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 "strvec.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 #include "promisor-remote.h"
  39
  40 #define IN_PACK(obj) oe_in_pack(&to_pack, obj)
  41 #define SIZE(obj) oe_size(&to_pack, obj)
  42 #define SET_SIZE(obj,size) oe_set_size(&to_pack, obj, size)
  43 #define DELTA_SIZE(obj) oe_delta_size(&to_pack, obj)
  44 #define DELTA(obj) oe_delta(&to_pack, obj)
  45 #define DELTA_CHILD(obj) oe_delta_child(&to_pack, obj)
  46 #define DELTA_SIBLING(obj) oe_delta_sibling(&to_pack, obj)
  47 #define SET_DELTA(obj, val) oe_set_delta(&to_pack, obj, val)
  48 #define SET_DELTA_EXT(obj, oid) oe_set_delta_ext(&to_pack, obj, oid)
  49 #define SET_DELTA_SIZE(obj, val) oe_set_delta_size(&to_pack, obj, val)
  50 #define SET_DELTA_CHILD(obj, val) oe_set_delta_child(&to_pack, obj, val)
  51 #define SET_DELTA_SIBLING(obj, val) oe_set_delta_sibling(&to_pack, obj, val)
  52
  53 static const char *pack_usage[] = {
  54         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
  55         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
  56         NULL
  57 };
  58
  59 /*
  60  * Objects we are going to pack are collected in the `to_pack` structure.
  61  * It contains an array (dynamically expanded) of the object data, and a map
  62  * that can resolve SHA1s to their position in the array.
  63  */
  64 static struct packing_data to_pack;
  65
  66 static struct pack_idx_entry **written_list;
  67 static uint32_t nr_result, nr_written, nr_seen;
  68 static struct bitmap_index *bitmap_git;
  69 static uint32_t write_layer;
  70
  71 static int non_empty;
  72 static int reuse_delta = 1, reuse_object = 1;
  73 static int keep_unreachable, unpack_unreachable, include_tag;
  74 static timestamp_t unpack_unreachable_expiration;
  75 static int pack_loose_unreachable;
  76 static int local;
  77 static int have_non_local_packs;
  78 static int incremental;
  79 static int ignore_packed_keep_on_disk;
  80 static int ignore_packed_keep_in_core;
  81 static int allow_ofs_delta;
  82 static struct pack_idx_option pack_idx_opts;
  83 static const char *base_name;
  84 static int progress = 1;
  85 static int window = 10;
  86 static unsigned long pack_size_limit;
  87 static int depth = 50;
  88 static int delta_search_threads;
  89 static int pack_to_stdout;
  90 static int sparse;
  91 static int thin;
  92 static int num_preferred_base;
  93 static struct progress *progress_state;
  94
  95 static struct packed_git *reuse_packfile;
  96 static uint32_t reuse_packfile_objects;
  97 static struct bitmap *reuse_packfile_bitmap;
  98
  99 static int use_bitmap_index_default = 1;
 100 static int use_bitmap_index = -1;
 101 static int allow_pack_reuse = 1;
 102 static enum {
 103         WRITE_BITMAP_FALSE = 0,
 104         WRITE_BITMAP_QUIET,
 105         WRITE_BITMAP_TRUE,
 106 } write_bitmap_index;
 107 static uint16_t write_bitmap_options = BITMAP_OPT_HASH_CACHE;
 108
 109 static int exclude_promisor_objects;
 110
 111 static int use_delta_islands;
 112
 113 static unsigned long delta_cache_size = 0;
 114 static unsigned long max_delta_cache_size = DEFAULT_DELTA_CACHE_SIZE;
 115 static unsigned long cache_max_small_delta_size = 1000;
 116
 117 static unsigned long window_memory_limit = 0;
 118
 119 static struct list_objects_filter_options filter_options;
 120
 121 static struct string_list uri_protocols = STRING_LIST_INIT_NODUP;
 122
 123 enum missing_action {
 124         MA_ERROR = 0,      /* fail if any missing objects are encountered */
 125         MA_ALLOW_ANY,      /* silently allow ALL missing objects */
 126         MA_ALLOW_PROMISOR, /* silently allow all missing PROMISOR objects */
 127 };
 128 static enum missing_action arg_missing_action;
 129 static show_object_fn fn_show_object;
 130
 131 struct configured_exclusion {
 132         struct oidmap_entry e;
 133         char *pack_hash_hex;
 134         char *uri;
 135 };
 136 static struct oidmap configured_exclusions;
 137
 138 static struct oidset excluded_by_config;
 139
 140 /*
 141  * stats
 142  */
 143 static uint32_t written, written_delta;
 144 static uint32_t reused, reused_delta;
 145
 146 /*
 147  * Indexed commits
 148  */
 149 static struct commit **indexed_commits;
 150 static unsigned int indexed_commits_nr;
 151 static unsigned int indexed_commits_alloc;
 152
 153 static void index_commit_for_bitmap(struct commit *commit)
 154 {
 155         if (indexed_commits_nr >= indexed_commits_alloc) {
 156                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
 157                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
 158         }
 159
 160         indexed_commits[indexed_commits_nr++] = commit;
 161 }
 162
 163 static void *get_delta(struct object_entry *entry)
 164 {
 165         unsigned long size, base_size, delta_size;
 166         void *buf, *base_buf, *delta_buf;
 167         enum object_type type;
 168
 169         buf = read_object_file(&entry->idx.oid, &type, &size);
 170         if (!buf)
 171                 die(_("unable to read %s"), oid_to_hex(&entry->idx.oid));
 172         base_buf = read_object_file(&DELTA(entry)->idx.oid, &type,
 173                                     &base_size);
 174         if (!base_buf)
 175                 die("unable to read %s",
 176                     oid_to_hex(&DELTA(entry)->idx.oid));
 177         delta_buf = diff_delta(base_buf, base_size,
 178                                buf, size, &delta_size, 0);
 179         /*
 180          * We successfully computed this delta once but dropped it for
 181          * memory reasons. Something is very wrong if this time we
 182          * recompute and create a different delta.
 183          */
 184         if (!delta_buf || delta_size != DELTA_SIZE(entry))
 185                 BUG("delta size changed");
 186         free(buf);
 187         free(base_buf);
 188         return delta_buf;
 189 }
 190
 191 static unsigned long do_compress(void **pptr, unsigned long size)
 192 {
 193         git_zstream stream;
 194         void *in, *out;
 195         unsigned long maxsize;
 196
 197         git_deflate_init(&stream, pack_compression_level);
 198         maxsize = git_deflate_bound(&stream, size);
 199
 200         in = *pptr;
 201         out = xmalloc(maxsize);
 202         *pptr = out;
 203
 204         stream.next_in = in;
 205         stream.avail_in = size;
 206         stream.next_out = out;
 207         stream.avail_out = maxsize;
 208         while (git_deflate(&stream, Z_FINISH) == Z_OK)
 209                 ; /* nothing */
 210         git_deflate_end(&stream);
 211
 212         free(in);
 213         return stream.total_out;
 214 }
 215
 216 static unsigned long write_large_blob_data(struct git_istream *st, struct hashfile *f,
 217                                            const struct object_id *oid)
 218 {
 219         git_zstream stream;
 220         unsigned char ibuf[1024 * 16];
 221         unsigned char obuf[1024 * 16];
 222         unsigned long olen = 0;
 223
 224         git_deflate_init(&stream, pack_compression_level);
 225
 226         for (;;) {
 227                 ssize_t readlen;
 228                 int zret = Z_OK;
 229                 readlen = read_istream(st, ibuf, sizeof(ibuf));
 230                 if (readlen == -1)
 231                         die(_("unable to read %s"), oid_to_hex(oid));
 232
 233                 stream.next_in = ibuf;
 234                 stream.avail_in = readlen;
 235                 while ((stream.avail_in || readlen == 0) &&
 236                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
 237                         stream.next_out = obuf;
 238                         stream.avail_out = sizeof(obuf);
 239                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 240                         hashwrite(f, obuf, stream.next_out - obuf);
 241                         olen += stream.next_out - obuf;
 242                 }
 243                 if (stream.avail_in)
 244                         die(_("deflate error (%d)"), zret);
 245                 if (readlen == 0) {
 246                         if (zret != Z_STREAM_END)
 247                                 die(_("deflate error (%d)"), zret);
 248                         break;
 249                 }
 250         }
 251         git_deflate_end(&stream);
 252         return olen;
 253 }
 254
 255 /*
 256  * we are going to reuse the existing object data as is.  make
 257  * sure it is not corrupt.
 258  */
 259 static int check_pack_inflate(struct packed_git *p,
 260                 struct pack_window **w_curs,
 261                 off_t offset,
 262                 off_t len,
 263                 unsigned long expect)
 264 {
 265         git_zstream stream;
 266         unsigned char fakebuf[4096], *in;
 267         int st;
 268
 269         memset(&stream, 0, sizeof(stream));
 270         git_inflate_init(&stream);
 271         do {
 272                 in = use_pack(p, w_curs, offset, &stream.avail_in);
 273                 stream.next_in = in;
 274                 stream.next_out = fakebuf;
 275                 stream.avail_out = sizeof(fakebuf);
 276                 st = git_inflate(&stream, Z_FINISH);
 277                 offset += stream.next_in - in;
 278         } while (st == Z_OK || st == Z_BUF_ERROR);
 279         git_inflate_end(&stream);
 280         return (st == Z_STREAM_END &&
 281                 stream.total_out == expect &&
 282                 stream.total_in == len) ? 0 : -1;
 283 }
 284
 285 static void copy_pack_data(struct hashfile *f,
 286                 struct packed_git *p,
 287                 struct pack_window **w_curs,
 288                 off_t offset,
 289                 off_t len)
 290 {
 291         unsigned char *in;
 292         unsigned long avail;
 293
 294         while (len) {
 295                 in = use_pack(p, w_curs, offset, &avail);
 296                 if (avail > len)
 297                         avail = (unsigned long)len;
 298                 hashwrite(f, in, avail);
 299                 offset += avail;
 300                 len -= avail;
 301         }
 302 }
 303
 304 /* Return 0 if we will bust the pack-size limit */
 305 static unsigned long write_no_reuse_object(struct hashfile *f, struct object_entry *entry,
 306                                            unsigned long limit, int usable_delta)
 307 {
 308         unsigned long size, datalen;
 309         unsigned char header[MAX_PACK_OBJECT_HEADER],
 310                       dheader[MAX_PACK_OBJECT_HEADER];
 311         unsigned hdrlen;
 312         enum object_type type;
 313         void *buf;
 314         struct git_istream *st = NULL;
 315         const unsigned hashsz = the_hash_algo->rawsz;
 316
 317         if (!usable_delta) {
 318                 if (oe_type(entry) == OBJ_BLOB &&
 319                     oe_size_greater_than(&to_pack, entry, big_file_threshold) &&
 320                     (st = open_istream(the_repository, &entry->idx.oid, &type,
 321                                        &size, NULL)) != NULL)
 322                         buf = NULL;
 323                 else {
 324                         buf = read_object_file(&entry->idx.oid, &type, &size);
 325                         if (!buf)
 326                                 die(_("unable to read %s"),
 327                                     oid_to_hex(&entry->idx.oid));
 328                 }
 329                 /*
 330                  * make sure no cached delta data remains from a
 331                  * previous attempt before a pack split occurred.
 332                  */
 333                 FREE_AND_NULL(entry->delta_data);
 334                 entry->z_delta_size = 0;
 335         } else if (entry->delta_data) {
 336                 size = DELTA_SIZE(entry);
 337                 buf = entry->delta_data;
 338                 entry->delta_data = NULL;
 339                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
 340                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 341         } else {
 342                 buf = get_delta(entry);
 343                 size = DELTA_SIZE(entry);
 344                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
 345                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 346         }
 347
 348         if (st) /* large blob case, just assume we don't compress well */
 349                 datalen = size;
 350         else if (entry->z_delta_size)
 351                 datalen = entry->z_delta_size;
 352         else
 353                 datalen = do_compress(&buf, size);
 354
 355         /*
 356          * The object header is a byte of 'type' followed by zero or
 357          * more bytes of length.
 358          */
 359         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 360                                               type, size);
 361
 362         if (type == OBJ_OFS_DELTA) {
 363                 /*
 364                  * Deltas with relative base contain an additional
 365                  * encoding of the relative offset for the delta
 366                  * base from this object's position in the pack.
 367                  */
 368                 off_t ofs = entry->idx.offset - DELTA(entry)->idx.offset;
 369                 unsigned pos = sizeof(dheader) - 1;
 370                 dheader[pos] = ofs & 127;
 371                 while (ofs >>= 7)
 372                         dheader[--pos] = 128 | (--ofs & 127);
 373                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + hashsz >= limit) {
 374                         if (st)
 375                                 close_istream(st);
 376                         free(buf);
 377                         return 0;
 378                 }
 379                 hashwrite(f, header, hdrlen);
 380                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
 381                 hdrlen += sizeof(dheader) - pos;
 382         } else if (type == OBJ_REF_DELTA) {
 383                 /*
 384                  * Deltas with a base reference contain
 385                  * additional bytes for the base object ID.
 386                  */
 387                 if (limit && hdrlen + hashsz + datalen + hashsz >= limit) {
 388                         if (st)
 389                                 close_istream(st);
 390                         free(buf);
 391                         return 0;
 392                 }
 393                 hashwrite(f, header, hdrlen);
 394                 hashwrite(f, DELTA(entry)->idx.oid.hash, hashsz);
 395                 hdrlen += hashsz;
 396         } else {
 397                 if (limit && hdrlen + datalen + hashsz >= limit) {
 398                         if (st)
 399                                 close_istream(st);
 400                         free(buf);
 401                         return 0;
 402                 }
 403                 hashwrite(f, header, hdrlen);
 404         }
 405         if (st) {
 406                 datalen = write_large_blob_data(st, f, &entry->idx.oid);
 407                 close_istream(st);
 408         } else {
 409                 hashwrite(f, buf, datalen);
 410                 free(buf);
 411         }
 412
 413         return hdrlen + datalen;
 414 }
 415
 416 /* Return 0 if we will bust the pack-size limit */
 417 static off_t write_reuse_object(struct hashfile *f, struct object_entry *entry,
 418                                 unsigned long limit, int usable_delta)
 419 {
 420         struct packed_git *p = IN_PACK(entry);
 421         struct pack_window *w_curs = NULL;
 422         struct revindex_entry *revidx;
 423         off_t offset;
 424         enum object_type type = oe_type(entry);
 425         off_t datalen;
 426         unsigned char header[MAX_PACK_OBJECT_HEADER],
 427                       dheader[MAX_PACK_OBJECT_HEADER];
 428         unsigned hdrlen;
 429         const unsigned hashsz = the_hash_algo->rawsz;
 430         unsigned long entry_size = SIZE(entry);
 431
 432         if (DELTA(entry))
 433                 type = (allow_ofs_delta && DELTA(entry)->idx.offset) ?
 434                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 435         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 436                                               type, entry_size);
 437
 438         offset = entry->in_pack_offset;
 439         revidx = find_pack_revindex(p, offset);
 440         datalen = revidx[1].offset - offset;
 441         if (!pack_to_stdout && p->index_version > 1 &&
 442             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 443                 error(_("bad packed object CRC for %s"),
 444                       oid_to_hex(&entry->idx.oid));
 445                 unuse_pack(&w_curs);
 446                 return write_no_reuse_object(f, entry, limit, usable_delta);
 447         }
 448
 449         offset += entry->in_pack_header_size;
 450         datalen -= entry->in_pack_header_size;
 451
 452         if (!pack_to_stdout && p->index_version == 1 &&
 453             check_pack_inflate(p, &w_curs, offset, datalen, entry_size)) {
 454                 error(_("corrupt packed object for %s"),
 455                       oid_to_hex(&entry->idx.oid));
 456                 unuse_pack(&w_curs);
 457                 return write_no_reuse_object(f, entry, limit, usable_delta);
 458         }
 459
 460         if (type == OBJ_OFS_DELTA) {
 461                 off_t ofs = entry->idx.offset - DELTA(entry)->idx.offset;
 462                 unsigned pos = sizeof(dheader) - 1;
 463                 dheader[pos] = ofs & 127;
 464                 while (ofs >>= 7)
 465                         dheader[--pos] = 128 | (--ofs & 127);
 466                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + hashsz >= limit) {
 467                         unuse_pack(&w_curs);
 468                         return 0;
 469                 }
 470                 hashwrite(f, header, hdrlen);
 471                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
 472                 hdrlen += sizeof(dheader) - pos;
 473                 reused_delta++;
 474         } else if (type == OBJ_REF_DELTA) {
 475                 if (limit && hdrlen + hashsz + datalen + hashsz >= limit) {
 476                         unuse_pack(&w_curs);
 477                         return 0;
 478                 }
 479                 hashwrite(f, header, hdrlen);
 480                 hashwrite(f, DELTA(entry)->idx.oid.hash, hashsz);
 481                 hdrlen += hashsz;
 482                 reused_delta++;
 483         } else {
 484                 if (limit && hdrlen + datalen + hashsz >= limit) {
 485                         unuse_pack(&w_curs);
 486                         return 0;
 487                 }
 488                 hashwrite(f, header, hdrlen);
 489         }
 490         copy_pack_data(f, p, &w_curs, offset, datalen);
 491         unuse_pack(&w_curs);
 492         reused++;
 493         return hdrlen + datalen;
 494 }
 495
 496 /* Return 0 if we will bust the pack-size limit */
 497 static off_t write_object(struct hashfile *f,
 498                           struct object_entry *entry,
 499                           off_t write_offset)
 500 {
 501         unsigned long limit;
 502         off_t len;
 503         int usable_delta, to_reuse;
 504
 505         if (!pack_to_stdout)
 506                 crc32_begin(f);
 507
 508         /* apply size limit if limited packsize and not first object */
 509         if (!pack_size_limit || !nr_written)
 510                 limit = 0;
 511         else if (pack_size_limit <= write_offset)
 512                 /*
 513                  * the earlier object did not fit the limit; avoid
 514                  * mistaking this with unlimited (i.e. limit = 0).
