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

   1 #include "builtin.h"
   2 #include "cache.h"
   3 #include "config.h"
   4 #include "attr.h"
   5 #include "object.h"
   6 #include "blob.h"
   7 #include "commit.h"
   8 #include "tag.h"
   9 #include "tree.h"
  10 #include "delta.h"
  11 #include "pack.h"
  12 #include "pack-revindex.h"
  13 #include "csum-file.h"
  14 #include "tree-walk.h"
  15 #include "diff.h"
  16 #include "revision.h"
  17 #include "list-objects.h"
  18 #include "list-objects-filter.h"
  19 #include "list-objects-filter-options.h"
  20 #include "pack-objects.h"
  21 #include "progress.h"
  22 #include "refs.h"
  23 #include "streaming.h"
  24 #include "thread-utils.h"
  25 #include "pack-bitmap.h"
  26 #include "reachable.h"
  27 #include "sha1-array.h"
  28 #include "argv-array.h"
  29 #include "list.h"
  30 #include "packfile.h"
  31
  32 static const char *pack_usage[] = {
  33         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
  34         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
  35         NULL
  36 };
  37
  38 /*
  39  * Objects we are going to pack are collected in the `to_pack` structure.
  40  * It contains an array (dynamically expanded) of the object data, and a map
  41  * that can resolve SHA1s to their position in the array.
  42  */
  43 static struct packing_data to_pack;
  44
  45 static struct pack_idx_entry **written_list;
  46 static uint32_t nr_result, nr_written;
  47
  48 static int non_empty;
  49 static int reuse_delta = 1, reuse_object = 1;
  50 static int keep_unreachable, unpack_unreachable, include_tag;
  51 static timestamp_t unpack_unreachable_expiration;
  52 static int pack_loose_unreachable;
  53 static int local;
  54 static int have_non_local_packs;
  55 static int incremental;
  56 static int ignore_packed_keep;
  57 static int allow_ofs_delta;
  58 static struct pack_idx_option pack_idx_opts;
  59 static const char *base_name;
  60 static int progress = 1;
  61 static int window = 10;
  62 static unsigned long pack_size_limit;
  63 static int depth = 50;
  64 static int delta_search_threads;
  65 static int pack_to_stdout;
  66 static int num_preferred_base;
  67 static struct progress *progress_state;
  68
  69 static struct packed_git *reuse_packfile;
  70 static uint32_t reuse_packfile_objects;
  71 static off_t reuse_packfile_offset;
  72
  73 static int use_bitmap_index_default = 1;
  74 static int use_bitmap_index = -1;
  75 static int write_bitmap_index;
  76 static uint16_t write_bitmap_options;
  77
  78 static int exclude_promisor_objects;
  79
  80 static unsigned long delta_cache_size = 0;
  81 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
  82 static unsigned long cache_max_small_delta_size = 1000;
  83
  84 static unsigned long window_memory_limit = 0;
  85
  86 static struct list_objects_filter_options filter_options;
  87
  88 enum missing_action {
  89         MA_ERROR = 0,      /* fail if any missing objects are encountered */
  90         MA_ALLOW_ANY,      /* silently allow ALL missing objects */
  91         MA_ALLOW_PROMISOR, /* silently allow all missing PROMISOR objects */
  92 };
  93 static enum missing_action arg_missing_action;
  94 static show_object_fn fn_show_object;
  95
  96 /*
  97  * stats
  98  */
  99 static uint32_t written, written_delta;
 100 static uint32_t reused, reused_delta;
 101
 102 /*
 103  * Indexed commits
 104  */
 105 static struct commit **indexed_commits;
 106 static unsigned int indexed_commits_nr;
 107 static unsigned int indexed_commits_alloc;
 108
 109 static void index_commit_for_bitmap(struct commit *commit)
 110 {
 111         if (indexed_commits_nr >= indexed_commits_alloc) {
 112                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
 113                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
 114         }
 115
 116         indexed_commits[indexed_commits_nr++] = commit;
 117 }
 118
 119 static void *get_delta(struct object_entry *entry)
 120 {
 121         unsigned long size, base_size, delta_size;
 122         void *buf, *base_buf, *delta_buf;
 123         enum object_type type;
 124
 125         buf = read_sha1_file(entry->idx.oid.hash, &type, &size);
 126         if (!buf)
 127                 die("unable to read %s", oid_to_hex(&entry->idx.oid));
 128         base_buf = read_sha1_file(entry->delta->idx.oid.hash, &type,
 129                                   &base_size);
 130         if (!base_buf)
 131                 die("unable to read %s",
 132                     oid_to_hex(&entry->delta->idx.oid));
 133         delta_buf = diff_delta(base_buf, base_size,
 134                                buf, size, &delta_size, 0);
 135         if (!delta_buf || delta_size != entry->delta_size)
 136                 die("delta size changed");
 137         free(buf);
 138         free(base_buf);
 139         return delta_buf;
 140 }
 141
 142 static unsigned long do_compress(void **pptr, unsigned long size)
 143 {
 144         git_zstream stream;
 145         void *in, *out;
 146         unsigned long maxsize;
 147
 148         git_deflate_init(&stream, pack_compression_level);
 149         maxsize = git_deflate_bound(&stream, size);
 150
 151         in = *pptr;
 152         out = xmalloc(maxsize);
 153         *pptr = out;
 154
 155         stream.next_in = in;
 156         stream.avail_in = size;
 157         stream.next_out = out;
 158         stream.avail_out = maxsize;
 159         while (git_deflate(&stream, Z_FINISH) == Z_OK)
 160                 ; /* nothing */
 161         git_deflate_end(&stream);
 162
 163         free(in);
 164         return stream.total_out;
 165 }
 166
 167 static unsigned long write_large_blob_data(struct git_istream *st, struct hashfile *f,
 168                                            const struct object_id *oid)
 169 {
 170         git_zstream stream;
 171         unsigned char ibuf[1024 * 16];
 172         unsigned char obuf[1024 * 16];
 173         unsigned long olen = 0;
 174
 175         git_deflate_init(&stream, pack_compression_level);
 176
 177         for (;;) {
 178                 ssize_t readlen;
 179                 int zret = Z_OK;
 180                 readlen = read_istream(st, ibuf, sizeof(ibuf));
 181                 if (readlen == -1)
 182                         die(_("unable to read %s"), oid_to_hex(oid));
 183
 184                 stream.next_in = ibuf;
 185                 stream.avail_in = readlen;
 186                 while ((stream.avail_in || readlen == 0) &&
 187                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
 188                         stream.next_out = obuf;
 189                         stream.avail_out = sizeof(obuf);
 190                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 191                         hashwrite(f, obuf, stream.next_out - obuf);
 192                         olen += stream.next_out - obuf;
 193                 }
 194                 if (stream.avail_in)
 195                         die(_("deflate error (%d)"), zret);
 196                 if (readlen == 0) {
 197                         if (zret != Z_STREAM_END)
 198                                 die(_("deflate error (%d)"), zret);
 199                         break;
 200                 }
 201         }
 202         git_deflate_end(&stream);
 203         return olen;
 204 }
 205
 206 /*
 207  * we are going to reuse the existing object data as is.  make
 208  * sure it is not corrupt.
 209  */
 210 static int check_pack_inflate(struct packed_git *p,
 211                 struct pack_window **w_curs,
 212                 off_t offset,
 213                 off_t len,
 214                 unsigned long expect)
 215 {
 216         git_zstream stream;
 217         unsigned char fakebuf[4096], *in;
 218         int st;
 219
 220         memset(&stream, 0, sizeof(stream));
 221         git_inflate_init(&stream);
 222         do {
 223                 in = use_pack(p, w_curs, offset, &stream.avail_in);
 224                 stream.next_in = in;
 225                 stream.next_out = fakebuf;
 226                 stream.avail_out = sizeof(fakebuf);
 227                 st = git_inflate(&stream, Z_FINISH);
 228                 offset += stream.next_in - in;
 229         } while (st == Z_OK || st == Z_BUF_ERROR);
 230         git_inflate_end(&stream);
 231         return (st == Z_STREAM_END &&
 232                 stream.total_out == expect &&
 233                 stream.total_in == len) ? 0 : -1;
 234 }
 235
 236 static void copy_pack_data(struct hashfile *f,
 237                 struct packed_git *p,
 238                 struct pack_window **w_curs,
 239                 off_t offset,
 240                 off_t len)
 241 {
 242         unsigned char *in;
 243         unsigned long avail;
 244
 245         while (len) {
 246                 in = use_pack(p, w_curs, offset, &avail);
 247                 if (avail > len)
 248                         avail = (unsigned long)len;
 249                 hashwrite(f, in, avail);
 250                 offset += avail;
 251                 len -= avail;
 252         }
 253 }
 254
 255 /* Return 0 if we will bust the pack-size limit */
 256 static unsigned long write_no_reuse_object(struct hashfile *f, struct object_entry *entry,
 257                                            unsigned long limit, int usable_delta)
 258 {
 259         unsigned long size, datalen;
 260         unsigned char header[MAX_PACK_OBJECT_HEADER],
 261                       dheader[MAX_PACK_OBJECT_HEADER];
 262         unsigned hdrlen;
 263         enum object_type type;
 264         void *buf;
 265         struct git_istream *st = NULL;
 266
 267         if (!usable_delta) {
 268                 if (entry->type == OBJ_BLOB &&
 269                     entry->size > big_file_threshold &&
 270                     (st = open_istream(entry->idx.oid.hash, &type, &size, NULL)) != NULL)
 271                         buf = NULL;
 272                 else {
 273                         buf = read_sha1_file(entry->idx.oid.hash, &type,
 274                                              &size);
 275                         if (!buf)
 276                                 die(_("unable to read %s"),
 277                                     oid_to_hex(&entry->idx.oid));
 278                 }
 279                 /*
 280                  * make sure no cached delta data remains from a
 281                  * previous attempt before a pack split occurred.
 282                  */
 283                 FREE_AND_NULL(entry->delta_data);
 284                 entry->z_delta_size = 0;
 285         } else if (entry->delta_data) {
 286                 size = entry->delta_size;
 287                 buf = entry->delta_data;
 288                 entry->delta_data = NULL;
 289                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 290                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 291         } else {
 292                 buf = get_delta(entry);
 293                 size = entry->delta_size;
 294                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 295                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 296         }
 297
 298         if (st) /* large blob case, just assume we don't compress well */
 299                 datalen = size;
 300         else if (entry->z_delta_size)
 301                 datalen = entry->z_delta_size;
 302         else
 303                 datalen = do_compress(&buf, size);
 304
 305         /*
 306          * The object header is a byte of 'type' followed by zero or
 307          * more bytes of length.
 308          */
 309         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 310                                               type, size);
 311
 312         if (type == OBJ_OFS_DELTA) {
 313                 /*
 314                  * Deltas with relative base contain an additional
 315                  * encoding of the relative offset for the delta
 316                  * base from this object's position in the pack.
 317                  */
 318                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 319                 unsigned pos = sizeof(dheader) - 1;
 320                 dheader[pos] = ofs & 127;
 321                 while (ofs >>= 7)
 322                         dheader[--pos] = 128 | (--ofs & 127);
 323                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 324                         if (st)
 325                                 close_istream(st);
 326                         free(buf);
 327                         return 0;
 328                 }
 329                 hashwrite(f, header, hdrlen);
 330                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
 331                 hdrlen += sizeof(dheader) - pos;
 332         } else if (type == OBJ_REF_DELTA) {
 333                 /*
 334                  * Deltas with a base reference contain
 335                  * an additional 20 bytes for the base sha1.
