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

   1 #include "builtin.h"
   2 #include "cache.h"
   3 #include "repository.h"
   4 #include "config.h"
   5 #include "attr.h"
   6 #include "object.h"
   7 #include "blob.h"
   8 #include "commit.h"
   9 #include "tag.h"
  10 #include "tree.h"
  11 #include "delta.h"
  12 #include "pack.h"
  13 #include "pack-revindex.h"
  14 #include "csum-file.h"
  15 #include "tree-walk.h"
  16 #include "diff.h"
  17 #include "revision.h"
  18 #include "list-objects.h"
  19 #include "list-objects-filter.h"
  20 #include "list-objects-filter-options.h"
  21 #include "pack-objects.h"
  22 #include "progress.h"
  23 #include "refs.h"
  24 #include "streaming.h"
  25 #include "thread-utils.h"
  26 #include "pack-bitmap.h"
  27 #include "reachable.h"
  28 #include "sha1-array.h"
  29 #include "argv-array.h"
  30 #include "list.h"
  31 #include "packfile.h"
  32 #include "object-store.h"
  33 #include "dir.h"
  34
  35 static const char *pack_usage[] = {
  36         N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
  37         N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
  38         NULL
  39 };
  40
  41 /*
  42  * Objects we are going to pack are collected in the `to_pack` structure.
  43  * It contains an array (dynamically expanded) of the object data, and a map
  44  * that can resolve SHA1s to their position in the array.
  45  */
  46 static struct packing_data to_pack;
  47
  48 static struct pack_idx_entry **written_list;
  49 static uint32_t nr_result, nr_written, nr_seen;
  50
  51 static int non_empty;
  52 static int reuse_delta = 1, reuse_object = 1;
  53 static int keep_unreachable, unpack_unreachable, include_tag;
  54 static timestamp_t unpack_unreachable_expiration;
  55 static int pack_loose_unreachable;
  56 static int local;
  57 static int have_non_local_packs;
  58 static int incremental;
  59 static int ignore_packed_keep_on_disk;
  60 static int ignore_packed_keep_in_core;
  61 static int allow_ofs_delta;
  62 static struct pack_idx_option pack_idx_opts;
  63 static const char *base_name;
  64 static int progress = 1;
  65 static int window = 10;
  66 static unsigned long pack_size_limit;
  67 static int depth = 50;
  68 static int delta_search_threads;
  69 static int pack_to_stdout;
  70 static int num_preferred_base;
  71 static struct progress *progress_state;
  72
  73 static struct packed_git *reuse_packfile;
  74 static uint32_t reuse_packfile_objects;
  75 static off_t reuse_packfile_offset;
  76
  77 static int use_bitmap_index_default = 1;
  78 static int use_bitmap_index = -1;
  79 static int write_bitmap_index;
  80 static uint16_t write_bitmap_options;
  81
  82 static int exclude_promisor_objects;
  83
  84 static unsigned long delta_cache_size = 0;
  85 static unsigned long max_delta_cache_size = DEFAULT_DELTA_CACHE_SIZE;
  86 static unsigned long cache_max_small_delta_size = 1000;
  87
  88 static unsigned long window_memory_limit = 0;
  89
  90 static struct list_objects_filter_options filter_options;
  91
  92 enum missing_action {
  93         MA_ERROR = 0,      /* fail if any missing objects are encountered */
  94         MA_ALLOW_ANY,      /* silently allow ALL missing objects */
  95         MA_ALLOW_PROMISOR, /* silently allow all missing PROMISOR objects */
  96 };
  97 static enum missing_action arg_missing_action;
  98 static show_object_fn fn_show_object;
  99
 100 /*
 101  * stats
 102  */
 103 static uint32_t written, written_delta;
 104 static uint32_t reused, reused_delta;
 105
 106 /*
 107  * Indexed commits
 108  */
 109 static struct commit **indexed_commits;
 110 static unsigned int indexed_commits_nr;
 111 static unsigned int indexed_commits_alloc;
 112
 113 static void index_commit_for_bitmap(struct commit *commit)
 114 {
 115         if (indexed_commits_nr >= indexed_commits_alloc) {
 116                 indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
 117                 REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
 118         }
 119
 120         indexed_commits[indexed_commits_nr++] = commit;
 121 }
 122
 123 static void *get_delta(struct object_entry *entry)
 124 {
 125         unsigned long size, base_size, delta_size;
 126         void *buf, *base_buf, *delta_buf;
 127         enum object_type type;
 128
 129         buf = read_object_file(&entry->idx.oid, &type, &size);
 130         if (!buf)
 131                 die("unable to read %s", oid_to_hex(&entry->idx.oid));
 132         base_buf = read_object_file(&entry->delta->idx.oid, &type, &base_size);
 133         if (!base_buf)
 134                 die("unable to read %s",
 135                     oid_to_hex(&entry->delta->idx.oid));
 136         delta_buf = diff_delta(base_buf, base_size,
 137                                buf, size, &delta_size, 0);
 138         if (!delta_buf || delta_size != entry->delta_size)
 139                 die("delta size changed");
 140         free(buf);
 141         free(base_buf);
 142         return delta_buf;
 143 }
 144
 145 static unsigned long do_compress(void **pptr, unsigned long size)
 146 {
 147         git_zstream stream;
 148         void *in, *out;
 149         unsigned long maxsize;
 150
 151         git_deflate_init(&stream, pack_compression_level);
 152         maxsize = git_deflate_bound(&stream, size);
 153
 154         in = *pptr;
 155         out = xmalloc(maxsize);
 156         *pptr = out;
 157
 158         stream.next_in = in;
 159         stream.avail_in = size;
 160         stream.next_out = out;
 161         stream.avail_out = maxsize;
 162         while (git_deflate(&stream, Z_FINISH) == Z_OK)
 163                 ; /* nothing */
 164         git_deflate_end(&stream);
 165
 166         free(in);
 167         return stream.total_out;
 168 }
 169
 170 static unsigned long write_large_blob_data(struct git_istream *st, struct hashfile *f,
 171                                            const struct object_id *oid)
 172 {
 173         git_zstream stream;
 174         unsigned char ibuf[1024 * 16];
 175         unsigned char obuf[1024 * 16];
 176         unsigned long olen = 0;
 177
 178         git_deflate_init(&stream, pack_compression_level);
 179
 180         for (;;) {
 181                 ssize_t readlen;
 182                 int zret = Z_OK;
 183                 readlen = read_istream(st, ibuf, sizeof(ibuf));
 184                 if (readlen == -1)
 185                         die(_("unable to read %s"), oid_to_hex(oid));
 186
 187                 stream.next_in = ibuf;
 188                 stream.avail_in = readlen;
 189                 while ((stream.avail_in || readlen == 0) &&
 190                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
 191                         stream.next_out = obuf;
 192                         stream.avail_out = sizeof(obuf);
 193                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 194                         hashwrite(f, obuf, stream.next_out - obuf);
 195                         olen += stream.next_out - obuf;
 196                 }
 197                 if (stream.avail_in)
 198                         die(_("deflate error (%d)"), zret);
 199                 if (readlen == 0) {
 200                         if (zret != Z_STREAM_END)
 201                                 die(_("deflate error (%d)"), zret);
 202                         break;
 203                 }
 204         }
 205         git_deflate_end(&stream);
 206         return olen;
 207 }
 208
 209 /*
 210  * we are going to reuse the existing object data as is.  make
 211  * sure it is not corrupt.
 212  */
 213 static int check_pack_inflate(struct packed_git *p,
 214                 struct pack_window **w_curs,
 215                 off_t offset,
 216                 off_t len,
 217                 unsigned long expect)
 218 {
 219         git_zstream stream;
 220         unsigned char fakebuf[4096], *in;
 221         int st;
 222
 223         memset(&stream, 0, sizeof(stream));
 224         git_inflate_init(&stream);
 225         do {
 226                 in = use_pack(p, w_curs, offset, &stream.avail_in);
 227                 stream.next_in = in;
 228                 stream.next_out = fakebuf;
 229                 stream.avail_out = sizeof(fakebuf);
 230                 st = git_inflate(&stream, Z_FINISH);
 231                 offset += stream.next_in - in;
 232         } while (st == Z_OK || st == Z_BUF_ERROR);
 233         git_inflate_end(&stream);
 234         return (st == Z_STREAM_END &&
 235                 stream.total_out == expect &&
 236                 stream.total_in == len) ? 0 : -1;
 237 }
 238
 239 static void copy_pack_data(struct hashfile *f,
 240                 struct packed_git *p,
 241                 struct pack_window **w_curs,
 242                 off_t offset,
 243                 off_t len)
 244 {
 245         unsigned char *in;
 246         unsigned long avail;
 247
 248         while (len) {
 249                 in = use_pack(p, w_curs, offset, &avail);
 250                 if (avail > len)
 251                         avail = (unsigned long)len;
 252                 hashwrite(f, in, avail);
 253                 offset += avail;
 254                 len -= avail;
 255         }
 256 }
 257
 258 /* Return 0 if we will bust the pack-size limit */
 259 static unsigned long write_no_reuse_object(struct hashfile *f, struct object_entry *entry,
 260                                            unsigned long limit, int usable_delta)
 261 {
 262         unsigned long size, datalen;
 263         unsigned char header[MAX_PACK_OBJECT_HEADER],
 264                       dheader[MAX_PACK_OBJECT_HEADER];
 265         unsigned hdrlen;
 266         enum object_type type;
 267         void *buf;
 268         struct git_istream *st = NULL;
 269
 270         if (!usable_delta) {
 271                 if (entry->type == OBJ_BLOB &&
 272                     entry->size > big_file_threshold &&
 273                     (st = open_istream(&entry->idx.oid, &type, &size, NULL)) != NULL)
 274                         buf = NULL;
 275                 else {
 276                         buf = read_object_file(&entry->idx.oid, &type, &size);
 277                         if (!buf)
 278                                 die(_("unable to read %s"),
 279                                     oid_to_hex(&entry->idx.oid));
 280                 }
 281                 /*
 282                  * make sure no cached delta data remains from a
 283                  * previous attempt before a pack split occurred.
 284                  */
 285                 FREE_AND_NULL(entry->delta_data);
 286                 entry->z_delta_size = 0;
 287         } else if (entry->delta_data) {
 288                 size = entry->delta_size;
 289                 buf = entry->delta_data;
 290                 entry->delta_data = NULL;
 291                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 292                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 293         } else {
 294                 buf = get_delta(entry);
 295                 size = entry->delta_size;
 296                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 297                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 298         }
 299
 300         if (st) /* large blob case, just assume we don't compress well */
 301                 datalen = size;
 302         else if (entry->z_delta_size)
 303                 datalen = entry->z_delta_size;
 304         else
 305                 datalen = do_compress(&buf, size);
 306
 307         /*
 308          * The object header is a byte of 'type' followed by zero or
 309          * more bytes of length.
 310          */
 311         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 312                                               type, size);
 313
 314         if (type == OBJ_OFS_DELTA) {
 315                 /*
 316                  * Deltas with relative base contain an additional
 317                  * encoding of the relative offset for the delta
 318                  * base from this object's position in the pack.