 515                  */
 516                 limit = 1;
 517         else
 518                 limit = pack_size_limit - write_offset;
 519
 520         if (!DELTA(entry))
 521                 usable_delta = 0;       /* no delta */
 522         else if (!pack_size_limit)
 523                usable_delta = 1;        /* unlimited packfile */
 524         else if (DELTA(entry)->idx.offset == (off_t)-1)
 525                 usable_delta = 0;       /* base was written to another pack */
 526         else if (DELTA(entry)->idx.offset)
 527                 usable_delta = 1;       /* base already exists in this pack */
 528         else
 529                 usable_delta = 0;       /* base could end up in another pack */
 530
 531         if (!reuse_object)
 532                 to_reuse = 0;   /* explicit */
 533         else if (!IN_PACK(entry))
 534                 to_reuse = 0;   /* can't reuse what we don't have */
 535         else if (oe_type(entry) == OBJ_REF_DELTA ||
 536                  oe_type(entry) == OBJ_OFS_DELTA)
 537                                 /* check_object() decided it for us ... */
 538                 to_reuse = usable_delta;
 539                                 /* ... but pack split may override that */
 540         else if (oe_type(entry) != entry->in_pack_type)
 541                 to_reuse = 0;   /* pack has delta which is unusable */
 542         else if (DELTA(entry))
 543                 to_reuse = 0;   /* we want to pack afresh */
 544         else
 545                 to_reuse = 1;   /* we have it in-pack undeltified,
 546                                  * and we do not need to deltify it.
 547                                  */
 548
 549         if (!to_reuse)
 550                 len = write_no_reuse_object(f, entry, limit, usable_delta);
 551         else
 552                 len = write_reuse_object(f, entry, limit, usable_delta);
 553         if (!len)
 554                 return 0;
 555
 556         if (usable_delta)
 557                 written_delta++;
 558         written++;
 559         if (!pack_to_stdout)
 560                 entry->idx.crc32 = crc32_end(f);
 561         return len;
 562 }
 563
 564 enum write_one_status {
 565         WRITE_ONE_SKIP = -1, /* already written */
 566         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 567         WRITE_ONE_WRITTEN = 1, /* normal */
 568         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 569 };
 570
 571 static enum write_one_status write_one(struct hashfile *f,
 572                                        struct object_entry *e,
 573                                        off_t *offset)
 574 {
 575         off_t size;
 576         int recursing;
 577
 578         /*
 579          * we set offset to 1 (which is an impossible value) to mark
 580          * the fact that this object is involved in "write its base
 581          * first before writing a deltified object" recursion.
 582          */
 583         recursing = (e->idx.offset == 1);
 584         if (recursing) {
 585                 warning(_("recursive delta detected for object %s"),
 586                         oid_to_hex(&e->idx.oid));
 587                 return WRITE_ONE_RECURSIVE;
 588         } else if (e->idx.offset || e->preferred_base) {
 589                 /* offset is non zero if object is written already. */
 590                 return WRITE_ONE_SKIP;
 591         }
 592
 593         /* if we are deltified, write out base object first. */
 594         if (DELTA(e)) {
 595                 e->idx.offset = 1; /* now recurse */
 596                 switch (write_one(f, DELTA(e), offset)) {
 597                 case WRITE_ONE_RECURSIVE:
 598                         /* we cannot depend on this one */
 599                         SET_DELTA(e, NULL);
 600                         break;
 601                 default:
 602                         break;
 603                 case WRITE_ONE_BREAK:
 604                         e->idx.offset = recursing;
 605                         return WRITE_ONE_BREAK;
 606                 }
 607         }
 608
 609         e->idx.offset = *offset;
 610         size = write_object(f, e, *offset);
 611         if (!size) {
 612                 e->idx.offset = recursing;
 613                 return WRITE_ONE_BREAK;
 614         }
 615         written_list[nr_written++] = &e->idx;
 616
 617         /* make sure off_t is sufficiently large not to wrap */
 618         if (signed_add_overflows(*offset, size))
 619                 die(_("pack too large for current definition of off_t"));
 620         *offset += size;
 621         return WRITE_ONE_WRITTEN;
 622 }
 623
 624 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
 625                        void *cb_data)
 626 {
 627         struct object_id peeled;
 628         struct object_entry *entry = packlist_find(&to_pack, oid);
 629
 630         if (entry)
 631                 entry->tagged = 1;
 632         if (!peel_ref(path, &peeled)) {
 633                 entry = packlist_find(&to_pack, &peeled);
 634                 if (entry)
 635                         entry->tagged = 1;
 636         }
 637         return 0;
 638 }
 639
 640 static inline void add_to_write_order(struct object_entry **wo,
 641                                unsigned int *endp,
 642                                struct object_entry *e)
 643 {
 644         if (e->filled || oe_layer(&to_pack, e) != write_layer)
 645                 return;
 646         wo[(*endp)++] = e;
 647         e->filled = 1;
 648 }
 649
 650 static void add_descendants_to_write_order(struct object_entry **wo,
 651                                            unsigned int *endp,
 652                                            struct object_entry *e)
 653 {
 654         int add_to_order = 1;
 655         while (e) {
 656                 if (add_to_order) {
 657                         struct object_entry *s;
 658                         /* add this node... */
 659                         add_to_write_order(wo, endp, e);
 660                         /* all its siblings... */
 661                         for (s = DELTA_SIBLING(e); s; s = DELTA_SIBLING(s)) {
 662                                 add_to_write_order(wo, endp, s);
 663                         }
 664                 }
 665                 /* drop down a level to add left subtree nodes if possible */
 666                 if (DELTA_CHILD(e)) {
 667                         add_to_order = 1;
 668                         e = DELTA_CHILD(e);
 669                 } else {
 670                         add_to_order = 0;
 671                         /* our sibling might have some children, it is next */
 672                         if (DELTA_SIBLING(e)) {
 673                                 e = DELTA_SIBLING(e);
 674                                 continue;
 675                         }
 676                         /* go back to our parent node */
 677                         e = DELTA(e);
 678                         while (e && !DELTA_SIBLING(e)) {
 679                                 /* we're on the right side of a subtree, keep
 680                                  * going up until we can go right again */
 681                                 e = DELTA(e);
 682                         }
 683                         if (!e) {
 684                                 /* done- we hit our original root node */
 685                                 return;
 686                         }
 687                         /* pass it off to sibling at this level */
 688                         e = DELTA_SIBLING(e);
 689                 }
 690         };
 691 }
 692
 693 static void add_family_to_write_order(struct object_entry **wo,
 694                                       unsigned int *endp,
 695                                       struct object_entry *e)
 696 {
 697         struct object_entry *root;
 698
 699         for (root = e; DELTA(root); root = DELTA(root))
 700                 ; /* nothing */
 701         add_descendants_to_write_order(wo, endp, root);
 702 }
 703
 704 static void compute_layer_order(struct object_entry **wo, unsigned int *wo_end)
 705 {
 706         unsigned int i, last_untagged;
 707         struct object_entry *objects = to_pack.objects;
 708
 709         for (i = 0; i < to_pack.nr_objects; i++) {
 710                 if (objects[i].tagged)
 711                         break;
 712                 add_to_write_order(wo, wo_end, &objects[i]);
 713         }
 714         last_untagged = i;
 715
 716         /*
 717          * Then fill all the tagged tips.
 718          */
 719         for (; i < to_pack.nr_objects; i++) {
 720                 if (objects[i].tagged)
 721                         add_to_write_order(wo, wo_end, &objects[i]);
 722         }
 723
 724         /*
 725          * And then all remaining commits and tags.
 726          */
 727         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 728                 if (oe_type(&objects[i]) != OBJ_COMMIT &&
 729                     oe_type(&objects[i]) != OBJ_TAG)
 730                         continue;
 731                 add_to_write_order(wo, wo_end, &objects[i]);
 732         }
 733
 734         /*
 735          * And then all the trees.
 736          */
 737         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 738                 if (oe_type(&objects[i]) != OBJ_TREE)
 739                         continue;
 740                 add_to_write_order(wo, wo_end, &objects[i]);
 741         }
 742
 743         /*
 744          * Finally all the rest in really tight order
 745          */
 746         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 747                 if (!objects[i].filled && oe_layer(&to_pack, &objects[i]) == write_layer)
 748                         add_family_to_write_order(wo, wo_end, &objects[i]);
 749         }
 750 }
 751
 752 static struct object_entry **compute_write_order(void)
 753 {
 754         uint32_t max_layers = 1;
 755         unsigned int i, wo_end;
 756
 757         struct object_entry **wo;
 758         struct object_entry *objects = to_pack.objects;
 759
 760         for (i = 0; i < to_pack.nr_objects; i++) {
 761                 objects[i].tagged = 0;
 762                 objects[i].filled = 0;
 763                 SET_DELTA_CHILD(&objects[i], NULL);
 764                 SET_DELTA_SIBLING(&objects[i], NULL);
 765         }
 766
 767         /*
 768          * Fully connect delta_child/delta_sibling network.
 769          * Make sure delta_sibling is sorted in the original
 770          * recency order.
 771          */
 772         for (i = to_pack.nr_objects; i > 0;) {
 773                 struct object_entry *e = &objects[--i];
 774                 if (!DELTA(e))
 775                         continue;
 776                 /* Mark me as the first child */
 777                 e->delta_sibling_idx = DELTA(e)->delta_child_idx;
 778                 SET_DELTA_CHILD(DELTA(e), e);
 779         }
 780
 781         /*
 782          * Mark objects that are at the tip of tags.
 783          */
 784         for_each_tag_ref(mark_tagged, NULL);
 785
 786         if (use_delta_islands)
 787                 max_layers = compute_pack_layers(&to_pack);
 788
 789         ALLOC_ARRAY(wo, to_pack.nr_objects);
 790         wo_end = 0;
 791
 792         for (; write_layer < max_layers; ++write_layer)
 793                 compute_layer_order(wo, &wo_end);
 794
 795         if (wo_end != to_pack.nr_objects)
 796                 die(_("ordered %u objects, expected %"PRIu32),
 797                     wo_end, to_pack.nr_objects);
 798
 799         return wo;
 800 }
 801
 802
 803 /*
 804  * A reused set of objects. All objects in a chunk have the same
 805  * relative position in the original packfile and the generated
 806  * packfile.
 807  */
 808
 809 static struct reused_chunk {
 810         /* The offset of the first object of this chunk in the original
 811          * packfile. */
 812         off_t original;
 813         /* The offset of the first object of this chunk in the generated
 814          * packfile minus "original". */
 815         off_t difference;
 816 } *reused_chunks;
 817 static int reused_chunks_nr;
 818 static int reused_chunks_alloc;
 819
 820 static void record_reused_object(off_t where, off_t offset)
 821 {
 822         if (reused_chunks_nr && reused_chunks[reused_chunks_nr-1].difference == offset)
 823                 return;
 824
 825         ALLOC_GROW(reused_chunks, reused_chunks_nr + 1,
 826                    reused_chunks_alloc);
 827         reused_chunks[reused_chunks_nr].original = where;
 828         reused_chunks[reused_chunks_nr].difference = offset;
 829         reused_chunks_nr++;
 830 }
 831
 832 /*
 833  * Binary search to find the chunk that "where" is in. Note
 834  * that we're not looking for an exact match, just the first
 835  * chunk that contains it (which implicitly ends at the start
 836  * of the next chunk.
 837  */
 838 static off_t find_reused_offset(off_t where)
 839 {
 840         int lo = 0, hi = reused_chunks_nr;
 841         while (lo < hi) {
 842                 int mi = lo + ((hi - lo) / 2);
 843                 if (where == reused_chunks[mi].original)
 844                         return reused_chunks[mi].difference;
 845                 if (where < reused_chunks[mi].original)
 846                         hi = mi;
 847                 else
 848                         lo = mi + 1;
 849         }
 850
 851         /*
 852          * The first chunk starts at zero, so we can't have gone below
 853          * there.
 854          */
 855         assert(lo);
 856         return reused_chunks[lo-1].difference;
 857 }
 858
 859 static void write_reused_pack_one(size_t pos, struct hashfile *out,
 860                                   struct pack_window **w_curs)
 861 {
 862         off_t offset, next, cur;
 863         enum object_type type;
 864         unsigned long size;
 865
 866         offset = reuse_packfile->revindex[pos].offset;
 867         next = reuse_packfile->revindex[pos + 1].offset;
 868
 869         record_reused_object(offset, offset - hashfile_total(out));
 870
 871         cur = offset;
 872         type = unpack_object_header(reuse_packfile, w_curs, &cur, &size);
 873         assert(type >= 0);
 874
 875         if (type == OBJ_OFS_DELTA) {
 876                 off_t base_offset;
 877                 off_t fixup;
 878
 879                 unsigned char header[MAX_PACK_OBJECT_HEADER];
 880                 unsigned len;
 881
 882                 base_offset = get_delta_base(reuse_packfile, w_curs, &cur, type, offset);
 883                 assert(base_offset != 0);
 884
 885                 /* Convert to REF_DELTA if we must... */
 886                 if (!allow_ofs_delta) {
 887                         int base_pos = find_revindex_position(reuse_packfile, base_offset);
 888                         struct object_id base_oid;
 889
 890                         nth_packed_object_id(&base_oid, reuse_packfile,
 891                                              reuse_packfile->revindex[base_pos].nr);
 892
 893                         len = encode_in_pack_object_header(header, sizeof(header),
 894                                                            OBJ_REF_DELTA, size);
 895                         hashwrite(out, header, len);
 896                         hashwrite(out, base_oid.hash, the_hash_algo->rawsz);
 897                         copy_pack_data(out, reuse_packfile, w_curs, cur, next - cur);
 898                         return;
 899                 }
 900
 901                 /* Otherwise see if we need to rewrite the offset... */
 902                 fixup = find_reused_offset(offset) -
 903                         find_reused_offset(base_offset);
 904                 if (fixup) {
 905                         unsigned char ofs_header[10];
 906                         unsigned i, ofs_len;
 907                         off_t ofs = offset - base_offset - fixup;
 908
 909                         len = encode_in_pack_object_header(header, sizeof(header),
 910                                                            OBJ_OFS_DELTA, size);
 911
 912                         i = sizeof(ofs_header) - 1;
 913                         ofs_header[i] = ofs & 127;
 914                         while (ofs >>= 7)
 915                                 ofs_header[--i] = 128 | (--ofs & 127);
 916
 917                         ofs_len = sizeof(ofs_header) - i;
 918
 919                         hashwrite(out, header, len);
 920                         hashwrite(out, ofs_header + sizeof(ofs_header) - ofs_len, ofs_len);
 921                         copy_pack_data(out, reuse_packfile, w_curs, cur, next - cur);
 922                         return;
 923                 }
 924
 925                 /* ...otherwise we have no fixup, and can write it verbatim */
 926         }
 927
 928         copy_pack_data(out, reuse_packfile, w_curs, offset, next - offset);
 929 }
 930
 931 static size_t write_reused_pack_verbatim(struct hashfile *out,
 932                                          struct pack_window **w_curs)
 933 {
 934         size_t pos = 0;
 935
 936         while (pos < reuse_packfile_bitmap->word_alloc &&
 937                         reuse_packfile_bitmap->words[pos] == (eword_t)~0)
 938                 pos++;
 939
 940         if (pos) {
 941                 off_t to_write;
 942
 943                 written = (pos * BITS_IN_EWORD);
 944                 to_write = reuse_packfile->revindex[written].offset
 945                         - sizeof(struct pack_header);
 946
 947                 /* We're recording one chunk, not one object. */
 948                 record_reused_object(sizeof(struct pack_header), 0);
 949                 hashflush(out);
 950                 copy_pack_data(out, reuse_packfile, w_curs,
 951                         sizeof(struct pack_header), to_write);
 952
 953                 display_progress(progress_state, written);
 954         }
 955         return pos;
 956 }
 957
 958 static void write_reused_pack(struct hashfile *f)
 959 {
 960         size_t i = 0;
 961         uint32_t offset;
 962         struct pack_window *w_curs = NULL;
 963
 964         if (allow_ofs_delta)
 965                 i = write_reused_pack_verbatim(f, &w_curs);
 966
 967         for (; i < reuse_packfile_bitmap->word_alloc; ++i) {
 968                 eword_t word = reuse_packfile_bitmap->words[i];
 969                 size_t pos = (i * BITS_IN_EWORD);
 970
 971                 for (offset = 0; offset < BITS_IN_EWORD; ++offset) {
 972                         if ((word >> offset) == 0)
 973                                 break;
 974
 975                         offset += ewah_bit_ctz64(word >> offset);
 976                         write_reused_pack_one(pos + offset, f, &w_curs);
 977                         display_progress(progress_state, ++written);
 978                 }
 979         }
 980
 981         unuse_pack(&w_curs);
 982 }
 983
 984 static void write_excluded_by_configs(void)
 985 {
 986         struct oidset_iter iter;
 987         const struct object_id *oid;
 988
 989         oidset_iter_init(&excluded_by_config, &iter);
 990         while ((oid = oidset_iter_next(&iter))) {
 991                 struct configured_exclusion *ex =
 992                         oidmap_get(&configured_exclusions, oid);
 993
 994                 if (!ex)
 995                         BUG("configured exclusion wasn't configured");
 996                 write_in_full(1, ex->pack_hash_hex, strlen(ex->pack_hash_hex));
 997                 write_in_full(1, " ", 1);
 998                 write_in_full(1, ex->uri, strlen(ex->uri));
 999                 write_in_full(1, "\n", 1);
1000         }
1001 }
1002
1003 static const char no_split_warning[] = N_(
1004 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
1005 );
1006
1007 static void write_pack_file(void)
1008 {
1009         uint32_t i = 0, j;
1010         struct hashfile *f;
1011         off_t offset;
1012         uint32_t nr_remaining = nr_result;
1013         time_t last_mtime = 0;
1014         struct object_entry **write_order;
1015
1016         if (progress > pack_to_stdout)
1017                 progress_state = start_progress(_("Writing objects"), nr_result);
1018         ALLOC_ARRAY(written_list, to_pack.nr_objects);
1019         write_order = compute_write_order();
1020
1021         do {
1022                 struct object_id oid;
1023                 char *pack_tmp_name = NULL;
1024
1025                 if (pack_to_stdout)
1026                         f = hashfd_throughput(1, "<stdout>", progress_state);
1027                 else
1028                         f = create_tmp_packfile(&pack_tmp_name);
1029
1030                 offset = write_pack_header(f, nr_remaining);
1031
1032                 if (reuse_packfile) {
1033                         assert(pack_to_stdout);
1034                         write_reused_pack(f);
1035                         offset = hashfile_total(f);
1036                 }
1037
1038                 nr_written = 0;
1039                 for (; i < to_pack.nr_objects; i++) {
1040                         struct object_entry *e = write_order[i];
1041                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
1042                                 break;
1043                         display_progress(progress_state, written);
1044                 }
1045
1046                 /*
1047                  * Did we write the wrong # entries in the header?