 336                  */
 337                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 338                         if (st)
 339                                 close_istream(st);
 340                         free(buf);
 341                         return 0;
 342                 }
 343                 hashwrite(f, header, hdrlen);
 344                 hashwrite(f, entry->delta->idx.oid.hash, 20);
 345                 hdrlen += 20;
 346         } else {
 347                 if (limit && hdrlen + datalen + 20 >= limit) {
 348                         if (st)
 349                                 close_istream(st);
 350                         free(buf);
 351                         return 0;
 352                 }
 353                 hashwrite(f, header, hdrlen);
 354         }
 355         if (st) {
 356                 datalen = write_large_blob_data(st, f, &entry->idx.oid);
 357                 close_istream(st);
 358         } else {
 359                 hashwrite(f, buf, datalen);
 360                 free(buf);
 361         }
 362
 363         return hdrlen + datalen;
 364 }
 365
 366 /* Return 0 if we will bust the pack-size limit */
 367 static off_t write_reuse_object(struct hashfile *f, struct object_entry *entry,
 368                                 unsigned long limit, int usable_delta)
 369 {
 370         struct packed_git *p = entry->in_pack;
 371         struct pack_window *w_curs = NULL;
 372         struct revindex_entry *revidx;
 373         off_t offset;
 374         enum object_type type = entry->type;
 375         off_t datalen;
 376         unsigned char header[MAX_PACK_OBJECT_HEADER],
 377                       dheader[MAX_PACK_OBJECT_HEADER];
 378         unsigned hdrlen;
 379
 380         if (entry->delta)
 381                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 382                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 383         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 384                                               type, entry->size);
 385
 386         offset = entry->in_pack_offset;
 387         revidx = find_pack_revindex(p, offset);
 388         datalen = revidx[1].offset - offset;
 389         if (!pack_to_stdout && p->index_version > 1 &&
 390             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 391                 error("bad packed object CRC for %s",
 392                       oid_to_hex(&entry->idx.oid));
 393                 unuse_pack(&w_curs);
 394                 return write_no_reuse_object(f, entry, limit, usable_delta);
 395         }
 396
 397         offset += entry->in_pack_header_size;
 398         datalen -= entry->in_pack_header_size;
 399
 400         if (!pack_to_stdout && p->index_version == 1 &&
 401             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 402                 error("corrupt packed object for %s",
 403                       oid_to_hex(&entry->idx.oid));
 404                 unuse_pack(&w_curs);
 405                 return write_no_reuse_object(f, entry, limit, usable_delta);
 406         }
 407
 408         if (type == OBJ_OFS_DELTA) {
 409                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 410                 unsigned pos = sizeof(dheader) - 1;
 411                 dheader[pos] = ofs & 127;
 412                 while (ofs >>= 7)
 413                         dheader[--pos] = 128 | (--ofs & 127);
 414                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 415                         unuse_pack(&w_curs);
 416                         return 0;
 417                 }
 418                 hashwrite(f, header, hdrlen);
 419                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
 420                 hdrlen += sizeof(dheader) - pos;
 421                 reused_delta++;
 422         } else if (type == OBJ_REF_DELTA) {
 423                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 424                         unuse_pack(&w_curs);
 425                         return 0;
 426                 }
 427                 hashwrite(f, header, hdrlen);
 428                 hashwrite(f, entry->delta->idx.oid.hash, 20);
 429                 hdrlen += 20;
 430                 reused_delta++;
 431         } else {
 432                 if (limit && hdrlen + datalen + 20 >= limit) {
 433                         unuse_pack(&w_curs);
 434                         return 0;
 435                 }
 436                 hashwrite(f, header, hdrlen);
 437         }
 438         copy_pack_data(f, p, &w_curs, offset, datalen);
 439         unuse_pack(&w_curs);
 440         reused++;
 441         return hdrlen + datalen;
 442 }
 443
 444 /* Return 0 if we will bust the pack-size limit */
 445 static off_t write_object(struct hashfile *f,
 446                           struct object_entry *entry,
 447                           off_t write_offset)
 448 {
 449         unsigned long limit;
 450         off_t len;
 451         int usable_delta, to_reuse;
 452
 453         if (!pack_to_stdout)
 454                 crc32_begin(f);
 455
 456         /* apply size limit if limited packsize and not first object */
 457         if (!pack_size_limit || !nr_written)
 458                 limit = 0;
 459         else if (pack_size_limit <= write_offset)
 460                 /*
 461                  * the earlier object did not fit the limit; avoid
 462                  * mistaking this with unlimited (i.e. limit = 0).
 463                  */
 464                 limit = 1;
 465         else
 466                 limit = pack_size_limit - write_offset;
 467
 468         if (!entry->delta)
 469                 usable_delta = 0;       /* no delta */
 470         else if (!pack_size_limit)
 471                usable_delta = 1;        /* unlimited packfile */
 472         else if (entry->delta->idx.offset == (off_t)-1)
 473                 usable_delta = 0;       /* base was written to another pack */
 474         else if (entry->delta->idx.offset)
 475                 usable_delta = 1;       /* base already exists in this pack */
 476         else
 477                 usable_delta = 0;       /* base could end up in another pack */
 478
 479         if (!reuse_object)
 480                 to_reuse = 0;   /* explicit */
 481         else if (!entry->in_pack)
 482                 to_reuse = 0;   /* can't reuse what we don't have */
 483         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
 484                                 /* check_object() decided it for us ... */
 485                 to_reuse = usable_delta;
 486                                 /* ... but pack split may override that */
 487         else if (entry->type != entry->in_pack_type)
 488                 to_reuse = 0;   /* pack has delta which is unusable */
 489         else if (entry->delta)
 490                 to_reuse = 0;   /* we want to pack afresh */
 491         else
 492                 to_reuse = 1;   /* we have it in-pack undeltified,
 493                                  * and we do not need to deltify it.
 494                                  */
 495
 496         if (!to_reuse)
 497                 len = write_no_reuse_object(f, entry, limit, usable_delta);
 498         else
 499                 len = write_reuse_object(f, entry, limit, usable_delta);
 500         if (!len)
 501                 return 0;
 502
 503         if (usable_delta)
 504                 written_delta++;
 505         written++;
 506         if (!pack_to_stdout)
 507                 entry->idx.crc32 = crc32_end(f);
 508         return len;
 509 }
 510
 511 enum write_one_status {
 512         WRITE_ONE_SKIP = -1, /* already written */
 513         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 514         WRITE_ONE_WRITTEN = 1, /* normal */
 515         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 516 };
 517
 518 static enum write_one_status write_one(struct hashfile *f,
 519                                        struct object_entry *e,
 520                                        off_t *offset)
 521 {
 522         off_t size;
 523         int recursing;
 524
 525         /*
 526          * we set offset to 1 (which is an impossible value) to mark
 527          * the fact that this object is involved in "write its base
 528          * first before writing a deltified object" recursion.
 529          */
 530         recursing = (e->idx.offset == 1);
 531         if (recursing) {
 532                 warning("recursive delta detected for object %s",
 533                         oid_to_hex(&e->idx.oid));
 534                 return WRITE_ONE_RECURSIVE;
 535         } else if (e->idx.offset || e->preferred_base) {
 536                 /* offset is non zero if object is written already. */
 537                 return WRITE_ONE_SKIP;
 538         }
 539
 540         /* if we are deltified, write out base object first. */
 541         if (e->delta) {
 542                 e->idx.offset = 1; /* now recurse */
 543                 switch (write_one(f, e->delta, offset)) {
 544                 case WRITE_ONE_RECURSIVE:
 545                         /* we cannot depend on this one */
 546                         e->delta = NULL;
 547                         break;
 548                 default:
 549                         break;
 550                 case WRITE_ONE_BREAK:
 551                         e->idx.offset = recursing;
 552                         return WRITE_ONE_BREAK;
 553                 }
 554         }
 555
 556         e->idx.offset = *offset;
 557         size = write_object(f, e, *offset);
 558         if (!size) {
 559                 e->idx.offset = recursing;
 560                 return WRITE_ONE_BREAK;
 561         }
 562         written_list[nr_written++] = &e->idx;
 563
 564         /* make sure off_t is sufficiently large not to wrap */
 565         if (signed_add_overflows(*offset, size))
 566                 die("pack too large for current definition of off_t");
 567         *offset += size;
 568         return WRITE_ONE_WRITTEN;
 569 }
 570
 571 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
 572                        void *cb_data)
 573 {
 574         struct object_id peeled;
 575         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
 576
 577         if (entry)
 578                 entry->tagged = 1;
 579         if (!peel_ref(path, &peeled)) {
 580                 entry = packlist_find(&to_pack, peeled.hash, NULL);
 581                 if (entry)
 582                         entry->tagged = 1;
 583         }
 584         return 0;
 585 }
 586
 587 static inline void add_to_write_order(struct object_entry **wo,
 588                                unsigned int *endp,
 589                                struct object_entry *e)
 590 {
 591         if (e->filled)
 592                 return;
 593         wo[(*endp)++] = e;
 594         e->filled = 1;
 595 }
 596
 597 static void add_descendants_to_write_order(struct object_entry **wo,
 598                                            unsigned int *endp,
 599                                            struct object_entry *e)
 600 {
 601         int add_to_order = 1;
 602         while (e) {
 603                 if (add_to_order) {
 604                         struct object_entry *s;
 605                         /* add this node... */
 606                         add_to_write_order(wo, endp, e);
 607                         /* all its siblings... */
 608                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
 609                                 add_to_write_order(wo, endp, s);
 610                         }
 611                 }
 612                 /* drop down a level to add left subtree nodes if possible */
 613                 if (e->delta_child) {
 614                         add_to_order = 1;
 615                         e = e->delta_child;
 616                 } else {
 617                         add_to_order = 0;
 618                         /* our sibling might have some children, it is next */
 619                         if (e->delta_sibling) {
 620                                 e = e->delta_sibling;
 621                                 continue;
 622                         }
 623                         /* go back to our parent node */
 624                         e = e->delta;
 625                         while (e && !e->delta_sibling) {
 626                                 /* we're on the right side of a subtree, keep
 627                                  * going up until we can go right again */
 628                                 e = e->delta;
 629                         }
 630                         if (!e) {
 631                                 /* done- we hit our original root node */
 632                                 return;
 633                         }
 634                         /* pass it off to sibling at this level */
 635                         e = e->delta_sibling;
 636                 }
 637         };
 638 }
 639
 640 static void add_family_to_write_order(struct object_entry **wo,
 641                                       unsigned int *endp,
 642                                       struct object_entry *e)
 643 {
 644         struct object_entry *root;
 645
 646         for (root = e; root->delta; root = root->delta)
 647                 ; /* nothing */
 648         add_descendants_to_write_order(wo, endp, root);
 649 }
 650
 651 static struct object_entry **compute_write_order(void)
 652 {
 653         unsigned int i, wo_end, last_untagged;
 654
 655         struct object_entry **wo;
 656         struct object_entry *objects = to_pack.objects;
 657
 658         for (i = 0; i < to_pack.nr_objects; i++) {
 659                 objects[i].tagged = 0;
 660                 objects[i].filled = 0;
 661                 objects[i].delta_child = NULL;
 662                 objects[i].delta_sibling = NULL;
 663         }
 664
 665         /*
 666          * Fully connect delta_child/delta_sibling network.
 667          * Make sure delta_sibling is sorted in the original
 668          * recency order.
 669          */
 670         for (i = to_pack.nr_objects; i > 0;) {
 671                 struct object_entry *e = &objects[--i];
 672                 if (!e->delta)
 673                         continue;
 674                 /* Mark me as the first child */
 675                 e->delta_sibling = e->delta->delta_child;
 676                 e->delta->delta_child = e;
 677         }
 678
 679         /*
 680          * Mark objects that are at the tip of tags.
 681          */
 682         for_each_tag_ref(mark_tagged, NULL);
 683
 684         /*
 685          * Give the objects in the original recency order until
 686          * we see a tagged tip.
 687          */
 688         ALLOC_ARRAY(wo, to_pack.nr_objects);
 689         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
 690                 if (objects[i].tagged)
 691                         break;
 692                 add_to_write_order(wo, &wo_end, &objects[i]);
 693         }
 694         last_untagged = i;
 695
 696         /*
 697          * Then fill all the tagged tips.
 698          */
 699         for (; i < to_pack.nr_objects; i++) {
 700                 if (objects[i].tagged)
 701                         add_to_write_order(wo, &wo_end, &objects[i]);
 702         }
 703
 704         /*
 705          * And then all remaining commits and tags.
 706          */
 707         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 708                 if (objects[i].type != OBJ_COMMIT &&
 709                     objects[i].type != OBJ_TAG)
 710                         continue;
 711                 add_to_write_order(wo, &wo_end, &objects[i]);
 712         }
 713
 714         /*
 715          * And then all the trees.
 716          */
 717         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 718                 if (objects[i].type != OBJ_TREE)
 719                         continue;
 720                 add_to_write_order(wo, &wo_end, &objects[i]);
 721         }
 722
 723         /*
 724          * Finally all the rest in really tight order
 725          */
 726         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 727                 if (!objects[i].filled)
 728                         add_family_to_write_order(wo, &wo_end, &objects[i]);
 729         }
 730
 731         if (wo_end != to_pack.nr_objects)
 732                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
 733
 734         return wo;
 735 }
 736
 737 static off_t write_reused_pack(struct hashfile *f)
 738 {
 739         unsigned char buffer[8192];
 740         off_t to_write, total;
 741         int fd;
 742
 743         if (!is_pack_valid(reuse_packfile))
 744                 die("packfile is invalid: %s", reuse_packfile->pack_name);
 745
 746         fd = git_open(reuse_packfile->pack_name);
 747         if (fd < 0)
 748                 die_errno("unable to open packfile for reuse: %s",
 749                           reuse_packfile->pack_name);
 750
 751         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
 752                 die_errno("unable to seek in reused packfile");
 753
 754         if (reuse_packfile_offset < 0)
 755                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
 756
 757         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
 758
 759         while (to_write) {
 760                 int read_pack = xread(fd, buffer, sizeof(buffer));
 761
 762                 if (read_pack <= 0)
 763                         die_errno("unable to read from reused packfile");
 764
 765                 if (read_pack > to_write)
 766                         read_pack = to_write;
 767
 768                 hashwrite(f, buffer, read_pack);
 769                 to_write -= read_pack;
 770
 771                 /*
 772                  * We don't know the actual number of objects written,
 773                  * only how many bytes written, how many bytes total, and
 774                  * how many objects total. So we can fake it by pretending all
 775                  * objects we are writing are the same size. This gives us a
 776                  * smooth progress meter, and at the end it matches the true
 777                  * answer.
