
1440 static void pbase_tree_put(struct pbase_tree_cache *cache)
1441 {
1442         if (!cache->temporary) {
1443                 cache->ref--;
1444                 return;
1445         }
1446         free(cache->tree_data);
1447         free(cache);
1448 }
1449
1450 static int name_cmp_len(const char *name)
1451 {
1452         int i;
1453         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1454                 ;
1455         return i;
1456 }
1457
1458 static void add_pbase_object(struct tree_desc *tree,
1459                              const char *name,
1460                              int cmplen,
1461                              const char *fullname)
1462 {
1463         struct name_entry entry;
1464         int cmp;
1465
1466         while (tree_entry(tree,&entry)) {
1467                 if (S_ISGITLINK(entry.mode))
1468                         continue;
1469                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1470                       memcmp(name, entry.path, cmplen);
1471                 if (cmp > 0)
1472                         continue;
1473                 if (cmp < 0)
1474                         return;
1475                 if (name[cmplen] != '/') {
1476                         add_object_entry(&entry.oid,
1477                                          object_type(entry.mode),
1478                                          fullname, 1);
1479                         return;
1480                 }
1481                 if (S_ISDIR(entry.mode)) {
1482                         struct tree_desc sub;
1483                         struct pbase_tree_cache *tree;
1484                         const char *down = name+cmplen+1;
1485                         int downlen = name_cmp_len(down);
1486
1487                         tree = pbase_tree_get(&entry.oid);
1488                         if (!tree)
1489                                 return;
1490                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1491
1492                         add_pbase_object(&sub, down, downlen, fullname);
1493                         pbase_tree_put(tree);
1494                 }
1495         }
1496 }
1497
1498 static unsigned *done_pbase_paths;
1499 static int done_pbase_paths_num;
1500 static int done_pbase_paths_alloc;
1501 static int done_pbase_path_pos(unsigned hash)
1502 {
1503         int lo = 0;
1504         int hi = done_pbase_paths_num;
1505         while (lo < hi) {
1506                 int mi = lo + (hi - lo) / 2;
1507                 if (done_pbase_paths[mi] == hash)
1508                         return mi;
1509                 if (done_pbase_paths[mi] < hash)
1510                         hi = mi;
1511                 else
1512                         lo = mi + 1;
1513         }
1514         return -lo-1;
1515 }
1516
1517 static int check_pbase_path(unsigned hash)
1518 {
1519         int pos = done_pbase_path_pos(hash);
1520         if (0 <= pos)
1521                 return 1;
1522         pos = -pos - 1;
1523         ALLOC_GROW(done_pbase_paths,
1524                    done_pbase_paths_num + 1,
1525                    done_pbase_paths_alloc);
1526         done_pbase_paths_num++;
1527         if (pos < done_pbase_paths_num)
1528                 MOVE_ARRAY(done_pbase_paths + pos + 1, done_pbase_paths + pos,
1529                            done_pbase_paths_num - pos - 1);
1530         done_pbase_paths[pos] = hash;
1531         return 0;
1532 }
1533
1534 static void add_preferred_base_object(const char *name)
1535 {
1536         struct pbase_tree *it;
1537         int cmplen;
1538         unsigned hash = pack_name_hash(name);
1539
1540         if (!num_preferred_base || check_pbase_path(hash))
1541                 return;
1542
1543         cmplen = name_cmp_len(name);
1544         for (it = pbase_tree; it; it = it->next) {
1545                 if (cmplen == 0) {
1546                         add_object_entry(&it->pcache.oid, OBJ_TREE, NULL, 1);
1547                 }
1548                 else {
1549                         struct tree_desc tree;
1550                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1551                         add_pbase_object(&tree, name, cmplen, name);
1552                 }
1553         }
1554 }
1555
1556 static void add_preferred_base(struct object_id *oid)
1557 {
1558         struct pbase_tree *it;
1559         void *data;
1560         unsigned long size;
1561         struct object_id tree_oid;
1562
1563         if (window <= num_preferred_base++)
1564                 return;
1565
1566         data = read_object_with_reference(the_repository, oid,
1567                                           tree_type, &size, &tree_oid);
1568         if (!data)
1569                 return;
1570
1571         for (it = pbase_tree; it; it = it->next) {
1572                 if (oideq(&it->pcache.oid, &tree_oid)) {
1573                         free(data);
1574                         return;
1575                 }
1576         }
1577
1578         it = xcalloc(1, sizeof(*it));
1579         it->next = pbase_tree;
1580         pbase_tree = it;
1581
1582         oidcpy(&it->pcache.oid, &tree_oid);
1583         it->pcache.tree_data = data;
1584         it->pcache.tree_size = size;
1585 }
1586
1587 static void cleanup_preferred_base(void)
1588 {
1589         struct pbase_tree *it;
1590         unsigned i;
1591
1592         it = pbase_tree;
1593         pbase_tree = NULL;
1594         while (it) {
1595                 struct pbase_tree *tmp = it;
1596                 it = tmp->next;
1597                 free(tmp->pcache.tree_data);
1598                 free(tmp);
1599         }
1600
1601         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1602                 if (!pbase_tree_cache[i])
1603                         continue;
1604                 free(pbase_tree_cache[i]->tree_data);
1605                 FREE_AND_NULL(pbase_tree_cache[i]);
1606         }
1607
1608         FREE_AND_NULL(done_pbase_paths);
1609         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1610 }
1611
1612 /*
1613  * Return 1 iff the object specified by "delta" can be sent
1614  * literally as a delta against the base in "base_sha1". If
1615  * so, then *base_out will point to the entry in our packing
1616  * list, or NULL if we must use the external-base list.
1617  *
1618  * Depth value does not matter - find_deltas() will
1619  * never consider reused delta as the base object to
1620  * deltify other objects against, in order to avoid
1621  * circular deltas.
1622  */
1623 static int can_reuse_delta(const struct object_id *base_oid,
1624                            struct object_entry *delta,
1625                            struct object_entry **base_out)
1626 {
1627         struct object_entry *base;
1628
1629         /*
1630          * First see if we're already sending the base (or it's explicitly in
1631          * our "excluded" list).
1632          */
1633         base = packlist_find(&to_pack, base_oid);
1634         if (base) {
1635                 if (!in_same_island(&delta->idx.oid, &base->idx.oid))
1636                         return 0;
1637                 *base_out = base;
1638                 return 1;
1639         }
1640
1641         /*
1642          * Otherwise, reachability bitmaps may tell us if the receiver has it,
1643          * even if it was buried too deep in history to make it into the
1644          * packing list.
1645          */
1646         if (thin && bitmap_has_oid_in_uninteresting(bitmap_git, base_oid)) {
1647                 if (use_delta_islands) {
1648                         if (!in_same_island(&delta->idx.oid, base_oid))
1649                                 return 0;
1650                 }
1651                 *base_out = NULL;
1652                 return 1;
1653         }
1654
1655         return 0;
1656 }
1657
1658 static void check_object(struct object_entry *entry)
1659 {
1660         unsigned long canonical_size;
1661
1662         if (IN_PACK(entry)) {
1663                 struct packed_git *p = IN_PACK(entry);
1664                 struct pack_window *w_curs = NULL;
1665                 int have_base = 0;
1666                 struct object_id base_ref;
1667                 struct object_entry *base_entry;
1668                 unsigned long used, used_0;
1669                 unsigned long avail;
1670                 off_t ofs;
1671                 unsigned char *buf, c;
1672                 enum object_type type;
1673                 unsigned long in_pack_size;
1674
1675                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1676
1677                 /*
1678                  * We want in_pack_type even if we do not reuse delta
1679                  * since non-delta representations could still be reused.
1680                  */
1681                 used = unpack_object_header_buffer(buf, avail,
1682                                                    &type,
1683                                                    &in_pack_size);
1684                 if (used == 0)
1685                         goto give_up;
1686
1687                 if (type < 0)
1688                         BUG("invalid type %d", type);
1689                 entry->in_pack_type = type;
1690
1691                 /*
1692                  * Determine if this is a delta and if so whether we can
1693                  * reuse it or not.  Otherwise let's find out as cheaply as
1694                  * possible what the actual type and size for this object is.
1695                  */
1696                 switch (entry->in_pack_type) {
1697                 default:
1698                         /* Not a delta hence we've already got all we need. */
1699                         oe_set_type(entry, entry->in_pack_type);
1700                         SET_SIZE(entry, in_pack_size);
1701                         entry->in_pack_header_size = used;
1702                         if (oe_type(entry) < OBJ_COMMIT || oe_type(entry) > OBJ_BLOB)
1703                                 goto give_up;
1704                         unuse_pack(&w_curs);
1705                         return;
1706                 case OBJ_REF_DELTA:
1707                         if (reuse_delta && !entry->preferred_base) {
1708                                 oidread(&base_ref,
1709                                         use_pack(p, &w_curs,
1710                                                  entry->in_pack_offset + used,
1711                                                  NULL));
1712                                 have_base = 1;
1713                         }
1714                         entry->in_pack_header_size = used + the_hash_algo->rawsz;
1715                         break;
1716                 case OBJ_OFS_DELTA:
1717                         buf = use_pack(p, &w_curs,
1718                                        entry->in_pack_offset + used, NULL);
1719                         used_0 = 0;
1720                         c = buf[used_0++];
1721                         ofs = c & 127;
1722                         while (c & 128) {
1723                                 ofs += 1;
1724                                 if (!ofs || MSB(ofs, 7)) {
1725                                         error(_("delta base offset overflow in pack for %s"),
1726                                               oid_to_hex(&entry->idx.oid));
1727                                         goto give_up;
1728                                 }
1729                                 c = buf[used_0++];
1730                                 ofs = (ofs << 7) + (c & 127);
1731                         }
1732                         ofs = entry->in_pack_offset - ofs;
1733                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1734                                 error(_("delta base offset out of bound for %s"),
1735                                       oid_to_hex(&entry->idx.oid));
1736                                 goto give_up;
1737                         }
1738                         if (reuse_delta && !entry->preferred_base) {
1739                                 struct revindex_entry *revidx;
1740                                 revidx = find_pack_revindex(p, ofs);
1741                                 if (!revidx)
1742                                         goto give_up;
1743                                 if (!nth_packed_object_id(&base_ref, p, revidx->nr))
1744                                         have_base = 1;
1745                         }
1746                         entry->in_pack_header_size = used + used_0;
1747                         break;
1748                 }
1749
1750                 if (have_base &&
1751                     can_reuse_delta(&base_ref, entry, &base_entry)) {
1752                         oe_set_type(entry, entry->in_pack_type);
1753                         SET_SIZE(entry, in_pack_size); /* delta size */
1754                         SET_DELTA_SIZE(entry, in_pack_size);
1755
1756                         if (base_entry) {
1757                                 SET_DELTA(entry, base_entry);
1758                                 entry->delta_sibling_idx = base_entry->delta_child_idx;
1759                                 SET_DELTA_CHILD(base_entry, entry);
1760                         } else {
1761                                 SET_DELTA_EXT(entry, &base_ref);
1762                         }
1763
1764                         unuse_pack(&w_curs);
1765                         return;
1766                 }
1767
1768                 if (oe_type(entry)) {
1769                         off_t delta_pos;
1770
1771                         /*
1772                          * This must be a delta and we already know what the
1773                          * final object type is.  Let's extract the actual
1774                          * object size from the delta header.