1048                  * If so, rewrite it like in fast-import
1049                  */
1050                 if (pack_to_stdout) {
1051                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_CLOSE);
1052                 } else if (nr_written == nr_remaining) {
1053                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE);
1054                 } else {
1055                         int fd = finalize_hashfile(f, oid.hash, 0);
1056                         fixup_pack_header_footer(fd, oid.hash, pack_tmp_name,
1057                                                  nr_written, oid.hash, offset);
1058                         close(fd);
1059                         if (write_bitmap_index) {
1060                                 if (write_bitmap_index != WRITE_BITMAP_QUIET)
1061                                         warning(_(no_split_warning));
1062                                 write_bitmap_index = 0;
1063                         }
1064                 }
1065
1066                 if (!pack_to_stdout) {
1067                         struct stat st;
1068                         struct strbuf tmpname = STRBUF_INIT;
1069
1070                         /*
1071                          * Packs are runtime accessed in their mtime
1072                          * order since newer packs are more likely to contain
1073                          * younger objects.  So if we are creating multiple
1074                          * packs then we should modify the mtime of later ones
1075                          * to preserve this property.
1076                          */
1077                         if (stat(pack_tmp_name, &st) < 0) {
1078                                 warning_errno(_("failed to stat %s"), pack_tmp_name);
1079                         } else if (!last_mtime) {
1080                                 last_mtime = st.st_mtime;
1081                         } else {
1082                                 struct utimbuf utb;
1083                                 utb.actime = st.st_atime;
1084                                 utb.modtime = --last_mtime;
1085                                 if (utime(pack_tmp_name, &utb) < 0)
1086                                         warning_errno(_("failed utime() on %s"), pack_tmp_name);
1087                         }
1088
1089                         strbuf_addf(&tmpname, "%s-", base_name);
1090
1091                         if (write_bitmap_index) {
1092                                 bitmap_writer_set_checksum(oid.hash);
1093                                 bitmap_writer_build_type_index(
1094                                         &to_pack, written_list, nr_written);
1095                         }
1096
1097                         finish_tmp_packfile(&tmpname, pack_tmp_name,
1098                                             written_list, nr_written,
1099                                             &pack_idx_opts, oid.hash);
1100
1101                         if (write_bitmap_index) {
1102                                 strbuf_addf(&tmpname, "%s.bitmap", oid_to_hex(&oid));
1103
1104                                 stop_progress(&progress_state);
1105
1106                                 bitmap_writer_show_progress(progress);
1107                                 bitmap_writer_reuse_bitmaps(&to_pack);
1108                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
1109                                 bitmap_writer_build(&to_pack);
1110                                 bitmap_writer_finish(written_list, nr_written,
1111                                                      tmpname.buf, write_bitmap_options);
1112                                 write_bitmap_index = 0;
1113                         }
1114
1115                         strbuf_release(&tmpname);
1116                         free(pack_tmp_name);
1117                         puts(oid_to_hex(&oid));
1118                 }
1119
1120                 /* mark written objects as written to previous pack */
1121                 for (j = 0; j < nr_written; j++) {
1122                         written_list[j]->offset = (off_t)-1;
1123                 }
1124                 nr_remaining -= nr_written;
1125         } while (nr_remaining && i < to_pack.nr_objects);
1126
1127         free(written_list);
1128         free(write_order);
1129         stop_progress(&progress_state);
1130         if (written != nr_result)
1131                 die(_("wrote %"PRIu32" objects while expecting %"PRIu32),
1132                     written, nr_result);
1133         trace2_data_intmax("pack-objects", the_repository,
1134                            "write_pack_file/wrote", nr_result);
1135 }
1136
1137 static int no_try_delta(const char *path)
1138 {
1139         static struct attr_check *check;
1140
1141         if (!check)
1142                 check = attr_check_initl("delta", NULL);
1143         git_check_attr(the_repository->index, path, check);
1144         if (ATTR_FALSE(check->items[0].value))
1145                 return 1;
1146         return 0;
1147 }
1148
1149 /*
1150  * When adding an object, check whether we have already added it
1151  * to our packing list. If so, we can skip. However, if we are
1152  * being asked to excludei t, but the previous mention was to include
1153  * it, make sure to adjust its flags and tweak our numbers accordingly.
1154  *
1155  * As an optimization, we pass out the index position where we would have
1156  * found the item, since that saves us from having to look it up again a
1157  * few lines later when we want to add the new entry.
1158  */
1159 static int have_duplicate_entry(const struct object_id *oid,
1160                                 int exclude)
1161 {
1162         struct object_entry *entry;
1163
1164         if (reuse_packfile_bitmap &&
1165             bitmap_walk_contains(bitmap_git, reuse_packfile_bitmap, oid))
1166                 return 1;
1167
1168         entry = packlist_find(&to_pack, oid);
1169         if (!entry)
1170                 return 0;
1171
1172         if (exclude) {
1173                 if (!entry->preferred_base)
1174                         nr_result--;
1175                 entry->preferred_base = 1;
1176         }
1177
1178         return 1;
1179 }
1180
1181 static int want_found_object(int exclude, struct packed_git *p)
1182 {
1183         if (exclude)
1184                 return 1;
1185         if (incremental)
1186                 return 0;
1187
1188         /*
1189          * When asked to do --local (do not include an object that appears in a
1190          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
1191          * an object that appears in a pack marked with .keep), finding a pack
1192          * that matches the criteria is sufficient for us to decide to omit it.
1193          * However, even if this pack does not satisfy the criteria, we need to
1194          * make sure no copy of this object appears in _any_ pack that makes us
1195          * to omit the object, so we need to check all the packs.
1196          *
1197          * We can however first check whether these options can possible matter;
1198          * if they do not matter we know we want the object in generated pack.
1199          * Otherwise, we signal "-1" at the end to tell the caller that we do
1200          * not know either way, and it needs to check more packs.
1201          */
1202         if (!ignore_packed_keep_on_disk &&
1203             !ignore_packed_keep_in_core &&
1204             (!local || !have_non_local_packs))
1205                 return 1;
1206
1207         if (local && !p->pack_local)
1208                 return 0;
1209         if (p->pack_local &&
1210             ((ignore_packed_keep_on_disk && p->pack_keep) ||
1211              (ignore_packed_keep_in_core && p->pack_keep_in_core)))
1212                 return 0;
1213
1214         /* we don't know yet; keep looking for more packs */
1215         return -1;
1216 }
1217
1218 /*
1219  * Check whether we want the object in the pack (e.g., we do not want
1220  * objects found in non-local stores if the "--local" option was used).
1221  *
1222  * If the caller already knows an existing pack it wants to take the object
1223  * from, that is passed in *found_pack and *found_offset; otherwise this
1224  * function finds if there is any pack that has the object and returns the pack
1225  * and its offset in these variables.
1226  */
1227 static int want_object_in_pack(const struct object_id *oid,
1228                                int exclude,
1229                                struct packed_git **found_pack,
1230                                off_t *found_offset)
1231 {
1232         int want;
1233         struct list_head *pos;
1234         struct multi_pack_index *m;
1235
1236         if (!exclude && local && has_loose_object_nonlocal(oid))
1237                 return 0;
1238
1239         /*
1240          * If we already know the pack object lives in, start checks from that
1241          * pack - in the usual case when neither --local was given nor .keep files
1242          * are present we will determine the answer right now.
1243          */
1244         if (*found_pack) {
1245                 want = want_found_object(exclude, *found_pack);
1246                 if (want != -1)
1247                         return want;
1248         }
1249
1250         for (m = get_multi_pack_index(the_repository); m; m = m->next) {
1251                 struct pack_entry e;
1252                 if (fill_midx_entry(the_repository, oid, &e, m)) {
1253                         struct packed_git *p = e.p;
1254                         off_t offset;
1255
1256                         if (p == *found_pack)
1257                                 offset = *found_offset;
1258                         else
1259                                 offset = find_pack_entry_one(oid->hash, p);
1260
1261                         if (offset) {
1262                                 if (!*found_pack) {
1263                                         if (!is_pack_valid(p))
1264                                                 continue;
1265                                         *found_offset = offset;
1266                                         *found_pack = p;
1267                                 }
1268                                 want = want_found_object(exclude, p);
1269                                 if (want != -1)
1270                                         return want;
1271                         }
1272                 }
1273         }
1274
1275         list_for_each(pos, get_packed_git_mru(the_repository)) {
1276                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
1277                 off_t offset;
1278
1279                 if (p == *found_pack)
1280                         offset = *found_offset;
1281                 else
1282                         offset = find_pack_entry_one(oid->hash, p);
1283
1284                 if (offset) {
1285                         if (!*found_pack) {
1286                                 if (!is_pack_valid(p))
1287                                         continue;
1288                                 *found_offset = offset;
1289                                 *found_pack = p;
1290                         }
1291                         want = want_found_object(exclude, p);
1292                         if (!exclude && want > 0)
1293                                 list_move(&p->mru,
1294                                           get_packed_git_mru(the_repository));
1295                         if (want != -1)
1296                                 return want;
1297                 }
1298         }
1299
1300         if (uri_protocols.nr) {
1301                 struct configured_exclusion *ex =
1302                         oidmap_get(&configured_exclusions, oid);
1303                 int i;
1304                 const char *p;
1305
1306                 if (ex) {
1307                         for (i = 0; i < uri_protocols.nr; i++) {
1308                                 if (skip_prefix(ex->uri,
1309                                                 uri_protocols.items[i].string,
1310                                                 &p) &&
1311                                     *p == ':') {
1312                                         oidset_insert(&excluded_by_config, oid);
1313                                         return 0;
1314                                 }
1315                         }
1316                 }
1317         }
1318
1319         return 1;
1320 }
1321
1322 static void create_object_entry(const struct object_id *oid,
1323                                 enum object_type type,
1324                                 uint32_t hash,
1325                                 int exclude,
1326                                 int no_try_delta,
1327                                 struct packed_git *found_pack,
1328                                 off_t found_offset)
1329 {
1330         struct object_entry *entry;
1331
1332         entry = packlist_alloc(&to_pack, oid);
1333         entry->hash = hash;
1334         oe_set_type(entry, type);
1335         if (exclude)
1336                 entry->preferred_base = 1;
1337         else
1338                 nr_result++;
1339         if (found_pack) {
1340                 oe_set_in_pack(&to_pack, entry, found_pack);
1341                 entry->in_pack_offset = found_offset;
1342         }
1343
1344         entry->no_try_delta = no_try_delta;
1345 }
1346
1347 static const char no_closure_warning[] = N_(
1348 "disabling bitmap writing, as some objects are not being packed"
1349 );
1350
1351 static int add_object_entry(const struct object_id *oid, enum object_type type,
1352                             const char *name, int exclude)
1353 {
1354         struct packed_git *found_pack = NULL;
1355         off_t found_offset = 0;
1356
1357         display_progress(progress_state, ++nr_seen);
1358
1359         if (have_duplicate_entry(oid, exclude))
1360                 return 0;
1361
1362         if (!want_object_in_pack(oid, exclude, &found_pack, &found_offset)) {
1363                 /* The pack is missing an object, so it will not have closure */
1364                 if (write_bitmap_index) {
1365                         if (write_bitmap_index != WRITE_BITMAP_QUIET)
1366                                 warning(_(no_closure_warning));
1367                         write_bitmap_index = 0;
1368                 }
1369                 return 0;
1370         }
1371
1372         create_object_entry(oid, type, pack_name_hash(name),
1373                             exclude, name && no_try_delta(name),
1374                             found_pack, found_offset);
1375         return 1;
1376 }
1377
1378 static int add_object_entry_from_bitmap(const struct object_id *oid,
1379                                         enum object_type type,
1380                                         int flags, uint32_t name_hash,
1381                                         struct packed_git *pack, off_t offset)
1382 {
1383         display_progress(progress_state, ++nr_seen);
1384
1385         if (have_duplicate_entry(oid, 0))
1386                 return 0;
1387
1388         if (!want_object_in_pack(oid, 0, &pack, &offset))
1389                 return 0;
1390
1391         create_object_entry(oid, type, name_hash, 0, 0, pack, offset);
1392         return 1;
1393 }
1394
1395 struct pbase_tree_cache {
1396         struct object_id oid;
1397         int ref;
1398         int temporary;
1399         void *tree_data;
1400         unsigned long tree_size;
1401 };
1402
1403 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1404 static int pbase_tree_cache_ix(const struct object_id *oid)
1405 {
1406         return oid->hash[0] % ARRAY_SIZE(pbase_tree_cache);
1407 }
1408 static int pbase_tree_cache_ix_incr(int ix)
1409 {
1410         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1411 }
1412
1413 static struct pbase_tree {
1414         struct pbase_tree *next;
1415         /* This is a phony "cache" entry; we are not
1416          * going to evict it or find it through _get()
1417          * mechanism -- this is for the toplevel node that
1418          * would almost always change with any commit.
1419          */
1420         struct pbase_tree_cache pcache;
1421 } *pbase_tree;
1422
1423 static struct pbase_tree_cache *pbase_tree_get(const struct object_id *oid)
1424 {
1425         struct pbase_tree_cache *ent, *nent;
1426         void *data;
1427         unsigned long size;
1428         enum object_type type;
1429         int neigh;
1430         int my_ix = pbase_tree_cache_ix(oid);
1431         int available_ix = -1;
1432
1433         /* pbase-tree-cache acts as a limited hashtable.
1434          * your object will be found at your index or within a few
1435          * slots after that slot if it is cached.