 778                  */
 779                 written = reuse_packfile_objects *
 780                                 (((double)(total - to_write)) / total);
 781                 display_progress(progress_state, written);
 782         }
 783
 784         close(fd);
 785         written = reuse_packfile_objects;
 786         display_progress(progress_state, written);
 787         return reuse_packfile_offset - sizeof(struct pack_header);
 788 }
 789
 790 static const char no_split_warning[] = N_(
 791 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
 792 );
 793
 794 static void write_pack_file(void)
 795 {
 796         uint32_t i = 0, j;
 797         struct hashfile *f;
 798         off_t offset;
 799         uint32_t nr_remaining = nr_result;
 800         time_t last_mtime = 0;
 801         struct object_entry **write_order;
 802
 803         if (progress > pack_to_stdout)
 804                 progress_state = start_progress(_("Writing objects"), nr_result);
 805         ALLOC_ARRAY(written_list, to_pack.nr_objects);
 806         write_order = compute_write_order();
 807
 808         do {
 809                 struct object_id oid;
 810                 char *pack_tmp_name = NULL;
 811
 812                 if (pack_to_stdout)
 813                         f = hashfd_throughput(1, "<stdout>", progress_state);
 814                 else
 815                         f = create_tmp_packfile(&pack_tmp_name);
 816
 817                 offset = write_pack_header(f, nr_remaining);
 818
 819                 if (reuse_packfile) {
 820                         off_t packfile_size;
 821                         assert(pack_to_stdout);
 822
 823                         packfile_size = write_reused_pack(f);
 824                         offset += packfile_size;
 825                 }
 826
 827                 nr_written = 0;
 828                 for (; i < to_pack.nr_objects; i++) {
 829                         struct object_entry *e = write_order[i];
 830                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 831                                 break;
 832                         display_progress(progress_state, written);
 833                 }
 834
 835                 /*
 836                  * Did we write the wrong # entries in the header?
 837                  * If so, rewrite it like in fast-import
 838                  */
 839                 if (pack_to_stdout) {
 840                         hashclose(f, oid.hash, CSUM_CLOSE);
 841                 } else if (nr_written == nr_remaining) {
 842                         hashclose(f, oid.hash, CSUM_FSYNC);
 843                 } else {
 844                         int fd = hashclose(f, oid.hash, 0);
 845                         fixup_pack_header_footer(fd, oid.hash, pack_tmp_name,
 846                                                  nr_written, oid.hash, offset);
 847                         close(fd);
 848                         if (write_bitmap_index) {
 849                                 warning(_(no_split_warning));
 850                                 write_bitmap_index = 0;
 851                         }
 852                 }
 853
 854                 if (!pack_to_stdout) {
 855                         struct stat st;
 856                         struct strbuf tmpname = STRBUF_INIT;
 857
 858                         /*
 859                          * Packs are runtime accessed in their mtime
 860                          * order since newer packs are more likely to contain
 861                          * younger objects.  So if we are creating multiple
 862                          * packs then we should modify the mtime of later ones
 863                          * to preserve this property.
 864                          */
 865                         if (stat(pack_tmp_name, &st) < 0) {
 866                                 warning_errno("failed to stat %s", pack_tmp_name);
 867                         } else if (!last_mtime) {
 868                                 last_mtime = st.st_mtime;
 869                         } else {
 870                                 struct utimbuf utb;
 871                                 utb.actime = st.st_atime;
 872                                 utb.modtime = --last_mtime;
 873                                 if (utime(pack_tmp_name, &utb) < 0)
 874                                         warning_errno("failed utime() on %s", pack_tmp_name);
 875                         }
 876
 877                         strbuf_addf(&tmpname, "%s-", base_name);
 878
 879                         if (write_bitmap_index) {
 880                                 bitmap_writer_set_checksum(oid.hash);
 881                                 bitmap_writer_build_type_index(written_list, nr_written);
 882                         }
 883
 884                         finish_tmp_packfile(&tmpname, pack_tmp_name,
 885                                             written_list, nr_written,
 886                                             &pack_idx_opts, oid.hash);
 887
 888                         if (write_bitmap_index) {
 889                                 strbuf_addf(&tmpname, "%s.bitmap", oid_to_hex(&oid));
 890
 891                                 stop_progress(&progress_state);
 892
 893                                 bitmap_writer_show_progress(progress);
 894                                 bitmap_writer_reuse_bitmaps(&to_pack);
 895                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
 896                                 bitmap_writer_build(&to_pack);
 897                                 bitmap_writer_finish(written_list, nr_written,
 898                                                      tmpname.buf, write_bitmap_options);
 899                                 write_bitmap_index = 0;
 900                         }
 901
 902                         strbuf_release(&tmpname);
 903                         free(pack_tmp_name);
 904                         puts(oid_to_hex(&oid));
 905                 }
 906
 907                 /* mark written objects as written to previous pack */
 908                 for (j = 0; j < nr_written; j++) {
 909                         written_list[j]->offset = (off_t)-1;
 910                 }
 911                 nr_remaining -= nr_written;
 912         } while (nr_remaining && i < to_pack.nr_objects);
 913
 914         free(written_list);
 915         free(write_order);
 916         stop_progress(&progress_state);
 917         if (written != nr_result)
 918                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
 919                         written, nr_result);
 920 }
 921
 922 static int no_try_delta(const char *path)
 923 {
 924         static struct attr_check *check;
 925
 926         if (!check)
 927                 check = attr_check_initl("delta", NULL);
 928         if (git_check_attr(path, check))
 929                 return 0;
 930         if (ATTR_FALSE(check->items[0].value))
 931                 return 1;
 932         return 0;
 933 }
 934
 935 /*
 936  * When adding an object, check whether we have already added it
 937  * to our packing list. If so, we can skip. However, if we are
 938  * being asked to excludei t, but the previous mention was to include
 939  * it, make sure to adjust its flags and tweak our numbers accordingly.
 940  *
 941  * As an optimization, we pass out the index position where we would have
 942  * found the item, since that saves us from having to look it up again a
 943  * few lines later when we want to add the new entry.
 944  */
 945 static int have_duplicate_entry(const struct object_id *oid,
 946                                 int exclude,
 947                                 uint32_t *index_pos)
 948 {
 949         struct object_entry *entry;
 950
 951         entry = packlist_find(&to_pack, oid->hash, index_pos);
 952         if (!entry)
 953                 return 0;
 954
 955         if (exclude) {
 956                 if (!entry->preferred_base)
 957                         nr_result--;
 958                 entry->preferred_base = 1;
 959         }
 960
 961         return 1;
 962 }
 963
 964 static int want_found_object(int exclude, struct packed_git *p)
 965 {
 966         if (exclude)
 967                 return 1;
 968         if (incremental)
 969                 return 0;
 970
 971         /*
 972          * When asked to do --local (do not include an object that appears in a
 973          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
 974          * an object that appears in a pack marked with .keep), finding a pack
 975          * that matches the criteria is sufficient for us to decide to omit it.
 976          * However, even if this pack does not satisfy the criteria, we need to
 977          * make sure no copy of this object appears in _any_ pack that makes us
 978          * to omit the object, so we need to check all the packs.
 979          *
 980          * We can however first check whether these options can possible matter;
 981          * if they do not matter we know we want the object in generated pack.
 982          * Otherwise, we signal "-1" at the end to tell the caller that we do
 983          * not know either way, and it needs to check more packs.
 984          */
 985         if (!ignore_packed_keep &&
 986             (!local || !have_non_local_packs))
 987                 return 1;
 988
 989         if (local && !p->pack_local)
 990                 return 0;
 991         if (ignore_packed_keep && p->pack_local && p->pack_keep)
 992                 return 0;
 993
 994         /* we don't know yet; keep looking for more packs */
 995         return -1;
 996 }
 997
 998 /*
 999  * Check whether we want the object in the pack (e.g., we do not want
1000  * objects found in non-local stores if the "--local" option was used).
1001  *
1002  * If the caller already knows an existing pack it wants to take the object
1003  * from, that is passed in *found_pack and *found_offset; otherwise this
1004  * function finds if there is any pack that has the object and returns the pack
1005  * and its offset in these variables.
1006  */
1007 static int want_object_in_pack(const struct object_id *oid,
1008                                int exclude,
1009                                struct packed_git **found_pack,
1010                                off_t *found_offset)
1011 {
1012         int want;
1013         struct list_head *pos;
1014
1015         if (!exclude && local && has_loose_object_nonlocal(oid->hash))
1016                 return 0;
1017
1018         /*
1019          * If we already know the pack object lives in, start checks from that
1020          * pack - in the usual case when neither --local was given nor .keep files
1021          * are present we will determine the answer right now.
1022          */
1023         if (*found_pack) {
1024                 want = want_found_object(exclude, *found_pack);
1025                 if (want != -1)
1026                         return want;
1027         }
1028
1029         list_for_each(pos, &packed_git_mru) {
1030                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
1031                 off_t offset;
1032
1033                 if (p == *found_pack)
1034                         offset = *found_offset;
1035                 else
1036                         offset = find_pack_entry_one(oid->hash, p);
1037
1038                 if (offset) {
1039                         if (!*found_pack) {
1040                                 if (!is_pack_valid(p))
1041                                         continue;
1042                                 *found_offset = offset;
1043                                 *found_pack = p;
1044                         }
1045                         want = want_found_object(exclude, p);
1046                         if (!exclude && want > 0)
1047                                 list_move(&p->mru, &packed_git_mru);
1048                         if (want != -1)
1049                                 return want;
1050                 }
1051         }
1052
1053         return 1;
1054 }
1055
1056 static void create_object_entry(const struct object_id *oid,
1057                                 enum object_type type,
1058                                 uint32_t hash,
1059                                 int exclude,
1060                                 int no_try_delta,
1061                                 uint32_t index_pos,
1062                                 struct packed_git *found_pack,
1063                                 off_t found_offset)
1064 {
1065         struct object_entry *entry;
1066
1067         entry = packlist_alloc(&to_pack, oid->hash, index_pos);
1068         entry->hash = hash;
1069         if (type)
1070                 entry->type = type;
1071         if (exclude)
1072                 entry->preferred_base = 1;
1073         else
1074                 nr_result++;
1075         if (found_pack) {
1076                 entry->in_pack = found_pack;
1077                 entry->in_pack_offset = found_offset;
1078         }
1079
1080         entry->no_try_delta = no_try_delta;
1081 }
1082
1083 static const char no_closure_warning[] = N_(
1084 "disabling bitmap writing, as some objects are not being packed"
1085 );
1086
1087 static int add_object_entry(const struct object_id *oid, enum object_type type,
1088                             const char *name, int exclude)
1089 {
1090         struct packed_git *found_pack = NULL;
1091         off_t found_offset = 0;
1092         uint32_t index_pos;
1093
1094         if (have_duplicate_entry(oid, exclude, &index_pos))
1095                 return 0;
1096
1097         if (!want_object_in_pack(oid, exclude, &found_pack, &found_offset)) {
1098                 /* The pack is missing an object, so it will not have closure */
1099                 if (write_bitmap_index) {
1100                         warning(_(no_closure_warning));
1101                         write_bitmap_index = 0;
1102                 }
1103                 return 0;
1104         }
1105
1106         create_object_entry(oid, type, pack_name_hash(name),
1107                             exclude, name && no_try_delta(name),
1108                             index_pos, found_pack, found_offset);
1109
1110         display_progress(progress_state, nr_result);
1111         return 1;
1112 }
1113
1114 static int add_object_entry_from_bitmap(const struct object_id *oid,
1115                                         enum object_type type,
1116                                         int flags, uint32_t name_hash,
1117                                         struct packed_git *pack, off_t offset)
1118 {
1119         uint32_t index_pos;
1120
1121         if (have_duplicate_entry(oid, 0, &index_pos))
1122                 return 0;
1123
1124         if (!want_object_in_pack(oid, 0, &pack, &offset))
1125                 return 0;
1126
1127         create_object_entry(oid, type, name_hash, 0, 0, index_pos, pack, offset);
1128
1129         display_progress(progress_state, nr_result);
1130         return 1;
1131 }
1132
1133 struct pbase_tree_cache {
1134         struct object_id oid;
1135         int ref;
1136         int temporary;
1137         void *tree_data;
1138         unsigned long tree_size;
1139 };
1140
1141 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1142 static int pbase_tree_cache_ix(const struct object_id *oid)
1143 {
1144         return oid->hash[0] % ARRAY_SIZE(pbase_tree_cache);
1145 }
1146 static int pbase_tree_cache_ix_incr(int ix)
1147 {
1148         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1149 }
1150
1151 static struct pbase_tree {
1152         struct pbase_tree *next;
1153         /* This is a phony "cache" entry; we are not
1154          * going to evict it or find it through _get()
1155          * mechanism -- this is for the toplevel node that
1156          * would almost always change with any commit.
1157          */
1158         struct pbase_tree_cache pcache;
1159 } *pbase_tree;
1160
1161 static struct pbase_tree_cache *pbase_tree_get(const struct object_id *oid)
1162 {
1163         struct pbase_tree_cache *ent, *nent;
1164         void *data;
1165         unsigned long size;
1166         enum object_type type;
1167         int neigh;
1168         int my_ix = pbase_tree_cache_ix(oid);
1169         int available_ix = -1;
1170
1171         /* pbase-tree-cache acts as a limited hashtable.
1172          * your object will be found at your index or within a few
1173          * slots after that slot if it is cached.
1174          */
1175         for (neigh = 0; neigh < 8; neigh++) {
1176                 ent = pbase_tree_cache[my_ix];
1177                 if (ent && !oidcmp(&ent->oid, oid)) {
1178                         ent->ref++;
1179                         return ent;
1180                 }
1181                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1182                          ((0 <= available_ix) &&
1183                           (!ent && pbase_tree_cache[available_ix])))
1184                         available_ix = my_ix;
1185                 if (!ent)
1186                         break;
1187                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1188         }
1189
1190         /* Did not find one.  Either we got a bogus request or
1191          * we need to read and perhaps cache.