 319                  */
 320                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 321                 unsigned pos = sizeof(dheader) - 1;
 322                 dheader[pos] = ofs & 127;
 323                 while (ofs >>= 7)
 324                         dheader[--pos] = 128 | (--ofs & 127);
 325                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 326                         if (st)
 327                                 close_istream(st);
 328                         free(buf);
 329                         return 0;
 330                 }
 331                 hashwrite(f, header, hdrlen);
 332                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
 333                 hdrlen += sizeof(dheader) - pos;
 334         } else if (type == OBJ_REF_DELTA) {
 335                 /*
 336                  * Deltas with a base reference contain
 337                  * an additional 20 bytes for the base sha1.
 338                  */
 339                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 340                         if (st)
 341                                 close_istream(st);
 342                         free(buf);
 343                         return 0;
 344                 }
 345                 hashwrite(f, header, hdrlen);
 346                 hashwrite(f, entry->delta->idx.oid.hash, 20);
 347                 hdrlen += 20;
 348         } else {
 349                 if (limit && hdrlen + datalen + 20 >= limit) {
 350                         if (st)
 351                                 close_istream(st);
 352                         free(buf);
 353                         return 0;
 354                 }
 355                 hashwrite(f, header, hdrlen);
 356         }
 357         if (st) {
 358                 datalen = write_large_blob_data(st, f, &entry->idx.oid);
 359                 close_istream(st);
 360         } else {
 361                 hashwrite(f, buf, datalen);
 362                 free(buf);
 363         }
 364
 365         return hdrlen + datalen;
 366 }
 367
 368 /* Return 0 if we will bust the pack-size limit */
 369 static off_t write_reuse_object(struct hashfile *f, struct object_entry *entry,
 370                                 unsigned long limit, int usable_delta)
 371 {
 372         struct packed_git *p = entry->in_pack;
 373         struct pack_window *w_curs = NULL;
 374         struct revindex_entry *revidx;
 375         off_t offset;
 376         enum object_type type = entry->type;
 377         off_t datalen;
 378         unsigned char header[MAX_PACK_OBJECT_HEADER],
 379                       dheader[MAX_PACK_OBJECT_HEADER];
 380         unsigned hdrlen;
 381
 382         if (entry->delta)
 383                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 384                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 385         hdrlen = encode_in_pack_object_header(header, sizeof(header),
 386                                               type, entry->size);
 387
 388         offset = entry->in_pack_offset;
 389         revidx = find_pack_revindex(p, offset);
 390         datalen = revidx[1].offset - offset;
 391         if (!pack_to_stdout && p->index_version > 1 &&
 392             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 393                 error("bad packed object CRC for %s",
 394                       oid_to_hex(&entry->idx.oid));
 395                 unuse_pack(&w_curs);
 396                 return write_no_reuse_object(f, entry, limit, usable_delta);
 397         }
 398
 399         offset += entry->in_pack_header_size;
 400         datalen -= entry->in_pack_header_size;
 401
 402         if (!pack_to_stdout && p->index_version == 1 &&
 403             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 404                 error("corrupt packed object for %s",
 405                       oid_to_hex(&entry->idx.oid));
 406                 unuse_pack(&w_curs);
 407                 return write_no_reuse_object(f, entry, limit, usable_delta);
 408         }
 409
 410         if (type == OBJ_OFS_DELTA) {
 411                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 412                 unsigned pos = sizeof(dheader) - 1;
 413                 dheader[pos] = ofs & 127;
 414                 while (ofs >>= 7)
 415                         dheader[--pos] = 128 | (--ofs & 127);
 416                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 417                         unuse_pack(&w_curs);
 418                         return 0;
 419                 }
 420                 hashwrite(f, header, hdrlen);
 421                 hashwrite(f, dheader + pos, sizeof(dheader) - pos);
 422                 hdrlen += sizeof(dheader) - pos;
 423                 reused_delta++;
 424         } else if (type == OBJ_REF_DELTA) {
 425                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 426                         unuse_pack(&w_curs);
 427                         return 0;
 428                 }
 429                 hashwrite(f, header, hdrlen);
 430                 hashwrite(f, entry->delta->idx.oid.hash, 20);
 431                 hdrlen += 20;
 432                 reused_delta++;
 433         } else {
 434                 if (limit && hdrlen + datalen + 20 >= limit) {
 435                         unuse_pack(&w_curs);
 436                         return 0;
 437                 }
 438                 hashwrite(f, header, hdrlen);
 439         }
 440         copy_pack_data(f, p, &w_curs, offset, datalen);
 441         unuse_pack(&w_curs);
 442         reused++;
 443         return hdrlen + datalen;
 444 }
 445
 446 /* Return 0 if we will bust the pack-size limit */
 447 static off_t write_object(struct hashfile *f,
 448                           struct object_entry *entry,
 449                           off_t write_offset)
 450 {
 451         unsigned long limit;
 452         off_t len;
 453         int usable_delta, to_reuse;
 454
 455         if (!pack_to_stdout)
 456                 crc32_begin(f);
 457
 458         /* apply size limit if limited packsize and not first object */
 459         if (!pack_size_limit || !nr_written)
 460                 limit = 0;
 461         else if (pack_size_limit <= write_offset)
 462                 /*
 463                  * the earlier object did not fit the limit; avoid
 464                  * mistaking this with unlimited (i.e. limit = 0).
 465                  */
 466                 limit = 1;
 467         else
 468                 limit = pack_size_limit - write_offset;
 469
 470         if (!entry->delta)
 471                 usable_delta = 0;       /* no delta */
 472         else if (!pack_size_limit)
 473                usable_delta = 1;        /* unlimited packfile */
 474         else if (entry->delta->idx.offset == (off_t)-1)
 475                 usable_delta = 0;       /* base was written to another pack */
 476         else if (entry->delta->idx.offset)
 477                 usable_delta = 1;       /* base already exists in this pack */
 478         else
 479                 usable_delta = 0;       /* base could end up in another pack */
 480
 481         if (!reuse_object)
 482                 to_reuse = 0;   /* explicit */
 483         else if (!entry->in_pack)
 484                 to_reuse = 0;   /* can't reuse what we don't have */
 485         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
 486                                 /* check_object() decided it for us ... */
 487                 to_reuse = usable_delta;
 488                                 /* ... but pack split may override that */
 489         else if (entry->type != entry->in_pack_type)
 490                 to_reuse = 0;   /* pack has delta which is unusable */
 491         else if (entry->delta)
 492                 to_reuse = 0;   /* we want to pack afresh */
 493         else
 494                 to_reuse = 1;   /* we have it in-pack undeltified,
 495                                  * and we do not need to deltify it.
 496                                  */
 497
 498         if (!to_reuse)
 499                 len = write_no_reuse_object(f, entry, limit, usable_delta);
 500         else
 501                 len = write_reuse_object(f, entry, limit, usable_delta);
 502         if (!len)
 503                 return 0;
 504
 505         if (usable_delta)
 506                 written_delta++;
 507         written++;
 508         if (!pack_to_stdout)
 509                 entry->idx.crc32 = crc32_end(f);
 510         return len;
 511 }
 512
 513 enum write_one_status {
 514         WRITE_ONE_SKIP = -1, /* already written */
 515         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 516         WRITE_ONE_WRITTEN = 1, /* normal */
 517         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 518 };
 519
 520 static enum write_one_status write_one(struct hashfile *f,
 521                                        struct object_entry *e,
 522                                        off_t *offset)
 523 {
 524         off_t size;
 525         int recursing;
 526
 527         /*
 528          * we set offset to 1 (which is an impossible value) to mark
 529          * the fact that this object is involved in "write its base
 530          * first before writing a deltified object" recursion.
 531          */
 532         recursing = (e->idx.offset == 1);
 533         if (recursing) {
 534                 warning("recursive delta detected for object %s",
 535                         oid_to_hex(&e->idx.oid));
 536                 return WRITE_ONE_RECURSIVE;
 537         } else if (e->idx.offset || e->preferred_base) {
 538                 /* offset is non zero if object is written already. */
 539                 return WRITE_ONE_SKIP;
 540         }
 541
 542         /* if we are deltified, write out base object first. */
 543         if (e->delta) {
 544                 e->idx.offset = 1; /* now recurse */
 545                 switch (write_one(f, e->delta, offset)) {
 546                 case WRITE_ONE_RECURSIVE:
 547                         /* we cannot depend on this one */
 548                         e->delta = NULL;
 549                         break;
 550                 default:
 551                         break;
 552                 case WRITE_ONE_BREAK:
 553                         e->idx.offset = recursing;
 554                         return WRITE_ONE_BREAK;
 555                 }
 556         }
 557
 558         e->idx.offset = *offset;
 559         size = write_object(f, e, *offset);
 560         if (!size) {
 561                 e->idx.offset = recursing;
 562                 return WRITE_ONE_BREAK;
 563         }
 564         written_list[nr_written++] = &e->idx;
 565
 566         /* make sure off_t is sufficiently large not to wrap */
 567         if (signed_add_overflows(*offset, size))
 568                 die("pack too large for current definition of off_t");
 569         *offset += size;
 570         return WRITE_ONE_WRITTEN;
 571 }
 572
 573 static int mark_tagged(const char *path, const struct object_id *oid, int flag,
 574                        void *cb_data)
 575 {
 576         struct object_id peeled;
 577         struct object_entry *entry = packlist_find(&to_pack, oid->hash, NULL);
 578
 579         if (entry)
 580                 entry->tagged = 1;
 581         if (!peel_ref(path, &peeled)) {
 582                 entry = packlist_find(&to_pack, peeled.hash, NULL);
 583                 if (entry)
 584                         entry->tagged = 1;
 585         }
 586         return 0;
 587 }
 588
 589 static inline void add_to_write_order(struct object_entry **wo,
 590                                unsigned int *endp,
 591                                struct object_entry *e)
 592 {
 593         if (e->filled)
 594                 return;
 595         wo[(*endp)++] = e;
 596         e->filled = 1;
 597 }
 598
 599 static void add_descendants_to_write_order(struct object_entry **wo,
 600                                            unsigned int *endp,
 601                                            struct object_entry *e)
 602 {
 603         int add_to_order = 1;
 604         while (e) {
 605                 if (add_to_order) {
 606                         struct object_entry *s;
 607                         /* add this node... */
 608                         add_to_write_order(wo, endp, e);
 609                         /* all its siblings... */
 610                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
 611                                 add_to_write_order(wo, endp, s);
 612                         }
 613                 }
 614                 /* drop down a level to add left subtree nodes if possible */
 615                 if (e->delta_child) {
 616                         add_to_order = 1;
 617                         e = e->delta_child;
 618                 } else {
 619                         add_to_order = 0;
 620                         /* our sibling might have some children, it is next */
 621                         if (e->delta_sibling) {
 622                                 e = e->delta_sibling;
 623                                 continue;
 624                         }
 625                         /* go back to our parent node */
 626                         e = e->delta;
 627                         while (e && !e->delta_sibling) {
 628                                 /* we're on the right side of a subtree, keep
 629                                  * going up until we can go right again */
 630                                 e = e->delta;
 631                         }
 632                         if (!e) {
 633                                 /* done- we hit our original root node */
 634                                 return;
 635                         }
 636                         /* pass it off to sibling at this level */
 637                         e = e->delta_sibling;
 638                 }
 639         };
 640 }
 641
 642 static void add_family_to_write_order(struct object_entry **wo,
 643                                       unsigned int *endp,
 644                                       struct object_entry *e)
 645 {
 646         struct object_entry *root;
 647
 648         for (root = e; root->delta; root = root->delta)
 649                 ; /* nothing */
 650         add_descendants_to_write_order(wo, endp, root);
 651 }
 652
 653 static struct object_entry **compute_write_order(void)
 654 {
 655         unsigned int i, wo_end, last_untagged;
 656
 657         struct object_entry **wo;
 658         struct object_entry *objects = to_pack.objects;
 659
 660         for (i = 0; i < to_pack.nr_objects; i++) {
 661                 objects[i].tagged = 0;
 662                 objects[i].filled = 0;
 663                 objects[i].delta_child = NULL;
 664                 objects[i].delta_sibling = NULL;
 665         }
 666
 667         /*
 668          * Fully connect delta_child/delta_sibling network.