1436          */
1437         for (neigh = 0; neigh < 8; neigh++) {
1438                 ent = pbase_tree_cache[my_ix];
1439                 if (ent && oideq(&ent->oid, oid)) {
1440                         ent->ref++;
1441                         return ent;
1442                 }
1443                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1444                          ((0 <= available_ix) &&
1445                           (!ent && pbase_tree_cache[available_ix])))
1446                         available_ix = my_ix;
1447                 if (!ent)
1448                         break;
1449                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1450         }
1451
1452         /* Did not find one.  Either we got a bogus request or
1453          * we need to read and perhaps cache.
1454          */
1455         data = read_object_file(oid, &type, &size);
1456         if (!data)
1457                 return NULL;
1458         if (type != OBJ_TREE) {
1459                 free(data);
1460                 return NULL;
1461         }
1462
1463         /* We need to either cache or return a throwaway copy */
1464
1465         if (available_ix < 0)
1466                 ent = NULL;
1467         else {
1468                 ent = pbase_tree_cache[available_ix];
1469                 my_ix = available_ix;
1470         }
1471
1472         if (!ent) {
1473                 nent = xmalloc(sizeof(*nent));
1474                 nent->temporary = (available_ix < 0);
1475         }
1476         else {
1477                 /* evict and reuse */
1478                 free(ent->tree_data);
1479                 nent = ent;
1480         }
1481         oidcpy(&nent->oid, oid);
1482         nent->tree_data = data;
1483         nent->tree_size = size;
1484         nent->ref = 1;
1485         if (!nent->temporary)
1486                 pbase_tree_cache[my_ix] = nent;
1487         return nent;
1488 }
1489
1490 static void pbase_tree_put(struct pbase_tree_cache *cache)
1491 {
1492         if (!cache->temporary) {
1493                 cache->ref--;
1494                 return;
1495         }
1496         free(cache->tree_data);
1497         free(cache);
1498 }
1499
1500 static int name_cmp_len(const char *name)
1501 {
1502         int i;
1503         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1504                 ;
1505         return i;
1506 }
1507
1508 static void add_pbase_object(struct tree_desc *tree,
1509                              const char *name,
1510                              int cmplen,
1511                              const char *fullname)
1512 {
1513         struct name_entry entry;
1514         int cmp;
1515
1516         while (tree_entry(tree,&entry)) {
1517                 if (S_ISGITLINK(entry.mode))
1518                         continue;
1519                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1520                       memcmp(name, entry.path, cmplen);
1521                 if (cmp > 0)
1522                         continue;
1523                 if (cmp < 0)
1524                         return;
1525                 if (name[cmplen] != '/') {
1526                         add_object_entry(&entry.oid,
1527                                          object_type(entry.mode),
1528                                          fullname, 1);
1529                         return;
1530                 }
1531                 if (S_ISDIR(entry.mode)) {
1532                         struct tree_desc sub;
1533                         struct pbase_tree_cache *tree;
1534                         const char *down = name+cmplen+1;
1535                         int downlen = name_cmp_len(down);
1536
1537                         tree = pbase_tree_get(&entry.oid);
1538                         if (!tree)
1539                                 return;
1540                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1541
1542                         add_pbase_object(&sub, down, downlen, fullname);
1543                         pbase_tree_put(tree);
1544                 }
1545         }
1546 }
1547
1548 static unsigned *done_pbase_paths;
1549 static int done_pbase_paths_num;
1550 static int done_pbase_paths_alloc;
1551 static int done_pbase_path_pos(unsigned hash)
1552 {
1553         int lo = 0;
1554         int hi = done_pbase_paths_num;
1555         while (lo < hi) {
1556                 int mi = lo + (hi - lo) / 2;
1557                 if (done_pbase_paths[mi] == hash)
1558                         return mi;
1559                 if (done_pbase_paths[mi] < hash)
1560                         hi = mi;
1561                 else
1562                         lo = mi + 1;
1563         }
1564         return -lo-1;
1565 }
1566
1567 static int check_pbase_path(unsigned hash)
1568 {
1569         int pos = done_pbase_path_pos(hash);
1570         if (0 <= pos)
1571                 return 1;
1572         pos = -pos - 1;
1573         ALLOC_GROW(done_pbase_paths,
1574                    done_pbase_paths_num + 1,
1575                    done_pbase_paths_alloc);
1576         done_pbase_paths_num++;
1577         if (pos < done_pbase_paths_num)
1578                 MOVE_ARRAY(done_pbase_paths + pos + 1, done_pbase_paths + pos,
1579                            done_pbase_paths_num - pos - 1);
1580         done_pbase_paths[pos] = hash;
1581         return 0;
1582 }
1583
1584 static void add_preferred_base_object(const char *name)
1585 {
1586         struct pbase_tree *it;
1587         int cmplen;
1588         unsigned hash = pack_name_hash(name);
1589
1590         if (!num_preferred_base || check_pbase_path(hash))
1591                 return;
1592
1593         cmplen = name_cmp_len(name);
1594         for (it = pbase_tree; it; it = it->next) {
1595                 if (cmplen == 0) {
1596                         add_object_entry(&it->pcache.oid, OBJ_TREE, NULL, 1);
1597                 }
1598                 else {
1599                         struct tree_desc tree;
1600                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1601                         add_pbase_object(&tree, name, cmplen, name);
1602                 }
1603         }
1604 }
1605
1606 static void add_preferred_base(struct object_id *oid)
1607 {
1608         struct pbase_tree *it;
1609         void *data;
1610         unsigned long size;
1611         struct object_id tree_oid;
1612
1613         if (window <= num_preferred_base++)
1614                 return;
1615
1616         data = read_object_with_reference(the_repository, oid,
1617                                           tree_type, &size, &tree_oid);
1618         if (!data)
1619                 return;
1620
1621         for (it = pbase_tree; it; it = it->next) {
1622                 if (oideq(&it->pcache.oid, &tree_oid)) {
1623                         free(data);
1624                         return;
1625                 }
1626         }
1627
1628         it = xcalloc(1, sizeof(*it));
1629         it->next = pbase_tree;
1630         pbase_tree = it;
1631
1632         oidcpy(&it->pcache.oid, &tree_oid);
1633         it->pcache.tree_data = data;
1634         it->pcache.tree_size = size;
1635 }
1636
1637 static void cleanup_preferred_base(void)
1638 {
1639         struct pbase_tree *it;
1640         unsigned i;
1641
1642         it = pbase_tree;
1643         pbase_tree = NULL;
1644         while (it) {
1645                 struct pbase_tree *tmp = it;
1646                 it = tmp->next;
1647                 free(tmp->pcache.tree_data);
1648                 free(tmp);
1649         }
1650
1651         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1652                 if (!pbase_tree_cache[i])
1653                         continue;
1654                 free(pbase_tree_cache[i]->tree_data);
1655                 FREE_AND_NULL(pbase_tree_cache[i]);
1656         }
1657
1658         FREE_AND_NULL(done_pbase_paths);
1659         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1660 }
1661
1662 /*
1663  * Return 1 iff the object specified by "delta" can be sent
1664  * literally as a delta against the base in "base_sha1". If
1665  * so, then *base_out will point to the entry in our packing
1666  * list, or NULL if we must use the external-base list.
1667  *
1668  * Depth value does not matter - find_deltas() will
1669  * never consider reused delta as the base object to
1670  * deltify other objects against, in order to avoid
1671  * circular deltas.
1672  */
1673 static int can_reuse_delta(const struct object_id *base_oid,
1674                            struct object_entry *delta,
1675                            struct object_entry **base_out)
1676 {
1677         struct object_entry *base;
1678
1679         /*
1680          * First see if we're already sending the base (or it's explicitly in
1681          * our "excluded" list).
1682          */
1683         base = packlist_find(&to_pack, base_oid);
1684         if (base) {
1685                 if (!in_same_island(&delta->idx.oid, &base->idx.oid))
1686                         return 0;
1687                 *base_out = base;
1688                 return 1;
1689         }
1690
1691         /*
1692          * Otherwise, reachability bitmaps may tell us if the receiver has it,
1693          * even if it was buried too deep in history to make it into the
1694          * packing list.
1695          */
1696         if (thin && bitmap_has_oid_in_uninteresting(bitmap_git, base_oid)) {
1697                 if (use_delta_islands) {
1698                         if (!in_same_island(&delta->idx.oid, base_oid))
1699                                 return 0;
1700                 }
1701                 *base_out = NULL;
1702                 return 1;
1703         }
1704
1705         return 0;
1706 }
1707
1708 static void prefetch_to_pack(uint32_t object_index_start) {
1709         struct oid_array to_fetch = OID_ARRAY_INIT;
1710         uint32_t i;
1711
1712         for (i = object_index_start; i < to_pack.nr_objects; i++) {
1713                 struct object_entry *entry = to_pack.objects + i;
1714
1715                 if (!oid_object_info_extended(the_repository,
1716                                               &entry->idx.oid,
1717                                               NULL,
1718                                               OBJECT_INFO_FOR_PREFETCH))
1719                         continue;
1720                 oid_array_append(&to_fetch, &entry->idx.oid);
1721         }
1722         promisor_remote_get_direct(the_repository,
1723                                    to_fetch.oid, to_fetch.nr);
1724         oid_array_clear(&to_fetch);
1725 }
1726
1727 static void check_object(struct object_entry *entry, uint32_t object_index)
1728 {
1729         unsigned long canonical_size;
1730         enum object_type type;
1731         struct object_info oi = {.typep = &type, .sizep = &canonical_size};
1732
1733         if (IN_PACK(entry)) {
1734                 struct packed_git *p = IN_PACK(entry);
1735                 struct pack_window *w_curs = NULL;
1736                 int have_base = 0;
1737                 struct object_id base_ref;
1738                 struct object_entry *base_entry;
1739                 unsigned long used, used_0;
1740                 unsigned long avail;
1741                 off_t ofs;
1742                 unsigned char *buf, c;
1743                 enum object_type type;
1744                 unsigned long in_pack_size;
1745
1746                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1747
1748                 /*
1749                  * We want in_pack_type even if we do not reuse delta
1750                  * since non-delta representations could still be reused.
1751                  */
1752                 used = unpack_object_header_buffer(buf, avail,
1753                                                    &type,
1754                                                    &in_pack_size);
1755                 if (used == 0)
1756                         goto give_up;
1757
1758                 if (type < 0)
1759                         BUG("invalid type %d", type);
1760                 entry->in_pack_type = type;
1761
1762                 /*
1763                  * Determine if this is a delta and if so whether we can
1764                  * reuse it or not.  Otherwise let's find out as cheaply as
1765                  * possible what the actual type and size for this object is.
1766                  */
1767                 switch (entry->in_pack_type) {
1768                 default:
1769                         /* Not a delta hence we've already got all we need. */
1770                         oe_set_type(entry, entry->in_pack_type);
1771                         SET_SIZE(entry, in_pack_size);
1772                         entry->in_pack_header_size = used;
1773                         if (oe_type(entry) < OBJ_COMMIT || oe_type(entry) > OBJ_BLOB)
1774                                 goto give_up;
1775                         unuse_pack(&w_curs);
1776                         return;
1777                 case OBJ_REF_DELTA:
1778                         if (reuse_delta && !entry->preferred_base) {
1779                                 oidread(&base_ref,
1780                                         use_pack(p, &w_curs,
1781                                                  entry->in_pack_offset + used,
1782                                                  NULL));
1783                                 have_base = 1;
1784                         }
1785                         entry->in_pack_header_size = used + the_hash_algo->rawsz;
1786                         break;
1787                 case OBJ_OFS_DELTA:
1788                         buf = use_pack(p, &w_curs,
1789                                        entry->in_pack_offset + used, NULL);
1790                         used_0 = 0;
1791                         c = buf[used_0++];
1792                         ofs = c & 127;
1793                         while (c & 128) {
1794                                 ofs += 1;
1795                                 if (!ofs || MSB(ofs, 7)) {
1796                                         error(_("delta base offset overflow in pack for %s"),
1797                                               oid_to_hex(&entry->idx.oid));
1798                                         goto give_up;
1799                                 }
1800                                 c = buf[used_0++];
1801                                 ofs = (ofs << 7) + (c & 127);
1802                         }
1803                         ofs = entry->in_pack_offset - ofs;
1804                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1805                                 error(_("delta base offset out of bound for %s"),
1806                                       oid_to_hex(&entry->idx.oid));
1807                                 goto give_up;
1808                         }
1809                         if (reuse_delta && !entry->preferred_base) {
1810                                 struct revindex_entry *revidx;
1811                                 revidx = find_pack_revindex(p, ofs);
1812                                 if (!revidx)
1813                                         goto give_up;
1814                                 if (!nth_packed_object_id(&base_ref, p, revidx->nr))
1815                                         have_base = 1;
1816                         }
1817                         entry->in_pack_header_size = used + used_0;
1818                         break;
1819                 }
1820
1821                 if (have_base &&
1822                     can_reuse_delta(&base_ref, entry, &base_entry)) {
1823                         oe_set_type(entry, entry->in_pack_type);
1824                         SET_SIZE(entry, in_pack_size); /* delta size */
1825                         SET_DELTA_SIZE(entry, in_pack_size);
1826
1827                         if (base_entry) {
1828                                 SET_DELTA(entry, base_entry);
1829                                 entry->delta_sibling_idx = base_entry->delta_child_idx;
1830                                 SET_DELTA_CHILD(base_entry, entry);
1831                         } else {
1832                                 SET_DELTA_EXT(entry, &base_ref);
1833                         }
1834
1835                         unuse_pack(&w_curs);
1836                         return;
1837                 }
1838
1839                 if (oe_type(entry)) {
1840                         off_t delta_pos;
1841
1842                         /*
1843                          * This must be a delta and we already know what the
1844                          * final object type is.  Let's extract the actual
1845                          * object size from the delta header.
1846                          */
1847                         delta_pos = entry->in_pack_offset + entry->in_pack_header_size;
1848                         canonical_size = get_size_from_delta(p, &w_curs, delta_pos);
1849                         if (canonical_size == 0)
1850                                 goto give_up;
1851                         SET_SIZE(entry, canonical_size);
1852                         unuse_pack(&w_curs);
1853                         return;
1854                 }
1855
1856                 /*
1857                  * No choice but to fall back to the recursive delta walk
1858                  * with oid_object_info() to find about the object type
1859                  * at this point...
1860                  */
1861                 give_up:
1862                 unuse_pack(&w_curs);
1863         }
1864
1865         if (oid_object_info_extended(the_repository, &entry->idx.oid, &oi,
1866                                      OBJECT_INFO_SKIP_FETCH_OBJECT | OBJECT_INFO_LOOKUP_REPLACE) < 0) {
1867                 if (has_promisor_remote()) {
1868                         prefetch_to_pack(object_index);
1869                         if (oid_object_info_extended(the_repository, &entry->idx.oid, &oi,
1870                                                      OBJECT_INFO_SKIP_FETCH_OBJECT | OBJECT_INFO_LOOKUP_REPLACE) < 0)
1871                                 type = -1;
1872                 } else {
1873                         type = -1;
1874                 }
1875         }
1876         oe_set_type(entry, type);
1877         if (entry->type_valid) {
1878                 SET_SIZE(entry, canonical_size);
1879         } else {
1880                 /*
1881                  * Bad object type is checked in prepare_pack().  This is
1882                  * to permit a missing preferred base object to be ignored
1883                  * as a preferred base.  Doing so can result in a larger
1884                  * pack file, but the transfer will still take place.
1885                  */
1886         }
1887 }
1888
1889 static int pack_offset_sort(const void *_a, const void *_b)
1890 {
1891         const struct object_entry *a = *(struct object_entry **)_a;
1892         const struct object_entry *b = *(struct object_entry **)_b;
1893         const struct packed_git *a_in_pack = IN_PACK(a);
1894         const struct packed_git *b_in_pack = IN_PACK(b);
1895
1896         /* avoid filesystem trashing with loose objects */
1897         if (!a_in_pack && !b_in_pack)
1898                 return oidcmp(&a->idx.oid, &b->idx.oid);
1899
1900         if (a_in_pack < b_in_pack)
1901                 return -1;
1902         if (a_in_pack > b_in_pack)
1903                 return 1;
1904         return a->in_pack_offset < b->in_pack_offset ? -1 :
1905                         (a->in_pack_offset > b->in_pack_offset);
1906 }
1907
1908 /*
1909  * Drop an on-disk delta we were planning to reuse. Naively, this would
1910  * just involve blanking out the "delta" field, but we have to deal
1911  * with some extra book-keeping:
1912  *
1913  *   1. Removing ourselves from the delta_sibling linked list.
1914  *
1915  *   2. Updating our size/type to the non-delta representation. These were
1916  *      either not recorded initially (size) or overwritten with the delta type
1917  *      (type) when check_object() decided to reuse the delta.
1918  *
1919  *   3. Resetting our delta depth, as we are now a base object.