1192          */
1193         data = read_sha1_file(oid->hash, &type, &size);
1194         if (!data)
1195                 return NULL;
1196         if (type != OBJ_TREE) {
1197                 free(data);
1198                 return NULL;
1199         }
1200
1201         /* We need to either cache or return a throwaway copy */
1202
1203         if (available_ix < 0)
1204                 ent = NULL;
1205         else {
1206                 ent = pbase_tree_cache[available_ix];
1207                 my_ix = available_ix;
1208         }
1209
1210         if (!ent) {
1211                 nent = xmalloc(sizeof(*nent));
1212                 nent->temporary = (available_ix < 0);
1213         }
1214         else {
1215                 /* evict and reuse */
1216                 free(ent->tree_data);
1217                 nent = ent;
1218         }
1219         oidcpy(&nent->oid, oid);
1220         nent->tree_data = data;
1221         nent->tree_size = size;
1222         nent->ref = 1;
1223         if (!nent->temporary)
1224                 pbase_tree_cache[my_ix] = nent;
1225         return nent;
1226 }
1227
1228 static void pbase_tree_put(struct pbase_tree_cache *cache)
1229 {
1230         if (!cache->temporary) {
1231                 cache->ref--;
1232                 return;
1233         }
1234         free(cache->tree_data);
1235         free(cache);
1236 }
1237
1238 static int name_cmp_len(const char *name)
1239 {
1240         int i;
1241         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1242                 ;
1243         return i;
1244 }
1245
1246 static void add_pbase_object(struct tree_desc *tree,
1247                              const char *name,
1248                              int cmplen,
1249                              const char *fullname)
1250 {
1251         struct name_entry entry;
1252         int cmp;
1253
1254         while (tree_entry(tree,&entry)) {
1255                 if (S_ISGITLINK(entry.mode))
1256                         continue;
1257                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1258                       memcmp(name, entry.path, cmplen);
1259                 if (cmp > 0)
1260                         continue;
1261                 if (cmp < 0)
1262                         return;
1263                 if (name[cmplen] != '/') {
1264                         add_object_entry(entry.oid,
1265                                          object_type(entry.mode),
1266                                          fullname, 1);
1267                         return;
1268                 }
1269                 if (S_ISDIR(entry.mode)) {
1270                         struct tree_desc sub;
1271                         struct pbase_tree_cache *tree;
1272                         const char *down = name+cmplen+1;
1273                         int downlen = name_cmp_len(down);
1274
1275                         tree = pbase_tree_get(entry.oid);
1276                         if (!tree)
1277                                 return;
1278                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1279
1280                         add_pbase_object(&sub, down, downlen, fullname);
1281                         pbase_tree_put(tree);
1282                 }
1283         }
1284 }
1285
1286 static unsigned *done_pbase_paths;
1287 static int done_pbase_paths_num;
1288 static int done_pbase_paths_alloc;
1289 static int done_pbase_path_pos(unsigned hash)
1290 {
1291         int lo = 0;
1292         int hi = done_pbase_paths_num;
1293         while (lo < hi) {
1294                 int mi = lo + (hi - lo) / 2;
1295                 if (done_pbase_paths[mi] == hash)
1296                         return mi;
1297                 if (done_pbase_paths[mi] < hash)
1298                         hi = mi;
1299                 else
1300                         lo = mi + 1;
1301         }
1302         return -lo-1;
1303 }
1304
1305 static int check_pbase_path(unsigned hash)
1306 {
1307         int pos = done_pbase_path_pos(hash);
1308         if (0 <= pos)
1309                 return 1;
1310         pos = -pos - 1;
1311         ALLOC_GROW(done_pbase_paths,
1312                    done_pbase_paths_num + 1,
1313                    done_pbase_paths_alloc);
1314         done_pbase_paths_num++;
1315         if (pos < done_pbase_paths_num)
1316                 MOVE_ARRAY(done_pbase_paths + pos + 1, done_pbase_paths + pos,
1317                            done_pbase_paths_num - pos - 1);
1318         done_pbase_paths[pos] = hash;
1319         return 0;
1320 }
1321
1322 static void add_preferred_base_object(const char *name)
1323 {
1324         struct pbase_tree *it;
1325         int cmplen;
1326         unsigned hash = pack_name_hash(name);
1327
1328         if (!num_preferred_base || check_pbase_path(hash))
1329                 return;
1330
1331         cmplen = name_cmp_len(name);
1332         for (it = pbase_tree; it; it = it->next) {
1333                 if (cmplen == 0) {
1334                         add_object_entry(&it->pcache.oid, OBJ_TREE, NULL, 1);
1335                 }
1336                 else {
1337                         struct tree_desc tree;
1338                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1339                         add_pbase_object(&tree, name, cmplen, name);
1340                 }
1341         }
1342 }
1343
1344 static void add_preferred_base(struct object_id *oid)
1345 {
1346         struct pbase_tree *it;
1347         void *data;
1348         unsigned long size;
1349         struct object_id tree_oid;
1350
1351         if (window <= num_preferred_base++)
1352                 return;
1353
1354         data = read_object_with_reference(oid->hash, tree_type, &size, tree_oid.hash);
1355         if (!data)
1356                 return;
1357
1358         for (it = pbase_tree; it; it = it->next) {
1359                 if (!oidcmp(&it->pcache.oid, &tree_oid)) {
1360                         free(data);
1361                         return;
1362                 }
1363         }
1364
1365         it = xcalloc(1, sizeof(*it));
1366         it->next = pbase_tree;
1367         pbase_tree = it;
1368
1369         oidcpy(&it->pcache.oid, &tree_oid);
1370         it->pcache.tree_data = data;
1371         it->pcache.tree_size = size;
1372 }
1373
1374 static void cleanup_preferred_base(void)
1375 {
1376         struct pbase_tree *it;
1377         unsigned i;
1378
1379         it = pbase_tree;
1380         pbase_tree = NULL;
1381         while (it) {
1382                 struct pbase_tree *this = it;
1383                 it = this->next;
1384                 free(this->pcache.tree_data);
1385                 free(this);
1386         }
1387
1388         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1389                 if (!pbase_tree_cache[i])
1390                         continue;
1391                 free(pbase_tree_cache[i]->tree_data);
1392                 FREE_AND_NULL(pbase_tree_cache[i]);
1393         }
1394
1395         FREE_AND_NULL(done_pbase_paths);
1396         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1397 }
1398
1399 static void check_object(struct object_entry *entry)
1400 {
1401         if (entry->in_pack) {
1402                 struct packed_git *p = entry->in_pack;
1403                 struct pack_window *w_curs = NULL;
1404                 const unsigned char *base_ref = NULL;
1405                 struct object_entry *base_entry;
1406                 unsigned long used, used_0;
1407                 unsigned long avail;
1408                 off_t ofs;
1409                 unsigned char *buf, c;
1410
1411                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1412
1413                 /*
1414                  * We want in_pack_type even if we do not reuse delta
1415                  * since non-delta representations could still be reused.
1416                  */
1417                 used = unpack_object_header_buffer(buf, avail,
1418                                                    &entry->in_pack_type,
1419                                                    &entry->size);
1420                 if (used == 0)
1421                         goto give_up;
1422
1423                 /*
1424                  * Determine if this is a delta and if so whether we can
1425                  * reuse it or not.  Otherwise let's find out as cheaply as
1426                  * possible what the actual type and size for this object is.
1427                  */
1428                 switch (entry->in_pack_type) {
1429                 default:
1430                         /* Not a delta hence we've already got all we need. */
1431                         entry->type = entry->in_pack_type;
1432                         entry->in_pack_header_size = used;
1433                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1434                                 goto give_up;
1435                         unuse_pack(&w_curs);
1436                         return;
1437                 case OBJ_REF_DELTA:
1438                         if (reuse_delta && !entry->preferred_base)
1439                                 base_ref = use_pack(p, &w_curs,
1440                                                 entry->in_pack_offset + used, NULL);
1441                         entry->in_pack_header_size = used + 20;
1442                         break;
1443                 case OBJ_OFS_DELTA:
1444                         buf = use_pack(p, &w_curs,
1445                                        entry->in_pack_offset + used, NULL);
1446                         used_0 = 0;
1447                         c = buf[used_0++];
1448                         ofs = c & 127;
1449                         while (c & 128) {
1450                                 ofs += 1;
1451                                 if (!ofs || MSB(ofs, 7)) {
1452                                         error("delta base offset overflow in pack for %s",
1453                                               oid_to_hex(&entry->idx.oid));
1454                                         goto give_up;
1455                                 }
1456                                 c = buf[used_0++];
1457                                 ofs = (ofs << 7) + (c & 127);
1458                         }
1459                         ofs = entry->in_pack_offset - ofs;
1460                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1461                                 error("delta base offset out of bound for %s",
1462                                       oid_to_hex(&entry->idx.oid));
1463                                 goto give_up;
1464                         }
1465                         if (reuse_delta && !entry->preferred_base) {
1466                                 struct revindex_entry *revidx;
1467                                 revidx = find_pack_revindex(p, ofs);
1468                                 if (!revidx)
1469                                         goto give_up;
1470                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1471                         }
1472                         entry->in_pack_header_size = used + used_0;
1473                         break;
1474                 }
1475
1476                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1477                         /*
1478                          * If base_ref was set above that means we wish to
1479                          * reuse delta data, and we even found that base
1480                          * in the list of objects we want to pack. Goodie!
1481                          *
1482                          * Depth value does not matter - find_deltas() will
1483                          * never consider reused delta as the base object to
1484                          * deltify other objects against, in order to avoid
1485                          * circular deltas.
1486                          */
1487                         entry->type = entry->in_pack_type;
1488                         entry->delta = base_entry;
1489                         entry->delta_size = entry->size;
1490                         entry->delta_sibling = base_entry->delta_child;
1491                         base_entry->delta_child = entry;
1492                         unuse_pack(&w_curs);
1493                         return;
1494                 }
1495
1496                 if (entry->type) {
1497                         /*
1498                          * This must be a delta and we already know what the
1499                          * final object type is.  Let's extract the actual
1500                          * object size from the delta header.
1501                          */
1502                         entry->size = get_size_from_delta(p, &w_curs,
1503                                         entry->in_pack_offset + entry->in_pack_header_size);
1504                         if (entry->size == 0)
1505                                 goto give_up;
1506                         unuse_pack(&w_curs);
1507                         return;
1508                 }
1509
1510                 /*
1511                  * No choice but to fall back to the recursive delta walk
1512                  * with sha1_object_info() to find about the object type
1513                  * at this point...
1514                  */
1515                 give_up:
1516                 unuse_pack(&w_curs);
1517         }
1518
1519         entry->type = sha1_object_info(entry->idx.oid.hash, &entry->size);
1520         /*
1521          * The error condition is checked in prepare_pack().  This is
1522          * to permit a missing preferred base object to be ignored
1523          * as a preferred base.  Doing so can result in a larger
1524          * pack file, but the transfer will still take place.
1525          */
1526 }
1527
1528 static int pack_offset_sort(const void *_a, const void *_b)
1529 {
1530         const struct object_entry *a = *(struct object_entry **)_a;
1531         const struct object_entry *b = *(struct object_entry **)_b;
1532
1533         /* avoid filesystem trashing with loose objects */
1534         if (!a->in_pack && !b->in_pack)
1535                 return oidcmp(&a->idx.oid, &b->idx.oid);
1536
1537         if (a->in_pack < b->in_pack)
1538                 return -1;
1539         if (a->in_pack > b->in_pack)
1540                 return 1;
1541         return a->in_pack_offset < b->in_pack_offset ? -1 :
1542                         (a->in_pack_offset > b->in_pack_offset);
1543 }
1544
1545 /*
1546  * Drop an on-disk delta we were planning to reuse. Naively, this would
1547  * just involve blanking out the "delta" field, but we have to deal
1548  * with some extra book-keeping:
1549  *
1550  *   1. Removing ourselves from the delta_sibling linked list.
1551  *
1552  *   2. Updating our size/type to the non-delta representation. These were
1553  *      either not recorded initially (size) or overwritten with the delta type
1554  *      (type) when check_object() decided to reuse the delta.
1555  *
1556  *   3. Resetting our delta depth, as we are now a base object.
1557  */
1558 static void drop_reused_delta(struct object_entry *entry)
1559 {
1560         struct object_entry **p = &entry->delta->delta_child;
1561         struct object_info oi = OBJECT_INFO_INIT;
1562
1563         while (*p) {
1564                 if (*p == entry)
1565                         *p = (*p)->delta_sibling;
1566                 else
1567                         p = &(*p)->delta_sibling;
1568         }
1569         entry->delta = NULL;
1570         entry->depth = 0;
1571
1572         oi.sizep = &entry->size;
1573         oi.typep = &entry->type;
1574         if (packed_object_info(entry->in_pack, entry->in_pack_offset, &oi) < 0) {
1575                 /*
1576                  * We failed to get the info from this pack for some reason;
1577                  * fall back to sha1_object_info, which may find another copy.
1578                  * And if that fails, the error will be recorded in entry->type
1579                  * and dealt with in prepare_pack().
1580                  */
1581                 entry->type = sha1_object_info(entry->idx.oid.hash,
1582                                                &entry->size);
1583         }
1584 }
1585
1586 /*
1587  * Follow the chain of deltas from this entry onward, throwing away any links
1588  * that cause us to hit a cycle (as determined by the DFS state flags in
1589  * the entries).
1590  *
1591  * We also detect too-long reused chains that would violate our --depth
1592  * limit.
1593  */
1594 static void break_delta_chains(struct object_entry *entry)
1595 {
1596         /*
1597          * The actual depth of each object we will write is stored as an int,
1598          * as it cannot exceed our int "depth" limit. But before we break
1599          * changes based no that limit, we may potentially go as deep as the
1600          * number of objects, which is elsewhere bounded to a uint32_t.
1601          */
1602         uint32_t total_depth;
1603         struct object_entry *cur, *next;
1604
1605         for (cur = entry, total_depth = 0;
1606              cur;
1607              cur = cur->delta, total_depth++) {
1608                 if (cur->dfs_state == DFS_DONE) {
1609                         /*
1610                          * We've already seen this object and know it isn't
1611                          * part of a cycle. We do need to append its depth
1612                          * to our count.