 669          * Make sure delta_sibling is sorted in the original
 670          * recency order.
 671          */
 672         for (i = to_pack.nr_objects; i > 0;) {
 673                 struct object_entry *e = &objects[--i];
 674                 if (!e->delta)
 675                         continue;
 676                 /* Mark me as the first child */
 677                 e->delta_sibling = e->delta->delta_child;
 678                 e->delta->delta_child = e;
 679         }
 680
 681         /*
 682          * Mark objects that are at the tip of tags.
 683          */
 684         for_each_tag_ref(mark_tagged, NULL);
 685
 686         /*
 687          * Give the objects in the original recency order until
 688          * we see a tagged tip.
 689          */
 690         ALLOC_ARRAY(wo, to_pack.nr_objects);
 691         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
 692                 if (objects[i].tagged)
 693                         break;
 694                 add_to_write_order(wo, &wo_end, &objects[i]);
 695         }
 696         last_untagged = i;
 697
 698         /*
 699          * Then fill all the tagged tips.
 700          */
 701         for (; i < to_pack.nr_objects; i++) {
 702                 if (objects[i].tagged)
 703                         add_to_write_order(wo, &wo_end, &objects[i]);
 704         }
 705
 706         /*
 707          * And then all remaining commits and tags.
 708          */
 709         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 710                 if (objects[i].type != OBJ_COMMIT &&
 711                     objects[i].type != OBJ_TAG)
 712                         continue;
 713                 add_to_write_order(wo, &wo_end, &objects[i]);
 714         }
 715
 716         /*
 717          * And then all the trees.
 718          */
 719         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 720                 if (objects[i].type != OBJ_TREE)
 721                         continue;
 722                 add_to_write_order(wo, &wo_end, &objects[i]);
 723         }
 724
 725         /*
 726          * Finally all the rest in really tight order
 727          */
 728         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 729                 if (!objects[i].filled)
 730                         add_family_to_write_order(wo, &wo_end, &objects[i]);
 731         }
 732
 733         if (wo_end != to_pack.nr_objects)
 734                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
 735
 736         return wo;
 737 }
 738
 739 static off_t write_reused_pack(struct hashfile *f)
 740 {
 741         unsigned char buffer[8192];
 742         off_t to_write, total;
 743         int fd;
 744
 745         if (!is_pack_valid(reuse_packfile))
 746                 die("packfile is invalid: %s", reuse_packfile->pack_name);
 747
 748         fd = git_open(reuse_packfile->pack_name);
 749         if (fd < 0)
 750                 die_errno("unable to open packfile for reuse: %s",
 751                           reuse_packfile->pack_name);
 752
 753         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
 754                 die_errno("unable to seek in reused packfile");
 755
 756         if (reuse_packfile_offset < 0)
 757                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
 758
 759         total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
 760
 761         while (to_write) {
 762                 int read_pack = xread(fd, buffer, sizeof(buffer));
 763
 764                 if (read_pack <= 0)
 765                         die_errno("unable to read from reused packfile");
 766
 767                 if (read_pack > to_write)
 768                         read_pack = to_write;
 769
 770                 hashwrite(f, buffer, read_pack);
 771                 to_write -= read_pack;
 772
 773                 /*
 774                  * We don't know the actual number of objects written,
 775                  * only how many bytes written, how many bytes total, and
 776                  * how many objects total. So we can fake it by pretending all
 777                  * objects we are writing are the same size. This gives us a
 778                  * smooth progress meter, and at the end it matches the true
 779                  * answer.
 780                  */
 781                 written = reuse_packfile_objects *
 782                                 (((double)(total - to_write)) / total);
 783                 display_progress(progress_state, written);
 784         }
 785
 786         close(fd);
 787         written = reuse_packfile_objects;
 788         display_progress(progress_state, written);
 789         return reuse_packfile_offset - sizeof(struct pack_header);
 790 }
 791
 792 static const char no_split_warning[] = N_(
 793 "disabling bitmap writing, packs are split due to pack.packSizeLimit"
 794 );
 795
 796 static void write_pack_file(void)
 797 {
 798         uint32_t i = 0, j;
 799         struct hashfile *f;
 800         off_t offset;
 801         uint32_t nr_remaining = nr_result;
 802         time_t last_mtime = 0;
 803         struct object_entry **write_order;
 804
 805         if (progress > pack_to_stdout)
 806                 progress_state = start_progress(_("Writing objects"), nr_result);
 807         ALLOC_ARRAY(written_list, to_pack.nr_objects);
 808         write_order = compute_write_order();
 809
 810         do {
 811                 struct object_id oid;
 812                 char *pack_tmp_name = NULL;
 813
 814                 if (pack_to_stdout)
 815                         f = hashfd_throughput(1, "<stdout>", progress_state);
 816                 else
 817                         f = create_tmp_packfile(&pack_tmp_name);
 818
 819                 offset = write_pack_header(f, nr_remaining);
 820
 821                 if (reuse_packfile) {
 822                         off_t packfile_size;
 823                         assert(pack_to_stdout);
 824
 825                         packfile_size = write_reused_pack(f);
 826                         offset += packfile_size;
 827                 }
 828
 829                 nr_written = 0;
 830                 for (; i < to_pack.nr_objects; i++) {
 831                         struct object_entry *e = write_order[i];
 832                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 833                                 break;
 834                         display_progress(progress_state, written);
 835                 }
 836
 837                 /*
 838                  * Did we write the wrong # entries in the header?
 839                  * If so, rewrite it like in fast-import
 840                  */
 841                 if (pack_to_stdout) {
 842                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_CLOSE);
 843                 } else if (nr_written == nr_remaining) {
 844                         finalize_hashfile(f, oid.hash, CSUM_HASH_IN_STREAM | CSUM_FSYNC | CSUM_CLOSE);
 845                 } else {
 846                         int fd = finalize_hashfile(f, oid.hash, 0);
 847                         fixup_pack_header_footer(fd, oid.hash, pack_tmp_name,
 848                                                  nr_written, oid.hash, offset);
 849                         close(fd);
 850                         if (write_bitmap_index) {
 851                                 warning(_(no_split_warning));
 852                                 write_bitmap_index = 0;
 853                         }
 854                 }
 855
 856                 if (!pack_to_stdout) {
 857                         struct stat st;
 858                         struct strbuf tmpname = STRBUF_INIT;
 859
 860                         /*
 861                          * Packs are runtime accessed in their mtime
 862                          * order since newer packs are more likely to contain
 863                          * younger objects.  So if we are creating multiple
 864                          * packs then we should modify the mtime of later ones
 865                          * to preserve this property.
 866                          */
 867                         if (stat(pack_tmp_name, &st) < 0) {
 868                                 warning_errno("failed to stat %s", pack_tmp_name);
 869                         } else if (!last_mtime) {
 870                                 last_mtime = st.st_mtime;
 871                         } else {
 872                                 struct utimbuf utb;
 873                                 utb.actime = st.st_atime;
 874                                 utb.modtime = --last_mtime;
 875                                 if (utime(pack_tmp_name, &utb) < 0)
 876                                         warning_errno("failed utime() on %s", pack_tmp_name);
 877                         }
 878
 879                         strbuf_addf(&tmpname, "%s-", base_name);
 880
 881                         if (write_bitmap_index) {
 882                                 bitmap_writer_set_checksum(oid.hash);
 883                                 bitmap_writer_build_type_index(written_list, nr_written);
 884                         }
 885
 886                         finish_tmp_packfile(&tmpname, pack_tmp_name,
 887                                             written_list, nr_written,
 888                                             &pack_idx_opts, oid.hash);
 889
 890                         if (write_bitmap_index) {
 891                                 strbuf_addf(&tmpname, "%s.bitmap", oid_to_hex(&oid));
 892
 893                                 stop_progress(&progress_state);
 894
 895                                 bitmap_writer_show_progress(progress);
 896                                 bitmap_writer_reuse_bitmaps(&to_pack);
 897                                 bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
 898                                 bitmap_writer_build(&to_pack);
 899                                 bitmap_writer_finish(written_list, nr_written,
 900                                                      tmpname.buf, write_bitmap_options);
 901                                 write_bitmap_index = 0;
 902                         }
 903
 904                         strbuf_release(&tmpname);
 905                         free(pack_tmp_name);
 906                         puts(oid_to_hex(&oid));
 907                 }
 908
 909                 /* mark written objects as written to previous pack */
 910                 for (j = 0; j < nr_written; j++) {
 911                         written_list[j]->offset = (off_t)-1;
 912                 }
 913                 nr_remaining -= nr_written;
 914         } while (nr_remaining && i < to_pack.nr_objects);
 915
 916         free(written_list);
 917         free(write_order);
 918         stop_progress(&progress_state);
 919         if (written != nr_result)
 920                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
 921                         written, nr_result);
 922 }
 923
 924 static int no_try_delta(const char *path)
 925 {
 926         static struct attr_check *check;
 927
 928         if (!check)
 929                 check = attr_check_initl("delta", NULL);
 930         if (git_check_attr(path, check))
 931                 return 0;
 932         if (ATTR_FALSE(check->items[0].value))
 933                 return 1;
 934         return 0;
 935 }
 936
 937 /*
 938  * When adding an object, check whether we have already added it
 939  * to our packing list. If so, we can skip. However, if we are
 940  * being asked to excludei t, but the previous mention was to include
 941  * it, make sure to adjust its flags and tweak our numbers accordingly.
 942  *
 943  * As an optimization, we pass out the index position where we would have
 944  * found the item, since that saves us from having to look it up again a
 945  * few lines later when we want to add the new entry.
 946  */
 947 static int have_duplicate_entry(const struct object_id *oid,
 948                                 int exclude,
 949                                 uint32_t *index_pos)
 950 {
 951         struct object_entry *entry;
 952
 953         entry = packlist_find(&to_pack, oid->hash, index_pos);
 954         if (!entry)
 955                 return 0;
 956
 957         if (exclude) {
 958                 if (!entry->preferred_base)
 959                         nr_result--;
 960                 entry->preferred_base = 1;
 961         }
 962
 963         return 1;
 964 }
 965
 966 static int want_found_object(int exclude, struct packed_git *p)
 967 {
 968         if (exclude)
 969                 return 1;
 970         if (incremental)
 971                 return 0;
 972
 973         /*
 974          * When asked to do --local (do not include an object that appears in a
 975          * pack we borrow from elsewhere) or --honor-pack-keep (do not include
 976          * an object that appears in a pack marked with .keep), finding a pack
 977          * that matches the criteria is sufficient for us to decide to omit it.
 978          * However, even if this pack does not satisfy the criteria, we need to
 979          * make sure no copy of this object appears in _any_ pack that makes us
 980          * to omit the object, so we need to check all the packs.
 981          *
 982          * We can however first check whether these options can possible matter;
 983          * if they do not matter we know we want the object in generated pack.