1775                          */
1776                         delta_pos = entry->in_pack_offset + entry->in_pack_header_size;
1777                         canonical_size = get_size_from_delta(p, &w_curs, delta_pos);
1778                         if (canonical_size == 0)
1779                                 goto give_up;
1780                         SET_SIZE(entry, canonical_size);
1781                         unuse_pack(&w_curs);
1782                         return;
1783                 }
1784
1785                 /*
1786                  * No choice but to fall back to the recursive delta walk
1787                  * with oid_object_info() to find about the object type
1788                  * at this point...
1789                  */
1790                 give_up:
1791                 unuse_pack(&w_curs);
1792         }
1793
1794         oe_set_type(entry,
1795                     oid_object_info(the_repository, &entry->idx.oid, &canonical_size));
1796         if (entry->type_valid) {
1797                 SET_SIZE(entry, canonical_size);
1798         } else {
1799                 /*
1800                  * Bad object type is checked in prepare_pack().  This is
1801                  * to permit a missing preferred base object to be ignored
1802                  * as a preferred base.  Doing so can result in a larger
1803                  * pack file, but the transfer will still take place.
1804                  */
1805         }
1806 }
1807
1808 static int pack_offset_sort(const void *_a, const void *_b)
1809 {
1810         const struct object_entry *a = *(struct object_entry **)_a;
1811         const struct object_entry *b = *(struct object_entry **)_b;
1812         const struct packed_git *a_in_pack = IN_PACK(a);
1813         const struct packed_git *b_in_pack = IN_PACK(b);
1814
1815         /* avoid filesystem trashing with loose objects */
1816         if (!a_in_pack && !b_in_pack)
1817                 return oidcmp(&a->idx.oid, &b->idx.oid);
1818
1819         if (a_in_pack < b_in_pack)
1820                 return -1;
1821         if (a_in_pack > b_in_pack)
1822                 return 1;
1823         return a->in_pack_offset < b->in_pack_offset ? -1 :
1824                         (a->in_pack_offset > b->in_pack_offset);
1825 }
1826
1827 /*
1828  * Drop an on-disk delta we were planning to reuse. Naively, this would
1829  * just involve blanking out the "delta" field, but we have to deal
1830  * with some extra book-keeping:
1831  *
1832  *   1. Removing ourselves from the delta_sibling linked list.
1833  *
1834  *   2. Updating our size/type to the non-delta representation. These were
1835  *      either not recorded initially (size) or overwritten with the delta type
1836  *      (type) when check_object() decided to reuse the delta.
1837  *
1838  *   3. Resetting our delta depth, as we are now a base object.
1839  */
1840 static void drop_reused_delta(struct object_entry *entry)
1841 {
1842         unsigned *idx = &to_pack.objects[entry->delta_idx - 1].delta_child_idx;
1843         struct object_info oi = OBJECT_INFO_INIT;
1844         enum object_type type;
1845         unsigned long size;
1846
1847         while (*idx) {
1848                 struct object_entry *oe = &to_pack.objects[*idx - 1];
1849
1850                 if (oe == entry)
1851                         *idx = oe->delta_sibling_idx;
1852                 else
1853                         idx = &oe->delta_sibling_idx;
1854         }
1855         SET_DELTA(entry, NULL);
1856         entry->depth = 0;
1857
1858         oi.sizep = &size;
1859         oi.typep = &type;
1860         if (packed_object_info(the_repository, IN_PACK(entry), entry->in_pack_offset, &oi) < 0) {
1861                 /*
1862                  * We failed to get the info from this pack for some reason;
1863                  * fall back to oid_object_info, which may find another copy.
1864                  * And if that fails, the error will be recorded in oe_type(entry)
1865                  * and dealt with in prepare_pack().
1866                  */
1867                 oe_set_type(entry,
1868                             oid_object_info(the_repository, &entry->idx.oid, &size));
1869         } else {
1870                 oe_set_type(entry, type);
1871         }
1872         SET_SIZE(entry, size);
1873 }
1874
1875 /*
1876  * Follow the chain of deltas from this entry onward, throwing away any links
1877  * that cause us to hit a cycle (as determined by the DFS state flags in
1878  * the entries).
1879  *
1880  * We also detect too-long reused chains that would violate our --depth
1881  * limit.
1882  */
1883 static void break_delta_chains(struct object_entry *entry)
1884 {
1885         /*
1886          * The actual depth of each object we will write is stored as an int,
1887          * as it cannot exceed our int "depth" limit. But before we break
1888          * changes based no that limit, we may potentially go as deep as the
1889          * number of objects, which is elsewhere bounded to a uint32_t.
1890          */
1891         uint32_t total_depth;
1892         struct object_entry *cur, *next;
1893
1894         for (cur = entry, total_depth = 0;
1895              cur;
1896              cur = DELTA(cur), total_depth++) {
1897                 if (cur->dfs_state == DFS_DONE) {
1898                         /*
1899                          * We've already seen this object and know it isn't
1900                          * part of a cycle. We do need to append its depth
1901                          * to our count.
1902                          */
1903                         total_depth += cur->depth;
1904                         break;
1905                 }
1906
1907                 /*
1908                  * We break cycles before looping, so an ACTIVE state (or any
1909                  * other cruft which made its way into the state variable)
1910                  * is a bug.
1911                  */
1912                 if (cur->dfs_state != DFS_NONE)
1913                         BUG("confusing delta dfs state in first pass: %d",
1914                             cur->dfs_state);
1915
1916                 /*
1917                  * Now we know this is the first time we've seen the object. If
1918                  * it's not a delta, we're done traversing, but we'll mark it
1919                  * done to save time on future traversals.
1920                  */
1921                 if (!DELTA(cur)) {
1922                         cur->dfs_state = DFS_DONE;
1923                         break;
1924                 }
1925
1926                 /*
1927                  * Mark ourselves as active and see if the next step causes
1928                  * us to cycle to another active object. It's important to do
1929                  * this _before_ we loop, because it impacts where we make the
1930                  * cut, and thus how our total_depth counter works.
1931                  * E.g., We may see a partial loop like:
1932                  *
1933                  *   A -> B -> C -> D -> B
1934                  *
1935                  * Cutting B->C breaks the cycle. But now the depth of A is
1936                  * only 1, and our total_depth counter is at 3. The size of the
1937                  * error is always one less than the size of the cycle we
1938                  * broke. Commits C and D were "lost" from A's chain.
1939                  *
1940                  * If we instead cut D->B, then the depth of A is correct at 3.
1941                  * We keep all commits in the chain that we examined.
1942                  */
1943                 cur->dfs_state = DFS_ACTIVE;
1944                 if (DELTA(cur)->dfs_state == DFS_ACTIVE) {
1945                         drop_reused_delta(cur);
1946                         cur->dfs_state = DFS_DONE;
1947                         break;
1948                 }
1949         }
1950
1951         /*
1952          * And now that we've gone all the way to the bottom of the chain, we
1953          * need to clear the active flags and set the depth fields as
1954          * appropriate. Unlike the loop above, which can quit when it drops a
1955          * delta, we need to keep going to look for more depth cuts. So we need
1956          * an extra "next" pointer to keep going after we reset cur->delta.
1957          */
1958         for (cur = entry; cur; cur = next) {
1959                 next = DELTA(cur);
1960
1961                 /*
1962                  * We should have a chain of zero or more ACTIVE states down to
1963                  * a final DONE. We can quit after the DONE, because either it
1964                  * has no bases, or we've already handled them in a previous
1965                  * call.
1966                  */
1967                 if (cur->dfs_state == DFS_DONE)
1968                         break;
1969                 else if (cur->dfs_state != DFS_ACTIVE)
1970                         BUG("confusing delta dfs state in second pass: %d",
1971                             cur->dfs_state);
1972
1973                 /*
1974                  * If the total_depth is more than depth, then we need to snip
1975                  * the chain into two or more smaller chains that don't exceed
1976                  * the maximum depth. Most of the resulting chains will contain
1977                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1978                  * the last chain (i.e., the one containing entry) will contain
1979                  * whatever entries are left over, namely
1980                  * (total_depth % (depth + 1)) of them.
1981                  *
1982                  * Since we are iterating towards decreasing depth, we need to
1983                  * decrement total_depth as we go, and we need to write to the
1984                  * entry what its final depth will be after all of the
1985                  * snipping. Since we're snipping into chains of length (depth
1986                  * + 1) entries, the final depth of an entry will be its
1987                  * original depth modulo (depth + 1). Any time we encounter an
1988                  * entry whose final depth is supposed to be zero, we snip it
1989                  * from its delta base, thereby making it so.