1613                          */
1614                         total_depth += cur->depth;
1615                         break;
1616                 }
1617
1618                 /*
1619                  * We break cycles before looping, so an ACTIVE state (or any
1620                  * other cruft which made its way into the state variable)
1621                  * is a bug.
1622                  */
1623                 if (cur->dfs_state != DFS_NONE)
1624                         die("BUG: confusing delta dfs state in first pass: %d",
1625                             cur->dfs_state);
1626
1627                 /*
1628                  * Now we know this is the first time we've seen the object. If
1629                  * it's not a delta, we're done traversing, but we'll mark it
1630                  * done to save time on future traversals.
1631                  */
1632                 if (!cur->delta) {
1633                         cur->dfs_state = DFS_DONE;
1634                         break;
1635                 }
1636
1637                 /*
1638                  * Mark ourselves as active and see if the next step causes
1639                  * us to cycle to another active object. It's important to do
1640                  * this _before_ we loop, because it impacts where we make the
1641                  * cut, and thus how our total_depth counter works.
1642                  * E.g., We may see a partial loop like:
1643                  *
1644                  *   A -> B -> C -> D -> B
1645                  *
1646                  * Cutting B->C breaks the cycle. But now the depth of A is
1647                  * only 1, and our total_depth counter is at 3. The size of the
1648                  * error is always one less than the size of the cycle we
1649                  * broke. Commits C and D were "lost" from A's chain.
1650                  *
1651                  * If we instead cut D->B, then the depth of A is correct at 3.
1652                  * We keep all commits in the chain that we examined.
1653                  */
1654                 cur->dfs_state = DFS_ACTIVE;
1655                 if (cur->delta->dfs_state == DFS_ACTIVE) {
1656                         drop_reused_delta(cur);
1657                         cur->dfs_state = DFS_DONE;
1658                         break;
1659                 }
1660         }
1661
1662         /*
1663          * And now that we've gone all the way to the bottom of the chain, we
1664          * need to clear the active flags and set the depth fields as
1665          * appropriate. Unlike the loop above, which can quit when it drops a
1666          * delta, we need to keep going to look for more depth cuts. So we need
1667          * an extra "next" pointer to keep going after we reset cur->delta.
1668          */
1669         for (cur = entry; cur; cur = next) {
1670                 next = cur->delta;
1671
1672                 /*
1673                  * We should have a chain of zero or more ACTIVE states down to
1674                  * a final DONE. We can quit after the DONE, because either it
1675                  * has no bases, or we've already handled them in a previous
1676                  * call.
1677                  */
1678                 if (cur->dfs_state == DFS_DONE)
1679                         break;
1680                 else if (cur->dfs_state != DFS_ACTIVE)
1681                         die("BUG: confusing delta dfs state in second pass: %d",
1682                             cur->dfs_state);
1683
1684                 /*
1685                  * If the total_depth is more than depth, then we need to snip
1686                  * the chain into two or more smaller chains that don't exceed
1687                  * the maximum depth. Most of the resulting chains will contain
1688                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1689                  * the last chain (i.e., the one containing entry) will contain
1690                  * whatever entries are left over, namely
1691                  * (total_depth % (depth + 1)) of them.
1692                  *
1693                  * Since we are iterating towards decreasing depth, we need to
1694                  * decrement total_depth as we go, and we need to write to the
1695                  * entry what its final depth will be after all of the
1696                  * snipping. Since we're snipping into chains of length (depth
1697                  * + 1) entries, the final depth of an entry will be its
1698                  * original depth modulo (depth + 1). Any time we encounter an
1699                  * entry whose final depth is supposed to be zero, we snip it
1700                  * from its delta base, thereby making it so.
1701                  */
1702                 cur->depth = (total_depth--) % (depth + 1);
1703                 if (!cur->depth)
1704                         drop_reused_delta(cur);
1705
1706                 cur->dfs_state = DFS_DONE;
1707         }
1708 }
1709
1710 static void get_object_details(void)
1711 {
1712         uint32_t i;
1713         struct object_entry **sorted_by_offset;
1714
1715         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1716         for (i = 0; i < to_pack.nr_objects; i++)
1717                 sorted_by_offset[i] = to_pack.objects + i;
1718         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1719
1720         for (i = 0; i < to_pack.nr_objects; i++) {
1721                 struct object_entry *entry = sorted_by_offset[i];
1722                 check_object(entry);
1723                 if (big_file_threshold < entry->size)
1724                         entry->no_try_delta = 1;
1725         }
1726
1727         /*
1728          * This must happen in a second pass, since we rely on the delta
1729          * information for the whole list being completed.
1730          */
1731         for (i = 0; i < to_pack.nr_objects; i++)
1732                 break_delta_chains(&to_pack.objects[i]);
1733
1734         free(sorted_by_offset);
1735 }
1736
1737 /*
1738  * We search for deltas in a list sorted by type, by filename hash, and then
1739  * by size, so that we see progressively smaller and smaller files.
1740  * That's because we prefer deltas to be from the bigger file
1741  * to the smaller -- deletes are potentially cheaper, but perhaps
1742  * more importantly, the bigger file is likely the more recent
1743  * one.  The deepest deltas are therefore the oldest objects which are
1744  * less susceptible to be accessed often.
1745  */
1746 static int type_size_sort(const void *_a, const void *_b)
1747 {
1748         const struct object_entry *a = *(struct object_entry **)_a;
1749         const struct object_entry *b = *(struct object_entry **)_b;
1750
1751         if (a->type > b->type)
1752                 return -1;
1753         if (a->type < b->type)
1754                 return 1;
1755         if (a->hash > b->hash)
1756                 return -1;
1757         if (a->hash < b->hash)
1758                 return 1;
1759         if (a->preferred_base > b->preferred_base)
1760                 return -1;
1761         if (a->preferred_base < b->preferred_base)
1762                 return 1;
1763         if (a->size > b->size)
1764                 return -1;
1765         if (a->size < b->size)
1766                 return 1;
1767         return a < b ? -1 : (a > b);  /* newest first */
1768 }
1769
1770 struct unpacked {
1771         struct object_entry *entry;
1772         void *data;
1773         struct delta_index *index;
1774         unsigned depth;
1775 };
1776
1777 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1778                            unsigned long delta_size)
1779 {
1780         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1781                 return 0;
1782
1783         if (delta_size < cache_max_small_delta_size)
1784                 return 1;
1785
1786         /* cache delta, if objects are large enough compared to delta size */
1787         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1788                 return 1;
1789
1790         return 0;
1791 }
1792
1793 #ifndef NO_PTHREADS
1794
1795 static pthread_mutex_t read_mutex;
1796 #define read_lock()             pthread_mutex_lock(&read_mutex)
1797 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1798
1799 static pthread_mutex_t cache_mutex;
1800 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1801 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1802
1803 static pthread_mutex_t progress_mutex;
1804 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1805 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1806
1807 #else
1808
1809 #define read_lock()             (void)0
1810 #define read_unlock()           (void)0
1811 #define cache_lock()            (void)0
1812 #define cache_unlock()          (void)0
1813 #define progress_lock()         (void)0
1814 #define progress_unlock()       (void)0
1815
1816 #endif
1817
1818 static int try_delta(struct unpacked *trg, struct unpacked *src,
1819                      unsigned max_depth, unsigned long *mem_usage)
1820 {
1821         struct object_entry *trg_entry = trg->entry;
1822         struct object_entry *src_entry = src->entry;
1823         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1824         unsigned ref_depth;
1825         enum object_type type;
1826         void *delta_buf;
1827
1828         /* Don't bother doing diffs between different types */
1829         if (trg_entry->type != src_entry->type)
1830                 return -1;
1831
1832         /*
1833          * We do not bother to try a delta that we discarded on an
1834          * earlier try, but only when reusing delta data.  Note that
1835          * src_entry that is marked as the preferred_base should always
1836          * be considered, as even if we produce a suboptimal delta against
1837          * it, we will still save the transfer cost, as we already know
1838          * the other side has it and we won't send src_entry at all.
1839          */
1840         if (reuse_delta && trg_entry->in_pack &&
1841             trg_entry->in_pack == src_entry->in_pack &&
1842             !src_entry->preferred_base &&
1843             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1844             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1845                 return 0;
1846
1847         /* Let's not bust the allowed depth. */
1848         if (src->depth >= max_depth)
1849                 return 0;
1850
1851         /* Now some size filtering heuristics. */
1852         trg_size = trg_entry->size;
1853         if (!trg_entry->delta) {
1854                 max_size = trg_size/2 - 20;
1855                 ref_depth = 1;
1856         } else {
1857                 max_size = trg_entry->delta_size;
1858                 ref_depth = trg->depth;
1859         }
1860         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1861                                                 (max_depth - ref_depth + 1);
1862         if (max_size == 0)
1863                 return 0;
1864         src_size = src_entry->size;
1865         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1866         if (sizediff >= max_size)
1867                 return 0;
1868         if (trg_size < src_size / 32)
1869                 return 0;
1870
1871         /* Load data if not already done */
1872         if (!trg->data) {
1873                 read_lock();
1874                 trg->data = read_sha1_file(trg_entry->idx.oid.hash, &type,
1875                                            &sz);
1876                 read_unlock();
1877                 if (!trg->data)
1878                         die("object %s cannot be read",
1879                             oid_to_hex(&trg_entry->idx.oid));
1880                 if (sz != trg_size)
1881                         die("object %s inconsistent object length (%lu vs %lu)",
1882                             oid_to_hex(&trg_entry->idx.oid), sz,
1883                             trg_size);
1884                 *mem_usage += sz;
1885         }
1886         if (!src->data) {
1887                 read_lock();
1888                 src->data = read_sha1_file(src_entry->idx.oid.hash, &type,
1889                                            &sz);
1890                 read_unlock();
1891                 if (!src->data) {
1892                         if (src_entry->preferred_base) {
1893                                 static int warned = 0;
1894                                 if (!warned++)
1895                                         warning("object %s cannot be read",
1896                                                 oid_to_hex(&src_entry->idx.oid));
1897                                 /*
1898                                  * Those objects are not included in the
1899                                  * resulting pack.  Be resilient and ignore
1900                                  * them if they can't be read, in case the
1901                                  * pack could be created nevertheless.
1902                                  */
1903                                 return 0;
1904                         }
1905                         die("object %s cannot be read",
1906                             oid_to_hex(&src_entry->idx.oid));
1907                 }
1908                 if (sz != src_size)
1909                         die("object %s inconsistent object length (%lu vs %lu)",
1910                             oid_to_hex(&src_entry->idx.oid), sz,
1911                             src_size);
1912                 *mem_usage += sz;
1913         }
1914         if (!src->index) {
1915                 src->index = create_delta_index(src->data, src_size);
1916                 if (!src->index) {
1917                         static int warned = 0;
1918                         if (!warned++)
1919                                 warning("suboptimal pack - out of memory");
1920                         return 0;
1921                 }
1922                 *mem_usage += sizeof_delta_index(src->index);
1923         }
1924
1925         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1926         if (!delta_buf)
1927                 return 0;
1928
1929         if (trg_entry->delta) {
1930                 /* Prefer only shallower same-sized deltas. */
1931                 if (delta_size == trg_entry->delta_size &&
1932                     src->depth + 1 >= trg->depth) {
1933                         free(delta_buf);
1934                         return 0;
1935                 }
1936         }
1937
1938         /*
1939          * Handle memory allocation outside of the cache
1940          * accounting lock.  Compiler will optimize the strangeness
1941          * away when NO_PTHREADS is defined.
1942          */
1943         free(trg_entry->delta_data);
1944         cache_lock();
1945         if (trg_entry->delta_data) {
1946                 delta_cache_size -= trg_entry->delta_size;
1947                 trg_entry->delta_data = NULL;
1948         }
1949         if (delta_cacheable(src_size, trg_size, delta_size)) {
1950                 delta_cache_size += delta_size;
1951                 cache_unlock();
1952                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1953         } else {
1954                 cache_unlock();
1955                 free(delta_buf);
1956         }
1957
1958         trg_entry->delta = src_entry;
1959         trg_entry->delta_size = delta_size;
1960         trg->depth = src->depth + 1;
1961
1962         return 1;
1963 }
1964
1965 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1966 {
1967         struct object_entry *child = me->delta_child;
1968         unsigned int m = n;
1969         while (child) {
1970                 unsigned int c = check_delta_limit(child, n + 1);
1971                 if (m < c)
1972                         m = c;
1973                 child = child->delta_sibling;
1974         }
1975         return m;
1976 }
1977
1978 static unsigned long free_unpacked(struct unpacked *n)
1979 {
1980         unsigned long freed_mem = sizeof_delta_index(n->index);
1981         free_delta_index(n->index);
1982         n->index = NULL;
1983         if (n->data) {
1984                 freed_mem += n->entry->size;
1985                 FREE_AND_NULL(n->data);
1986         }
1987         n->entry = NULL;
1988         n->depth = 0;
1989         return freed_mem;
1990 }
1991
1992 static void find_deltas(struct object_entry **list, unsigned *list_size,
1993                         int window, int depth, unsigned *processed)
1994 {
1995         uint32_t i, idx = 0, count = 0;
1996         struct unpacked *array;
1997         unsigned long mem_usage = 0;
1998
1999         array = xcalloc(window, sizeof(struct unpacked));
2000
2001         for (;;) {
2002                 struct object_entry *entry;
2003                 struct unpacked *n = array + idx;
2004                 int j, max_depth, best_base = -1;
2005
2006                 progress_lock();
2007                 if (!*list_size) {
2008                         progress_unlock();
2009                         break;
2010                 }
2011                 entry = *list++;
2012                 (*list_size)--;
2013                 if (!entry->preferred_base) {
2014                         (*processed)++;
2015                         display_progress(progress_state, *processed);
2016                 }
2017                 progress_unlock();
2018
2019                 mem_usage -= free_unpacked(n);
2020                 n->entry = entry;
2021
2022                 while (window_memory_limit &&
2023                        mem_usage > window_memory_limit &&
2024                        count > 1) {
2025                         uint32_t tail = (idx + window - count) % window;
2026                         mem_usage -= free_unpacked(array + tail);
2027                         count--;
2028                 }
2029
2030                 /* We do not compute delta to *create* objects we are not
2031                  * going to pack.