1920  */
1921 static void drop_reused_delta(struct object_entry *entry)
1922 {
1923         unsigned *idx = &to_pack.objects[entry->delta_idx - 1].delta_child_idx;
1924         struct object_info oi = OBJECT_INFO_INIT;
1925         enum object_type type;
1926         unsigned long size;
1927
1928         while (*idx) {
1929                 struct object_entry *oe = &to_pack.objects[*idx - 1];
1930
1931                 if (oe == entry)
1932                         *idx = oe->delta_sibling_idx;
1933                 else
1934                         idx = &oe->delta_sibling_idx;
1935         }
1936         SET_DELTA(entry, NULL);
1937         entry->depth = 0;
1938
1939         oi.sizep = &size;
1940         oi.typep = &type;
1941         if (packed_object_info(the_repository, IN_PACK(entry), entry->in_pack_offset, &oi) < 0) {
1942                 /*
1943                  * We failed to get the info from this pack for some reason;
1944                  * fall back to oid_object_info, which may find another copy.
1945                  * And if that fails, the error will be recorded in oe_type(entry)
1946                  * and dealt with in prepare_pack().
1947                  */
1948                 oe_set_type(entry,
1949                             oid_object_info(the_repository, &entry->idx.oid, &size));
1950         } else {
1951                 oe_set_type(entry, type);
1952         }
1953         SET_SIZE(entry, size);
1954 }
1955
1956 /*
1957  * Follow the chain of deltas from this entry onward, throwing away any links
1958  * that cause us to hit a cycle (as determined by the DFS state flags in
1959  * the entries).
1960  *
1961  * We also detect too-long reused chains that would violate our --depth
1962  * limit.
1963  */
1964 static void break_delta_chains(struct object_entry *entry)
1965 {
1966         /*
1967          * The actual depth of each object we will write is stored as an int,
1968          * as it cannot exceed our int "depth" limit. But before we break
1969          * changes based no that limit, we may potentially go as deep as the
1970          * number of objects, which is elsewhere bounded to a uint32_t.
1971          */
1972         uint32_t total_depth;
1973         struct object_entry *cur, *next;
1974
1975         for (cur = entry, total_depth = 0;
1976              cur;
1977              cur = DELTA(cur), total_depth++) {
1978                 if (cur->dfs_state == DFS_DONE) {
1979                         /*
1980                          * We've already seen this object and know it isn't
1981                          * part of a cycle. We do need to append its depth
1982                          * to our count.
1983                          */
1984                         total_depth += cur->depth;
1985                         break;
1986                 }
1987
1988                 /*
1989                  * We break cycles before looping, so an ACTIVE state (or any
1990                  * other cruft which made its way into the state variable)
1991                  * is a bug.
1992                  */
1993                 if (cur->dfs_state != DFS_NONE)
1994                         BUG("confusing delta dfs state in first pass: %d",
1995                             cur->dfs_state);
1996
1997                 /*
1998                  * Now we know this is the first time we've seen the object. If
1999                  * it's not a delta, we're done traversing, but we'll mark it
2000                  * done to save time on future traversals.
2001                  */
2002                 if (!DELTA(cur)) {
2003                         cur->dfs_state = DFS_DONE;
2004                         break;
2005                 }
2006
2007                 /*
2008                  * Mark ourselves as active and see if the next step causes
2009                  * us to cycle to another active object. It's important to do
2010                  * this _before_ we loop, because it impacts where we make the
2011                  * cut, and thus how our total_depth counter works.
2012                  * E.g., We may see a partial loop like:
2013                  *
2014                  *   A -> B -> C -> D -> B
2015                  *
2016                  * Cutting B->C breaks the cycle. But now the depth of A is
2017                  * only 1, and our total_depth counter is at 3. The size of the
2018                  * error is always one less than the size of the cycle we
2019                  * broke. Commits C and D were "lost" from A's chain.
2020                  *
2021                  * If we instead cut D->B, then the depth of A is correct at 3.
2022                  * We keep all commits in the chain that we examined.
2023                  */
2024                 cur->dfs_state = DFS_ACTIVE;
2025                 if (DELTA(cur)->dfs_state == DFS_ACTIVE) {
2026                         drop_reused_delta(cur);
2027                         cur->dfs_state = DFS_DONE;
2028                         break;
2029                 }
2030         }
2031
2032         /*
2033          * And now that we've gone all the way to the bottom of the chain, we
2034          * need to clear the active flags and set the depth fields as
2035          * appropriate. Unlike the loop above, which can quit when it drops a
2036          * delta, we need to keep going to look for more depth cuts. So we need
2037          * an extra "next" pointer to keep going after we reset cur->delta.
2038          */
2039         for (cur = entry; cur; cur = next) {
2040                 next = DELTA(cur);
2041
2042                 /*
2043                  * We should have a chain of zero or more ACTIVE states down to
2044                  * a final DONE. We can quit after the DONE, because either it
2045                  * has no bases, or we've already handled them in a previous
2046                  * call.
2047                  */
2048                 if (cur->dfs_state == DFS_DONE)
2049                         break;
2050                 else if (cur->dfs_state != DFS_ACTIVE)
2051                         BUG("confusing delta dfs state in second pass: %d",
2052                             cur->dfs_state);
2053
2054                 /*
2055                  * If the total_depth is more than depth, then we need to snip
2056                  * the chain into two or more smaller chains that don't exceed
2057                  * the maximum depth. Most of the resulting chains will contain
2058                  * (depth + 1) entries (i.e., depth deltas plus one base), and
2059                  * the last chain (i.e., the one containing entry) will contain
2060                  * whatever entries are left over, namely
2061                  * (total_depth % (depth + 1)) of them.
2062                  *
2063                  * Since we are iterating towards decreasing depth, we need to
2064                  * decrement total_depth as we go, and we need to write to the
2065                  * entry what its final depth will be after all of the
2066                  * snipping. Since we're snipping into chains of length (depth
2067                  * + 1) entries, the final depth of an entry will be its
2068                  * original depth modulo (depth + 1). Any time we encounter an
2069                  * entry whose final depth is supposed to be zero, we snip it
2070                  * from its delta base, thereby making it so.
2071                  */
2072                 cur->depth = (total_depth--) % (depth + 1);
2073                 if (!cur->depth)
2074                         drop_reused_delta(cur);
2075
2076                 cur->dfs_state = DFS_DONE;
2077         }
2078 }
2079
2080 static void get_object_details(void)
2081 {
2082         uint32_t i;
2083         struct object_entry **sorted_by_offset;
2084
2085         if (progress)
2086                 progress_state = start_progress(_("Counting objects"),
2087                                                 to_pack.nr_objects);
2088
2089         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
2090         for (i = 0; i < to_pack.nr_objects; i++)
2091                 sorted_by_offset[i] = to_pack.objects + i;
2092         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
2093
2094         for (i = 0; i < to_pack.nr_objects; i++) {
2095                 struct object_entry *entry = sorted_by_offset[i];
2096                 check_object(entry, i);
2097                 if (entry->type_valid &&
2098                     oe_size_greater_than(&to_pack, entry, big_file_threshold))
2099                         entry->no_try_delta = 1;
2100                 display_progress(progress_state, i + 1);
2101         }
2102         stop_progress(&progress_state);
2103
2104         /*
2105          * This must happen in a second pass, since we rely on the delta
2106          * information for the whole list being completed.
2107          */
2108         for (i = 0; i < to_pack.nr_objects; i++)
2109                 break_delta_chains(&to_pack.objects[i]);
2110
2111         free(sorted_by_offset);
2112 }
2113
2114 /*
2115  * We search for deltas in a list sorted by type, by filename hash, and then
2116  * by size, so that we see progressively smaller and smaller files.
2117  * That's because we prefer deltas to be from the bigger file
2118  * to the smaller -- deletes are potentially cheaper, but perhaps
2119  * more importantly, the bigger file is likely the more recent
2120  * one.  The deepest deltas are therefore the oldest objects which are
2121  * less susceptible to be accessed often.
2122  */
2123 static int type_size_sort(const void *_a, const void *_b)
2124 {
2125         const struct object_entry *a = *(struct object_entry **)_a;
2126         const struct object_entry *b = *(struct object_entry **)_b;
2127         const enum object_type a_type = oe_type(a);
2128         const enum object_type b_type = oe_type(b);
2129         const unsigned long a_size = SIZE(a);
2130         const unsigned long b_size = SIZE(b);
2131
2132         if (a_type > b_type)
2133                 return -1;
2134         if (a_type < b_type)
2135                 return 1;
2136         if (a->hash > b->hash)
2137                 return -1;
2138         if (a->hash < b->hash)
2139                 return 1;
2140         if (a->preferred_base > b->preferred_base)
2141                 return -1;
2142         if (a->preferred_base < b->preferred_base)
2143                 return 1;
2144         if (use_delta_islands) {
2145                 const int island_cmp = island_delta_cmp(&a->idx.oid, &b->idx.oid);
2146                 if (island_cmp)
2147                         return island_cmp;
2148         }
2149         if (a_size > b_size)
2150                 return -1;
2151         if (a_size < b_size)
2152                 return 1;
2153         return a < b ? -1 : (a > b);  /* newest first */
2154 }
2155
2156 struct unpacked {
2157         struct object_entry *entry;
2158         void *data;
2159         struct delta_index *index;
2160         unsigned depth;
2161 };
2162
2163 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
2164                            unsigned long delta_size)
2165 {
2166         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
2167                 return 0;
2168
2169         if (delta_size < cache_max_small_delta_size)
2170                 return 1;
2171
2172         /* cache delta, if objects are large enough compared to delta size */
2173         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
2174                 return 1;
2175
2176         return 0;
2177 }
2178
2179 /* Protect delta_cache_size */
2180 static pthread_mutex_t cache_mutex;
2181 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
2182 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
2183
2184 /*
2185  * Protect object list partitioning (e.g. struct thread_param) and
2186  * progress_state
2187  */
2188 static pthread_mutex_t progress_mutex;
2189 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
2190 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
2191
2192 /*
2193  * Access to struct object_entry is unprotected since each thread owns
2194  * a portion of the main object list. Just don't access object entries
2195  * ahead in the list because they can be stolen and would need
2196  * progress_mutex for protection.
2197  */
2198
2199 /*
2200  * Return the size of the object without doing any delta
2201  * reconstruction (so non-deltas are true object sizes, but deltas
2202  * return the size of the delta data).
2203  */
2204 unsigned long oe_get_size_slow(struct packing_data *pack,
2205                                const struct object_entry *e)
2206 {
2207         struct packed_git *p;
2208         struct pack_window *w_curs;
2209         unsigned char *buf;
2210         enum object_type type;
2211         unsigned long used, avail, size;
2212
2213         if (e->type_ != OBJ_OFS_DELTA && e->type_ != OBJ_REF_DELTA) {
2214                 packing_data_lock(&to_pack);
2215                 if (oid_object_info(the_repository, &e->idx.oid, &size) < 0)
2216                         die(_("unable to get size of %s"),
2217                             oid_to_hex(&e->idx.oid));
2218                 packing_data_unlock(&to_pack);
2219                 return size;
2220         }
2221
2222         p = oe_in_pack(pack, e);
2223         if (!p)
2224                 BUG("when e->type is a delta, it must belong to a pack");
2225
2226         packing_data_lock(&to_pack);
2227         w_curs = NULL;
2228         buf = use_pack(p, &w_curs, e->in_pack_offset, &avail);
2229         used = unpack_object_header_buffer(buf, avail, &type, &size);
2230         if (used == 0)
2231                 die(_("unable to parse object header of %s"),
2232                     oid_to_hex(&e->idx.oid));
2233
2234         unuse_pack(&w_curs);
2235         packing_data_unlock(&to_pack);
2236         return size;
2237 }
2238
2239 static int try_delta(struct unpacked *trg, struct unpacked *src,
2240                      unsigned max_depth, unsigned long *mem_usage)
2241 {
2242         struct object_entry *trg_entry = trg->entry;
2243         struct object_entry *src_entry = src->entry;
2244         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
2245         unsigned ref_depth;
2246         enum object_type type;
2247         void *delta_buf;
2248
2249         /* Don't bother doing diffs between different types */
2250         if (oe_type(trg_entry) != oe_type(src_entry))
2251                 return -1;
2252
2253         /*
2254          * We do not bother to try a delta that we discarded on an
2255          * earlier try, but only when reusing delta data.  Note that
2256          * src_entry that is marked as the preferred_base should always
2257          * be considered, as even if we produce a suboptimal delta against
2258          * it, we will still save the transfer cost, as we already know
2259          * the other side has it and we won't send src_entry at all.
2260          */
2261         if (reuse_delta && IN_PACK(trg_entry) &&
2262             IN_PACK(trg_entry) == IN_PACK(src_entry) &&
2263             !src_entry->preferred_base &&
2264             trg_entry->in_pack_type != OBJ_REF_DELTA &&
2265             trg_entry->in_pack_type != OBJ_OFS_DELTA)
2266                 return 0;
2267
2268         /* Let's not bust the allowed depth. */
2269         if (src->depth >= max_depth)
2270                 return 0;
2271
2272         /* Now some size filtering heuristics. */
2273         trg_size = SIZE(trg_entry);
2274         if (!DELTA(trg_entry)) {
2275                 max_size = trg_size/2 - the_hash_algo->rawsz;
2276                 ref_depth = 1;
2277         } else {
2278                 max_size = DELTA_SIZE(trg_entry);
2279                 ref_depth = trg->depth;
2280         }
2281         max_size = (uint64_t)max_size * (max_depth - src->depth) /
2282                                                 (max_depth - ref_depth + 1);
2283         if (max_size == 0)
2284                 return 0;
2285         src_size = SIZE(src_entry);
2286         sizediff = src_size < trg_size ? trg_size - src_size : 0;
2287         if (sizediff >= max_size)
2288                 return 0;
2289         if (trg_size < src_size / 32)
2290                 return 0;
2291
2292         if (!in_same_island(&trg->entry->idx.oid, &src->entry->idx.oid))
2293                 return 0;
2294
2295         /* Load data if not already done */
2296         if (!trg->data) {
2297                 packing_data_lock(&to_pack);
2298                 trg->data = read_object_file(&trg_entry->idx.oid, &type, &sz);
2299                 packing_data_unlock(&to_pack);
2300                 if (!trg->data)
2301                         die(_("object %s cannot be read"),
2302                             oid_to_hex(&trg_entry->idx.oid));
2303                 if (sz != trg_size)
2304                         die(_("object %s inconsistent object length (%"PRIuMAX" vs %"PRIuMAX")"),
2305                             oid_to_hex(&trg_entry->idx.oid), (uintmax_t)sz,
2306                             (uintmax_t)trg_size);
2307                 *mem_usage += sz;
2308         }
2309         if (!src->data) {
2310                 packing_data_lock(&to_pack);
2311                 src->data = read_object_file(&src_entry->idx.oid, &type, &sz);
2312                 packing_data_unlock(&to_pack);
2313                 if (!src->data) {
2314                         if (src_entry->preferred_base) {
2315                                 static int warned = 0;
2316                                 if (!warned++)
2317                                         warning(_("object %s cannot be read"),
2318                                                 oid_to_hex(&src_entry->idx.oid));
2319                                 /*
2320                                  * Those objects are not included in the
2321                                  * resulting pack.  Be resilient and ignore
2322                                  * them if they can't be read, in case the
2323                                  * pack could be created nevertheless.
2324                                  */
2325                                 return 0;
2326                         }
2327                         die(_("object %s cannot be read"),
2328                             oid_to_hex(&src_entry->idx.oid));
2329                 }
2330                 if (sz != src_size)
2331                         die(_("object %s inconsistent object length (%"PRIuMAX" vs %"PRIuMAX")"),
2332                             oid_to_hex(&src_entry->idx.oid), (uintmax_t)sz,
2333                             (uintmax_t)src_size);
2334                 *mem_usage += sz;
2335         }
2336         if (!src->index) {
2337                 src->index = create_delta_index(src->data, src_size);
2338                 if (!src->index) {
2339                         static int warned = 0;
2340                         if (!warned++)
2341                                 warning(_("suboptimal pack - out of memory"));
2342                         return 0;
2343                 }
2344                 *mem_usage += sizeof_delta_index(src->index);
2345         }
2346
2347         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
2348         if (!delta_buf)
2349                 return 0;
2350
2351         if (DELTA(trg_entry)) {
2352                 /* Prefer only shallower same-sized deltas. */
2353                 if (delta_size == DELTA_SIZE(trg_entry) &&
2354                     src->depth + 1 >= trg->depth) {
2355                         free(delta_buf);
2356                         return 0;
2357                 }
2358         }
2359
2360         /*
2361          * Handle memory allocation outside of the cache
2362          * accounting lock.  Compiler will optimize the strangeness
2363          * away when NO_PTHREADS is defined.