 984          * Otherwise, we signal "-1" at the end to tell the caller that we do
 985          * not know either way, and it needs to check more packs.
 986          */
 987         if (!ignore_packed_keep_on_disk &&
 988             !ignore_packed_keep_in_core &&
 989             (!local || !have_non_local_packs))
 990                 return 1;
 991
 992         if (local && !p->pack_local)
 993                 return 0;
 994         if (p->pack_local &&
 995             ((ignore_packed_keep_on_disk && p->pack_keep) ||
 996              (ignore_packed_keep_in_core && p->pack_keep_in_core)))
 997                 return 0;
 998
 999         /* we don't know yet; keep looking for more packs */
1000         return -1;
1001 }
1002
1003 /*
1004  * Check whether we want the object in the pack (e.g., we do not want
1005  * objects found in non-local stores if the "--local" option was used).
1006  *
1007  * If the caller already knows an existing pack it wants to take the object
1008  * from, that is passed in *found_pack and *found_offset; otherwise this
1009  * function finds if there is any pack that has the object and returns the pack
1010  * and its offset in these variables.
1011  */
1012 static int want_object_in_pack(const struct object_id *oid,
1013                                int exclude,
1014                                struct packed_git **found_pack,
1015                                off_t *found_offset)
1016 {
1017         int want;
1018         struct list_head *pos;
1019
1020         if (!exclude && local && has_loose_object_nonlocal(oid->hash))
1021                 return 0;
1022
1023         /*
1024          * If we already know the pack object lives in, start checks from that
1025          * pack - in the usual case when neither --local was given nor .keep files
1026          * are present we will determine the answer right now.
1027          */
1028         if (*found_pack) {
1029                 want = want_found_object(exclude, *found_pack);
1030                 if (want != -1)
1031                         return want;
1032         }
1033         list_for_each(pos, get_packed_git_mru(the_repository)) {
1034                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
1035                 off_t offset;
1036
1037                 if (p == *found_pack)
1038                         offset = *found_offset;
1039                 else
1040                         offset = find_pack_entry_one(oid->hash, p);
1041
1042                 if (offset) {
1043                         if (!*found_pack) {
1044                                 if (!is_pack_valid(p))
1045                                         continue;
1046                                 *found_offset = offset;
1047                                 *found_pack = p;
1048                         }
1049                         want = want_found_object(exclude, p);
1050                         if (!exclude && want > 0)
1051                                 list_move(&p->mru,
1052                                           get_packed_git_mru(the_repository));
1053                         if (want != -1)
1054                                 return want;
1055                 }
1056         }
1057
1058         return 1;
1059 }
1060
1061 static void create_object_entry(const struct object_id *oid,
1062                                 enum object_type type,
1063                                 uint32_t hash,
1064                                 int exclude,
1065                                 int no_try_delta,
1066                                 uint32_t index_pos,
1067                                 struct packed_git *found_pack,
1068                                 off_t found_offset)
1069 {
1070         struct object_entry *entry;
1071
1072         entry = packlist_alloc(&to_pack, oid->hash, index_pos);
1073         entry->hash = hash;
1074         if (type)
1075                 entry->type = type;
1076         if (exclude)
1077                 entry->preferred_base = 1;
1078         else
1079                 nr_result++;
1080         if (found_pack) {
1081                 entry->in_pack = found_pack;
1082                 entry->in_pack_offset = found_offset;
1083         }
1084
1085         entry->no_try_delta = no_try_delta;
1086 }
1087
1088 static const char no_closure_warning[] = N_(
1089 "disabling bitmap writing, as some objects are not being packed"
1090 );
1091
1092 static int add_object_entry(const struct object_id *oid, enum object_type type,
1093                             const char *name, int exclude)
1094 {
1095         struct packed_git *found_pack = NULL;
1096         off_t found_offset = 0;
1097         uint32_t index_pos;
1098
1099         display_progress(progress_state, ++nr_seen);
1100
1101         if (have_duplicate_entry(oid, exclude, &index_pos))
1102                 return 0;
1103
1104         if (!want_object_in_pack(oid, exclude, &found_pack, &found_offset)) {
1105                 /* The pack is missing an object, so it will not have closure */
1106                 if (write_bitmap_index) {
1107                         warning(_(no_closure_warning));
1108                         write_bitmap_index = 0;
1109                 }
1110                 return 0;
1111         }
1112
1113         create_object_entry(oid, type, pack_name_hash(name),
1114                             exclude, name && no_try_delta(name),
1115                             index_pos, found_pack, found_offset);
1116         return 1;
1117 }
1118
1119 static int add_object_entry_from_bitmap(const struct object_id *oid,
1120                                         enum object_type type,
1121                                         int flags, uint32_t name_hash,
1122                                         struct packed_git *pack, off_t offset)
1123 {
1124         uint32_t index_pos;
1125
1126         display_progress(progress_state, ++nr_seen);
1127
1128         if (have_duplicate_entry(oid, 0, &index_pos))
1129                 return 0;
1130
1131         if (!want_object_in_pack(oid, 0, &pack, &offset))
1132                 return 0;
1133
1134         create_object_entry(oid, type, name_hash, 0, 0, index_pos, pack, offset);
1135         return 1;
1136 }
1137
1138 struct pbase_tree_cache {
1139         struct object_id oid;
1140         int ref;
1141         int temporary;
1142         void *tree_data;
1143         unsigned long tree_size;
1144 };
1145
1146 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1147 static int pbase_tree_cache_ix(const struct object_id *oid)
1148 {
1149         return oid->hash[0] % ARRAY_SIZE(pbase_tree_cache);
1150 }
1151 static int pbase_tree_cache_ix_incr(int ix)
1152 {
1153         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1154 }
1155
1156 static struct pbase_tree {
1157         struct pbase_tree *next;
1158         /* This is a phony "cache" entry; we are not
1159          * going to evict it or find it through _get()
1160          * mechanism -- this is for the toplevel node that
1161          * would almost always change with any commit.
1162          */
1163         struct pbase_tree_cache pcache;
1164 } *pbase_tree;
1165
1166 static struct pbase_tree_cache *pbase_tree_get(const struct object_id *oid)
1167 {
1168         struct pbase_tree_cache *ent, *nent;
1169         void *data;
1170         unsigned long size;
1171         enum object_type type;
1172         int neigh;
1173         int my_ix = pbase_tree_cache_ix(oid);
1174         int available_ix = -1;
1175
1176         /* pbase-tree-cache acts as a limited hashtable.
1177          * your object will be found at your index or within a few
1178          * slots after that slot if it is cached.
1179          */
1180         for (neigh = 0; neigh < 8; neigh++) {
1181                 ent = pbase_tree_cache[my_ix];
1182                 if (ent && !oidcmp(&ent->oid, oid)) {
1183                         ent->ref++;
1184                         return ent;
1185                 }
1186                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1187                          ((0 <= available_ix) &&
1188                           (!ent && pbase_tree_cache[available_ix])))
1189                         available_ix = my_ix;
1190                 if (!ent)
1191                         break;
1192                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1193         }
1194
1195         /* Did not find one.  Either we got a bogus request or
1196          * we need to read and perhaps cache.
1197          */
1198         data = read_object_file(oid, &type, &size);
1199         if (!data)
1200                 return NULL;
1201         if (type != OBJ_TREE) {
1202                 free(data);
1203                 return NULL;
1204         }
1205
1206         /* We need to either cache or return a throwaway copy */
1207
1208         if (available_ix < 0)
1209                 ent = NULL;
1210         else {
1211                 ent = pbase_tree_cache[available_ix];
1212                 my_ix = available_ix;
1213         }
1214
1215         if (!ent) {
1216                 nent = xmalloc(sizeof(*nent));
1217                 nent->temporary = (available_ix < 0);
1218         }
1219         else {
1220                 /* evict and reuse */
1221                 free(ent->tree_data);
1222                 nent = ent;
1223         }
1224         oidcpy(&nent->oid, oid);
1225         nent->tree_data = data;
1226         nent->tree_size = size;
1227         nent->ref = 1;
1228         if (!nent->temporary)
1229                 pbase_tree_cache[my_ix] = nent;
1230         return nent;
1231 }
1232
1233 static void pbase_tree_put(struct pbase_tree_cache *cache)
1234 {
1235         if (!cache->temporary) {
1236                 cache->ref--;
1237                 return;
1238         }
1239         free(cache->tree_data);
1240         free(cache);
1241 }
1242
1243 static int name_cmp_len(const char *name)
1244 {
1245         int i;
1246         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1247                 ;
1248         return i;
1249 }
1250
1251 static void add_pbase_object(struct tree_desc *tree,
1252                              const char *name,
1253                              int cmplen,
1254                              const char *fullname)
1255 {
1256         struct name_entry entry;
1257         int cmp;
1258
1259         while (tree_entry(tree,&entry)) {
1260                 if (S_ISGITLINK(entry.mode))
1261                         continue;
1262                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1263                       memcmp(name, entry.path, cmplen);
1264                 if (cmp > 0)
1265                         continue;
1266                 if (cmp < 0)
1267                         return;
1268                 if (name[cmplen] != '/') {
1269                         add_object_entry(entry.oid,
1270                                          object_type(entry.mode),
1271                                          fullname, 1);
1272                         return;
1273                 }
1274                 if (S_ISDIR(entry.mode)) {
1275                         struct tree_desc sub;
1276                         struct pbase_tree_cache *tree;
1277                         const char *down = name+cmplen+1;
1278                         int downlen = name_cmp_len(down);
1279
1280                         tree = pbase_tree_get(entry.oid);
1281                         if (!tree)
1282                                 return;
1283                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1284
1285                         add_pbase_object(&sub, down, downlen, fullname);
1286                         pbase_tree_put(tree);
1287                 }
1288         }
1289 }
1290
1291 static unsigned *done_pbase_paths;
1292 static int done_pbase_paths_num;
1293 static int done_pbase_paths_alloc;
1294 static int done_pbase_path_pos(unsigned hash)
1295 {
1296         int lo = 0;
1297         int hi = done_pbase_paths_num;
1298         while (lo < hi) {
1299                 int mi = lo + (hi - lo) / 2;
1300                 if (done_pbase_paths[mi] == hash)
1301                         return mi;
1302                 if (done_pbase_paths[mi] < hash)
1303                         hi = mi;
1304                 else
1305                         lo = mi + 1;
1306         }
1307         return -lo-1;
1308 }
1309
1310 static int check_pbase_path(unsigned hash)
1311 {
1312         int pos = done_pbase_path_pos(hash);
1313         if (0 <= pos)
1314                 return 1;
1315         pos = -pos - 1;
1316         ALLOC_GROW(done_pbase_paths,
1317                    done_pbase_paths_num + 1,
1318                    done_pbase_paths_alloc);
1319         done_pbase_paths_num++;
1320         if (pos < done_pbase_paths_num)
1321                 MOVE_ARRAY(done_pbase_paths + pos + 1, done_pbase_paths + pos,
1322                            done_pbase_paths_num - pos - 1);
1323         done_pbase_paths[pos] = hash;
1324         return 0;
1325 }
1326
1327 static void add_preferred_base_object(const char *name)
1328 {
1329         struct pbase_tree *it;
1330         int cmplen;
1331         unsigned hash = pack_name_hash(name);
1332
1333         if (!num_preferred_base || check_pbase_path(hash))
1334                 return;
1335
1336         cmplen = name_cmp_len(name);
1337         for (it = pbase_tree; it; it = it->next) {
1338                 if (cmplen == 0) {
1339                         add_object_entry(&it->pcache.oid, OBJ_TREE, NULL, 1);
1340                 }
1341                 else {
1342                         struct tree_desc tree;
1343                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1344                         add_pbase_object(&tree, name, cmplen, name);
1345                 }
1346         }
1347 }
1348
1349 static void add_preferred_base(struct object_id *oid)
1350 {
1351         struct pbase_tree *it;
1352         void *data;
1353         unsigned long size;
1354         struct object_id tree_oid;
1355
1356         if (window <= num_preferred_base++)
1357                 return;
1358
1359         data = read_object_with_reference(oid, tree_type, &size, &tree_oid);
1360         if (!data)
1361                 return;
1362
1363         for (it = pbase_tree; it; it = it->next) {
1364                 if (!oidcmp(&it->pcache.oid, &tree_oid)) {
1365                         free(data);
1366                         return;
1367                 }
1368         }
1369
1370         it = xcalloc(1, sizeof(*it));
1371         it->next = pbase_tree;
1372         pbase_tree = it;
1373
1374         oidcpy(&it->pcache.oid, &tree_oid);
1375         it->pcache.tree_data = data;
1376         it->pcache.tree_size = size;
1377 }
1378
1379 static void cleanup_preferred_base(void)
1380 {
1381         struct pbase_tree *it;
1382         unsigned i;
1383
1384         it = pbase_tree;
1385         pbase_tree = NULL;
1386         while (it) {
1387                 struct pbase_tree *tmp = it;
1388                 it = tmp->next;
1389                 free(tmp->pcache.tree_data);
1390                 free(tmp);
1391         }
1392
1393         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1394                 if (!pbase_tree_cache[i])
1395                         continue;
1396                 free(pbase_tree_cache[i]->tree_data);
1397                 FREE_AND_NULL(pbase_tree_cache[i]);
1398         }
1399
1400         FREE_AND_NULL(done_pbase_paths);
1401         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1402 }
1403
1404 static void check_object(struct object_entry *entry)
1405 {
1406         if (entry->in_pack) {
1407                 struct packed_git *p = entry->in_pack;
1408                 struct pack_window *w_curs = NULL;
1409                 const unsigned char *base_ref = NULL;
1410                 struct object_entry *base_entry;
1411                 unsigned long used, used_0;
1412                 unsigned long avail;
1413                 off_t ofs;
1414                 unsigned char *buf, c;
1415
1416                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1417
1418                 /*
1419                  * We want in_pack_type even if we do not reuse delta
1420                  * since non-delta representations could still be reused.