2032                  */
2033                 if (entry->preferred_base)
2034                         goto next;
2035
2036                 /*
2037                  * If the current object is at pack edge, take the depth the
2038                  * objects that depend on the current object into account
2039                  * otherwise they would become too deep.
2040                  */
2041                 max_depth = depth;
2042                 if (entry->delta_child) {
2043                         max_depth -= check_delta_limit(entry, 0);
2044                         if (max_depth <= 0)
2045                                 goto next;
2046                 }
2047
2048                 j = window;
2049                 while (--j > 0) {
2050                         int ret;
2051                         uint32_t other_idx = idx + j;
2052                         struct unpacked *m;
2053                         if (other_idx >= window)
2054                                 other_idx -= window;
2055                         m = array + other_idx;
2056                         if (!m->entry)
2057                                 break;
2058                         ret = try_delta(n, m, max_depth, &mem_usage);
2059                         if (ret < 0)
2060                                 break;
2061                         else if (ret > 0)
2062                                 best_base = other_idx;
2063                 }
2064
2065                 /*
2066                  * If we decided to cache the delta data, then it is best
2067                  * to compress it right away.  First because we have to do
2068                  * it anyway, and doing it here while we're threaded will
2069                  * save a lot of time in the non threaded write phase,
2070                  * as well as allow for caching more deltas within
2071                  * the same cache size limit.
2072                  * ...
2073                  * But only if not writing to stdout, since in that case
2074                  * the network is most likely throttling writes anyway,
2075                  * and therefore it is best to go to the write phase ASAP
2076                  * instead, as we can afford spending more time compressing
2077                  * between writes at that moment.
2078                  */
2079                 if (entry->delta_data && !pack_to_stdout) {
2080                         entry->z_delta_size = do_compress(&entry->delta_data,
2081                                                           entry->delta_size);
2082                         cache_lock();
2083                         delta_cache_size -= entry->delta_size;
2084                         delta_cache_size += entry->z_delta_size;
2085                         cache_unlock();
2086                 }
2087
2088                 /* if we made n a delta, and if n is already at max
2089                  * depth, leaving it in the window is pointless.  we
2090                  * should evict it first.
2091                  */
2092                 if (entry->delta && max_depth <= n->depth)
2093                         continue;
2094
2095                 /*
2096                  * Move the best delta base up in the window, after the
2097                  * currently deltified object, to keep it longer.  It will
2098                  * be the first base object to be attempted next.
2099                  */
2100                 if (entry->delta) {
2101                         struct unpacked swap = array[best_base];
2102                         int dist = (window + idx - best_base) % window;
2103                         int dst = best_base;
2104                         while (dist--) {
2105                                 int src = (dst + 1) % window;
2106                                 array[dst] = array[src];
2107                                 dst = src;
2108                         }
2109                         array[dst] = swap;
2110                 }
2111
2112                 next:
2113                 idx++;
2114                 if (count + 1 < window)
2115                         count++;
2116                 if (idx >= window)
2117                         idx = 0;
2118         }
2119
2120         for (i = 0; i < window; ++i) {
2121                 free_delta_index(array[i].index);
2122                 free(array[i].data);
2123         }
2124         free(array);
2125 }
2126
2127 #ifndef NO_PTHREADS
2128
2129 static void try_to_free_from_threads(size_t size)
2130 {
2131         read_lock();
2132         release_pack_memory(size);
2133         read_unlock();
2134 }
2135
2136 static try_to_free_t old_try_to_free_routine;
2137
2138 /*
2139  * The main thread waits on the condition that (at least) one of the workers
2140  * has stopped working (which is indicated in the .working member of
2141  * struct thread_params).
2142  * When a work thread has completed its work, it sets .working to 0 and
2143  * signals the main thread and waits on the condition that .data_ready
2144  * becomes 1.
2145  */
2146
2147 struct thread_params {
2148         pthread_t thread;
2149         struct object_entry **list;
2150         unsigned list_size;
2151         unsigned remaining;
2152         int window;
2153         int depth;
2154         int working;
2155         int data_ready;
2156         pthread_mutex_t mutex;
2157         pthread_cond_t cond;
2158         unsigned *processed;
2159 };
2160
2161 static pthread_cond_t progress_cond;
2162
2163 /*
2164  * Mutex and conditional variable can't be statically-initialized on Windows.
2165  */
2166 static void init_threaded_search(void)
2167 {
2168         init_recursive_mutex(&read_mutex);
2169         pthread_mutex_init(&cache_mutex, NULL);
2170         pthread_mutex_init(&progress_mutex, NULL);
2171         pthread_cond_init(&progress_cond, NULL);
2172         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2173 }
2174
2175 static void cleanup_threaded_search(void)
2176 {
2177         set_try_to_free_routine(old_try_to_free_routine);
2178         pthread_cond_destroy(&progress_cond);
2179         pthread_mutex_destroy(&read_mutex);
2180         pthread_mutex_destroy(&cache_mutex);
2181         pthread_mutex_destroy(&progress_mutex);
2182 }
2183
2184 static void *threaded_find_deltas(void *arg)
2185 {
2186         struct thread_params *me = arg;
2187
2188         progress_lock();
2189         while (me->remaining) {
2190                 progress_unlock();
2191
2192                 find_deltas(me->list, &me->remaining,
2193                             me->window, me->depth, me->processed);
2194
2195                 progress_lock();
2196                 me->working = 0;
2197                 pthread_cond_signal(&progress_cond);
2198                 progress_unlock();
2199
2200                 /*
2201                  * We must not set ->data_ready before we wait on the
2202                  * condition because the main thread may have set it to 1
2203                  * before we get here. In order to be sure that new
2204                  * work is available if we see 1 in ->data_ready, it
2205                  * was initialized to 0 before this thread was spawned
2206                  * and we reset it to 0 right away.
2207                  */
2208                 pthread_mutex_lock(&me->mutex);
2209                 while (!me->data_ready)
2210                         pthread_cond_wait(&me->cond, &me->mutex);
2211                 me->data_ready = 0;
2212                 pthread_mutex_unlock(&me->mutex);
2213
2214                 progress_lock();
2215         }
2216         progress_unlock();
2217         /* leave ->working 1 so that this doesn't get more work assigned */
2218         return NULL;
2219 }
2220
2221 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2222                            int window, int depth, unsigned *processed)
2223 {
2224         struct thread_params *p;
2225         int i, ret, active_threads = 0;
2226
2227         init_threaded_search();
2228
2229         if (delta_search_threads <= 1) {
2230                 find_deltas(list, &list_size, window, depth, processed);
2231                 cleanup_threaded_search();
2232                 return;
2233         }
2234         if (progress > pack_to_stdout)
2235                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2236                                 delta_search_threads);
2237         p = xcalloc(delta_search_threads, sizeof(*p));
2238
2239         /* Partition the work amongst work threads. */
2240         for (i = 0; i < delta_search_threads; i++) {
2241                 unsigned sub_size = list_size / (delta_search_threads - i);
2242
2243                 /* don't use too small segments or no deltas will be found */
2244                 if (sub_size < 2*window && i+1 < delta_search_threads)
2245                         sub_size = 0;
2246
2247                 p[i].window = window;
2248                 p[i].depth = depth;
2249                 p[i].processed = processed;
2250                 p[i].working = 1;
2251                 p[i].data_ready = 0;
2252
2253                 /* try to split chunks on "path" boundaries */
2254                 while (sub_size && sub_size < list_size &&
2255                        list[sub_size]->hash &&
2256                        list[sub_size]->hash == list[sub_size-1]->hash)
2257                         sub_size++;
2258
2259                 p[i].list = list;
2260                 p[i].list_size = sub_size;
2261                 p[i].remaining = sub_size;
2262
2263                 list += sub_size;
2264                 list_size -= sub_size;
2265         }
2266
2267         /* Start work threads. */
2268         for (i = 0; i < delta_search_threads; i++) {
2269                 if (!p[i].list_size)
2270                         continue;
2271                 pthread_mutex_init(&p[i].mutex, NULL);
2272                 pthread_cond_init(&p[i].cond, NULL);
2273                 ret = pthread_create(&p[i].thread, NULL,
2274                                      threaded_find_deltas, &p[i]);
2275                 if (ret)
2276                         die("unable to create thread: %s", strerror(ret));
2277                 active_threads++;
2278         }
2279
2280         /*
2281          * Now let's wait for work completion.  Each time a thread is done
2282          * with its work, we steal half of the remaining work from the
2283          * thread with the largest number of unprocessed objects and give
2284          * it to that newly idle thread.  This ensure good load balancing
2285          * until the remaining object list segments are simply too short
2286          * to be worth splitting anymore.
2287          */
2288         while (active_threads) {
2289                 struct thread_params *target = NULL;
2290                 struct thread_params *victim = NULL;
2291                 unsigned sub_size = 0;
2292
2293                 progress_lock();
2294                 for (;;) {
2295                         for (i = 0; !target && i < delta_search_threads; i++)
2296                                 if (!p[i].working)
2297                                         target = &p[i];
2298                         if (target)
2299                                 break;
2300                         pthread_cond_wait(&progress_cond, &progress_mutex);
2301                 }
2302
2303                 for (i = 0; i < delta_search_threads; i++)
2304                         if (p[i].remaining > 2*window &&
2305                             (!victim || victim->remaining < p[i].remaining))
2306                                 victim = &p[i];
2307                 if (victim) {
2308                         sub_size = victim->remaining / 2;
2309                         list = victim->list + victim->list_size - sub_size;
2310                         while (sub_size && list[0]->hash &&
2311                                list[0]->hash == list[-1]->hash) {
2312                                 list++;
2313                                 sub_size--;
2314                         }
2315                         if (!sub_size) {
2316                                 /*
2317                                  * It is possible for some "paths" to have
2318                                  * so many objects that no hash boundary
2319                                  * might be found.  Let's just steal the
2320                                  * exact half in that case.
2321                                  */
2322                                 sub_size = victim->remaining / 2;
2323                                 list -= sub_size;
2324                         }
2325                         target->list = list;
2326                         victim->list_size -= sub_size;
2327                         victim->remaining -= sub_size;
2328                 }
2329                 target->list_size = sub_size;
2330                 target->remaining = sub_size;
2331                 target->working = 1;
2332                 progress_unlock();
2333
2334                 pthread_mutex_lock(&target->mutex);
2335                 target->data_ready = 1;
2336                 pthread_cond_signal(&target->cond);
2337                 pthread_mutex_unlock(&target->mutex);
2338
2339                 if (!sub_size) {
2340                         pthread_join(target->thread, NULL);
2341                         pthread_cond_destroy(&target->cond);
2342                         pthread_mutex_destroy(&target->mutex);
2343                         active_threads--;
2344                 }
2345         }
2346         cleanup_threaded_search();
2347         free(p);
2348 }
2349
2350 #else
2351 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2352 #endif
2353
2354 static void add_tag_chain(const struct object_id *oid)
2355 {
2356         struct tag *tag;
2357
2358         /*
2359          * We catch duplicates already in add_object_entry(), but we'd
2360          * prefer to do this extra check to avoid having to parse the
2361          * tag at all if we already know that it's being packed (e.g., if
2362          * it was included via bitmaps, we would not have parsed it
2363          * previously).
2364          */
2365         if (packlist_find(&to_pack, oid->hash, NULL))
2366                 return;
2367
2368         tag = lookup_tag(oid);
2369         while (1) {
2370                 if (!tag || parse_tag(tag) || !tag->tagged)
2371                         die("unable to pack objects reachable from tag %s",
2372                             oid_to_hex(oid));
2373
2374                 add_object_entry(&tag->object.oid, OBJ_TAG, NULL, 0);
2375
2376                 if (tag->tagged->type != OBJ_TAG)
2377                         return;
2378
2379                 tag = (struct tag *)tag->tagged;
2380         }
2381 }
2382
2383 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2384 {
2385         struct object_id peeled;
2386
2387         if (starts_with(path, "refs/tags/") && /* is a tag? */
2388             !peel_ref(path, &peeled)    && /* peelable? */
2389             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2390                 add_tag_chain(oid);
2391         return 0;
2392 }
2393
2394 static void prepare_pack(int window, int depth)
2395 {
2396         struct object_entry **delta_list;
2397         uint32_t i, nr_deltas;
2398         unsigned n;
2399
2400         get_object_details();
2401
2402         /*
2403          * If we're locally repacking then we need to be doubly careful
2404          * from now on in order to make sure no stealth corruption gets
2405          * propagated to the new pack.  Clients receiving streamed packs
2406          * should validate everything they get anyway so no need to incur
2407          * the additional cost here in that case.