1990                  */
1991                 cur->depth = (total_depth--) % (depth + 1);
1992                 if (!cur->depth)
1993                         drop_reused_delta(cur);
1994
1995                 cur->dfs_state = DFS_DONE;
1996         }
1997 }
1998
1999 static void get_object_details(void)
2000 {
2001         uint32_t i;
2002         struct object_entry **sorted_by_offset;
2003
2004         if (progress)
2005                 progress_state = start_progress(_("Counting objects"),
2006                                                 to_pack.nr_objects);
2007
2008         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
2009         for (i = 0; i < to_pack.nr_objects; i++)
2010                 sorted_by_offset[i] = to_pack.objects + i;
2011         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
2012
2013         for (i = 0; i < to_pack.nr_objects; i++) {
2014                 struct object_entry *entry = sorted_by_offset[i];
2015                 check_object(entry);
2016                 if (entry->type_valid &&
2017                     oe_size_greater_than(&to_pack, entry, big_file_threshold))
2018                         entry->no_try_delta = 1;
2019                 display_progress(progress_state, i + 1);
2020         }
2021         stop_progress(&progress_state);
2022
2023         /*
2024          * This must happen in a second pass, since we rely on the delta
2025          * information for the whole list being completed.
2026          */
2027         for (i = 0; i < to_pack.nr_objects; i++)
2028                 break_delta_chains(&to_pack.objects[i]);
2029
2030         free(sorted_by_offset);
2031 }
2032
2033 /*
2034  * We search for deltas in a list sorted by type, by filename hash, and then
2035  * by size, so that we see progressively smaller and smaller files.
2036  * That's because we prefer deltas to be from the bigger file
2037  * to the smaller -- deletes are potentially cheaper, but perhaps
2038  * more importantly, the bigger file is likely the more recent
2039  * one.  The deepest deltas are therefore the oldest objects which are
2040  * less susceptible to be accessed often.
2041  */
2042 static int type_size_sort(const void *_a, const void *_b)
2043 {
2044         const struct object_entry *a = *(struct object_entry **)_a;
2045         const struct object_entry *b = *(struct object_entry **)_b;
2046         const enum object_type a_type = oe_type(a);
2047         const enum object_type b_type = oe_type(b);
2048         const unsigned long a_size = SIZE(a);
2049         const unsigned long b_size = SIZE(b);
2050
2051         if (a_type > b_type)
2052                 return -1;
2053         if (a_type < b_type)
2054                 return 1;
2055         if (a->hash > b->hash)
2056                 return -1;
2057         if (a->hash < b->hash)
2058                 return 1;
2059         if (a->preferred_base > b->preferred_base)
2060                 return -1;
2061         if (a->preferred_base < b->preferred_base)
2062                 return 1;
2063         if (use_delta_islands) {
2064                 const int island_cmp = island_delta_cmp(&a->idx.oid, &b->idx.oid);
2065                 if (island_cmp)
2066                         return island_cmp;
2067         }
2068         if (a_size > b_size)
2069                 return -1;
2070         if (a_size < b_size)
2071                 return 1;
2072         return a < b ? -1 : (a > b);  /* newest first */
2073 }
2074
2075 struct unpacked {
2076         struct object_entry *entry;
2077         void *data;
2078         struct delta_index *index;
2079         unsigned depth;
2080 };
2081
2082 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
2083                            unsigned long delta_size)
2084 {
2085         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
2086                 return 0;
2087
2088         if (delta_size < cache_max_small_delta_size)
2089                 return 1;
2090
2091         /* cache delta, if objects are large enough compared to delta size */
2092         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
2093                 return 1;
2094
2095         return 0;
2096 }
2097
2098 /* Protect delta_cache_size */
2099 static pthread_mutex_t cache_mutex;
2100 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
2101 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
2102
2103 /*
2104  * Protect object list partitioning (e.g. struct thread_param) and
2105  * progress_state
2106  */
2107 static pthread_mutex_t progress_mutex;
2108 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
2109 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
2110
2111 /*
2112  * Access to struct object_entry is unprotected since each thread owns
2113  * a portion of the main object list. Just don't access object entries
2114  * ahead in the list because they can be stolen and would need
2115  * progress_mutex for protection.
2116  */
2117
2118 /*
2119  * Return the size of the object without doing any delta
2120  * reconstruction (so non-deltas are true object sizes, but deltas
2121  * return the size of the delta data).
2122  */
2123 unsigned long oe_get_size_slow(struct packing_data *pack,
2124                                const struct object_entry *e)
2125 {
2126         struct packed_git *p;
2127         struct pack_window *w_curs;
2128         unsigned char *buf;
2129         enum object_type type;
2130         unsigned long used, avail, size;
2131
2132         if (e->type_ != OBJ_OFS_DELTA && e->type_ != OBJ_REF_DELTA) {
2133                 packing_data_lock(&to_pack);
2134                 if (oid_object_info(the_repository, &e->idx.oid, &size) < 0)
2135                         die(_("unable to get size of %s"),
2136                             oid_to_hex(&e->idx.oid));
2137                 packing_data_unlock(&to_pack);
2138                 return size;
2139         }
2140
2141         p = oe_in_pack(pack, e);
2142         if (!p)
2143                 BUG("when e->type is a delta, it must belong to a pack");
2144
2145         packing_data_lock(&to_pack);
2146         w_curs = NULL;
2147         buf = use_pack(p, &w_curs, e->in_pack_offset, &avail);
2148         used = unpack_object_header_buffer(buf, avail, &type, &size);
2149         if (used == 0)
2150                 die(_("unable to parse object header of %s"),
2151                     oid_to_hex(&e->idx.oid));
2152
2153         unuse_pack(&w_curs);
2154         packing_data_unlock(&to_pack);
2155         return size;
2156 }
2157
2158 static int try_delta(struct unpacked *trg, struct unpacked *src,
2159                      unsigned max_depth, unsigned long *mem_usage)
2160 {
2161         struct object_entry *trg_entry = trg->entry;
2162         struct object_entry *src_entry = src->entry;
2163         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
2164         unsigned ref_depth;
2165         enum object_type type;
2166         void *delta_buf;
2167
2168         /* Don't bother doing diffs between different types */
2169         if (oe_type(trg_entry) != oe_type(src_entry))
2170                 return -1;
2171
2172         /*
2173          * We do not bother to try a delta that we discarded on an
2174          * earlier try, but only when reusing delta data.  Note that
2175          * src_entry that is marked as the preferred_base should always
2176          * be considered, as even if we produce a suboptimal delta against
2177          * it, we will still save the transfer cost, as we already know
2178          * the other side has it and we won't send src_entry at all.
2179          */
2180         if (reuse_delta && IN_PACK(trg_entry) &&
2181             IN_PACK(trg_entry) == IN_PACK(src_entry) &&
2182             !src_entry->preferred_base &&
2183             trg_entry->in_pack_type != OBJ_REF_DELTA &&
2184             trg_entry->in_pack_type != OBJ_OFS_DELTA)
2185                 return 0;
2186
2187         /* Let's not bust the allowed depth. */
2188         if (src->depth >= max_depth)
2189                 return 0;
2190
2191         /* Now some size filtering heuristics. */
2192         trg_size = SIZE(trg_entry);
2193         if (!DELTA(trg_entry)) {
2194                 max_size = trg_size/2 - the_hash_algo->rawsz;
2195                 ref_depth = 1;
2196         } else {
2197                 max_size = DELTA_SIZE(trg_entry);
2198                 ref_depth = trg->depth;
2199         }
2200         max_size = (uint64_t)max_size * (max_depth - src->depth) /
2201                                                 (max_depth - ref_depth + 1);
2202         if (max_size == 0)
2203                 return 0;
2204         src_size = SIZE(src_entry);
2205         sizediff = src_size < trg_size ? trg_size - src_size : 0;
2206         if (sizediff >= max_size)
2207                 return 0;
2208         if (trg_size < src_size / 32)
2209                 return 0;
2210
2211         if (!in_same_island(&trg->entry->idx.oid, &src->entry->idx.oid))
2212                 return 0;
2213
2214         /* Load data if not already done */
2215         if (!trg->data) {
2216                 packing_data_lock(&to_pack);
2217                 trg->data = read_object_file(&trg_entry->idx.oid, &type, &sz);
2218                 packing_data_unlock(&to_pack);
2219                 if (!trg->data)
2220                         die(_("object %s cannot be read"),
2221                             oid_to_hex(&trg_entry->idx.oid));
2222                 if (sz != trg_size)
2223                         die(_("object %s inconsistent object length (%"PRIuMAX" vs %"PRIuMAX")"),
2224                             oid_to_hex(&trg_entry->idx.oid), (uintmax_t)sz,
2225                             (uintmax_t)trg_size);
2226                 *mem_usage += sz;
2227         }
2228         if (!src->data) {
2229                 packing_data_lock(&to_pack);
2230                 src->data = read_object_file(&src_entry->idx.oid, &type, &sz);
2231                 packing_data_unlock(&to_pack);
2232                 if (!src->data) {
2233                         if (src_entry->preferred_base) {
2234                                 static int warned = 0;
2235                                 if (!warned++)
2236                                         warning(_("object %s cannot be read"),
2237                                                 oid_to_hex(&src_entry->idx.oid));
2238                                 /*
2239                                  * Those objects are not included in the
2240                                  * resulting pack.  Be resilient and ignore
2241                                  * them if they can't be read, in case the
2242                                  * pack could be created nevertheless.
2243                                  */
2244                                 return 0;
2245                         }
2246                         die(_("object %s cannot be read"),
2247                             oid_to_hex(&src_entry->idx.oid));
2248                 }
2249                 if (sz != src_size)
2250                         die(_("object %s inconsistent object length (%"PRIuMAX" vs %"PRIuMAX")"),
2251                             oid_to_hex(&src_entry->idx.oid), (uintmax_t)sz,
2252                             (uintmax_t)src_size);
2253                 *mem_usage += sz;
2254         }
2255         if (!src->index) {
2256                 src->index = create_delta_index(src->data, src_size);
2257                 if (!src->index) {
2258                         static int warned = 0;
2259                         if (!warned++)
2260                                 warning(_("suboptimal pack - out of memory"));
2261                         return 0;
2262                 }
2263                 *mem_usage += sizeof_delta_index(src->index);
2264         }
2265
2266         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
2267         if (!delta_buf)
2268                 return 0;
2269
2270         if (DELTA(trg_entry)) {
2271                 /* Prefer only shallower same-sized deltas. */
2272                 if (delta_size == DELTA_SIZE(trg_entry) &&
2273                     src->depth + 1 >= trg->depth) {
2274                         free(delta_buf);
2275                         return 0;
2276                 }
2277         }
2278
2279         /*
2280          * Handle memory allocation outside of the cache
2281          * accounting lock.  Compiler will optimize the strangeness
2282          * away when NO_PTHREADS is defined.