1421                  */
1422                 used = unpack_object_header_buffer(buf, avail,
1423                                                    &entry->in_pack_type,
1424                                                    &entry->size);
1425                 if (used == 0)
1426                         goto give_up;
1427
1428                 /*
1429                  * Determine if this is a delta and if so whether we can
1430                  * reuse it or not.  Otherwise let's find out as cheaply as
1431                  * possible what the actual type and size for this object is.
1432                  */
1433                 switch (entry->in_pack_type) {
1434                 default:
1435                         /* Not a delta hence we've already got all we need. */
1436                         entry->type = entry->in_pack_type;
1437                         entry->in_pack_header_size = used;
1438                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1439                                 goto give_up;
1440                         unuse_pack(&w_curs);
1441                         return;
1442                 case OBJ_REF_DELTA:
1443                         if (reuse_delta && !entry->preferred_base)
1444                                 base_ref = use_pack(p, &w_curs,
1445                                                 entry->in_pack_offset + used, NULL);
1446                         entry->in_pack_header_size = used + 20;
1447                         break;
1448                 case OBJ_OFS_DELTA:
1449                         buf = use_pack(p, &w_curs,
1450                                        entry->in_pack_offset + used, NULL);
1451                         used_0 = 0;
1452                         c = buf[used_0++];
1453                         ofs = c & 127;
1454                         while (c & 128) {
1455                                 ofs += 1;
1456                                 if (!ofs || MSB(ofs, 7)) {
1457                                         error("delta base offset overflow in pack for %s",
1458                                               oid_to_hex(&entry->idx.oid));
1459                                         goto give_up;
1460                                 }
1461                                 c = buf[used_0++];
1462                                 ofs = (ofs << 7) + (c & 127);
1463                         }
1464                         ofs = entry->in_pack_offset - ofs;
1465                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1466                                 error("delta base offset out of bound for %s",
1467                                       oid_to_hex(&entry->idx.oid));
1468                                 goto give_up;
1469                         }
1470                         if (reuse_delta && !entry->preferred_base) {
1471                                 struct revindex_entry *revidx;
1472                                 revidx = find_pack_revindex(p, ofs);
1473                                 if (!revidx)
1474                                         goto give_up;
1475                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1476                         }
1477                         entry->in_pack_header_size = used + used_0;
1478                         break;
1479                 }
1480
1481                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1482                         /*
1483                          * If base_ref was set above that means we wish to
1484                          * reuse delta data, and we even found that base
1485                          * in the list of objects we want to pack. Goodie!
1486                          *
1487                          * Depth value does not matter - find_deltas() will
1488                          * never consider reused delta as the base object to
1489                          * deltify other objects against, in order to avoid
1490                          * circular deltas.
1491                          */
1492                         entry->type = entry->in_pack_type;
1493                         entry->delta = base_entry;
1494                         entry->delta_size = entry->size;
1495                         entry->delta_sibling = base_entry->delta_child;
1496                         base_entry->delta_child = entry;
1497                         unuse_pack(&w_curs);
1498                         return;
1499                 }
1500
1501                 if (entry->type) {
1502                         /*
1503                          * This must be a delta and we already know what the
1504                          * final object type is.  Let's extract the actual
1505                          * object size from the delta header.
1506                          */
1507                         entry->size = get_size_from_delta(p, &w_curs,
1508                                         entry->in_pack_offset + entry->in_pack_header_size);
1509                         if (entry->size == 0)
1510                                 goto give_up;
1511                         unuse_pack(&w_curs);
1512                         return;
1513                 }
1514
1515                 /*
1516                  * No choice but to fall back to the recursive delta walk
1517                  * with sha1_object_info() to find about the object type
1518                  * at this point...
1519                  */
1520                 give_up:
1521                 unuse_pack(&w_curs);
1522         }
1523
1524         entry->type = oid_object_info(the_repository, &entry->idx.oid,
1525                                       &entry->size);
1526         /*
1527          * The error condition is checked in prepare_pack().  This is
1528          * to permit a missing preferred base object to be ignored
1529          * as a preferred base.  Doing so can result in a larger
1530          * pack file, but the transfer will still take place.
1531          */
1532 }
1533
1534 static int pack_offset_sort(const void *_a, const void *_b)
1535 {
1536         const struct object_entry *a = *(struct object_entry **)_a;
1537         const struct object_entry *b = *(struct object_entry **)_b;
1538
1539         /* avoid filesystem trashing with loose objects */
1540         if (!a->in_pack && !b->in_pack)
1541                 return oidcmp(&a->idx.oid, &b->idx.oid);
1542
1543         if (a->in_pack < b->in_pack)
1544                 return -1;
1545         if (a->in_pack > b->in_pack)
1546                 return 1;
1547         return a->in_pack_offset < b->in_pack_offset ? -1 :
1548                         (a->in_pack_offset > b->in_pack_offset);
1549 }
1550
1551 /*
1552  * Drop an on-disk delta we were planning to reuse. Naively, this would
1553  * just involve blanking out the "delta" field, but we have to deal
1554  * with some extra book-keeping:
1555  *
1556  *   1. Removing ourselves from the delta_sibling linked list.
1557  *
1558  *   2. Updating our size/type to the non-delta representation. These were
1559  *      either not recorded initially (size) or overwritten with the delta type
1560  *      (type) when check_object() decided to reuse the delta.
1561  *
1562  *   3. Resetting our delta depth, as we are now a base object.
1563  */
1564 static void drop_reused_delta(struct object_entry *entry)
1565 {
1566         struct object_entry **p = &entry->delta->delta_child;
1567         struct object_info oi = OBJECT_INFO_INIT;
1568
1569         while (*p) {
1570                 if (*p == entry)
1571                         *p = (*p)->delta_sibling;
1572                 else
1573                         p = &(*p)->delta_sibling;
1574         }
1575         entry->delta = NULL;
1576         entry->depth = 0;
1577
1578         oi.sizep = &entry->size;
1579         oi.typep = &entry->type;
1580         if (packed_object_info(the_repository, entry->in_pack,
1581                                entry->in_pack_offset, &oi) < 0) {
1582                 /*
1583                  * We failed to get the info from this pack for some reason;
1584                  * fall back to sha1_object_info, which may find another copy.
1585                  * And if that fails, the error will be recorded in entry->type
1586                  * and dealt with in prepare_pack().
1587                  */
1588                 entry->type = oid_object_info(the_repository, &entry->idx.oid,
1589                                               &entry->size);
1590         }
1591 }
1592
1593 /*
1594  * Follow the chain of deltas from this entry onward, throwing away any links
1595  * that cause us to hit a cycle (as determined by the DFS state flags in
1596  * the entries).
1597  *
1598  * We also detect too-long reused chains that would violate our --depth
1599  * limit.
1600  */
1601 static void break_delta_chains(struct object_entry *entry)
1602 {
1603         /*
1604          * The actual depth of each object we will write is stored as an int,
1605          * as it cannot exceed our int "depth" limit. But before we break
1606          * changes based no that limit, we may potentially go as deep as the
1607          * number of objects, which is elsewhere bounded to a uint32_t.
1608          */
1609         uint32_t total_depth;
1610         struct object_entry *cur, *next;
1611
1612         for (cur = entry, total_depth = 0;
1613              cur;
1614              cur = cur->delta, total_depth++) {
1615                 if (cur->dfs_state == DFS_DONE) {
1616                         /*
1617                          * We've already seen this object and know it isn't
1618                          * part of a cycle. We do need to append its depth
1619                          * to our count.
1620                          */
1621                         total_depth += cur->depth;
1622                         break;
1623                 }
1624
1625                 /*
1626                  * We break cycles before looping, so an ACTIVE state (or any
1627                  * other cruft which made its way into the state variable)
1628                  * is a bug.
1629                  */
1630                 if (cur->dfs_state != DFS_NONE)
1631                         die("BUG: confusing delta dfs state in first pass: %d",
1632                             cur->dfs_state);
1633
1634                 /*
1635                  * Now we know this is the first time we've seen the object. If
1636                  * it's not a delta, we're done traversing, but we'll mark it
1637                  * done to save time on future traversals.
1638                  */
1639                 if (!cur->delta) {
1640                         cur->dfs_state = DFS_DONE;
1641                         break;
1642                 }
1643
1644                 /*
1645                  * Mark ourselves as active and see if the next step causes
1646                  * us to cycle to another active object. It's important to do
1647                  * this _before_ we loop, because it impacts where we make the
1648                  * cut, and thus how our total_depth counter works.
1649                  * E.g., We may see a partial loop like:
1650                  *
1651                  *   A -> B -> C -> D -> B
1652                  *
1653                  * Cutting B->C breaks the cycle. But now the depth of A is
1654                  * only 1, and our total_depth counter is at 3. The size of the
1655                  * error is always one less than the size of the cycle we
1656                  * broke. Commits C and D were "lost" from A's chain.