2408          */
2409         if (!pack_to_stdout)
2410                 do_check_packed_object_crc = 1;
2411
2412         if (!to_pack.nr_objects || !window || !depth)
2413                 return;
2414
2415         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2416         nr_deltas = n = 0;
2417
2418         for (i = 0; i < to_pack.nr_objects; i++) {
2419                 struct object_entry *entry = to_pack.objects + i;
2420
2421                 if (entry->delta)
2422                         /* This happens if we decided to reuse existing
2423                          * delta from a pack.  "reuse_delta &&" is implied.
2424                          */
2425                         continue;
2426
2427                 if (entry->size < 50)
2428                         continue;
2429
2430                 if (entry->no_try_delta)
2431                         continue;
2432
2433                 if (!entry->preferred_base) {
2434                         nr_deltas++;
2435                         if (entry->type < 0)
2436                                 die("unable to get type of object %s",
2437                                     oid_to_hex(&entry->idx.oid));
2438                 } else {
2439                         if (entry->type < 0) {
2440                                 /*
2441                                  * This object is not found, but we
2442                                  * don't have to include it anyway.
2443                                  */
2444                                 continue;
2445                         }
2446                 }
2447
2448                 delta_list[n++] = entry;
2449         }
2450
2451         if (nr_deltas && n > 1) {
2452                 unsigned nr_done = 0;
2453                 if (progress)
2454                         progress_state = start_progress(_("Compressing objects"),
2455                                                         nr_deltas);
2456                 QSORT(delta_list, n, type_size_sort);
2457                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2458                 stop_progress(&progress_state);
2459                 if (nr_done != nr_deltas)
2460                         die("inconsistency with delta count");
2461         }
2462         free(delta_list);
2463 }
2464
2465 static int git_pack_config(const char *k, const char *v, void *cb)
2466 {
2467         if (!strcmp(k, "pack.window")) {
2468                 window = git_config_int(k, v);
2469                 return 0;
2470         }
2471         if (!strcmp(k, "pack.windowmemory")) {
2472                 window_memory_limit = git_config_ulong(k, v);
2473                 return 0;
2474         }
2475         if (!strcmp(k, "pack.depth")) {
2476                 depth = git_config_int(k, v);
2477                 return 0;
2478         }
2479         if (!strcmp(k, "pack.deltacachesize")) {
2480                 max_delta_cache_size = git_config_int(k, v);
2481                 return 0;
2482         }
2483         if (!strcmp(k, "pack.deltacachelimit")) {
2484                 cache_max_small_delta_size = git_config_int(k, v);
2485                 return 0;
2486         }
2487         if (!strcmp(k, "pack.writebitmaphashcache")) {
2488                 if (git_config_bool(k, v))
2489                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2490                 else
2491                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2492         }
2493         if (!strcmp(k, "pack.usebitmaps")) {
2494                 use_bitmap_index_default = git_config_bool(k, v);
2495                 return 0;
2496         }
2497         if (!strcmp(k, "pack.threads")) {
2498                 delta_search_threads = git_config_int(k, v);
2499                 if (delta_search_threads < 0)
2500                         die("invalid number of threads specified (%d)",
2501                             delta_search_threads);
2502 #ifdef NO_PTHREADS
2503                 if (delta_search_threads != 1) {
2504                         warning("no threads support, ignoring %s", k);
2505                         delta_search_threads = 0;
2506                 }
2507 #endif
2508                 return 0;
2509         }
2510         if (!strcmp(k, "pack.indexversion")) {
2511                 pack_idx_opts.version = git_config_int(k, v);
2512                 if (pack_idx_opts.version > 2)
2513                         die("bad pack.indexversion=%"PRIu32,
2514                             pack_idx_opts.version);
2515                 return 0;
2516         }
2517         return git_default_config(k, v, cb);
2518 }
2519
2520 static void read_object_list_from_stdin(void)
2521 {
2522         char line[GIT_MAX_HEXSZ + 1 + PATH_MAX + 2];
2523         struct object_id oid;
2524         const char *p;
2525
2526         for (;;) {
2527                 if (!fgets(line, sizeof(line), stdin)) {
2528                         if (feof(stdin))
2529                                 break;
2530                         if (!ferror(stdin))
2531                                 die("fgets returned NULL, not EOF, not error!");
2532                         if (errno != EINTR)
2533                                 die_errno("fgets");
2534                         clearerr(stdin);
2535                         continue;
2536                 }
2537                 if (line[0] == '-') {
2538                         if (get_oid_hex(line+1, &oid))
2539                                 die("expected edge object ID, got garbage:\n %s",
2540                                     line);
2541                         add_preferred_base(&oid);
2542                         continue;
2543                 }
2544                 if (parse_oid_hex(line, &oid, &p))
2545                         die("expected object ID, got garbage:\n %s", line);
2546
2547                 add_preferred_base_object(p + 1);
2548                 add_object_entry(&oid, 0, p + 1, 0);
2549         }
2550 }
2551
2552 #define OBJECT_ADDED (1u<<20)
2553
2554 static void show_commit(struct commit *commit, void *data)
2555 {
2556         add_object_entry(&commit->object.oid, OBJ_COMMIT, NULL, 0);
2557         commit->object.flags |= OBJECT_ADDED;
2558
2559         if (write_bitmap_index)
2560                 index_commit_for_bitmap(commit);
2561 }
2562
2563 static void show_object(struct object *obj, const char *name, void *data)
2564 {
2565         add_preferred_base_object(name);
2566         add_object_entry(&obj->oid, obj->type, name, 0);
2567         obj->flags |= OBJECT_ADDED;
2568 }
2569
2570 static void show_object__ma_allow_any(struct object *obj, const char *name, void *data)
2571 {
2572         assert(arg_missing_action == MA_ALLOW_ANY);
2573
2574         /*
2575          * Quietly ignore ALL missing objects.  This avoids problems with
2576          * staging them now and getting an odd error later.
2577          */
2578         if (!has_object_file(&obj->oid))
2579                 return;
2580
2581         show_object(obj, name, data);
2582 }
2583
2584 static void show_object__ma_allow_promisor(struct object *obj, const char *name, void *data)
2585 {
2586         assert(arg_missing_action == MA_ALLOW_PROMISOR);
2587
2588         /*
2589          * Quietly ignore EXPECTED missing objects.  This avoids problems with
2590          * staging them now and getting an odd error later.
2591          */
2592         if (!has_object_file(&obj->oid) && is_promisor_object(&obj->oid))
2593                 return;
2594
2595         show_object(obj, name, data);
2596 }
2597
2598 static int option_parse_missing_action(const struct option *opt,
2599                                        const char *arg, int unset)
2600 {
2601         assert(arg);
2602         assert(!unset);
2603
2604         if (!strcmp(arg, "error")) {
2605                 arg_missing_action = MA_ERROR;
2606                 fn_show_object = show_object;
2607                 return 0;
2608         }
2609
2610         if (!strcmp(arg, "allow-any")) {
2611                 arg_missing_action = MA_ALLOW_ANY;
2612                 fetch_if_missing = 0;
2613                 fn_show_object = show_object__ma_allow_any;
2614                 return 0;
2615         }
2616
2617         if (!strcmp(arg, "allow-promisor")) {
2618                 arg_missing_action = MA_ALLOW_PROMISOR;
2619                 fetch_if_missing = 0;
2620                 fn_show_object = show_object__ma_allow_promisor;
2621                 return 0;
2622         }
2623
2624         die(_("invalid value for --missing"));
2625         return 0;
2626 }
2627
2628 static void show_edge(struct commit *commit)
2629 {
2630         add_preferred_base(&commit->object.oid);
2631 }
2632
2633 struct in_pack_object {
2634         off_t offset;
2635         struct object *object;
2636 };
2637
2638 struct in_pack {
2639         unsigned int alloc;
2640         unsigned int nr;
2641         struct in_pack_object *array;
2642 };
2643
2644 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2645 {
2646         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2647         in_pack->array[in_pack->nr].object = object;
2648         in_pack->nr++;
2649 }
2650
2651 /*
2652  * Compare the objects in the offset order, in order to emulate the
2653  * "git rev-list --objects" output that produced the pack originally.
2654  */
2655 static int ofscmp(const void *a_, const void *b_)
2656 {
2657         struct in_pack_object *a = (struct in_pack_object *)a_;
2658         struct in_pack_object *b = (struct in_pack_object *)b_;
2659
2660         if (a->offset < b->offset)
2661                 return -1;
2662         else if (a->offset > b->offset)
2663                 return 1;
2664         else
2665                 return oidcmp(&a->object->oid, &b->object->oid);
2666 }
2667
2668 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2669 {
2670         struct packed_git *p;
2671         struct in_pack in_pack;
2672         uint32_t i;
2673
2674         memset(&in_pack, 0, sizeof(in_pack));
2675
2676         for (p = packed_git; p; p = p->next) {
2677                 struct object_id oid;
2678                 struct object *o;
2679
2680                 if (!p->pack_local || p->pack_keep)
2681                         continue;
2682                 if (open_pack_index(p))
2683                         die("cannot open pack index");
2684
2685                 ALLOC_GROW(in_pack.array,
2686                            in_pack.nr + p->num_objects,
2687                            in_pack.alloc);
2688
2689                 for (i = 0; i < p->num_objects; i++) {
2690                         nth_packed_object_oid(&oid, p, i);
2691                         o = lookup_unknown_object(oid.hash);
2692                         if (!(o->flags & OBJECT_ADDED))
2693                                 mark_in_pack_object(o, p, &in_pack);
2694                         o->flags |= OBJECT_ADDED;
2695                 }
2696         }
2697
2698         if (in_pack.nr) {
2699                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2700                 for (i = 0; i < in_pack.nr; i++) {
2701                         struct object *o = in_pack.array[i].object;
2702                         add_object_entry(&o->oid, o->type, "", 0);
2703                 }
2704         }
2705         free(in_pack.array);
2706 }
2707
2708 static int add_loose_object(const struct object_id *oid, const char *path,
2709                             void *data)
2710 {
2711         enum object_type type = sha1_object_info(oid->hash, NULL);
2712
2713         if (type < 0) {
2714                 warning("loose object at %s could not be examined", path);
2715                 return 0;
2716         }
2717
2718         add_object_entry(oid, type, "", 0);
2719         return 0;
2720 }
2721
2722 /*
2723  * We actually don't even have to worry about reachability here.
2724  * add_object_entry will weed out duplicates, so we just add every
2725  * loose object we find.
2726  */
2727 static void add_unreachable_loose_objects(void)
2728 {
2729         for_each_loose_file_in_objdir(get_object_directory(),
2730                                       add_loose_object,
2731                                       NULL, NULL, NULL);
2732 }
2733
2734 static int has_sha1_pack_kept_or_nonlocal(const struct object_id *oid)
2735 {
2736         static struct packed_git *last_found = (void *)1;
2737         struct packed_git *p;
2738
2739         p = (last_found != (void *)1) ? last_found : packed_git;
2740
2741         while (p) {
2742                 if ((!p->pack_local || p->pack_keep) &&
2743                         find_pack_entry_one(oid->hash, p)) {
2744                         last_found = p;
2745                         return 1;
2746                 }
2747                 if (p == last_found)
2748                         p = packed_git;
2749                 else
2750                         p = p->next;
2751                 if (p == last_found)
2752                         p = p->next;
2753         }
2754         return 0;
2755 }
2756
2757 /*
2758  * Store a list of sha1s that are should not be discarded
2759  * because they are either written too recently, or are
2760  * reachable from another object that was.
2761  *
2762  * This is filled by get_object_list.
2763  */
2764 static struct oid_array recent_objects;
2765
2766 static int loosened_object_can_be_discarded(const struct object_id *oid,
2767                                             timestamp_t mtime)
2768 {
2769         if (!unpack_unreachable_expiration)
2770                 return 0;
2771         if (mtime > unpack_unreachable_expiration)
2772                 return 0;
2773         if (oid_array_lookup(&recent_objects, oid) >= 0)
2774                 return 0;
2775         return 1;
2776 }
2777
2778 static void loosen_unused_packed_objects(struct rev_info *revs)
2779 {
2780         struct packed_git *p;
2781         uint32_t i;
2782         struct object_id oid;
2783
2784         for (p = packed_git; p; p = p->next) {
2785                 if (!p->pack_local || p->pack_keep)
2786                         continue;
2787
2788                 if (open_pack_index(p))
2789                         die("cannot open pack index");
2790
2791                 for (i = 0; i < p->num_objects; i++) {
2792                         nth_packed_object_oid(&oid, p, i);
2793                         if (!packlist_find(&to_pack, oid.hash, NULL) &&
2794                             !has_sha1_pack_kept_or_nonlocal(&oid) &&
2795                             !loosened_object_can_be_discarded(&oid, p->mtime))
2796                                 if (force_object_loose(&oid, p->mtime))
2797                                         die("unable to force loose object");
2798                 }
2799         }
2800 }
2801
2802 /*
2803  * This tracks any options which pack-reuse code expects to be on, or which a
2804  * reader of the pack might not understand, and which would therefore prevent
2805  * blind reuse of what we have on disk.