2283          */
2284         free(trg_entry->delta_data);
2285         cache_lock();
2286         if (trg_entry->delta_data) {
2287                 delta_cache_size -= DELTA_SIZE(trg_entry);
2288                 trg_entry->delta_data = NULL;
2289         }
2290         if (delta_cacheable(src_size, trg_size, delta_size)) {
2291                 delta_cache_size += delta_size;
2292                 cache_unlock();
2293                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
2294         } else {
2295                 cache_unlock();
2296                 free(delta_buf);
2297         }
2298
2299         SET_DELTA(trg_entry, src_entry);
2300         SET_DELTA_SIZE(trg_entry, delta_size);
2301         trg->depth = src->depth + 1;
2302
2303         return 1;
2304 }
2305
2306 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
2307 {
2308         struct object_entry *child = DELTA_CHILD(me);
2309         unsigned int m = n;
2310         while (child) {
2311                 const unsigned int c = check_delta_limit(child, n + 1);
2312                 if (m < c)
2313                         m = c;
2314                 child = DELTA_SIBLING(child);
2315         }
2316         return m;
2317 }
2318
2319 static unsigned long free_unpacked(struct unpacked *n)
2320 {
2321         unsigned long freed_mem = sizeof_delta_index(n->index);
2322         free_delta_index(n->index);
2323         n->index = NULL;
2324         if (n->data) {
2325                 freed_mem += SIZE(n->entry);
2326                 FREE_AND_NULL(n->data);
2327         }
2328         n->entry = NULL;
2329         n->depth = 0;
2330         return freed_mem;
2331 }
2332
2333 static void find_deltas(struct object_entry **list, unsigned *list_size,
2334                         int window, int depth, unsigned *processed)
2335 {
2336         uint32_t i, idx = 0, count = 0;
2337         struct unpacked *array;
2338         unsigned long mem_usage = 0;
2339
2340         array = xcalloc(window, sizeof(struct unpacked));
2341
2342         for (;;) {
2343                 struct object_entry *entry;
2344                 struct unpacked *n = array + idx;
2345                 int j, max_depth, best_base = -1;
2346
2347                 progress_lock();
2348                 if (!*list_size) {
2349                         progress_unlock();
2350                         break;
2351                 }
2352                 entry = *list++;
2353                 (*list_size)--;
2354                 if (!entry->preferred_base) {
2355                         (*processed)++;
2356                         display_progress(progress_state, *processed);
2357                 }
2358                 progress_unlock();
2359
2360                 mem_usage -= free_unpacked(n);
2361                 n->entry = entry;
2362
2363                 while (window_memory_limit &&
2364                        mem_usage > window_memory_limit &&
2365                        count > 1) {
2366                         const uint32_t tail = (idx + window - count) % window;
2367                         mem_usage -= free_unpacked(array + tail);
2368                         count--;
2369                 }
2370
2371                 /* We do not compute delta to *create* objects we are not
2372                  * going to pack.
2373                  */
2374                 if (entry->preferred_base)
2375                         goto next;
2376
2377                 /*
2378                  * If the current object is at pack edge, take the depth the
2379                  * objects that depend on the current object into account
2380                  * otherwise they would become too deep.
2381                  */
2382                 max_depth = depth;
2383                 if (DELTA_CHILD(entry)) {
2384                         max_depth -= check_delta_limit(entry, 0);
2385                         if (max_depth <= 0)
2386                                 goto next;
2387                 }
2388
2389                 j = window;
2390                 while (--j > 0) {
2391                         int ret;
2392                         uint32_t other_idx = idx + j;
2393                         struct unpacked *m;
2394                         if (other_idx >= window)
2395                                 other_idx -= window;
2396                         m = array + other_idx;
2397                         if (!m->entry)
2398                                 break;
2399                         ret = try_delta(n, m, max_depth, &mem_usage);
2400                         if (ret < 0)
2401                                 break;
2402                         else if (ret > 0)
2403                                 best_base = other_idx;
2404                 }
2405
2406                 /*
2407                  * If we decided to cache the delta data, then it is best
2408                  * to compress it right away.  First because we have to do
2409                  * it anyway, and doing it here while we're threaded will
2410                  * save a lot of time in the non threaded write phase,
2411                  * as well as allow for caching more deltas within
2412                  * the same cache size limit.
2413                  * ...
2414                  * But only if not writing to stdout, since in that case
2415                  * the network is most likely throttling writes anyway,
2416                  * and therefore it is best to go to the write phase ASAP
2417                  * instead, as we can afford spending more time compressing
2418                  * between writes at that moment.
2419                  */
2420                 if (entry->delta_data && !pack_to_stdout) {
2421                         unsigned long size;
2422
2423                         size = do_compress(&entry->delta_data, DELTA_SIZE(entry));
2424                         if (size < (1U << OE_Z_DELTA_BITS)) {
2425                                 entry->z_delta_size = size;
2426                                 cache_lock();
2427                                 delta_cache_size -= DELTA_SIZE(entry);
2428                                 delta_cache_size += entry->z_delta_size;
2429                                 cache_unlock();
2430                         } else {
2431                                 FREE_AND_NULL(entry->delta_data);
2432                                 entry->z_delta_size = 0;
2433                         }
2434                 }
2435
2436                 /* if we made n a delta, and if n is already at max
2437                  * depth, leaving it in the window is pointless.  we
2438                  * should evict it first.
2439                  */
2440                 if (DELTA(entry) && max_depth <= n->depth)
2441                         continue;
2442
2443                 /*
2444                  * Move the best delta base up in the window, after the
2445                  * currently deltified object, to keep it longer.  It will
2446                  * be the first base object to be attempted next.
2447                  */
2448                 if (DELTA(entry)) {
2449                         struct unpacked swap = array[best_base];
2450                         int dist = (window + idx - best_base) % window;
2451                         int dst = best_base;
2452                         while (dist--) {
2453                                 int src = (dst + 1) % window;
2454                                 array[dst] = array[src];
2455                                 dst = src;
2456                         }
2457                         array[dst] = swap;
2458                 }
2459
2460                 next:
2461                 idx++;
2462                 if (count + 1 < window)
2463                         count++;
2464                 if (idx >= window)
2465                         idx = 0;
2466         }
2467
2468         for (i = 0; i < window; ++i) {
2469                 free_delta_index(array[i].index);
2470                 free(array[i].data);
2471         }
2472         free(array);
2473 }
2474
2475 /*
2476  * The main object list is split into smaller lists, each is handed to
2477  * one worker.
2478  *
2479  * The main thread waits on the condition that (at least) one of the workers
2480  * has stopped working (which is indicated in the .working member of
2481  * struct thread_params).
2482  *
2483  * When a work thread has completed its work, it sets .working to 0 and
2484  * signals the main thread and waits on the condition that .data_ready
2485  * becomes 1.
2486  *
2487  * The main thread steals half of the work from the worker that has
2488  * most work left to hand it to the idle worker.
2489  */
2490
2491 struct thread_params {
2492         pthread_t thread;
2493         struct object_entry **list;
2494         unsigned list_size;
2495         unsigned remaining;
2496         int window;
2497         int depth;
2498         int working;
2499         int data_ready;
2500         pthread_mutex_t mutex;
2501         pthread_cond_t cond;
2502         unsigned *processed;
2503 };
2504
2505 static pthread_cond_t progress_cond;
2506
2507 /*
2508  * Mutex and conditional variable can't be statically-initialized on Windows.
2509  */
2510 static void init_threaded_search(void)
2511 {
2512         pthread_mutex_init(&cache_mutex, NULL);
2513         pthread_mutex_init(&progress_mutex, NULL);
2514         pthread_cond_init(&progress_cond, NULL);
2515 }
2516
2517 static void cleanup_threaded_search(void)
2518 {
2519         pthread_cond_destroy(&progress_cond);
2520         pthread_mutex_destroy(&cache_mutex);
2521         pthread_mutex_destroy(&progress_mutex);
2522 }
2523
2524 static void *threaded_find_deltas(void *arg)
2525 {
2526         struct thread_params *me = arg;
2527
2528         progress_lock();
2529         while (me->remaining) {
2530                 progress_unlock();
2531
2532                 find_deltas(me->list, &me->remaining,
2533                             me->window, me->depth, me->processed);
2534
2535                 progress_lock();
2536                 me->working = 0;
2537                 pthread_cond_signal(&progress_cond);
2538                 progress_unlock();
2539
2540                 /*
2541                  * We must not set ->data_ready before we wait on the
2542                  * condition because the main thread may have set it to 1
2543                  * before we get here. In order to be sure that new
2544                  * work is available if we see 1 in ->data_ready, it
2545                  * was initialized to 0 before this thread was spawned
2546                  * and we reset it to 0 right away.