1657                  *
1658                  * If we instead cut D->B, then the depth of A is correct at 3.
1659                  * We keep all commits in the chain that we examined.
1660                  */
1661                 cur->dfs_state = DFS_ACTIVE;
1662                 if (cur->delta->dfs_state == DFS_ACTIVE) {
1663                         drop_reused_delta(cur);
1664                         cur->dfs_state = DFS_DONE;
1665                         break;
1666                 }
1667         }
1668
1669         /*
1670          * And now that we've gone all the way to the bottom of the chain, we
1671          * need to clear the active flags and set the depth fields as
1672          * appropriate. Unlike the loop above, which can quit when it drops a
1673          * delta, we need to keep going to look for more depth cuts. So we need
1674          * an extra "next" pointer to keep going after we reset cur->delta.
1675          */
1676         for (cur = entry; cur; cur = next) {
1677                 next = cur->delta;
1678
1679                 /*
1680                  * We should have a chain of zero or more ACTIVE states down to
1681                  * a final DONE. We can quit after the DONE, because either it
1682                  * has no bases, or we've already handled them in a previous
1683                  * call.
1684                  */
1685                 if (cur->dfs_state == DFS_DONE)
1686                         break;
1687                 else if (cur->dfs_state != DFS_ACTIVE)
1688                         die("BUG: confusing delta dfs state in second pass: %d",
1689                             cur->dfs_state);
1690
1691                 /*
1692                  * If the total_depth is more than depth, then we need to snip
1693                  * the chain into two or more smaller chains that don't exceed
1694                  * the maximum depth. Most of the resulting chains will contain
1695                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1696                  * the last chain (i.e., the one containing entry) will contain
1697                  * whatever entries are left over, namely
1698                  * (total_depth % (depth + 1)) of them.
1699                  *
1700                  * Since we are iterating towards decreasing depth, we need to
1701                  * decrement total_depth as we go, and we need to write to the
1702                  * entry what its final depth will be after all of the
1703                  * snipping. Since we're snipping into chains of length (depth
1704                  * + 1) entries, the final depth of an entry will be its
1705                  * original depth modulo (depth + 1). Any time we encounter an
1706                  * entry whose final depth is supposed to be zero, we snip it
1707                  * from its delta base, thereby making it so.
1708                  */
1709                 cur->depth = (total_depth--) % (depth + 1);
1710                 if (!cur->depth)
1711                         drop_reused_delta(cur);
1712
1713                 cur->dfs_state = DFS_DONE;
1714         }
1715 }
1716
1717 static void get_object_details(void)
1718 {
1719         uint32_t i;
1720         struct object_entry **sorted_by_offset;
1721
1722         if (progress)
1723                 progress_state = start_progress(_("Counting objects"),
1724                                                 to_pack.nr_objects);
1725
1726         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1727         for (i = 0; i < to_pack.nr_objects; i++)
1728                 sorted_by_offset[i] = to_pack.objects + i;
1729         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1730
1731         for (i = 0; i < to_pack.nr_objects; i++) {
1732                 struct object_entry *entry = sorted_by_offset[i];
1733                 check_object(entry);
1734                 if (big_file_threshold < entry->size)
1735                         entry->no_try_delta = 1;
1736                 display_progress(progress_state, i + 1);
1737         }
1738         stop_progress(&progress_state);
1739
1740         /*
1741          * This must happen in a second pass, since we rely on the delta
1742          * information for the whole list being completed.
1743          */
1744         for (i = 0; i < to_pack.nr_objects; i++)
1745                 break_delta_chains(&to_pack.objects[i]);
1746
1747         free(sorted_by_offset);
1748 }
1749
1750 /*
1751  * We search for deltas in a list sorted by type, by filename hash, and then
1752  * by size, so that we see progressively smaller and smaller files.
1753  * That's because we prefer deltas to be from the bigger file
1754  * to the smaller -- deletes are potentially cheaper, but perhaps
1755  * more importantly, the bigger file is likely the more recent
1756  * one.  The deepest deltas are therefore the oldest objects which are
1757  * less susceptible to be accessed often.
1758  */
1759 static int type_size_sort(const void *_a, const void *_b)
1760 {
1761         const struct object_entry *a = *(struct object_entry **)_a;
1762         const struct object_entry *b = *(struct object_entry **)_b;
1763
1764         if (a->type > b->type)
1765                 return -1;
1766         if (a->type < b->type)
1767                 return 1;
1768         if (a->hash > b->hash)
1769                 return -1;
1770         if (a->hash < b->hash)
1771                 return 1;
1772         if (a->preferred_base > b->preferred_base)
1773                 return -1;
1774         if (a->preferred_base < b->preferred_base)
1775                 return 1;
1776         if (a->size > b->size)
1777                 return -1;
1778         if (a->size < b->size)
1779                 return 1;
1780         return a < b ? -1 : (a > b);  /* newest first */
1781 }
1782
1783 struct unpacked {
1784         struct object_entry *entry;
1785         void *data;
1786         struct delta_index *index;
1787         unsigned depth;
1788 };
1789
1790 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1791                            unsigned long delta_size)
1792 {
1793         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1794                 return 0;
1795
1796         if (delta_size < cache_max_small_delta_size)
1797                 return 1;
1798
1799         /* cache delta, if objects are large enough compared to delta size */
1800         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1801                 return 1;
1802
1803         return 0;
1804 }
1805
1806 #ifndef NO_PTHREADS
1807
1808 static pthread_mutex_t read_mutex;
1809 #define read_lock()             pthread_mutex_lock(&read_mutex)
1810 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1811
1812 static pthread_mutex_t cache_mutex;
1813 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1814 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1815
1816 static pthread_mutex_t progress_mutex;
1817 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1818 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1819
1820 #else
1821
1822 #define read_lock()             (void)0
1823 #define read_unlock()           (void)0
1824 #define cache_lock()            (void)0
1825 #define cache_unlock()          (void)0
1826 #define progress_lock()         (void)0
1827 #define progress_unlock()       (void)0
1828
1829 #endif
1830
1831 static int try_delta(struct unpacked *trg, struct unpacked *src,
1832                      unsigned max_depth, unsigned long *mem_usage)
1833 {
1834         struct object_entry *trg_entry = trg->entry;
1835         struct object_entry *src_entry = src->entry;
1836         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1837         unsigned ref_depth;
1838         enum object_type type;
1839         void *delta_buf;
1840
1841         /* Don't bother doing diffs between different types */
1842         if (trg_entry->type != src_entry->type)
1843                 return -1;
1844
1845         /*
1846          * We do not bother to try a delta that we discarded on an
1847          * earlier try, but only when reusing delta data.  Note that
1848          * src_entry that is marked as the preferred_base should always
1849          * be considered, as even if we produce a suboptimal delta against
1850          * it, we will still save the transfer cost, as we already know
1851          * the other side has it and we won't send src_entry at all.
1852          */
1853         if (reuse_delta && trg_entry->in_pack &&
1854             trg_entry->in_pack == src_entry->in_pack &&
1855             !src_entry->preferred_base &&
1856             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1857             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1858                 return 0;
1859
1860         /* Let's not bust the allowed depth. */
1861         if (src->depth >= max_depth)
1862                 return 0;
1863
1864         /* Now some size filtering heuristics. */
1865         trg_size = trg_entry->size;
1866         if (!trg_entry->delta) {
1867                 max_size = trg_size/2 - 20;
1868                 ref_depth = 1;
1869         } else {
1870                 max_size = trg_entry->delta_size;
1871                 ref_depth = trg->depth;
1872         }
1873         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1874                                                 (max_depth - ref_depth + 1);
1875         if (max_size == 0)
1876                 return 0;
1877         src_size = src_entry->size;
1878         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1879         if (sizediff >= max_size)
1880                 return 0;
1881         if (trg_size < src_size / 32)
1882                 return 0;
1883
1884         /* Load data if not already done */
1885         if (!trg->data) {
1886                 read_lock();
1887                 trg->data = read_object_file(&trg_entry->idx.oid, &type, &sz);
1888                 read_unlock();
1889                 if (!trg->data)
1890                         die("object %s cannot be read",
1891                             oid_to_hex(&trg_entry->idx.oid));
1892                 if (sz != trg_size)
1893                         die("object %s inconsistent object length (%lu vs %lu)",
1894                             oid_to_hex(&trg_entry->idx.oid), sz,
1895                             trg_size);
1896                 *mem_usage += sz;
1897         }
1898         if (!src->data) {
1899                 read_lock();
1900                 src->data = read_object_file(&src_entry->idx.oid, &type, &sz);
1901                 read_unlock();
1902                 if (!src->data) {
1903                         if (src_entry->preferred_base) {
1904                                 static int warned = 0;
1905                                 if (!warned++)
1906                                         warning("object %s cannot be read",
1907                                                 oid_to_hex(&src_entry->idx.oid));
1908                                 /*
1909                                  * Those objects are not included in the
1910                                  * resulting pack.  Be resilient and ignore
1911                                  * them if they can't be read, in case the
1912                                  * pack could be created nevertheless.
1913                                  */
1914                                 return 0;
1915                         }
1916                         die("object %s cannot be read",
1917                             oid_to_hex(&src_entry->idx.oid));
1918                 }
1919                 if (sz != src_size)
1920                         die("object %s inconsistent object length (%lu vs %lu)",
1921                             oid_to_hex(&src_entry->idx.oid), sz,
1922                             src_size);
1923                 *mem_usage += sz;
1924         }
1925         if (!src->index) {
1926                 src->index = create_delta_index(src->data, src_size);
1927                 if (!src->index) {
1928                         static int warned = 0;
1929                         if (!warned++)
1930                                 warning("suboptimal pack - out of memory");
1931                         return 0;
1932                 }
1933                 *mem_usage += sizeof_delta_index(src->index);
1934         }
1935
1936         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1937         if (!delta_buf)
1938                 return 0;
1939
1940         if (trg_entry->delta) {
1941                 /* Prefer only shallower same-sized deltas. */
1942                 if (delta_size == trg_entry->delta_size &&
1943                     src->depth + 1 >= trg->depth) {
1944                         free(delta_buf);
1945                         return 0;
1946                 }
1947         }
1948
1949         /*
1950          * Handle memory allocation outside of the cache
1951          * accounting lock.  Compiler will optimize the strangeness
1952          * away when NO_PTHREADS is defined.