2364          */
2365         free(trg_entry->delta_data);
2366         cache_lock();
2367         if (trg_entry->delta_data) {
2368                 delta_cache_size -= DELTA_SIZE(trg_entry);
2369                 trg_entry->delta_data = NULL;
2370         }
2371         if (delta_cacheable(src_size, trg_size, delta_size)) {
2372                 delta_cache_size += delta_size;
2373                 cache_unlock();
2374                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
2375         } else {
2376                 cache_unlock();
2377                 free(delta_buf);
2378         }
2379
2380         SET_DELTA(trg_entry, src_entry);
2381         SET_DELTA_SIZE(trg_entry, delta_size);
2382         trg->depth = src->depth + 1;
2383
2384         return 1;
2385 }
2386
2387 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
2388 {
2389         struct object_entry *child = DELTA_CHILD(me);
2390         unsigned int m = n;
2391         while (child) {
2392                 const unsigned int c = check_delta_limit(child, n + 1);
2393                 if (m < c)
2394                         m = c;
2395                 child = DELTA_SIBLING(child);
2396         }
2397         return m;
2398 }
2399
2400 static unsigned long free_unpacked(struct unpacked *n)
2401 {
2402         unsigned long freed_mem = sizeof_delta_index(n->index);
2403         free_delta_index(n->index);
2404         n->index = NULL;
2405         if (n->data) {
2406                 freed_mem += SIZE(n->entry);
2407                 FREE_AND_NULL(n->data);
2408         }
2409         n->entry = NULL;
2410         n->depth = 0;
2411         return freed_mem;
2412 }
2413
2414 static void find_deltas(struct object_entry **list, unsigned *list_size,
2415                         int window, int depth, unsigned *processed)
2416 {
2417         uint32_t i, idx = 0, count = 0;
2418         struct unpacked *array;
2419         unsigned long mem_usage = 0;
2420
2421         array = xcalloc(window, sizeof(struct unpacked));
2422
2423         for (;;) {
2424                 struct object_entry *entry;
2425                 struct unpacked *n = array + idx;
2426                 int j, max_depth, best_base = -1;
2427
2428                 progress_lock();
2429                 if (!*list_size) {
2430                         progress_unlock();
2431                         break;
2432                 }
2433                 entry = *list++;
2434                 (*list_size)--;
2435                 if (!entry->preferred_base) {
2436                         (*processed)++;
2437                         display_progress(progress_state, *processed);
2438                 }
2439                 progress_unlock();
2440
2441                 mem_usage -= free_unpacked(n);
2442                 n->entry = entry;
2443
2444                 while (window_memory_limit &&
2445                        mem_usage > window_memory_limit &&
2446                        count > 1) {
2447                         const uint32_t tail = (idx + window - count) % window;
2448                         mem_usage -= free_unpacked(array + tail);
2449                         count--;
2450                 }
2451
2452                 /* We do not compute delta to *create* objects we are not
2453                  * going to pack.
2454                  */
2455                 if (entry->preferred_base)
2456                         goto next;
2457
2458                 /*
2459                  * If the current object is at pack edge, take the depth the
2460                  * objects that depend on the current object into account
2461                  * otherwise they would become too deep.
2462                  */
2463                 max_depth = depth;
2464                 if (DELTA_CHILD(entry)) {
2465                         max_depth -= check_delta_limit(entry, 0);
2466                         if (max_depth <= 0)
2467                                 goto next;
2468                 }
2469
2470                 j = window;
2471                 while (--j > 0) {
2472                         int ret;
2473                         uint32_t other_idx = idx + j;
2474                         struct unpacked *m;
2475                         if (other_idx >= window)
2476                                 other_idx -= window;
2477                         m = array + other_idx;
2478                         if (!m->entry)
2479                                 break;
2480                         ret = try_delta(n, m, max_depth, &mem_usage);
2481                         if (ret < 0)
2482                                 break;
2483                         else if (ret > 0)
2484                                 best_base = other_idx;
2485                 }
2486
2487                 /*
2488                  * If we decided to cache the delta data, then it is best
2489                  * to compress it right away.  First because we have to do
2490                  * it anyway, and doing it here while we're threaded will
2491                  * save a lot of time in the non threaded write phase,
2492                  * as well as allow for caching more deltas within
2493                  * the same cache size limit.
2494                  * ...
2495                  * But only if not writing to stdout, since in that case
2496                  * the network is most likely throttling writes anyway,
2497                  * and therefore it is best to go to the write phase ASAP
2498                  * instead, as we can afford spending more time compressing
2499                  * between writes at that moment.
2500                  */
2501                 if (entry->delta_data && !pack_to_stdout) {
2502                         unsigned long size;
2503
2504                         size = do_compress(&entry->delta_data, DELTA_SIZE(entry));
2505                         if (size < (1U << OE_Z_DELTA_BITS)) {
2506                                 entry->z_delta_size = size;
2507                                 cache_lock();
2508                                 delta_cache_size -= DELTA_SIZE(entry);
2509                                 delta_cache_size += entry->z_delta_size;
2510                                 cache_unlock();
2511                         } else {
2512                                 FREE_AND_NULL(entry->delta_data);
2513                                 entry->z_delta_size = 0;
2514                         }
2515                 }
2516
2517                 /* if we made n a delta, and if n is already at max
2518                  * depth, leaving it in the window is pointless.  we
2519                  * should evict it first.
2520                  */
2521                 if (DELTA(entry) && max_depth <= n->depth)
2522                         continue;
2523
2524                 /*
2525                  * Move the best delta base up in the window, after the
2526                  * currently deltified object, to keep it longer.  It will
2527                  * be the first base object to be attempted next.
2528                  */
2529                 if (DELTA(entry)) {
2530                         struct unpacked swap = array[best_base];
2531                         int dist = (window + idx - best_base) % window;
2532                         int dst = best_base;
2533                         while (dist--) {
2534                                 int src = (dst + 1) % window;
2535                                 array[dst] = array[src];
2536                                 dst = src;
2537                         }
2538                         array[dst] = swap;
2539                 }
2540
2541                 next:
2542                 idx++;
2543                 if (count + 1 < window)
2544                         count++;
2545                 if (idx >= window)
2546                         idx = 0;
2547         }
2548
2549         for (i = 0; i < window; ++i) {
2550                 free_delta_index(array[i].index);
2551                 free(array[i].data);
2552         }
2553         free(array);
2554 }
2555
2556 /*
2557  * The main object list is split into smaller lists, each is handed to
2558  * one worker.
2559  *
2560  * The main thread waits on the condition that (at least) one of the workers
2561  * has stopped working (which is indicated in the .working member of
2562  * struct thread_params).
2563  *
2564  * When a work thread has completed its work, it sets .working to 0 and
2565  * signals the main thread and waits on the condition that .data_ready
2566  * becomes 1.
2567  *
2568  * The main thread steals half of the work from the worker that has
2569  * most work left to hand it to the idle worker.
2570  */
2571
2572 struct thread_params {
2573         pthread_t thread;
2574         struct object_entry **list;
2575         unsigned list_size;
2576         unsigned remaining;
2577         int window;
2578         int depth;
2579         int working;
2580         int data_ready;
2581         pthread_mutex_t mutex;
2582         pthread_cond_t cond;
2583         unsigned *processed;
2584 };
2585
2586 static pthread_cond_t progress_cond;
2587
2588 /*
2589  * Mutex and conditional variable can't be statically-initialized on Windows.
2590  */
2591 static void init_threaded_search(void)
2592 {
2593         pthread_mutex_init(&cache_mutex, NULL);
2594         pthread_mutex_init(&progress_mutex, NULL);
2595         pthread_cond_init(&progress_cond, NULL);
2596 }
2597
2598 static void cleanup_threaded_search(void)
2599 {
2600         pthread_cond_destroy(&progress_cond);
2601         pthread_mutex_destroy(&cache_mutex);
2602         pthread_mutex_destroy(&progress_mutex);
2603 }
2604
2605 static void *threaded_find_deltas(void *arg)
2606 {
2607         struct thread_params *me = arg;
2608
2609         progress_lock();
2610         while (me->remaining) {
2611                 progress_unlock();
2612
2613                 find_deltas(me->list, &me->remaining,
2614                             me->window, me->depth, me->processed);
2615
2616                 progress_lock();
2617                 me->working = 0;
2618                 pthread_cond_signal(&progress_cond);
2619                 progress_unlock();
2620
2621                 /*
2622                  * We must not set ->data_ready before we wait on the
2623                  * condition because the main thread may have set it to 1
2624                  * before we get here. In order to be sure that new
2625                  * work is available if we see 1 in ->data_ready, it
2626                  * was initialized to 0 before this thread was spawned
2627                  * and we reset it to 0 right away.
2628                  */
2629                 pthread_mutex_lock(&me->mutex);
2630                 while (!me->data_ready)
2631                         pthread_cond_wait(&me->cond, &me->mutex);
2632                 me->data_ready = 0;
2633                 pthread_mutex_unlock(&me->mutex);
2634
2635                 progress_lock();
2636         }
2637         progress_unlock();
2638         /* leave ->working 1 so that this doesn't get more work assigned */
2639         return NULL;
2640 }
2641
2642 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2643                            int window, int depth, unsigned *processed)
2644 {
2645         struct thread_params *p;
2646         int i, ret, active_threads = 0;
2647
2648         init_threaded_search();
2649
2650         if (delta_search_threads <= 1) {
2651                 find_deltas(list, &list_size, window, depth, processed);
2652                 cleanup_threaded_search();
2653                 return;
2654         }
2655         if (progress > pack_to_stdout)
2656                 fprintf_ln(stderr, _("Delta compression using up to %d threads"),
2657                            delta_search_threads);
2658         p = xcalloc(delta_search_threads, sizeof(*p));
2659
2660         /* Partition the work amongst work threads. */
2661         for (i = 0; i < delta_search_threads; i++) {
2662                 unsigned sub_size = list_size / (delta_search_threads - i);
2663
2664                 /* don't use too small segments or no deltas will be found */
2665                 if (sub_size < 2*window && i+1 < delta_search_threads)
2666                         sub_size = 0;
2667
2668                 p[i].window = window;
2669                 p[i].depth = depth;
2670                 p[i].processed = processed;
2671                 p[i].working = 1;
2672                 p[i].data_ready = 0;
2673
2674                 /* try to split chunks on "path" boundaries */
2675                 while (sub_size && sub_size < list_size &&
2676                        list[sub_size]->hash &&
2677                        list[sub_size]->hash == list[sub_size-1]->hash)
2678                         sub_size++;
2679
2680                 p[i].list = list;
2681                 p[i].list_size = sub_size;
2682                 p[i].remaining = sub_size;
2683
2684                 list += sub_size;
2685                 list_size -= sub_size;
2686         }
2687
2688         /* Start work threads. */
2689         for (i = 0; i < delta_search_threads; i++) {
2690                 if (!p[i].list_size)
2691                         continue;
2692                 pthread_mutex_init(&p[i].mutex, NULL);
2693                 pthread_cond_init(&p[i].cond, NULL);
2694                 ret = pthread_create(&p[i].thread, NULL,
2695                                      threaded_find_deltas, &p[i]);
2696                 if (ret)
2697                         die(_("unable to create thread: %s"), strerror(ret));
2698                 active_threads++;
2699         }
2700
2701         /*
2702          * Now let's wait for work completion.  Each time a thread is done
2703          * with its work, we steal half of the remaining work from the
2704          * thread with the largest number of unprocessed objects and give
2705          * it to that newly idle thread.  This ensure good load balancing
2706          * until the remaining object list segments are simply too short
2707          * to be worth splitting anymore.
2708          */
2709         while (active_threads) {
2710                 struct thread_params *target = NULL;
2711                 struct thread_params *victim = NULL;
2712                 unsigned sub_size = 0;
2713
2714                 progress_lock();
2715                 for (;;) {
2716                         for (i = 0; !target && i < delta_search_threads; i++)
2717                                 if (!p[i].working)
2718                                         target = &p[i];
2719                         if (target)
2720                                 break;
2721                         pthread_cond_wait(&progress_cond, &progress_mutex);
2722                 }
2723
2724                 for (i = 0; i < delta_search_threads; i++)
2725                         if (p[i].remaining > 2*window &&
2726                             (!victim || victim->remaining < p[i].remaining))
2727                                 victim = &p[i];
2728                 if (victim) {
2729                         sub_size = victim->remaining / 2;
2730                         list = victim->list + victim->list_size - sub_size;
2731                         while (sub_size && list[0]->hash &&
2732                                list[0]->hash == list[-1]->hash) {
2733                                 list++;
2734                                 sub_size--;
2735                         }
2736                         if (!sub_size) {
2737                                 /*
2738                                  * It is possible for some "paths" to have
2739                                  * so many objects that no hash boundary
2740                                  * might be found.  Let's just steal the
2741                                  * exact half in that case.
2742                                  */
2743                                 sub_size = victim->remaining / 2;
2744                                 list -= sub_size;
2745                         }
2746                         target->list = list;
2747                         victim->list_size -= sub_size;
2748                         victim->remaining -= sub_size;
2749                 }
2750                 target->list_size = sub_size;
2751                 target->remaining = sub_size;
2752                 target->working = 1;
2753                 progress_unlock();
2754
2755                 pthread_mutex_lock(&target->mutex);
2756                 target->data_ready = 1;
2757                 pthread_cond_signal(&target->cond);
2758                 pthread_mutex_unlock(&target->mutex);
2759
2760                 if (!sub_size) {
2761                         pthread_join(target->thread, NULL);
2762                         pthread_cond_destroy(&target->cond);
2763                         pthread_mutex_destroy(&target->mutex);
2764                         active_threads--;
2765                 }
2766         }
2767         cleanup_threaded_search();
2768         free(p);
2769 }
2770
2771 static int obj_is_packed(const struct object_id *oid)
2772 {
2773         return packlist_find(&to_pack, oid) ||
2774                 (reuse_packfile_bitmap &&
2775                  bitmap_walk_contains(bitmap_git, reuse_packfile_bitmap, oid));
2776 }
2777
2778 static void add_tag_chain(const struct object_id *oid)
2779 {
2780         struct tag *tag;
2781
2782         /*
2783          * We catch duplicates already in add_object_entry(), but we'd
2784          * prefer to do this extra check to avoid having to parse the
2785          * tag at all if we already know that it's being packed (e.g., if
2786          * it was included via bitmaps, we would not have parsed it
2787          * previously).
2788          */
2789         if (obj_is_packed(oid))
2790                 return;
2791
2792         tag = lookup_tag(the_repository, oid);
2793         while (1) {
2794                 if (!tag || parse_tag(tag) || !tag->tagged)
2795                         die(_("unable to pack objects reachable from tag %s"),
2796                             oid_to_hex(oid));
2797
2798                 add_object_entry(&tag->object.oid, OBJ_TAG, NULL, 0);
2799
2800                 if (tag->tagged->type != OBJ_TAG)
2801                         return;
2802
2803                 tag = (struct tag *)tag->tagged;
2804         }
2805 }
2806
2807 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2808 {
2809         struct object_id peeled;
2810
2811         if (starts_with(path, "refs/tags/") && /* is a tag? */
2812             !peel_ref(path, &peeled)    && /* peelable? */
2813             obj_is_packed(&peeled)) /* object packed? */
2814                 add_tag_chain(oid);
2815         return 0;
2816 }
2817
2818 static void prepare_pack(int window, int depth)
2819 {
2820         struct object_entry **delta_list;
2821         uint32_t i, nr_deltas;
2822         unsigned n;
2823
2824         if (use_delta_islands)
2825                 resolve_tree_islands(the_repository, progress, &to_pack);
2826
2827         get_object_details();
2828
2829         /*
2830          * If we're locally repacking then we need to be doubly careful
2831          * from now on in order to make sure no stealth corruption gets
2832          * propagated to the new pack.  Clients receiving streamed packs
2833          * should validate everything they get anyway so no need to incur
2834          * the additional cost here in that case.
2835          */
2836         if (!pack_to_stdout)
2837                 do_check_packed_object_crc = 1;
2838
2839         if (!to_pack.nr_objects || !window || !depth)
2840                 return;
2841
2842         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2843         nr_deltas = n = 0;
2844
2845         for (i = 0; i < to_pack.nr_objects; i++) {
2846                 struct object_entry *entry = to_pack.objects + i;
2847
2848                 if (DELTA(entry))
2849                         /* This happens if we decided to reuse existing
2850                          * delta from a pack.  "reuse_delta &&" is implied.
2851                          */
2852                         continue;
2853
2854                 if (!entry->type_valid ||
2855                     oe_size_less_than(&to_pack, entry, 50))
2856                         continue;
2857
2858                 if (entry->no_try_delta)
2859                         continue;
2860
2861                 if (!entry->preferred_base) {
2862                         nr_deltas++;
2863                         if (oe_type(entry) < 0)
2864                                 die(_("unable to get type of object %s"),
2865                                     oid_to_hex(&entry->idx.oid));
2866                 } else {
2867                         if (oe_type(entry) < 0) {
2868                                 /*
2869                                  * This object is not found, but we
2870                                  * don't have to include it anyway.