2547                  */
2548                 pthread_mutex_lock(&me->mutex);
2549                 while (!me->data_ready)
2550                         pthread_cond_wait(&me->cond, &me->mutex);
2551                 me->data_ready = 0;
2552                 pthread_mutex_unlock(&me->mutex);
2553
2554                 progress_lock();
2555         }
2556         progress_unlock();
2557         /* leave ->working 1 so that this doesn't get more work assigned */
2558         return NULL;
2559 }
2560
2561 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2562                            int window, int depth, unsigned *processed)
2563 {
2564         struct thread_params *p;
2565         int i, ret, active_threads = 0;
2566
2567         init_threaded_search();
2568
2569         if (delta_search_threads <= 1) {
2570                 find_deltas(list, &list_size, window, depth, processed);
2571                 cleanup_threaded_search();
2572                 return;
2573         }
2574         if (progress > pack_to_stdout)
2575                 fprintf_ln(stderr, _("Delta compression using up to %d threads"),
2576                            delta_search_threads);
2577         p = xcalloc(delta_search_threads, sizeof(*p));
2578
2579         /* Partition the work amongst work threads. */
2580         for (i = 0; i < delta_search_threads; i++) {
2581                 unsigned sub_size = list_size / (delta_search_threads - i);
2582
2583                 /* don't use too small segments or no deltas will be found */
2584                 if (sub_size < 2*window && i+1 < delta_search_threads)
2585                         sub_size = 0;
2586
2587                 p[i].window = window;
2588                 p[i].depth = depth;
2589                 p[i].processed = processed;
2590                 p[i].working = 1;
2591                 p[i].data_ready = 0;
2592
2593                 /* try to split chunks on "path" boundaries */
2594                 while (sub_size && sub_size < list_size &&
2595                        list[sub_size]->hash &&
2596                        list[sub_size]->hash == list[sub_size-1]->hash)
2597                         sub_size++;
2598
2599                 p[i].list = list;
2600                 p[i].list_size = sub_size;
2601                 p[i].remaining = sub_size;
2602
2603                 list += sub_size;
2604                 list_size -= sub_size;
2605         }
2606
2607         /* Start work threads. */
2608         for (i = 0; i < delta_search_threads; i++) {
2609                 if (!p[i].list_size)
2610                         continue;
2611                 pthread_mutex_init(&p[i].mutex, NULL);
2612                 pthread_cond_init(&p[i].cond, NULL);
2613                 ret = pthread_create(&p[i].thread, NULL,
2614                                      threaded_find_deltas, &p[i]);
2615                 if (ret)
2616                         die(_("unable to create thread: %s"), strerror(ret));
2617                 active_threads++;
2618         }
2619
2620         /*
2621          * Now let's wait for work completion.  Each time a thread is done
2622          * with its work, we steal half of the remaining work from the
2623          * thread with the largest number of unprocessed objects and give
2624          * it to that newly idle thread.  This ensure good load balancing
2625          * until the remaining object list segments are simply too short
2626          * to be worth splitting anymore.
2627          */
2628         while (active_threads) {
2629                 struct thread_params *target = NULL;
2630                 struct thread_params *victim = NULL;
2631                 unsigned sub_size = 0;
2632
2633                 progress_lock();
2634                 for (;;) {
2635                         for (i = 0; !target && i < delta_search_threads; i++)
2636                                 if (!p[i].working)
2637                                         target = &p[i];
2638                         if (target)
2639                                 break;
2640                         pthread_cond_wait(&progress_cond, &progress_mutex);
2641                 }
2642
2643                 for (i = 0; i < delta_search_threads; i++)
2644                         if (p[i].remaining > 2*window &&
2645                             (!victim || victim->remaining < p[i].remaining))
2646                                 victim = &p[i];
2647                 if (victim) {
2648                         sub_size = victim->remaining / 2;
2649                         list = victim->list + victim->list_size - sub_size;
2650                         while (sub_size && list[0]->hash &&
2651                                list[0]->hash == list[-1]->hash) {
2652                                 list++;
2653                                 sub_size--;
2654                         }
2655                         if (!sub_size) {
2656                                 /*
2657                                  * It is possible for some "paths" to have
2658                                  * so many objects that no hash boundary
2659                                  * might be found.  Let's just steal the
2660                                  * exact half in that case.
2661                                  */
2662                                 sub_size = victim->remaining / 2;
2663                                 list -= sub_size;
2664                         }
2665                         target->list = list;
2666                         victim->list_size -= sub_size;
2667                         victim->remaining -= sub_size;
2668                 }
2669                 target->list_size = sub_size;
2670                 target->remaining = sub_size;
2671                 target->working = 1;
2672                 progress_unlock();
2673
2674                 pthread_mutex_lock(&target->mutex);
2675                 target->data_ready = 1;
2676                 pthread_cond_signal(&target->cond);
2677                 pthread_mutex_unlock(&target->mutex);
2678
2679                 if (!sub_size) {
2680                         pthread_join(target->thread, NULL);
2681                         pthread_cond_destroy(&target->cond);
2682                         pthread_mutex_destroy(&target->mutex);
2683                         active_threads--;
2684                 }
2685         }
2686         cleanup_threaded_search();
2687         free(p);
2688 }
2689
2690 static int obj_is_packed(const struct object_id *oid)
2691 {
2692         return packlist_find(&to_pack, oid) ||
2693                 (reuse_packfile_bitmap &&
2694                  bitmap_walk_contains(bitmap_git, reuse_packfile_bitmap, oid));
2695 }
2696
2697 static void add_tag_chain(const struct object_id *oid)
2698 {
2699         struct tag *tag;
2700
2701         /*
2702          * We catch duplicates already in add_object_entry(), but we'd
2703          * prefer to do this extra check to avoid having to parse the
2704          * tag at all if we already know that it's being packed (e.g., if
2705          * it was included via bitmaps, we would not have parsed it
2706          * previously).
2707          */
2708         if (obj_is_packed(oid))
2709                 return;
2710
2711         tag = lookup_tag(the_repository, oid);
2712         while (1) {
2713                 if (!tag || parse_tag(tag) || !tag->tagged)
2714                         die(_("unable to pack objects reachable from tag %s"),
2715                             oid_to_hex(oid));
2716
2717                 add_object_entry(&tag->object.oid, OBJ_TAG, NULL, 0);
2718
2719                 if (tag->tagged->type != OBJ_TAG)
2720                         return;
2721
2722                 tag = (struct tag *)tag->tagged;
2723         }
2724 }
2725
2726 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2727 {
2728         struct object_id peeled;
2729
2730         if (starts_with(path, "refs/tags/") && /* is a tag? */
2731             !peel_ref(path, &peeled)    && /* peelable? */
2732             obj_is_packed(&peeled)) /* object packed? */
2733                 add_tag_chain(oid);
2734         return 0;
2735 }
2736
2737 static void prepare_pack(int window, int depth)
2738 {
2739         struct object_entry **delta_list;
2740         uint32_t i, nr_deltas;
2741         unsigned n;
2742
2743         if (use_delta_islands)
2744                 resolve_tree_islands(the_repository, progress, &to_pack);
2745
2746         get_object_details();
2747
2748         /*
2749          * If we're locally repacking then we need to be doubly careful
2750          * from now on in order to make sure no stealth corruption gets
2751          * propagated to the new pack.  Clients receiving streamed packs
2752          * should validate everything they get anyway so no need to incur
2753          * the additional cost here in that case.
2754          */
2755         if (!pack_to_stdout)
2756                 do_check_packed_object_crc = 1;
2757
2758         if (!to_pack.nr_objects || !window || !depth)
2759                 return;
2760
2761         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2762         nr_deltas = n = 0;
2763
2764         for (i = 0; i < to_pack.nr_objects; i++) {
2765                 struct object_entry *entry = to_pack.objects + i;
2766
2767                 if (DELTA(entry))
2768                         /* This happens if we decided to reuse existing
2769                          * delta from a pack.  "reuse_delta &&" is implied.
2770                          */
2771                         continue;
2772
2773                 if (!entry->type_valid ||
2774                     oe_size_less_than(&to_pack, entry, 50))
2775                         continue;
2776
2777                 if (entry->no_try_delta)
2778                         continue;
2779
2780                 if (!entry->preferred_base) {
2781                         nr_deltas++;
2782                         if (oe_type(entry) < 0)
2783                                 die(_("unable to get type of object %s"),
2784                                     oid_to_hex(&entry->idx.oid));
2785                 } else {
2786                         if (oe_type(entry) < 0) {
2787                                 /*
2788                                  * This object is not found, but we
2789                                  * don't have to include it anyway.