2806  */
2807 static int pack_options_allow_reuse(void)
2808 {
2809         return pack_to_stdout &&
2810                allow_ofs_delta &&
2811                !ignore_packed_keep &&
2812                (!local || !have_non_local_packs) &&
2813                !incremental;
2814 }
2815
2816 static int get_object_list_from_bitmap(struct rev_info *revs)
2817 {
2818         if (prepare_bitmap_walk(revs) < 0)
2819                 return -1;
2820
2821         if (pack_options_allow_reuse() &&
2822             !reuse_partial_packfile_from_bitmap(
2823                         &reuse_packfile,
2824                         &reuse_packfile_objects,
2825                         &reuse_packfile_offset)) {
2826                 assert(reuse_packfile_objects);
2827                 nr_result += reuse_packfile_objects;
2828                 display_progress(progress_state, nr_result);
2829         }
2830
2831         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2832         return 0;
2833 }
2834
2835 static void record_recent_object(struct object *obj,
2836                                  const char *name,
2837                                  void *data)
2838 {
2839         oid_array_append(&recent_objects, &obj->oid);
2840 }
2841
2842 static void record_recent_commit(struct commit *commit, void *data)
2843 {
2844         oid_array_append(&recent_objects, &commit->object.oid);
2845 }
2846
2847 static void get_object_list(int ac, const char **av)
2848 {
2849         struct rev_info revs;
2850         char line[1000];
2851         int flags = 0;
2852
2853         init_revisions(&revs, NULL);
2854         save_commit_buffer = 0;
2855         setup_revisions(ac, av, &revs, NULL);
2856
2857         /* make sure shallows are read */
2858         is_repository_shallow();
2859
2860         while (fgets(line, sizeof(line), stdin) != NULL) {
2861                 int len = strlen(line);
2862                 if (len && line[len - 1] == '\n')
2863                         line[--len] = 0;
2864                 if (!len)
2865                         break;
2866                 if (*line == '-') {
2867                         if (!strcmp(line, "--not")) {
2868                                 flags ^= UNINTERESTING;
2869                                 write_bitmap_index = 0;
2870                                 continue;
2871                         }
2872                         if (starts_with(line, "--shallow ")) {
2873                                 struct object_id oid;
2874                                 if (get_oid_hex(line + 10, &oid))
2875                                         die("not an SHA-1 '%s'", line + 10);
2876                                 register_shallow(&oid);
2877                                 use_bitmap_index = 0;
2878                                 continue;
2879                         }
2880                         die("not a rev '%s'", line);
2881                 }
2882                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2883                         die("bad revision '%s'", line);
2884         }
2885
2886         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2887                 return;
2888
2889         if (prepare_revision_walk(&revs))
2890                 die("revision walk setup failed");
2891         mark_edges_uninteresting(&revs, show_edge);
2892
2893         if (!fn_show_object)
2894                 fn_show_object = show_object;
2895         traverse_commit_list_filtered(&filter_options, &revs,
2896                                       show_commit, fn_show_object, NULL,
2897                                       NULL);
2898
2899         if (unpack_unreachable_expiration) {
2900                 revs.ignore_missing_links = 1;
2901                 if (add_unseen_recent_objects_to_traversal(&revs,
2902                                 unpack_unreachable_expiration))
2903                         die("unable to add recent objects");
2904                 if (prepare_revision_walk(&revs))
2905                         die("revision walk setup failed");
2906                 traverse_commit_list(&revs, record_recent_commit,
2907                                      record_recent_object, NULL);
2908         }
2909
2910         if (keep_unreachable)
2911                 add_objects_in_unpacked_packs(&revs);
2912         if (pack_loose_unreachable)
2913                 add_unreachable_loose_objects();
2914         if (unpack_unreachable)
2915                 loosen_unused_packed_objects(&revs);
2916
2917         oid_array_clear(&recent_objects);
2918 }
2919
2920 static int option_parse_index_version(const struct option *opt,
2921                                       const char *arg, int unset)
2922 {
2923         char *c;
2924         const char *val = arg;
2925         pack_idx_opts.version = strtoul(val, &c, 10);
2926         if (pack_idx_opts.version > 2)
2927                 die(_("unsupported index version %s"), val);
2928         if (*c == ',' && c[1])
2929                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2930         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2931                 die(_("bad index version '%s'"), val);
2932         return 0;
2933 }
2934
2935 static int option_parse_unpack_unreachable(const struct option *opt,
2936                                            const char *arg, int unset)
2937 {
2938         if (unset) {
2939                 unpack_unreachable = 0;
2940                 unpack_unreachable_expiration = 0;
2941         }
2942         else {
2943                 unpack_unreachable = 1;
2944                 if (arg)
2945                         unpack_unreachable_expiration = approxidate(arg);
2946         }
2947         return 0;
2948 }
2949
2950 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2951 {
2952         int use_internal_rev_list = 0;
2953         int thin = 0;
2954         int shallow = 0;
2955         int all_progress_implied = 0;
2956         struct argv_array rp = ARGV_ARRAY_INIT;
2957         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2958         int rev_list_index = 0;
2959         struct option pack_objects_options[] = {
2960                 OPT_SET_INT('q', "quiet", &progress,
2961                             N_("do not show progress meter"), 0),
2962                 OPT_SET_INT(0, "progress", &progress,
2963                             N_("show progress meter"), 1),
2964                 OPT_SET_INT(0, "all-progress", &progress,
2965                             N_("show progress meter during object writing phase"), 2),
2966                 OPT_BOOL(0, "all-progress-implied",
2967                          &all_progress_implied,
2968                          N_("similar to --all-progress when progress meter is shown")),
2969                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2970                   N_("write the pack index file in the specified idx format version"),
2971                   0, option_parse_index_version },
2972                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2973                               N_("maximum size of each output pack file")),
2974                 OPT_BOOL(0, "local", &local,
2975                          N_("ignore borrowed objects from alternate object store")),
2976                 OPT_BOOL(0, "incremental", &incremental,
2977                          N_("ignore packed objects")),
2978                 OPT_INTEGER(0, "window", &window,
2979                             N_("limit pack window by objects")),
2980                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2981                               N_("limit pack window by memory in addition to object limit")),
2982                 OPT_INTEGER(0, "depth", &depth,
2983                             N_("maximum length of delta chain allowed in the resulting pack")),
2984                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2985                          N_("reuse existing deltas")),
2986                 OPT_BOOL(0, "reuse-object", &reuse_object,
2987                          N_("reuse existing objects")),
2988                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2989                          N_("use OFS_DELTA objects")),
2990                 OPT_INTEGER(0, "threads", &delta_search_threads,
2991                             N_("use threads when searching for best delta matches")),
2992                 OPT_BOOL(0, "non-empty", &non_empty,
2993                          N_("do not create an empty pack output")),
2994                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2995                          N_("read revision arguments from standard input")),
2996                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2997                   N_("limit the objects to those that are not yet packed"),
2998                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2999                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
3000                   N_("include objects reachable from any reference"),
3001                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
3002                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
3003                   N_("include objects referred by reflog entries"),
3004                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
3005                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
3006                   N_("include objects referred to by the index"),
3007                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
3008                 OPT_BOOL(0, "stdout", &pack_to_stdout,
3009                          N_("output pack to stdout")),
3010                 OPT_BOOL(0, "include-tag", &include_tag,
3011                          N_("include tag objects that refer to objects to be packed")),
3012                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
3013                          N_("keep unreachable objects")),
3014                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
3015                          N_("pack loose unreachable objects")),
3016                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
3017                   N_("unpack unreachable objects newer than <time>"),
3018                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
3019                 OPT_BOOL(0, "thin", &thin,
3020                          N_("create thin packs")),
3021                 OPT_BOOL(0, "shallow", &shallow,
3022                          N_("create packs suitable for shallow fetches")),
3023                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
3024                          N_("ignore packs that have companion .keep file")),
3025                 OPT_INTEGER(0, "compression", &pack_compression_level,
3026                             N_("pack compression level")),
3027                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
3028                             N_("do not hide commits by grafts"), 0),
3029                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
3030                          N_("use a bitmap index if available to speed up counting objects")),
3031                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
3032                          N_("write a bitmap index together with the pack index")),
3033                 OPT_PARSE_LIST_OBJECTS_FILTER(&filter_options),
3034                 { OPTION_CALLBACK, 0, "missing", NULL, N_("action"),
3035                   N_("handling for missing objects"), PARSE_OPT_NONEG,
3036                   option_parse_missing_action },
3037                 OPT_BOOL(0, "exclude-promisor-objects", &exclude_promisor_objects,
3038                          N_("do not pack objects in promisor packfiles")),
3039                 OPT_END(),
3040         };
3041
3042         check_replace_refs = 0;
3043
3044         reset_pack_idx_option(&pack_idx_opts);
3045         git_config(git_pack_config, NULL);
3046
3047         progress = isatty(2);
3048         argc = parse_options(argc, argv, prefix, pack_objects_options,
3049                              pack_usage, 0);
3050
3051         if (argc) {
3052                 base_name = argv[0];
3053                 argc--;
3054         }
3055         if (pack_to_stdout != !base_name || argc)
3056                 usage_with_options(pack_usage, pack_objects_options);
3057
3058         argv_array_push(&rp, "pack-objects");
3059         if (thin) {
3060                 use_internal_rev_list = 1;
3061                 argv_array_push(&rp, shallow
3062                                 ? "--objects-edge-aggressive"
3063                                 : "--objects-edge");
3064         } else
3065                 argv_array_push(&rp, "--objects");
3066
3067         if (rev_list_all) {
3068                 use_internal_rev_list = 1;
3069                 argv_array_push(&rp, "--all");
3070         }
3071         if (rev_list_reflog) {
3072                 use_internal_rev_list = 1;
3073                 argv_array_push(&rp, "--reflog");
3074         }
3075         if (rev_list_index) {
3076                 use_internal_rev_list = 1;
3077                 argv_array_push(&rp, "--indexed-objects");
3078         }
3079         if (rev_list_unpacked) {
3080                 use_internal_rev_list = 1;
3081                 argv_array_push(&rp, "--unpacked");
3082         }
3083
3084         if (exclude_promisor_objects) {
3085                 use_internal_rev_list = 1;
3086                 fetch_if_missing = 0;
3087                 argv_array_push(&rp, "--exclude-promisor-objects");
3088         }
3089
3090         if (!reuse_object)
3091                 reuse_delta = 0;
3092         if (pack_compression_level == -1)
3093                 pack_compression_level = Z_DEFAULT_COMPRESSION;
3094         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
3095                 die("bad pack compression level %d", pack_compression_level);
3096
3097         if (!delta_search_threads)      /* --threads=0 means autodetect */
3098                 delta_search_threads = online_cpus();
3099
3100 #ifdef NO_PTHREADS
3101         if (delta_search_threads != 1)
3102                 warning("no threads support, ignoring --threads");
3103 #endif
3104         if (!pack_to_stdout && !pack_size_limit)
3105                 pack_size_limit = pack_size_limit_cfg;
3106         if (pack_to_stdout && pack_size_limit)
3107                 die("--max-pack-size cannot be used to build a pack for transfer.");
3108         if (pack_size_limit && pack_size_limit < 1024*1024) {
3109                 warning("minimum pack size limit is 1 MiB");
3110                 pack_size_limit = 1024*1024;
3111         }
3112
3113         if (!pack_to_stdout && thin)
3114                 die("--thin cannot be used to build an indexable pack.");
3115
3116         if (keep_unreachable && unpack_unreachable)
3117                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
3118         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3119                 unpack_unreachable_expiration = 0;
3120
3121         if (filter_options.choice) {
3122                 if (!pack_to_stdout)
3123                         die("cannot use --filter without --stdout.");
3124                 use_bitmap_index = 0;
3125         }
3126
3127         /*
3128          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3129          *
3130          * - to produce good pack (with bitmap index not-yet-packed objects are
3131          *   packed in suboptimal order).
3132          *
3133          * - to use more robust pack-generation codepath (avoiding possible
3134          *   bugs in bitmap code and possible bitmap index corruption).
3135          */
3136         if (!pack_to_stdout)
3137                 use_bitmap_index_default = 0;
3138
3139         if (use_bitmap_index < 0)
3140                 use_bitmap_index = use_bitmap_index_default;
3141
3142         /* "hard" reasons not to use bitmaps; these just won't work at all */
3143         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
3144                 use_bitmap_index = 0;
3145
3146         if (pack_to_stdout || !rev_list_all)
3147                 write_bitmap_index = 0;
3148
3149         if (progress && all_progress_implied)
3150                 progress = 2;
3151
3152         prepare_packed_git();
3153         if (ignore_packed_keep) {
3154                 struct packed_git *p;
3155                 for (p = packed_git; p; p = p->next)
3156                         if (p->pack_local && p->pack_keep)
3157                                 break;
3158                 if (!p) /* no keep-able packs found */
3159                         ignore_packed_keep = 0;
3160         }
3161         if (local) {
3162                 /*
3163                  * unlike ignore_packed_keep above, we do not want to
3164                  * unset "local" based on looking at packs, as it
3165                  * also covers non-local objects
3166                  */
3167                 struct packed_git *p;
3168                 for (p = packed_git; p; p = p->next) {
3169                         if (!p->pack_local) {
3170                                 have_non_local_packs = 1;
3171                                 break;
3172                         }
3173                 }
3174         }
3175
3176         if (progress)
3177                 progress_state = start_progress(_("Counting objects"), 0);
3178         if (!use_internal_rev_list)
3179                 read_object_list_from_stdin();
3180         else {
3181                 get_object_list(rp.argc, rp.argv);
3182                 argv_array_clear(&rp);
3183         }
3184         cleanup_preferred_base();
3185         if (include_tag && nr_result)
3186                 for_each_ref(add_ref_tag, NULL);
3187         stop_progress(&progress_state);
3188
3189         if (non_empty && !nr_result)
3190                 return 0;
3191         if (nr_result)
3192                 prepare_pack(window, depth);
3193         write_pack_file();
3194         if (progress)
3195                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
3196                         " reused %"PRIu32" (delta %"PRIu32")\n",
3197                         written, written_delta, reused, reused_delta);
3198         return 0;
3199 }