1953          */
1954         free(trg_entry->delta_data);
1955         cache_lock();
1956         if (trg_entry->delta_data) {
1957                 delta_cache_size -= trg_entry->delta_size;
1958                 trg_entry->delta_data = NULL;
1959         }
1960         if (delta_cacheable(src_size, trg_size, delta_size)) {
1961                 delta_cache_size += delta_size;
1962                 cache_unlock();
1963                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1964         } else {
1965                 cache_unlock();
1966                 free(delta_buf);
1967         }
1968
1969         trg_entry->delta = src_entry;
1970         trg_entry->delta_size = delta_size;
1971         trg->depth = src->depth + 1;
1972
1973         return 1;
1974 }
1975
1976 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1977 {
1978         struct object_entry *child = me->delta_child;
1979         unsigned int m = n;
1980         while (child) {
1981                 unsigned int c = check_delta_limit(child, n + 1);
1982                 if (m < c)
1983                         m = c;
1984                 child = child->delta_sibling;
1985         }
1986         return m;
1987 }
1988
1989 static unsigned long free_unpacked(struct unpacked *n)
1990 {
1991         unsigned long freed_mem = sizeof_delta_index(n->index);
1992         free_delta_index(n->index);
1993         n->index = NULL;
1994         if (n->data) {
1995                 freed_mem += n->entry->size;
1996                 FREE_AND_NULL(n->data);
1997         }
1998         n->entry = NULL;
1999         n->depth = 0;
2000         return freed_mem;
2001 }
2002
2003 static void find_deltas(struct object_entry **list, unsigned *list_size,
2004                         int window, int depth, unsigned *processed)
2005 {
2006         uint32_t i, idx = 0, count = 0;
2007         struct unpacked *array;
2008         unsigned long mem_usage = 0;
2009
2010         array = xcalloc(window, sizeof(struct unpacked));
2011
2012         for (;;) {
2013                 struct object_entry *entry;
2014                 struct unpacked *n = array + idx;
2015                 int j, max_depth, best_base = -1;
2016
2017                 progress_lock();
2018                 if (!*list_size) {
2019                         progress_unlock();
2020                         break;
2021                 }
2022                 entry = *list++;
2023                 (*list_size)--;
2024                 if (!entry->preferred_base) {
2025                         (*processed)++;
2026                         display_progress(progress_state, *processed);
2027                 }
2028                 progress_unlock();
2029
2030                 mem_usage -= free_unpacked(n);
2031                 n->entry = entry;
2032
2033                 while (window_memory_limit &&
2034                        mem_usage > window_memory_limit &&
2035                        count > 1) {
2036                         uint32_t tail = (idx + window - count) % window;
2037                         mem_usage -= free_unpacked(array + tail);
2038                         count--;
2039                 }
2040
2041                 /* We do not compute delta to *create* objects we are not
2042                  * going to pack.
2043                  */
2044                 if (entry->preferred_base)
2045                         goto next;
2046
2047                 /*
2048                  * If the current object is at pack edge, take the depth the
2049                  * objects that depend on the current object into account
2050                  * otherwise they would become too deep.
2051                  */
2052                 max_depth = depth;
2053                 if (entry->delta_child) {
2054                         max_depth -= check_delta_limit(entry, 0);
2055                         if (max_depth <= 0)
2056                                 goto next;
2057                 }
2058
2059                 j = window;
2060                 while (--j > 0) {
2061                         int ret;
2062                         uint32_t other_idx = idx + j;
2063                         struct unpacked *m;
2064                         if (other_idx >= window)
2065                                 other_idx -= window;
2066                         m = array + other_idx;
2067                         if (!m->entry)
2068                                 break;
2069                         ret = try_delta(n, m, max_depth, &mem_usage);
2070                         if (ret < 0)
2071                                 break;
2072                         else if (ret > 0)
2073                                 best_base = other_idx;
2074                 }
2075
2076                 /*
2077                  * If we decided to cache the delta data, then it is best
2078                  * to compress it right away.  First because we have to do
2079                  * it anyway, and doing it here while we're threaded will
2080                  * save a lot of time in the non threaded write phase,
2081                  * as well as allow for caching more deltas within
2082                  * the same cache size limit.
2083                  * ...
2084                  * But only if not writing to stdout, since in that case
2085                  * the network is most likely throttling writes anyway,
2086                  * and therefore it is best to go to the write phase ASAP
2087                  * instead, as we can afford spending more time compressing
2088                  * between writes at that moment.
2089                  */
2090                 if (entry->delta_data && !pack_to_stdout) {
2091                         entry->z_delta_size = do_compress(&entry->delta_data,
2092                                                           entry->delta_size);
2093                         cache_lock();
2094                         delta_cache_size -= entry->delta_size;
2095                         delta_cache_size += entry->z_delta_size;
2096                         cache_unlock();
2097                 }
2098
2099                 /* if we made n a delta, and if n is already at max
2100                  * depth, leaving it in the window is pointless.  we
2101                  * should evict it first.
2102                  */
2103                 if (entry->delta && max_depth <= n->depth)
2104                         continue;
2105
2106                 /*
2107                  * Move the best delta base up in the window, after the
2108                  * currently deltified object, to keep it longer.  It will
2109                  * be the first base object to be attempted next.
2110                  */
2111                 if (entry->delta) {
2112                         struct unpacked swap = array[best_base];
2113                         int dist = (window + idx - best_base) % window;
2114                         int dst = best_base;
2115                         while (dist--) {
2116                                 int src = (dst + 1) % window;
2117                                 array[dst] = array[src];
2118                                 dst = src;
2119                         }
2120                         array[dst] = swap;
2121                 }
2122
2123                 next:
2124                 idx++;
2125                 if (count + 1 < window)
2126                         count++;
2127                 if (idx >= window)
2128                         idx = 0;
2129         }
2130
2131         for (i = 0; i < window; ++i) {
2132                 free_delta_index(array[i].index);
2133                 free(array[i].data);
2134         }
2135         free(array);
2136 }
2137
2138 #ifndef NO_PTHREADS
2139
2140 static void try_to_free_from_threads(size_t size)
2141 {
2142         read_lock();
2143         release_pack_memory(size);
2144         read_unlock();
2145 }
2146
2147 static try_to_free_t old_try_to_free_routine;
2148
2149 /*
2150  * The main thread waits on the condition that (at least) one of the workers
2151  * has stopped working (which is indicated in the .working member of
2152  * struct thread_params).
2153  * When a work thread has completed its work, it sets .working to 0 and
2154  * signals the main thread and waits on the condition that .data_ready
2155  * becomes 1.
2156  */
2157
2158 struct thread_params {
2159         pthread_t thread;
2160         struct object_entry **list;
2161         unsigned list_size;
2162         unsigned remaining;
2163         int window;
2164         int depth;
2165         int working;
2166         int data_ready;
2167         pthread_mutex_t mutex;
2168         pthread_cond_t cond;
2169         unsigned *processed;
2170 };
2171
2172 static pthread_cond_t progress_cond;
2173
2174 /*
2175  * Mutex and conditional variable can't be statically-initialized on Windows.
2176  */
2177 static void init_threaded_search(void)
2178 {
2179         init_recursive_mutex(&read_mutex);
2180         pthread_mutex_init(&cache_mutex, NULL);
2181         pthread_mutex_init(&progress_mutex, NULL);
2182         pthread_cond_init(&progress_cond, NULL);
2183         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2184 }
2185
2186 static void cleanup_threaded_search(void)
2187 {
2188         set_try_to_free_routine(old_try_to_free_routine);
2189         pthread_cond_destroy(&progress_cond);
2190         pthread_mutex_destroy(&read_mutex);
2191         pthread_mutex_destroy(&cache_mutex);
2192         pthread_mutex_destroy(&progress_mutex);
2193 }
2194
2195 static void *threaded_find_deltas(void *arg)
2196 {
2197         struct thread_params *me = arg;
2198
2199         progress_lock();
2200         while (me->remaining) {
2201                 progress_unlock();
2202
2203                 find_deltas(me->list, &me->remaining,
2204                             me->window, me->depth, me->processed);
2205
2206                 progress_lock();
2207                 me->working = 0;
2208                 pthread_cond_signal(&progress_cond);
2209                 progress_unlock();
2210
2211                 /*
2212                  * We must not set ->data_ready before we wait on the
2213                  * condition because the main thread may have set it to 1
2214                  * before we get here. In order to be sure that new
2215                  * work is available if we see 1 in ->data_ready, it
2216                  * was initialized to 0 before this thread was spawned
2217                  * and we reset it to 0 right away.
2218                  */
2219                 pthread_mutex_lock(&me->mutex);
2220                 while (!me->data_ready)
2221                         pthread_cond_wait(&me->cond, &me->mutex);
2222                 me->data_ready = 0;
2223                 pthread_mutex_unlock(&me->mutex);
2224
2225                 progress_lock();
2226         }
2227         progress_unlock();
2228         /* leave ->working 1 so that this doesn't get more work assigned */
2229         return NULL;
2230 }
2231
2232 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2233                            int window, int depth, unsigned *processed)
2234 {
2235         struct thread_params *p;
2236         int i, ret, active_threads = 0;
2237
2238         init_threaded_search();
2239
2240         if (delta_search_threads <= 1) {
2241                 find_deltas(list, &list_size, window, depth, processed);
2242                 cleanup_threaded_search();
2243                 return;
2244         }
2245         if (progress > pack_to_stdout)
2246                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2247                                 delta_search_threads);
2248         p = xcalloc(delta_search_threads, sizeof(*p));
2249
2250         /* Partition the work amongst work threads. */
2251         for (i = 0; i < delta_search_threads; i++) {
2252                 unsigned sub_size = list_size / (delta_search_threads - i);
2253
2254                 /* don't use too small segments or no deltas will be found */
2255                 if (sub_size < 2*window && i+1 < delta_search_threads)
2256                         sub_size = 0;
2257
2258                 p[i].window = window;
2259                 p[i].depth = depth;
2260                 p[i].processed = processed;
2261                 p[i].working = 1;
2262                 p[i].data_ready = 0;
2263
2264                 /* try to split chunks on "path" boundaries */
2265                 while (sub_size && sub_size < list_size &&
2266                        list[sub_size]->hash &&
2267                        list[sub_size]->hash == list[sub_size-1]->hash)
2268                         sub_size++;
2269
2270                 p[i].list = list;
2271                 p[i].list_size = sub_size;
2272                 p[i].remaining = sub_size;
2273
2274                 list += sub_size;
2275                 list_size -= sub_size;
2276         }
2277
2278         /* Start work threads. */
2279         for (i = 0; i < delta_search_threads; i++) {
2280                 if (!p[i].list_size)
2281                         continue;
2282                 pthread_mutex_init(&p[i].mutex, NULL);
2283                 pthread_cond_init(&p[i].cond, NULL);
2284                 ret = pthread_create(&p[i].thread, NULL,
2285                                      threaded_find_deltas, &p[i]);
2286                 if (ret)
2287                         die("unable to create thread: %s", strerror(ret));
2288                 active_threads++;
2289         }
2290
2291         /*
2292          * Now let's wait for work completion.  Each time a thread is done
2293          * with its work, we steal half of the remaining work from the
2294          * thread with the largest number of unprocessed objects and give
2295          * it to that newly idle thread.  This ensure good load balancing
2296          * until the remaining object list segments are simply too short
2297          * to be worth splitting anymore.
2298          */
2299         while (active_threads) {
2300                 struct thread_params *target = NULL;
2301                 struct thread_params *victim = NULL;
2302                 unsigned sub_size = 0;
2303
2304                 progress_lock();
2305                 for (;;) {
2306                         for (i = 0; !target && i < delta_search_threads; i++)
2307                                 if (!p[i].working)
2308                                         target = &p[i];
2309                         if (target)
2310                                 break;
2311                         pthread_cond_wait(&progress_cond, &progress_mutex);
2312                 }
2313
2314                 for (i = 0; i < delta_search_threads; i++)
2315                         if (p[i].remaining > 2*window &&
2316                             (!victim || victim->remaining < p[i].remaining))
2317                                 victim = &p[i];
2318                 if (victim) {
2319                         sub_size = victim->remaining / 2;
2320                         list = victim->list + victim->list_size - sub_size;
2321                         while (sub_size && list[0]->hash &&
2322                                list[0]->hash == list[-1]->hash) {
2323                                 list++;
2324                                 sub_size--;
2325                         }
2326                         if (!sub_size) {
2327                                 /*
2328                                  * It is possible for some "paths" to have
2329                                  * so many objects that no hash boundary
2330                                  * might be found.  Let's just steal the
2331                                  * exact half in that case.