2790                                  */
2791                                 continue;
2792                         }
2793                 }
2794
2795                 delta_list[n++] = entry;
2796         }
2797
2798         if (nr_deltas && n > 1) {
2799                 unsigned nr_done = 0;
2800
2801                 if (progress)
2802                         progress_state = start_progress(_("Compressing objects"),
2803                                                         nr_deltas);
2804                 QSORT(delta_list, n, type_size_sort);
2805                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2806                 stop_progress(&progress_state);
2807                 if (nr_done != nr_deltas)
2808                         die(_("inconsistency with delta count"));
2809         }
2810         free(delta_list);
2811 }
2812
2813 static int git_pack_config(const char *k, const char *v, void *cb)
2814 {
2815         if (!strcmp(k, "pack.window")) {
2816                 window = git_config_int(k, v);
2817                 return 0;
2818         }
2819         if (!strcmp(k, "pack.windowmemory")) {
2820                 window_memory_limit = git_config_ulong(k, v);
2821                 return 0;
2822         }
2823         if (!strcmp(k, "pack.depth")) {
2824                 depth = git_config_int(k, v);
2825                 return 0;
2826         }
2827         if (!strcmp(k, "pack.deltacachesize")) {
2828                 max_delta_cache_size = git_config_int(k, v);
2829                 return 0;
2830         }
2831         if (!strcmp(k, "pack.deltacachelimit")) {
2832                 cache_max_small_delta_size = git_config_int(k, v);
2833                 return 0;
2834         }
2835         if (!strcmp(k, "pack.writebitmaphashcache")) {
2836                 if (git_config_bool(k, v))
2837                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2838                 else
2839                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2840         }
2841         if (!strcmp(k, "pack.usebitmaps")) {
2842                 use_bitmap_index_default = git_config_bool(k, v);
2843                 return 0;
2844         }
2845         if (!strcmp(k, "pack.allowpackreuse")) {
2846                 allow_pack_reuse = git_config_bool(k, v);
2847                 return 0;
2848         }
2849         if (!strcmp(k, "pack.threads")) {
2850                 delta_search_threads = git_config_int(k, v);
2851                 if (delta_search_threads < 0)
2852                         die(_("invalid number of threads specified (%d)"),
2853                             delta_search_threads);
2854                 if (!HAVE_THREADS && delta_search_threads != 1) {
2855                         warning(_("no threads support, ignoring %s"), k);
2856                         delta_search_threads = 0;
2857                 }
2858                 return 0;
2859         }
2860         if (!strcmp(k, "pack.indexversion")) {
2861                 pack_idx_opts.version = git_config_int(k, v);
2862                 if (pack_idx_opts.version > 2)
2863                         die(_("bad pack.indexversion=%"PRIu32),
2864                             pack_idx_opts.version);
2865                 return 0;
2866         }
2867         return git_default_config(k, v, cb);
2868 }
2869
2870 static void read_object_list_from_stdin(void)
2871 {
2872         char line[GIT_MAX_HEXSZ + 1 + PATH_MAX + 2];
2873         struct object_id oid;
2874         const char *p;
2875
2876         for (;;) {
2877                 if (!fgets(line, sizeof(line), stdin)) {
2878                         if (feof(stdin))
2879                                 break;
2880                         if (!ferror(stdin))
2881                                 die("BUG: fgets returned NULL, not EOF, not error!");
2882                         if (errno != EINTR)
2883                                 die_errno("fgets");
2884                         clearerr(stdin);
2885                         continue;
2886                 }
2887                 if (line[0] == '-') {
2888                         if (get_oid_hex(line+1, &oid))
2889                                 die(_("expected edge object ID, got garbage:\n %s"),
2890                                     line);
2891                         add_preferred_base(&oid);
2892                         continue;
2893                 }
2894                 if (parse_oid_hex(line, &oid, &p))
2895                         die(_("expected object ID, got garbage:\n %s"), line);
2896
2897                 add_preferred_base_object(p + 1);
2898                 add_object_entry(&oid, OBJ_NONE, p + 1, 0);
2899         }
2900 }
2901
2902 /* Remember to update object flag allocation in object.h */
2903 #define OBJECT_ADDED (1u<<20)
2904
2905 static void show_commit(struct commit *commit, void *data)
2906 {
2907         add_object_entry(&commit->object.oid, OBJ_COMMIT, NULL, 0);
2908         commit->object.flags |= OBJECT_ADDED;
2909
2910         if (write_bitmap_index)
2911                 index_commit_for_bitmap(commit);
2912
2913         if (use_delta_islands)
2914                 propagate_island_marks(commit);
2915 }
2916
2917 static void show_object(struct object *obj, const char *name, void *data)
2918 {
2919         add_preferred_base_object(name);
2920         add_object_entry(&obj->oid, obj->type, name, 0);
2921         obj->flags |= OBJECT_ADDED;
2922
2923         if (use_delta_islands) {
2924                 const char *p;
2925                 unsigned depth;
2926                 struct object_entry *ent;
2927
2928                 /* the empty string is a root tree, which is depth 0 */
2929                 depth = *name ? 1 : 0;
2930                 for (p = strchr(name, '/'); p; p = strchr(p + 1, '/'))
2931                         depth++;
2932
2933                 ent = packlist_find(&to_pack, &obj->oid);
2934                 if (ent && depth > oe_tree_depth(&to_pack, ent))
2935                         oe_set_tree_depth(&to_pack, ent, depth);
2936         }
2937 }
2938
2939 static void show_object__ma_allow_any(struct object *obj, const char *name, void *data)
2940 {
2941         assert(arg_missing_action == MA_ALLOW_ANY);
2942
2943         /*
2944          * Quietly ignore ALL missing objects.  This avoids problems with
2945          * staging them now and getting an odd error later.
2946          */
2947         if (!has_object_file(&obj->oid))
2948                 return;
2949
2950         show_object(obj, name, data);
2951 }
2952
2953 static void show_object__ma_allow_promisor(struct object *obj, const char *name, void *data)
2954 {
2955         assert(arg_missing_action == MA_ALLOW_PROMISOR);
2956
2957         /*
2958          * Quietly ignore EXPECTED missing objects.  This avoids problems with
2959          * staging them now and getting an odd error later.
2960          */
2961         if (!has_object_file(&obj->oid) && is_promisor_object(&obj->oid))
2962                 return;
2963
2964         show_object(obj, name, data);
2965 }
2966
2967 static int option_parse_missing_action(const struct option *opt,
2968                                        const char *arg, int unset)
2969 {
2970         assert(arg);
2971         assert(!unset);
2972
2973         if (!strcmp(arg, "error")) {
2974                 arg_missing_action = MA_ERROR;
2975                 fn_show_object = show_object;
2976                 return 0;
2977         }
2978
2979         if (!strcmp(arg, "allow-any")) {
2980                 arg_missing_action = MA_ALLOW_ANY;
2981                 fetch_if_missing = 0;
2982                 fn_show_object = show_object__ma_allow_any;
2983                 return 0;
2984         }
2985
2986         if (!strcmp(arg, "allow-promisor")) {
2987                 arg_missing_action = MA_ALLOW_PROMISOR;
2988                 fetch_if_missing = 0;
2989                 fn_show_object = show_object__ma_allow_promisor;
2990                 return 0;
2991         }
2992
2993         die(_("invalid value for --missing"));
2994         return 0;
2995 }
2996
2997 static void show_edge(struct commit *commit)
2998 {
2999         add_preferred_base(&commit->object.oid);
3000 }
3001
3002 struct in_pack_object {
3003         off_t offset;
3004         struct object *object;
3005 };
3006
3007 struct in_pack {
3008         unsigned int alloc;
3009         unsigned int nr;
3010         struct in_pack_object *array;
3011 };
3012
3013 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
3014 {
3015         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
3016         in_pack->array[in_pack->nr].object = object;
3017         in_pack->nr++;
3018 }
3019
3020 /*
3021  * Compare the objects in the offset order, in order to emulate the
3022  * "git rev-list --objects" output that produced the pack originally.
3023  */
3024 static int ofscmp(const void *a_, const void *b_)
3025 {
3026         struct in_pack_object *a = (struct in_pack_object *)a_;
3027         struct in_pack_object *b = (struct in_pack_object *)b_;
3028
3029         if (a->offset < b->offset)
3030                 return -1;
3031         else if (a->offset > b->offset)
3032                 return 1;
3033         else
3034                 return oidcmp(&a->object->oid, &b->object->oid);
3035 }
3036
3037 static void add_objects_in_unpacked_packs(void)
3038 {
3039         struct packed_git *p;
3040         struct in_pack in_pack;
3041         uint32_t i;
3042
3043         memset(&in_pack, 0, sizeof(in_pack));
3044
3045         for (p = get_all_packs(the_repository); p; p = p->next) {
3046                 struct object_id oid;
3047                 struct object *o;
3048
3049                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
3050                         continue;
3051                 if (open_pack_index(p))
3052                         die(_("cannot open pack index"));
3053
3054                 ALLOC_GROW(in_pack.array,
3055                            in_pack.nr + p->num_objects,
3056                            in_pack.alloc);
3057
3058                 for (i = 0; i < p->num_objects; i++) {
3059                         nth_packed_object_id(&oid, p, i);
3060                         o = lookup_unknown_object(&oid);
3061                         if (!(o->flags & OBJECT_ADDED))
3062                                 mark_in_pack_object(o, p, &in_pack);
3063                         o->flags |= OBJECT_ADDED;
3064                 }
3065         }
3066
3067         if (in_pack.nr) {
3068                 QSORT(in_pack.array, in_pack.nr, ofscmp);
3069                 for (i = 0; i < in_pack.nr; i++) {
3070                         struct object *o = in_pack.array[i].object;
3071                         add_object_entry(&o->oid, o->type, "", 0);
3072                 }
3073         }
3074         free(in_pack.array);
3075 }
3076
3077 static int add_loose_object(const struct object_id *oid, const char *path,
3078                             void *data)
3079 {
3080         enum object_type type = oid_object_info(the_repository, oid, NULL);
3081
3082         if (type < 0) {
3083                 warning(_("loose object at %s could not be examined"), path);
3084                 return 0;
3085         }
3086
3087         add_object_entry(oid, type, "", 0);
3088         return 0;
3089 }
3090
3091 /*
3092  * We actually don't even have to worry about reachability here.
3093  * add_object_entry will weed out duplicates, so we just add every
3094  * loose object we find.
3095  */
3096 static void add_unreachable_loose_objects(void)
3097 {
3098         for_each_loose_file_in_objdir(get_object_directory(),
3099                                       add_loose_object,
3100                                       NULL, NULL, NULL);
3101 }
3102
3103 static int has_sha1_pack_kept_or_nonlocal(const struct object_id *oid)
3104 {
3105         static struct packed_git *last_found = (void *)1;
3106         struct packed_git *p;
3107
3108         p = (last_found != (void *)1) ? last_found :
3109                                         get_all_packs(the_repository);
3110
3111         while (p) {
3112                 if ((!p->pack_local || p->pack_keep ||
3113                                 p->pack_keep_in_core) &&
3114                         find_pack_entry_one(oid->hash, p)) {
3115                         last_found = p;
3116                         return 1;
3117                 }
3118                 if (p == last_found)
3119                         p = get_all_packs(the_repository);
3120                 else
3121                         p = p->next;
3122                 if (p == last_found)
3123                         p = p->next;
3124         }
3125         return 0;
3126 }
3127
3128 /*
3129  * Store a list of sha1s that are should not be discarded
3130  * because they are either written too recently, or are
3131  * reachable from another object that was.
3132  *
3133  * This is filled by get_object_list.
3134  */
3135 static struct oid_array recent_objects;
3136
3137 static int loosened_object_can_be_discarded(const struct object_id *oid,
3138                                             timestamp_t mtime)
3139 {
3140         if (!unpack_unreachable_expiration)
3141                 return 0;
3142         if (mtime > unpack_unreachable_expiration)
3143                 return 0;
3144         if (oid_array_lookup(&recent_objects, oid) >= 0)
3145                 return 0;
3146         return 1;
3147 }
3148
3149 static void loosen_unused_packed_objects(void)
3150 {
3151         struct packed_git *p;
3152         uint32_t i;
3153         struct object_id oid;
3154
3155         for (p = get_all_packs(the_repository); p; p = p->next) {
3156                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
3157                         continue;
3158
3159                 if (open_pack_index(p))
3160                         die(_("cannot open pack index"));
3161
3162                 for (i = 0; i < p->num_objects; i++) {
3163                         nth_packed_object_id(&oid, p, i);
3164                         if (!packlist_find(&to_pack, &oid) &&
3165                             !has_sha1_pack_kept_or_nonlocal(&oid) &&
3166                             !loosened_object_can_be_discarded(&oid, p->mtime))
3167                                 if (force_object_loose(&oid, p->mtime))
3168                                         die(_("unable to force loose object"));
3169                 }
3170         }
3171 }
3172
3173 /*
3174  * This tracks any options which pack-reuse code expects to be on, or which a
3175  * reader of the pack might not understand, and which would therefore prevent
3176  * blind reuse of what we have on disk.