2871                                  */
2872                                 continue;
2873                         }
2874                 }
2875
2876                 delta_list[n++] = entry;
2877         }
2878
2879         if (nr_deltas && n > 1) {
2880                 unsigned nr_done = 0;
2881
2882                 if (progress)
2883                         progress_state = start_progress(_("Compressing objects"),
2884                                                         nr_deltas);
2885                 QSORT(delta_list, n, type_size_sort);
2886                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2887                 stop_progress(&progress_state);
2888                 if (nr_done != nr_deltas)
2889                         die(_("inconsistency with delta count"));
2890         }
2891         free(delta_list);
2892 }
2893
2894 static int git_pack_config(const char *k, const char *v, void *cb)
2895 {
2896         if (!strcmp(k, "pack.window")) {
2897                 window = git_config_int(k, v);
2898                 return 0;
2899         }
2900         if (!strcmp(k, "pack.windowmemory")) {
2901                 window_memory_limit = git_config_ulong(k, v);
2902                 return 0;
2903         }
2904         if (!strcmp(k, "pack.depth")) {
2905                 depth = git_config_int(k, v);
2906                 return 0;
2907         }
2908         if (!strcmp(k, "pack.deltacachesize")) {
2909                 max_delta_cache_size = git_config_int(k, v);
2910                 return 0;
2911         }
2912         if (!strcmp(k, "pack.deltacachelimit")) {
2913                 cache_max_small_delta_size = git_config_int(k, v);
2914                 return 0;
2915         }
2916         if (!strcmp(k, "pack.writebitmaphashcache")) {
2917                 if (git_config_bool(k, v))
2918                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2919                 else
2920                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2921         }
2922         if (!strcmp(k, "pack.usebitmaps")) {
2923                 use_bitmap_index_default = git_config_bool(k, v);
2924                 return 0;
2925         }
2926         if (!strcmp(k, "pack.allowpackreuse")) {
2927                 allow_pack_reuse = git_config_bool(k, v);
2928                 return 0;
2929         }
2930         if (!strcmp(k, "pack.threads")) {
2931                 delta_search_threads = git_config_int(k, v);
2932                 if (delta_search_threads < 0)
2933                         die(_("invalid number of threads specified (%d)"),
2934                             delta_search_threads);
2935                 if (!HAVE_THREADS && delta_search_threads != 1) {
2936                         warning(_("no threads support, ignoring %s"), k);
2937                         delta_search_threads = 0;
2938                 }
2939                 return 0;
2940         }
2941         if (!strcmp(k, "pack.indexversion")) {
2942                 pack_idx_opts.version = git_config_int(k, v);
2943                 if (pack_idx_opts.version > 2)
2944                         die(_("bad pack.indexversion=%"PRIu32),
2945                             pack_idx_opts.version);
2946                 return 0;
2947         }
2948         if (!strcmp(k, "uploadpack.blobpackfileuri")) {
2949                 struct configured_exclusion *ex = xmalloc(sizeof(*ex));
2950                 const char *oid_end, *pack_end;
2951                 /*
2952                  * Stores the pack hash. This is not a true object ID, but is
2953                  * of the same form.
2954                  */
2955                 struct object_id pack_hash;
2956
2957                 if (parse_oid_hex(v, &ex->e.oid, &oid_end) ||
2958                     *oid_end != ' ' ||
2959                     parse_oid_hex(oid_end + 1, &pack_hash, &pack_end) ||
2960                     *pack_end != ' ')
2961                         die(_("value of uploadpack.blobpackfileuri must be "
2962                               "of the form '<object-hash> <pack-hash> <uri>' (got '%s')"), v);
2963                 if (oidmap_get(&configured_exclusions, &ex->e.oid))
2964                         die(_("object already configured in another "
2965                               "uploadpack.blobpackfileuri (got '%s')"), v);
2966                 ex->pack_hash_hex = xcalloc(1, pack_end - oid_end);
2967                 memcpy(ex->pack_hash_hex, oid_end + 1, pack_end - oid_end - 1);
2968                 ex->uri = xstrdup(pack_end + 1);
2969                 oidmap_put(&configured_exclusions, ex);
2970         }
2971         return git_default_config(k, v, cb);
2972 }
2973
2974 static void read_object_list_from_stdin(void)
2975 {
2976         char line[GIT_MAX_HEXSZ + 1 + PATH_MAX + 2];
2977         struct object_id oid;
2978         const char *p;
2979
2980         for (;;) {
2981                 if (!fgets(line, sizeof(line), stdin)) {
2982                         if (feof(stdin))
2983                                 break;
2984                         if (!ferror(stdin))
2985                                 die("BUG: fgets returned NULL, not EOF, not error!");
2986                         if (errno != EINTR)
2987                                 die_errno("fgets");
2988                         clearerr(stdin);
2989                         continue;
2990                 }
2991                 if (line[0] == '-') {
2992                         if (get_oid_hex(line+1, &oid))
2993                                 die(_("expected edge object ID, got garbage:\n %s"),
2994                                     line);
2995                         add_preferred_base(&oid);
2996                         continue;
2997                 }
2998                 if (parse_oid_hex(line, &oid, &p))
2999                         die(_("expected object ID, got garbage:\n %s"), line);
3000
3001                 add_preferred_base_object(p + 1);
3002                 add_object_entry(&oid, OBJ_NONE, p + 1, 0);
3003         }
3004 }
3005
3006 /* Remember to update object flag allocation in object.h */
3007 #define OBJECT_ADDED (1u<<20)
3008
3009 static void show_commit(struct commit *commit, void *data)
3010 {
3011         add_object_entry(&commit->object.oid, OBJ_COMMIT, NULL, 0);
3012         commit->object.flags |= OBJECT_ADDED;
3013
3014         if (write_bitmap_index)
3015                 index_commit_for_bitmap(commit);
3016
3017         if (use_delta_islands)
3018                 propagate_island_marks(commit);
3019 }
3020
3021 static void show_object(struct object *obj, const char *name, void *data)
3022 {
3023         add_preferred_base_object(name);
3024         add_object_entry(&obj->oid, obj->type, name, 0);
3025         obj->flags |= OBJECT_ADDED;
3026
3027         if (use_delta_islands) {
3028                 const char *p;
3029                 unsigned depth;
3030                 struct object_entry *ent;
3031
3032                 /* the empty string is a root tree, which is depth 0 */
3033                 depth = *name ? 1 : 0;
3034                 for (p = strchr(name, '/'); p; p = strchr(p + 1, '/'))
3035                         depth++;
3036
3037                 ent = packlist_find(&to_pack, &obj->oid);
3038                 if (ent && depth > oe_tree_depth(&to_pack, ent))
3039                         oe_set_tree_depth(&to_pack, ent, depth);
3040         }
3041 }
3042
3043 static void show_object__ma_allow_any(struct object *obj, const char *name, void *data)
3044 {
3045         assert(arg_missing_action == MA_ALLOW_ANY);
3046
3047         /*
3048          * Quietly ignore ALL missing objects.  This avoids problems with
3049          * staging them now and getting an odd error later.
3050          */
3051         if (!has_object(the_repository, &obj->oid, 0))
3052                 return;
3053
3054         show_object(obj, name, data);
3055 }
3056
3057 static void show_object__ma_allow_promisor(struct object *obj, const char *name, void *data)
3058 {
3059         assert(arg_missing_action == MA_ALLOW_PROMISOR);
3060
3061         /*
3062          * Quietly ignore EXPECTED missing objects.  This avoids problems with
3063          * staging them now and getting an odd error later.
3064          */
3065         if (!has_object(the_repository, &obj->oid, 0) && is_promisor_object(&obj->oid))
3066                 return;
3067
3068         show_object(obj, name, data);
3069 }
3070
3071 static int option_parse_missing_action(const struct option *opt,
3072                                        const char *arg, int unset)
3073 {
3074         assert(arg);
3075         assert(!unset);
3076
3077         if (!strcmp(arg, "error")) {
3078                 arg_missing_action = MA_ERROR;
3079                 fn_show_object = show_object;
3080                 return 0;
3081         }
3082
3083         if (!strcmp(arg, "allow-any")) {
3084                 arg_missing_action = MA_ALLOW_ANY;
3085                 fetch_if_missing = 0;
3086                 fn_show_object = show_object__ma_allow_any;
3087                 return 0;
3088         }
3089
3090         if (!strcmp(arg, "allow-promisor")) {
3091                 arg_missing_action = MA_ALLOW_PROMISOR;
3092                 fetch_if_missing = 0;
3093                 fn_show_object = show_object__ma_allow_promisor;
3094                 return 0;
3095         }
3096
3097         die(_("invalid value for --missing"));
3098         return 0;
3099 }
3100
3101 static void show_edge(struct commit *commit)
3102 {
3103         add_preferred_base(&commit->object.oid);
3104 }
3105
3106 struct in_pack_object {
3107         off_t offset;
3108         struct object *object;
3109 };
3110
3111 struct in_pack {
3112         unsigned int alloc;
3113         unsigned int nr;
3114         struct in_pack_object *array;
3115 };
3116
3117 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
3118 {
3119         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
3120         in_pack->array[in_pack->nr].object = object;
3121         in_pack->nr++;
3122 }
3123
3124 /*
3125  * Compare the objects in the offset order, in order to emulate the
3126  * "git rev-list --objects" output that produced the pack originally.
3127  */
3128 static int ofscmp(const void *a_, const void *b_)
3129 {
3130         struct in_pack_object *a = (struct in_pack_object *)a_;
3131         struct in_pack_object *b = (struct in_pack_object *)b_;
3132
3133         if (a->offset < b->offset)
3134                 return -1;
3135         else if (a->offset > b->offset)
3136                 return 1;
3137         else
3138                 return oidcmp(&a->object->oid, &b->object->oid);
3139 }
3140
3141 static void add_objects_in_unpacked_packs(void)
3142 {
3143         struct packed_git *p;
3144         struct in_pack in_pack;
3145         uint32_t i;
3146
3147         memset(&in_pack, 0, sizeof(in_pack));
3148
3149         for (p = get_all_packs(the_repository); p; p = p->next) {
3150                 struct object_id oid;
3151                 struct object *o;
3152
3153                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
3154                         continue;
3155                 if (open_pack_index(p))
3156                         die(_("cannot open pack index"));
3157
3158                 ALLOC_GROW(in_pack.array,
3159                            in_pack.nr + p->num_objects,
3160                            in_pack.alloc);
3161
3162                 for (i = 0; i < p->num_objects; i++) {
3163                         nth_packed_object_id(&oid, p, i);
3164                         o = lookup_unknown_object(&oid);
3165                         if (!(o->flags & OBJECT_ADDED))
3166                                 mark_in_pack_object(o, p, &in_pack);
3167                         o->flags |= OBJECT_ADDED;
3168                 }
3169         }
3170
3171         if (in_pack.nr) {
3172                 QSORT(in_pack.array, in_pack.nr, ofscmp);
3173                 for (i = 0; i < in_pack.nr; i++) {
3174                         struct object *o = in_pack.array[i].object;
3175                         add_object_entry(&o->oid, o->type, "", 0);
3176                 }
3177         }
3178         free(in_pack.array);
3179 }
3180
3181 static int add_loose_object(const struct object_id *oid, const char *path,
3182                             void *data)
3183 {
3184         enum object_type type = oid_object_info(the_repository, oid, NULL);
3185
3186         if (type < 0) {
3187                 warning(_("loose object at %s could not be examined"), path);
3188                 return 0;
3189         }
3190
3191         add_object_entry(oid, type, "", 0);
3192         return 0;
3193 }
3194
3195 /*
3196  * We actually don't even have to worry about reachability here.
3197  * add_object_entry will weed out duplicates, so we just add every
3198  * loose object we find.
3199  */
3200 static void add_unreachable_loose_objects(void)
3201 {
3202         for_each_loose_file_in_objdir(get_object_directory(),
3203                                       add_loose_object,
3204                                       NULL, NULL, NULL);
3205 }
3206
3207 static int has_sha1_pack_kept_or_nonlocal(const struct object_id *oid)
3208 {
3209         static struct packed_git *last_found = (void *)1;
3210         struct packed_git *p;
3211
3212         p = (last_found != (void *)1) ? last_found :
3213                                         get_all_packs(the_repository);
3214
3215         while (p) {
3216                 if ((!p->pack_local || p->pack_keep ||
3217                                 p->pack_keep_in_core) &&
3218                         find_pack_entry_one(oid->hash, p)) {
3219                         last_found = p;
3220                         return 1;
3221                 }
3222                 if (p == last_found)
3223                         p = get_all_packs(the_repository);
3224                 else
3225                         p = p->next;
3226                 if (p == last_found)
3227                         p = p->next;
3228         }
3229         return 0;
3230 }
3231
3232 /*
3233  * Store a list of sha1s that are should not be discarded
3234  * because they are either written too recently, or are
3235  * reachable from another object that was.
3236  *
3237  * This is filled by get_object_list.
3238  */
3239 static struct oid_array recent_objects;
3240
3241 static int loosened_object_can_be_discarded(const struct object_id *oid,
3242                                             timestamp_t mtime)
3243 {
3244         if (!unpack_unreachable_expiration)
3245                 return 0;
3246         if (mtime > unpack_unreachable_expiration)
3247                 return 0;
3248         if (oid_array_lookup(&recent_objects, oid) >= 0)
3249                 return 0;
3250         return 1;
3251 }
3252
3253 static void loosen_unused_packed_objects(void)
3254 {
3255         struct packed_git *p;
3256         uint32_t i;
3257         struct object_id oid;
3258
3259         for (p = get_all_packs(the_repository); p; p = p->next) {
3260                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
3261                         continue;
3262
3263                 if (open_pack_index(p))
3264                         die(_("cannot open pack index"));
3265
3266                 for (i = 0; i < p->num_objects; i++) {
3267                         nth_packed_object_id(&oid, p, i);
3268                         if (!packlist_find(&to_pack, &oid) &&
3269                             !has_sha1_pack_kept_or_nonlocal(&oid) &&
3270                             !loosened_object_can_be_discarded(&oid, p->mtime))
3271                                 if (force_object_loose(&oid, p->mtime))
3272                                         die(_("unable to force loose object"));
3273                 }
3274         }
3275 }
3276
3277 /*
3278  * This tracks any options which pack-reuse code expects to be on, or which a
3279  * reader of the pack might not understand, and which would therefore prevent
3280  * blind reuse of what we have on disk.