3177  */
3178 static int pack_options_allow_reuse(void)
3179 {
3180         return allow_pack_reuse &&
3181                pack_to_stdout &&
3182                !ignore_packed_keep_on_disk &&
3183                !ignore_packed_keep_in_core &&
3184                (!local || !have_non_local_packs) &&
3185                !incremental;
3186 }
3187
3188 static int get_object_list_from_bitmap(struct rev_info *revs)
3189 {
3190         if (!(bitmap_git = prepare_bitmap_walk(revs, &filter_options)))
3191                 return -1;
3192
3193         if (pack_options_allow_reuse() &&
3194             !reuse_partial_packfile_from_bitmap(
3195                         bitmap_git,
3196                         &reuse_packfile,
3197                         &reuse_packfile_objects,
3198                         &reuse_packfile_bitmap)) {
3199                 assert(reuse_packfile_objects);
3200                 nr_result += reuse_packfile_objects;
3201                 display_progress(progress_state, nr_result);
3202         }
3203
3204         traverse_bitmap_commit_list(bitmap_git, revs,
3205                                     &add_object_entry_from_bitmap);
3206         return 0;
3207 }
3208
3209 static void record_recent_object(struct object *obj,
3210                                  const char *name,
3211                                  void *data)
3212 {
3213         oid_array_append(&recent_objects, &obj->oid);
3214 }
3215
3216 static void record_recent_commit(struct commit *commit, void *data)
3217 {
3218         oid_array_append(&recent_objects, &commit->object.oid);
3219 }
3220
3221 static void get_object_list(int ac, const char **av)
3222 {
3223         struct rev_info revs;
3224         struct setup_revision_opt s_r_opt = {
3225                 .allow_exclude_promisor_objects = 1,
3226         };
3227         char line[1000];
3228         int flags = 0;
3229         int save_warning;
3230
3231         repo_init_revisions(the_repository, &revs, NULL);
3232         save_commit_buffer = 0;
3233         setup_revisions(ac, av, &revs, &s_r_opt);
3234
3235         /* make sure shallows are read */
3236         is_repository_shallow(the_repository);
3237
3238         save_warning = warn_on_object_refname_ambiguity;
3239         warn_on_object_refname_ambiguity = 0;
3240
3241         while (fgets(line, sizeof(line), stdin) != NULL) {
3242                 int len = strlen(line);
3243                 if (len && line[len - 1] == '\n')
3244                         line[--len] = 0;
3245                 if (!len)
3246                         break;
3247                 if (*line == '-') {
3248                         if (!strcmp(line, "--not")) {
3249                                 flags ^= UNINTERESTING;
3250                                 write_bitmap_index = 0;
3251                                 continue;
3252                         }
3253                         if (starts_with(line, "--shallow ")) {
3254                                 struct object_id oid;
3255                                 if (get_oid_hex(line + 10, &oid))
3256                                         die("not an SHA-1 '%s'", line + 10);
3257                                 register_shallow(the_repository, &oid);
3258                                 use_bitmap_index = 0;
3259                                 continue;
3260                         }
3261                         die(_("not a rev '%s'"), line);
3262                 }
3263                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
3264                         die(_("bad revision '%s'"), line);
3265         }
3266
3267         warn_on_object_refname_ambiguity = save_warning;
3268
3269         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
3270                 return;
3271
3272         if (use_delta_islands)
3273                 load_delta_islands(the_repository, progress);
3274
3275         if (prepare_revision_walk(&revs))
3276                 die(_("revision walk setup failed"));
3277         mark_edges_uninteresting(&revs, show_edge, sparse);
3278
3279         if (!fn_show_object)
3280                 fn_show_object = show_object;
3281         traverse_commit_list_filtered(&filter_options, &revs,
3282                                       show_commit, fn_show_object, NULL,
3283                                       NULL);
3284
3285         if (unpack_unreachable_expiration) {
3286                 revs.ignore_missing_links = 1;
3287                 if (add_unseen_recent_objects_to_traversal(&revs,
3288                                 unpack_unreachable_expiration))
3289                         die(_("unable to add recent objects"));
3290                 if (prepare_revision_walk(&revs))
3291                         die(_("revision walk setup failed"));
3292                 traverse_commit_list(&revs, record_recent_commit,
3293                                      record_recent_object, NULL);
3294         }
3295
3296         if (keep_unreachable)
3297                 add_objects_in_unpacked_packs();
3298         if (pack_loose_unreachable)
3299                 add_unreachable_loose_objects();
3300         if (unpack_unreachable)
3301                 loosen_unused_packed_objects();
3302
3303         oid_array_clear(&recent_objects);
3304 }
3305
3306 static void add_extra_kept_packs(const struct string_list *names)
3307 {
3308         struct packed_git *p;
3309
3310         if (!names->nr)
3311                 return;
3312
3313         for (p = get_all_packs(the_repository); p; p = p->next) {
3314                 const char *name = basename(p->pack_name);
3315                 int i;
3316
3317                 if (!p->pack_local)
3318                         continue;
3319
3320                 for (i = 0; i < names->nr; i++)
3321                         if (!fspathcmp(name, names->items[i].string))
3322                                 break;
3323
3324                 if (i < names->nr) {
3325                         p->pack_keep_in_core = 1;
3326                         ignore_packed_keep_in_core = 1;
3327                         continue;
3328                 }
3329         }
3330 }
3331
3332 static int option_parse_index_version(const struct option *opt,
3333                                       const char *arg, int unset)
3334 {
3335         char *c;
3336         const char *val = arg;
3337
3338         BUG_ON_OPT_NEG(unset);
3339
3340         pack_idx_opts.version = strtoul(val, &c, 10);
3341         if (pack_idx_opts.version > 2)
3342                 die(_("unsupported index version %s"), val);
3343         if (*c == ',' && c[1])
3344                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
3345         if (*c || pack_idx_opts.off32_limit & 0x80000000)
3346                 die(_("bad index version '%s'"), val);
3347         return 0;
3348 }
3349
3350 static int option_parse_unpack_unreachable(const struct option *opt,
3351                                            const char *arg, int unset)
3352 {
3353         if (unset) {
3354                 unpack_unreachable = 0;
3355                 unpack_unreachable_expiration = 0;
3356         }
3357         else {
3358                 unpack_unreachable = 1;
3359                 if (arg)
3360                         unpack_unreachable_expiration = approxidate(arg);
3361         }
3362         return 0;
3363 }
3364
3365 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
3366 {
3367         int use_internal_rev_list = 0;
3368         int shallow = 0;
3369         int all_progress_implied = 0;
3370         struct argv_array rp = ARGV_ARRAY_INIT;
3371         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
3372         int rev_list_index = 0;
3373         struct string_list keep_pack_list = STRING_LIST_INIT_NODUP;
3374         struct option pack_objects_options[] = {
3375                 OPT_SET_INT('q', "quiet", &progress,
3376                             N_("do not show progress meter"), 0),
3377                 OPT_SET_INT(0, "progress", &progress,
3378                             N_("show progress meter"), 1),
3379                 OPT_SET_INT(0, "all-progress", &progress,
3380                             N_("show progress meter during object writing phase"), 2),
3381                 OPT_BOOL(0, "all-progress-implied",
3382                          &all_progress_implied,
3383                          N_("similar to --all-progress when progress meter is shown")),
3384                 OPT_CALLBACK_F(0, "index-version", NULL, N_("<version>[,<offset>]"),
3385                   N_("write the pack index file in the specified idx format version"),
3386                   PARSE_OPT_NONEG, option_parse_index_version),
3387                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
3388                               N_("maximum size of each output pack file")),
3389                 OPT_BOOL(0, "local", &local,
3390                          N_("ignore borrowed objects from alternate object store")),
3391                 OPT_BOOL(0, "incremental", &incremental,
3392                          N_("ignore packed objects")),
3393                 OPT_INTEGER(0, "window", &window,
3394                             N_("limit pack window by objects")),
3395                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
3396                               N_("limit pack window by memory in addition to object limit")),
3397                 OPT_INTEGER(0, "depth", &depth,
3398                             N_("maximum length of delta chain allowed in the resulting pack")),
3399                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
3400                          N_("reuse existing deltas")),
3401                 OPT_BOOL(0, "reuse-object", &reuse_object,
3402                          N_("reuse existing objects")),
3403                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
3404                          N_("use OFS_DELTA objects")),
3405                 OPT_INTEGER(0, "threads", &delta_search_threads,
3406                             N_("use threads when searching for best delta matches")),
3407                 OPT_BOOL(0, "non-empty", &non_empty,
3408                          N_("do not create an empty pack output")),
3409                 OPT_BOOL(0, "revs", &use_internal_rev_list,
3410                          N_("read revision arguments from standard input")),
3411                 OPT_SET_INT_F(0, "unpacked", &rev_list_unpacked,
3412                               N_("limit the objects to those that are not yet packed"),
3413                               1, PARSE_OPT_NONEG),
3414                 OPT_SET_INT_F(0, "all", &rev_list_all,
3415                               N_("include objects reachable from any reference"),
3416                               1, PARSE_OPT_NONEG),
3417                 OPT_SET_INT_F(0, "reflog", &rev_list_reflog,
3418                               N_("include objects referred by reflog entries"),
3419                               1, PARSE_OPT_NONEG),
3420                 OPT_SET_INT_F(0, "indexed-objects", &rev_list_index,
3421                               N_("include objects referred to by the index"),
3422                               1, PARSE_OPT_NONEG),
3423                 OPT_BOOL(0, "stdout", &pack_to_stdout,
3424                          N_("output pack to stdout")),
3425                 OPT_BOOL(0, "include-tag", &include_tag,
3426                          N_("include tag objects that refer to objects to be packed")),
3427                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
3428                          N_("keep unreachable objects")),
3429                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
3430                          N_("pack loose unreachable objects")),
3431                 OPT_CALLBACK_F(0, "unpack-unreachable", NULL, N_("time"),
3432                   N_("unpack unreachable objects newer than <time>"),
3433                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable),
3434                 OPT_BOOL(0, "sparse", &sparse,
3435                          N_("use the sparse reachability algorithm")),
3436                 OPT_BOOL(0, "thin", &thin,
3437                          N_("create thin packs")),
3438                 OPT_BOOL(0, "shallow", &shallow,
3439                          N_("create packs suitable for shallow fetches")),
3440                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep_on_disk,
3441                          N_("ignore packs that have companion .keep file")),
3442                 OPT_STRING_LIST(0, "keep-pack", &keep_pack_list, N_("name"),
3443                                 N_("ignore this pack")),
3444                 OPT_INTEGER(0, "compression", &pack_compression_level,
3445                             N_("pack compression level")),
3446                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
3447                             N_("do not hide commits by grafts"), 0),
3448                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
3449                          N_("use a bitmap index if available to speed up counting objects")),
3450                 OPT_SET_INT(0, "write-bitmap-index", &write_bitmap_index,
3451                             N_("write a bitmap index together with the pack index"),
3452                             WRITE_BITMAP_TRUE),
3453                 OPT_SET_INT_F(0, "write-bitmap-index-quiet",
3454                               &write_bitmap_index,
3455                               N_("write a bitmap index if possible"),
3456                               WRITE_BITMAP_QUIET, PARSE_OPT_HIDDEN),
3457                 OPT_PARSE_LIST_OBJECTS_FILTER(&filter_options),
3458                 OPT_CALLBACK_F(0, "missing", NULL, N_("action"),
3459                   N_("handling for missing objects"), PARSE_OPT_NONEG,
3460                   option_parse_missing_action),
3461                 OPT_BOOL(0, "exclude-promisor-objects", &exclude_promisor_objects,
3462                          N_("do not pack objects in promisor packfiles")),
3463                 OPT_BOOL(0, "delta-islands", &use_delta_islands,
3464                          N_("respect islands during delta compression")),
3465                 OPT_END(),
3466         };
3467
3468         if (DFS_NUM_STATES > (1 << OE_DFS_STATE_BITS))
3469                 BUG("too many dfs states, increase OE_DFS_STATE_BITS");
3470
3471         read_replace_refs = 0;
3472
3473         sparse = git_env_bool("GIT_TEST_PACK_SPARSE", -1);
3474         prepare_repo_settings(the_repository);
3475         if (sparse < 0)
3476                 sparse = the_repository->settings.pack_use_sparse;
3477
3478         reset_pack_idx_option(&pack_idx_opts);
3479         git_config(git_pack_config, NULL);
3480
3481         progress = isatty(2);
3482         argc = parse_options(argc, argv, prefix, pack_objects_options,
3483                              pack_usage, 0);
3484
3485         if (argc) {
3486                 base_name = argv[0];
3487                 argc--;
3488         }
3489         if (pack_to_stdout != !base_name || argc)
3490                 usage_with_options(pack_usage, pack_objects_options);
3491
3492         if (depth >= (1 << OE_DEPTH_BITS)) {
3493                 warning(_("delta chain depth %d is too deep, forcing %d"),
3494                         depth, (1 << OE_DEPTH_BITS) - 1);
3495                 depth = (1 << OE_DEPTH_BITS) - 1;
3496         }
3497         if (cache_max_small_delta_size >= (1U << OE_Z_DELTA_BITS)) {
3498                 warning(_("pack.deltaCacheLimit is too high, forcing %d"),
3499                         (1U << OE_Z_DELTA_BITS) - 1);
3500                 cache_max_small_delta_size = (1U << OE_Z_DELTA_BITS) - 1;
3501         }
3502
3503         argv_array_push(&rp, "pack-objects");
3504         if (thin) {
3505                 use_internal_rev_list = 1;
3506                 argv_array_push(&rp, shallow
3507                                 ? "--objects-edge-aggressive"
3508                                 : "--objects-edge");
3509         } else
3510                 argv_array_push(&rp, "--objects");
3511
3512         if (rev_list_all) {
3513                 use_internal_rev_list = 1;
3514                 argv_array_push(&rp, "--all");
3515         }
3516         if (rev_list_reflog) {
3517                 use_internal_rev_list = 1;
3518                 argv_array_push(&rp, "--reflog");
3519         }
3520         if (rev_list_index) {
3521                 use_internal_rev_list = 1;
3522                 argv_array_push(&rp, "--indexed-objects");
3523         }
3524         if (rev_list_unpacked) {
3525                 use_internal_rev_list = 1;
3526                 argv_array_push(&rp, "--unpacked");
3527         }
3528
3529         if (exclude_promisor_objects) {
3530                 use_internal_rev_list = 1;
3531                 fetch_if_missing = 0;
3532                 argv_array_push(&rp, "--exclude-promisor-objects");
3533         }
3534         if (unpack_unreachable || keep_unreachable || pack_loose_unreachable)
3535                 use_internal_rev_list = 1;
3536
3537         if (!reuse_object)
3538                 reuse_delta = 0;
3539         if (pack_compression_level == -1)
3540                 pack_compression_level = Z_DEFAULT_COMPRESSION;
3541         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
3542                 die(_("bad pack compression level %d"), pack_compression_level);
3543
3544         if (!delta_search_threads)      /* --threads=0 means autodetect */
3545                 delta_search_threads = online_cpus();
3546
3547         if (!HAVE_THREADS && delta_search_threads != 1)
3548                 warning(_("no threads support, ignoring --threads"));
3549         if (!pack_to_stdout && !pack_size_limit)
3550                 pack_size_limit = pack_size_limit_cfg;
3551         if (pack_to_stdout && pack_size_limit)
3552                 die(_("--max-pack-size cannot be used to build a pack for transfer"));
3553         if (pack_size_limit && pack_size_limit < 1024*1024) {
3554                 warning(_("minimum pack size limit is 1 MiB"));
3555                 pack_size_limit = 1024*1024;
3556         }
3557
3558         if (!pack_to_stdout && thin)
3559                 die(_("--thin cannot be used to build an indexable pack"));
3560
3561         if (keep_unreachable && unpack_unreachable)
3562                 die(_("--keep-unreachable and --unpack-unreachable are incompatible"));
3563         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3564                 unpack_unreachable_expiration = 0;
3565
3566         if (filter_options.choice) {
3567                 if (!pack_to_stdout)
3568                         die(_("cannot use --filter without --stdout"));
3569         }
3570
3571         /*
3572          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3573          *
3574          * - to produce good pack (with bitmap index not-yet-packed objects are
3575          *   packed in suboptimal order).
3576          *
3577          * - to use more robust pack-generation codepath (avoiding possible
3578          *   bugs in bitmap code and possible bitmap index corruption).
3579          */
3580         if (!pack_to_stdout)
3581                 use_bitmap_index_default = 0;
3582
3583         if (use_bitmap_index < 0)
3584                 use_bitmap_index = use_bitmap_index_default;
3585
3586         /* "hard" reasons not to use bitmaps; these just won't work at all */
3587         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow(the_repository))
3588                 use_bitmap_index = 0;
3589
3590         if (pack_to_stdout || !rev_list_all)
3591                 write_bitmap_index = 0;
3592
3593         if (use_delta_islands)
3594                 argv_array_push(&rp, "--topo-order");
3595
3596         if (progress && all_progress_implied)
3597                 progress = 2;
3598
3599         add_extra_kept_packs(&keep_pack_list);
3600         if (ignore_packed_keep_on_disk) {
3601                 struct packed_git *p;
3602                 for (p = get_all_packs(the_repository); p; p = p->next)
3603                         if (p->pack_local && p->pack_keep)
3604                                 break;
3605                 if (!p) /* no keep-able packs found */
3606                         ignore_packed_keep_on_disk = 0;
3607         }
3608         if (local) {
3609                 /*
3610                  * unlike ignore_packed_keep_on_disk above, we do not
3611                  * want to unset "local" based on looking at packs, as
3612                  * it also covers non-local objects
3613                  */
3614                 struct packed_git *p;
3615                 for (p = get_all_packs(the_repository); p; p = p->next) {
3616                         if (!p->pack_local) {
3617                                 have_non_local_packs = 1;
3618                                 break;
3619                         }
3620                 }
3621         }
3622
3623         trace2_region_enter("pack-objects", "enumerate-objects",
3624                             the_repository);
3625         prepare_packing_data(the_repository, &to_pack);
3626
3627         if (progress)
3628                 progress_state = start_progress(_("Enumerating objects"), 0);
3629         if (!use_internal_rev_list)
3630                 read_object_list_from_stdin();
3631         else {
3632                 get_object_list(rp.argc, rp.argv);
3633                 argv_array_clear(&rp);
3634         }
3635         cleanup_preferred_base();
3636         if (include_tag && nr_result)
3637                 for_each_ref(add_ref_tag, NULL);
3638         stop_progress(&progress_state);
3639         trace2_region_leave("pack-objects", "enumerate-objects",
3640                             the_repository);
3641
3642         if (non_empty && !nr_result)
3643                 return 0;
3644         if (nr_result) {
3645                 trace2_region_enter("pack-objects", "prepare-pack",
3646                                     the_repository);
3647                 prepare_pack(window, depth);
3648                 trace2_region_leave("pack-objects", "prepare-pack",
3649                                     the_repository);
3650         }
3651
3652         trace2_region_enter("pack-objects", "write-pack-file", the_repository);
3653         write_pack_file();
3654         trace2_region_leave("pack-objects", "write-pack-file", the_repository);
3655
3656         if (progress)
3657                 fprintf_ln(stderr,
3658                            _("Total %"PRIu32" (delta %"PRIu32"),"
3659                              " reused %"PRIu32" (delta %"PRIu32"),"
3660                              " pack-reused %"PRIu32),
3661                            written, written_delta, reused, reused_delta,
3662                            reuse_packfile_objects);
3663         return 0;
3664 }