3281  */
3282 static int pack_options_allow_reuse(void)
3283 {
3284         return allow_pack_reuse &&
3285                pack_to_stdout &&
3286                !ignore_packed_keep_on_disk &&
3287                !ignore_packed_keep_in_core &&
3288                (!local || !have_non_local_packs) &&
3289                !incremental;
3290 }
3291
3292 static int get_object_list_from_bitmap(struct rev_info *revs)
3293 {
3294         if (!(bitmap_git = prepare_bitmap_walk(revs, &filter_options)))
3295                 return -1;
3296
3297         if (pack_options_allow_reuse() &&
3298             !reuse_partial_packfile_from_bitmap(
3299                         bitmap_git,
3300                         &reuse_packfile,
3301                         &reuse_packfile_objects,
3302                         &reuse_packfile_bitmap)) {
3303                 assert(reuse_packfile_objects);
3304                 nr_result += reuse_packfile_objects;
3305                 display_progress(progress_state, nr_result);
3306         }
3307
3308         traverse_bitmap_commit_list(bitmap_git, revs,
3309                                     &add_object_entry_from_bitmap);
3310         return 0;
3311 }
3312
3313 static void record_recent_object(struct object *obj,
3314                                  const char *name,
3315                                  void *data)
3316 {
3317         oid_array_append(&recent_objects, &obj->oid);
3318 }
3319
3320 static void record_recent_commit(struct commit *commit, void *data)
3321 {
3322         oid_array_append(&recent_objects, &commit->object.oid);
3323 }
3324
3325 static void get_object_list(int ac, const char **av)
3326 {
3327         struct rev_info revs;
3328         struct setup_revision_opt s_r_opt = {
3329                 .allow_exclude_promisor_objects = 1,
3330         };
3331         char line[1000];
3332         int flags = 0;
3333         int save_warning;
3334
3335         repo_init_revisions(the_repository, &revs, NULL);
3336         save_commit_buffer = 0;
3337         setup_revisions(ac, av, &revs, &s_r_opt);
3338
3339         /* make sure shallows are read */
3340         is_repository_shallow(the_repository);
3341
3342         save_warning = warn_on_object_refname_ambiguity;
3343         warn_on_object_refname_ambiguity = 0;
3344
3345         while (fgets(line, sizeof(line), stdin) != NULL) {
3346                 int len = strlen(line);
3347                 if (len && line[len - 1] == '\n')
3348                         line[--len] = 0;
3349                 if (!len)
3350                         break;
3351                 if (*line == '-') {
3352                         if (!strcmp(line, "--not")) {
3353                                 flags ^= UNINTERESTING;
3354                                 write_bitmap_index = 0;
3355                                 continue;
3356                         }
3357                         if (starts_with(line, "--shallow ")) {
3358                                 struct object_id oid;
3359                                 if (get_oid_hex(line + 10, &oid))
3360                                         die("not an object name '%s'", line + 10);
3361                                 register_shallow(the_repository, &oid);
3362                                 use_bitmap_index = 0;
3363                                 continue;
3364                         }
3365                         die(_("not a rev '%s'"), line);
3366                 }
3367                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
3368                         die(_("bad revision '%s'"), line);
3369         }
3370
3371         warn_on_object_refname_ambiguity = save_warning;
3372
3373         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
3374                 return;
3375
3376         if (use_delta_islands)
3377                 load_delta_islands(the_repository, progress);
3378
3379         if (prepare_revision_walk(&revs))
3380                 die(_("revision walk setup failed"));
3381         mark_edges_uninteresting(&revs, show_edge, sparse);
3382
3383         if (!fn_show_object)
3384                 fn_show_object = show_object;
3385         traverse_commit_list_filtered(&filter_options, &revs,
3386                                       show_commit, fn_show_object, NULL,
3387                                       NULL);
3388
3389         if (unpack_unreachable_expiration) {
3390                 revs.ignore_missing_links = 1;
3391                 if (add_unseen_recent_objects_to_traversal(&revs,
3392                                 unpack_unreachable_expiration))
3393                         die(_("unable to add recent objects"));
3394                 if (prepare_revision_walk(&revs))
3395                         die(_("revision walk setup failed"));
3396                 traverse_commit_list(&revs, record_recent_commit,
3397                                      record_recent_object, NULL);
3398         }
3399
3400         if (keep_unreachable)
3401                 add_objects_in_unpacked_packs();
3402         if (pack_loose_unreachable)
3403                 add_unreachable_loose_objects();
3404         if (unpack_unreachable)
3405                 loosen_unused_packed_objects();
3406
3407         oid_array_clear(&recent_objects);
3408 }
3409
3410 static void add_extra_kept_packs(const struct string_list *names)
3411 {
3412         struct packed_git *p;
3413
3414         if (!names->nr)
3415                 return;
3416
3417         for (p = get_all_packs(the_repository); p; p = p->next) {
3418                 const char *name = basename(p->pack_name);
3419                 int i;
3420
3421                 if (!p->pack_local)
3422                         continue;
3423
3424                 for (i = 0; i < names->nr; i++)
3425                         if (!fspathcmp(name, names->items[i].string))
3426                                 break;
3427
3428                 if (i < names->nr) {
3429                         p->pack_keep_in_core = 1;
3430                         ignore_packed_keep_in_core = 1;
3431                         continue;
3432                 }
3433         }
3434 }
3435
3436 static int option_parse_index_version(const struct option *opt,
3437                                       const char *arg, int unset)
3438 {
3439         char *c;
3440         const char *val = arg;
3441
3442         BUG_ON_OPT_NEG(unset);
3443
3444         pack_idx_opts.version = strtoul(val, &c, 10);
3445         if (pack_idx_opts.version > 2)
3446                 die(_("unsupported index version %s"), val);
3447         if (*c == ',' && c[1])
3448                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
3449         if (*c || pack_idx_opts.off32_limit & 0x80000000)
3450                 die(_("bad index version '%s'"), val);
3451         return 0;
3452 }
3453
3454 static int option_parse_unpack_unreachable(const struct option *opt,
3455                                            const char *arg, int unset)
3456 {
3457         if (unset) {
3458                 unpack_unreachable = 0;
3459                 unpack_unreachable_expiration = 0;
3460         }
3461         else {
3462                 unpack_unreachable = 1;
3463                 if (arg)
3464                         unpack_unreachable_expiration = approxidate(arg);
3465         }
3466         return 0;
3467 }
3468
3469 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
3470 {
3471         int use_internal_rev_list = 0;
3472         int shallow = 0;
3473         int all_progress_implied = 0;
3474         struct strvec rp = STRVEC_INIT;
3475         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
3476         int rev_list_index = 0;
3477         struct string_list keep_pack_list = STRING_LIST_INIT_NODUP;
3478         struct option pack_objects_options[] = {
3479                 OPT_SET_INT('q', "quiet", &progress,
3480                             N_("do not show progress meter"), 0),
3481                 OPT_SET_INT(0, "progress", &progress,
3482                             N_("show progress meter"), 1),
3483                 OPT_SET_INT(0, "all-progress", &progress,
3484                             N_("show progress meter during object writing phase"), 2),
3485                 OPT_BOOL(0, "all-progress-implied",
3486                          &all_progress_implied,
3487                          N_("similar to --all-progress when progress meter is shown")),
3488                 OPT_CALLBACK_F(0, "index-version", NULL, N_("<version>[,<offset>]"),
3489                   N_("write the pack index file in the specified idx format version"),
3490                   PARSE_OPT_NONEG, option_parse_index_version),
3491                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
3492                               N_("maximum size of each output pack file")),
3493                 OPT_BOOL(0, "local", &local,
3494                          N_("ignore borrowed objects from alternate object store")),
3495                 OPT_BOOL(0, "incremental", &incremental,
3496                          N_("ignore packed objects")),
3497                 OPT_INTEGER(0, "window", &window,
3498                             N_("limit pack window by objects")),
3499                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
3500                               N_("limit pack window by memory in addition to object limit")),
3501                 OPT_INTEGER(0, "depth", &depth,
3502                             N_("maximum length of delta chain allowed in the resulting pack")),
3503                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
3504                          N_("reuse existing deltas")),
3505                 OPT_BOOL(0, "reuse-object", &reuse_object,
3506                          N_("reuse existing objects")),
3507                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
3508                          N_("use OFS_DELTA objects")),
3509                 OPT_INTEGER(0, "threads", &delta_search_threads,
3510                             N_("use threads when searching for best delta matches")),
3511                 OPT_BOOL(0, "non-empty", &non_empty,
3512                          N_("do not create an empty pack output")),
3513                 OPT_BOOL(0, "revs", &use_internal_rev_list,
3514                          N_("read revision arguments from standard input")),
3515                 OPT_SET_INT_F(0, "unpacked", &rev_list_unpacked,
3516                               N_("limit the objects to those that are not yet packed"),
3517                               1, PARSE_OPT_NONEG),
3518                 OPT_SET_INT_F(0, "all", &rev_list_all,
3519                               N_("include objects reachable from any reference"),
3520                               1, PARSE_OPT_NONEG),
3521                 OPT_SET_INT_F(0, "reflog", &rev_list_reflog,
3522                               N_("include objects referred by reflog entries"),
3523                               1, PARSE_OPT_NONEG),
3524                 OPT_SET_INT_F(0, "indexed-objects", &rev_list_index,
3525                               N_("include objects referred to by the index"),
3526                               1, PARSE_OPT_NONEG),
3527                 OPT_BOOL(0, "stdout", &pack_to_stdout,
3528                          N_("output pack to stdout")),
3529                 OPT_BOOL(0, "include-tag", &include_tag,
3530                          N_("include tag objects that refer to objects to be packed")),
3531                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
3532                          N_("keep unreachable objects")),
3533                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
3534                          N_("pack loose unreachable objects")),
3535                 OPT_CALLBACK_F(0, "unpack-unreachable", NULL, N_("time"),
3536                   N_("unpack unreachable objects newer than <time>"),
3537                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable),
3538                 OPT_BOOL(0, "sparse", &sparse,
3539                          N_("use the sparse reachability algorithm")),
3540                 OPT_BOOL(0, "thin", &thin,
3541                          N_("create thin packs")),
3542                 OPT_BOOL(0, "shallow", &shallow,
3543                          N_("create packs suitable for shallow fetches")),
3544                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep_on_disk,
3545                          N_("ignore packs that have companion .keep file")),
3546                 OPT_STRING_LIST(0, "keep-pack", &keep_pack_list, N_("name"),
3547                                 N_("ignore this pack")),
3548                 OPT_INTEGER(0, "compression", &pack_compression_level,
3549                             N_("pack compression level")),
3550                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
3551                             N_("do not hide commits by grafts"), 0),
3552                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
3553                          N_("use a bitmap index if available to speed up counting objects")),
3554                 OPT_SET_INT(0, "write-bitmap-index", &write_bitmap_index,
3555                             N_("write a bitmap index together with the pack index"),
3556                             WRITE_BITMAP_TRUE),
3557                 OPT_SET_INT_F(0, "write-bitmap-index-quiet",
3558                               &write_bitmap_index,
3559                               N_("write a bitmap index if possible"),
3560                               WRITE_BITMAP_QUIET, PARSE_OPT_HIDDEN),
3561                 OPT_PARSE_LIST_OBJECTS_FILTER(&filter_options),
3562                 OPT_CALLBACK_F(0, "missing", NULL, N_("action"),
3563                   N_("handling for missing objects"), PARSE_OPT_NONEG,
3564                   option_parse_missing_action),
3565                 OPT_BOOL(0, "exclude-promisor-objects", &exclude_promisor_objects,
3566                          N_("do not pack objects in promisor packfiles")),
3567                 OPT_BOOL(0, "delta-islands", &use_delta_islands,
3568                          N_("respect islands during delta compression")),
3569                 OPT_STRING_LIST(0, "uri-protocol", &uri_protocols,
3570                                 N_("protocol"),
3571                                 N_("exclude any configured uploadpack.blobpackfileuri with this protocol")),
3572                 OPT_END(),
3573         };
3574
3575         if (DFS_NUM_STATES > (1 << OE_DFS_STATE_BITS))
3576                 BUG("too many dfs states, increase OE_DFS_STATE_BITS");
3577
3578         read_replace_refs = 0;
3579
3580         sparse = git_env_bool("GIT_TEST_PACK_SPARSE", -1);
3581         prepare_repo_settings(the_repository);
3582         if (sparse < 0)
3583                 sparse = the_repository->settings.pack_use_sparse;
3584
3585         reset_pack_idx_option(&pack_idx_opts);
3586         git_config(git_pack_config, NULL);
3587
3588         progress = isatty(2);
3589         argc = parse_options(argc, argv, prefix, pack_objects_options,
3590                              pack_usage, 0);
3591
3592         if (argc) {
3593                 base_name = argv[0];
3594                 argc--;
3595         }
3596         if (pack_to_stdout != !base_name || argc)
3597                 usage_with_options(pack_usage, pack_objects_options);
3598
3599         if (depth >= (1 << OE_DEPTH_BITS)) {
3600                 warning(_("delta chain depth %d is too deep, forcing %d"),
3601                         depth, (1 << OE_DEPTH_BITS) - 1);
3602                 depth = (1 << OE_DEPTH_BITS) - 1;
3603         }
3604         if (cache_max_small_delta_size >= (1U << OE_Z_DELTA_BITS)) {
3605                 warning(_("pack.deltaCacheLimit is too high, forcing %d"),
3606                         (1U << OE_Z_DELTA_BITS) - 1);
3607                 cache_max_small_delta_size = (1U << OE_Z_DELTA_BITS) - 1;
3608         }
3609
3610         strvec_push(&rp, "pack-objects");
3611         if (thin) {
3612                 use_internal_rev_list = 1;
3613                 strvec_push(&rp, shallow
3614                                 ? "--objects-edge-aggressive"
3615                                 : "--objects-edge");
3616         } else
3617                 strvec_push(&rp, "--objects");
3618
3619         if (rev_list_all) {
3620                 use_internal_rev_list = 1;
3621                 strvec_push(&rp, "--all");
3622         }
3623         if (rev_list_reflog) {
3624                 use_internal_rev_list = 1;
3625                 strvec_push(&rp, "--reflog");
3626         }
3627         if (rev_list_index) {
3628                 use_internal_rev_list = 1;
3629                 strvec_push(&rp, "--indexed-objects");
3630         }
3631         if (rev_list_unpacked) {
3632                 use_internal_rev_list = 1;
3633                 strvec_push(&rp, "--unpacked");
3634         }
3635
3636         if (exclude_promisor_objects) {
3637                 use_internal_rev_list = 1;
3638                 fetch_if_missing = 0;
3639                 strvec_push(&rp, "--exclude-promisor-objects");
3640         }
3641         if (unpack_unreachable || keep_unreachable || pack_loose_unreachable)
3642                 use_internal_rev_list = 1;
3643
3644         if (!reuse_object)
3645                 reuse_delta = 0;
3646         if (pack_compression_level == -1)
3647                 pack_compression_level = Z_DEFAULT_COMPRESSION;
3648         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
3649                 die(_("bad pack compression level %d"), pack_compression_level);
3650
3651         if (!delta_search_threads)      /* --threads=0 means autodetect */
3652                 delta_search_threads = online_cpus();
3653
3654         if (!HAVE_THREADS && delta_search_threads != 1)
3655                 warning(_("no threads support, ignoring --threads"));
3656         if (!pack_to_stdout && !pack_size_limit)
3657                 pack_size_limit = pack_size_limit_cfg;
3658         if (pack_to_stdout && pack_size_limit)
3659                 die(_("--max-pack-size cannot be used to build a pack for transfer"));
3660         if (pack_size_limit && pack_size_limit < 1024*1024) {
3661                 warning(_("minimum pack size limit is 1 MiB"));
3662                 pack_size_limit = 1024*1024;
3663         }
3664
3665         if (!pack_to_stdout && thin)
3666                 die(_("--thin cannot be used to build an indexable pack"));
3667
3668         if (keep_unreachable && unpack_unreachable)
3669                 die(_("--keep-unreachable and --unpack-unreachable are incompatible"));
3670         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3671                 unpack_unreachable_expiration = 0;
3672
3673         if (filter_options.choice) {
3674                 if (!pack_to_stdout)
3675                         die(_("cannot use --filter without --stdout"));
3676         }
3677
3678         /*
3679          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3680          *
3681          * - to produce good pack (with bitmap index not-yet-packed objects are
3682          *   packed in suboptimal order).
3683          *
3684          * - to use more robust pack-generation codepath (avoiding possible
3685          *   bugs in bitmap code and possible bitmap index corruption).
3686          */
3687         if (!pack_to_stdout)
3688                 use_bitmap_index_default = 0;
3689
3690         if (use_bitmap_index < 0)
3691                 use_bitmap_index = use_bitmap_index_default;
3692
3693         /* "hard" reasons not to use bitmaps; these just won't work at all */
3694         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow(the_repository))
3695                 use_bitmap_index = 0;
3696
3697         if (pack_to_stdout || !rev_list_all)
3698                 write_bitmap_index = 0;
3699
3700         if (use_delta_islands)
3701                 strvec_push(&rp, "--topo-order");
3702
3703         if (progress && all_progress_implied)
3704                 progress = 2;
3705
3706         add_extra_kept_packs(&keep_pack_list);
3707         if (ignore_packed_keep_on_disk) {
3708                 struct packed_git *p;
3709                 for (p = get_all_packs(the_repository); p; p = p->next)
3710                         if (p->pack_local && p->pack_keep)
3711                                 break;
3712                 if (!p) /* no keep-able packs found */
3713                         ignore_packed_keep_on_disk = 0;
3714         }
3715         if (local) {
3716                 /*
3717                  * unlike ignore_packed_keep_on_disk above, we do not
3718                  * want to unset "local" based on looking at packs, as
3719                  * it also covers non-local objects
3720                  */
3721                 struct packed_git *p;
3722                 for (p = get_all_packs(the_repository); p; p = p->next) {
3723                         if (!p->pack_local) {
3724                                 have_non_local_packs = 1;
3725                                 break;
3726                         }
3727                 }
3728         }
3729
3730         trace2_region_enter("pack-objects", "enumerate-objects",
3731                             the_repository);
3732         prepare_packing_data(the_repository, &to_pack);
3733
3734         if (progress)
3735                 progress_state = start_progress(_("Enumerating objects"), 0);
3736         if (!use_internal_rev_list)
3737                 read_object_list_from_stdin();
3738         else {
3739                 get_object_list(rp.nr, rp.v);
3740                 strvec_clear(&rp);
3741         }
3742         cleanup_preferred_base();
3743         if (include_tag && nr_result)
3744                 for_each_ref(add_ref_tag, NULL);
3745         stop_progress(&progress_state);
3746         trace2_region_leave("pack-objects", "enumerate-objects",
3747                             the_repository);
3748
3749         if (non_empty && !nr_result)
3750                 return 0;
3751         if (nr_result) {
3752                 trace2_region_enter("pack-objects", "prepare-pack",
3753                                     the_repository);
3754                 prepare_pack(window, depth);
3755                 trace2_region_leave("pack-objects", "prepare-pack",
3756                                     the_repository);
3757         }
3758
3759         trace2_region_enter("pack-objects", "write-pack-file", the_repository);
3760         write_excluded_by_configs();
3761         write_pack_file();
3762         trace2_region_leave("pack-objects", "write-pack-file", the_repository);
3763
3764         if (progress)
3765                 fprintf_ln(stderr,
3766                            _("Total %"PRIu32" (delta %"PRIu32"),"
3767                              " reused %"PRIu32" (delta %"PRIu32"),"
3768                              " pack-reused %"PRIu32),
3769                            written, written_delta, reused, reused_delta,
3770                            reuse_packfile_objects);
3771         return 0;
3772 }