2332                                  */
2333                                 sub_size = victim->remaining / 2;
2334                                 list -= sub_size;
2335                         }
2336                         target->list = list;
2337                         victim->list_size -= sub_size;
2338                         victim->remaining -= sub_size;
2339                 }
2340                 target->list_size = sub_size;
2341                 target->remaining = sub_size;
2342                 target->working = 1;
2343                 progress_unlock();
2344
2345                 pthread_mutex_lock(&target->mutex);
2346                 target->data_ready = 1;
2347                 pthread_cond_signal(&target->cond);
2348                 pthread_mutex_unlock(&target->mutex);
2349
2350                 if (!sub_size) {
2351                         pthread_join(target->thread, NULL);
2352                         pthread_cond_destroy(&target->cond);
2353                         pthread_mutex_destroy(&target->mutex);
2354                         active_threads--;
2355                 }
2356         }
2357         cleanup_threaded_search();
2358         free(p);
2359 }
2360
2361 #else
2362 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2363 #endif
2364
2365 static void add_tag_chain(const struct object_id *oid)
2366 {
2367         struct tag *tag;
2368
2369         /*
2370          * We catch duplicates already in add_object_entry(), but we'd
2371          * prefer to do this extra check to avoid having to parse the
2372          * tag at all if we already know that it's being packed (e.g., if
2373          * it was included via bitmaps, we would not have parsed it
2374          * previously).
2375          */
2376         if (packlist_find(&to_pack, oid->hash, NULL))
2377                 return;
2378
2379         tag = lookup_tag(oid);
2380         while (1) {
2381                 if (!tag || parse_tag(tag) || !tag->tagged)
2382                         die("unable to pack objects reachable from tag %s",
2383                             oid_to_hex(oid));
2384
2385                 add_object_entry(&tag->object.oid, OBJ_TAG, NULL, 0);
2386
2387                 if (tag->tagged->type != OBJ_TAG)
2388                         return;
2389
2390                 tag = (struct tag *)tag->tagged;
2391         }
2392 }
2393
2394 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2395 {
2396         struct object_id peeled;
2397
2398         if (starts_with(path, "refs/tags/") && /* is a tag? */
2399             !peel_ref(path, &peeled)    && /* peelable? */
2400             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2401                 add_tag_chain(oid);
2402         return 0;
2403 }
2404
2405 static void prepare_pack(int window, int depth)
2406 {
2407         struct object_entry **delta_list;
2408         uint32_t i, nr_deltas;
2409         unsigned n;
2410
2411         get_object_details();
2412
2413         /*
2414          * If we're locally repacking then we need to be doubly careful
2415          * from now on in order to make sure no stealth corruption gets
2416          * propagated to the new pack.  Clients receiving streamed packs
2417          * should validate everything they get anyway so no need to incur
2418          * the additional cost here in that case.
2419          */
2420         if (!pack_to_stdout)
2421                 do_check_packed_object_crc = 1;
2422
2423         if (!to_pack.nr_objects || !window || !depth)
2424                 return;
2425
2426         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2427         nr_deltas = n = 0;
2428
2429         for (i = 0; i < to_pack.nr_objects; i++) {
2430                 struct object_entry *entry = to_pack.objects + i;
2431
2432                 if (entry->delta)
2433                         /* This happens if we decided to reuse existing
2434                          * delta from a pack.  "reuse_delta &&" is implied.
2435                          */
2436                         continue;
2437
2438                 if (entry->size < 50)
2439                         continue;
2440
2441                 if (entry->no_try_delta)
2442                         continue;
2443
2444                 if (!entry->preferred_base) {
2445                         nr_deltas++;
2446                         if (entry->type < 0)
2447                                 die("unable to get type of object %s",
2448                                     oid_to_hex(&entry->idx.oid));
2449                 } else {
2450                         if (entry->type < 0) {
2451                                 /*
2452                                  * This object is not found, but we
2453                                  * don't have to include it anyway.
2454                                  */
2455                                 continue;
2456                         }
2457                 }
2458
2459                 delta_list[n++] = entry;
2460         }
2461
2462         if (nr_deltas && n > 1) {
2463                 unsigned nr_done = 0;
2464                 if (progress)
2465                         progress_state = start_progress(_("Compressing objects"),
2466                                                         nr_deltas);
2467                 QSORT(delta_list, n, type_size_sort);
2468                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2469                 stop_progress(&progress_state);
2470                 if (nr_done != nr_deltas)
2471                         die("inconsistency with delta count");
2472         }
2473         free(delta_list);
2474 }
2475
2476 static int git_pack_config(const char *k, const char *v, void *cb)
2477 {
2478         if (!strcmp(k, "pack.window")) {
2479                 window = git_config_int(k, v);
2480                 return 0;
2481         }
2482         if (!strcmp(k, "pack.windowmemory")) {
2483                 window_memory_limit = git_config_ulong(k, v);
2484                 return 0;
2485         }
2486         if (!strcmp(k, "pack.depth")) {
2487                 depth = git_config_int(k, v);
2488                 return 0;
2489         }
2490         if (!strcmp(k, "pack.deltacachesize")) {
2491                 max_delta_cache_size = git_config_int(k, v);
2492                 return 0;
2493         }
2494         if (!strcmp(k, "pack.deltacachelimit")) {
2495                 cache_max_small_delta_size = git_config_int(k, v);
2496                 return 0;
2497         }
2498         if (!strcmp(k, "pack.writebitmaphashcache")) {
2499                 if (git_config_bool(k, v))
2500                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2501                 else
2502                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2503         }
2504         if (!strcmp(k, "pack.usebitmaps")) {
2505                 use_bitmap_index_default = git_config_bool(k, v);
2506                 return 0;
2507         }
2508         if (!strcmp(k, "pack.threads")) {
2509                 delta_search_threads = git_config_int(k, v);
2510                 if (delta_search_threads < 0)
2511                         die("invalid number of threads specified (%d)",
2512                             delta_search_threads);
2513 #ifdef NO_PTHREADS
2514                 if (delta_search_threads != 1) {
2515                         warning("no threads support, ignoring %s", k);
2516                         delta_search_threads = 0;
2517                 }
2518 #endif
2519                 return 0;
2520         }
2521         if (!strcmp(k, "pack.indexversion")) {
2522                 pack_idx_opts.version = git_config_int(k, v);
2523                 if (pack_idx_opts.version > 2)
2524                         die("bad pack.indexversion=%"PRIu32,
2525                             pack_idx_opts.version);
2526                 return 0;
2527         }
2528         return git_default_config(k, v, cb);
2529 }
2530
2531 static void read_object_list_from_stdin(void)
2532 {
2533         char line[GIT_MAX_HEXSZ + 1 + PATH_MAX + 2];
2534         struct object_id oid;
2535         const char *p;
2536
2537         for (;;) {
2538                 if (!fgets(line, sizeof(line), stdin)) {
2539                         if (feof(stdin))
2540                                 break;
2541                         if (!ferror(stdin))
2542                                 die("fgets returned NULL, not EOF, not error!");
2543                         if (errno != EINTR)
2544                                 die_errno("fgets");
2545                         clearerr(stdin);
2546                         continue;
2547                 }
2548                 if (line[0] == '-') {
2549                         if (get_oid_hex(line+1, &oid))
2550                                 die("expected edge object ID, got garbage:\n %s",
2551                                     line);
2552                         add_preferred_base(&oid);
2553                         continue;
2554                 }
2555                 if (parse_oid_hex(line, &oid, &p))
2556                         die("expected object ID, got garbage:\n %s", line);
2557
2558                 add_preferred_base_object(p + 1);
2559                 add_object_entry(&oid, 0, p + 1, 0);
2560         }
2561 }
2562
2563 /* Remember to update object flag allocation in object.h */
2564 #define OBJECT_ADDED (1u<<20)
2565
2566 static void show_commit(struct commit *commit, void *data)
2567 {
2568         add_object_entry(&commit->object.oid, OBJ_COMMIT, NULL, 0);
2569         commit->object.flags |= OBJECT_ADDED;
2570
2571         if (write_bitmap_index)
2572                 index_commit_for_bitmap(commit);
2573 }
2574
2575 static void show_object(struct object *obj, const char *name, void *data)
2576 {
2577         add_preferred_base_object(name);
2578         add_object_entry(&obj->oid, obj->type, name, 0);
2579         obj->flags |= OBJECT_ADDED;
2580 }
2581
2582 static void show_object__ma_allow_any(struct object *obj, const char *name, void *data)
2583 {
2584         assert(arg_missing_action == MA_ALLOW_ANY);
2585
2586         /*
2587          * Quietly ignore ALL missing objects.  This avoids problems with
2588          * staging them now and getting an odd error later.
2589          */
2590         if (!has_object_file(&obj->oid))
2591                 return;
2592
2593         show_object(obj, name, data);
2594 }
2595
2596 static void show_object__ma_allow_promisor(struct object *obj, const char *name, void *data)
2597 {
2598         assert(arg_missing_action == MA_ALLOW_PROMISOR);
2599
2600         /*
2601          * Quietly ignore EXPECTED missing objects.  This avoids problems with
2602          * staging them now and getting an odd error later.
2603          */
2604         if (!has_object_file(&obj->oid) && is_promisor_object(&obj->oid))
2605                 return;
2606
2607         show_object(obj, name, data);
2608 }
2609
2610 static int option_parse_missing_action(const struct option *opt,
2611                                        const char *arg, int unset)
2612 {
2613         assert(arg);
2614         assert(!unset);
2615
2616         if (!strcmp(arg, "error")) {
2617                 arg_missing_action = MA_ERROR;
2618                 fn_show_object = show_object;
2619                 return 0;
2620         }
2621
2622         if (!strcmp(arg, "allow-any")) {
2623                 arg_missing_action = MA_ALLOW_ANY;
2624                 fetch_if_missing = 0;
2625                 fn_show_object = show_object__ma_allow_any;
2626                 return 0;
2627         }
2628
2629         if (!strcmp(arg, "allow-promisor")) {
2630                 arg_missing_action = MA_ALLOW_PROMISOR;
2631                 fetch_if_missing = 0;
2632                 fn_show_object = show_object__ma_allow_promisor;
2633                 return 0;
2634         }
2635
2636         die(_("invalid value for --missing"));
2637         return 0;
2638 }
2639
2640 static void show_edge(struct commit *commit)
2641 {
2642         add_preferred_base(&commit->object.oid);
2643 }
2644
2645 struct in_pack_object {
2646         off_t offset;
2647         struct object *object;
2648 };
2649
2650 struct in_pack {
2651         unsigned int alloc;
2652         unsigned int nr;
2653         struct in_pack_object *array;
2654 };
2655
2656 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2657 {
2658         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2659         in_pack->array[in_pack->nr].object = object;
2660         in_pack->nr++;
2661 }
2662
2663 /*
2664  * Compare the objects in the offset order, in order to emulate the
2665  * "git rev-list --objects" output that produced the pack originally.
2666  */
2667 static int ofscmp(const void *a_, const void *b_)
2668 {
2669         struct in_pack_object *a = (struct in_pack_object *)a_;
2670         struct in_pack_object *b = (struct in_pack_object *)b_;
2671
2672         if (a->offset < b->offset)
2673                 return -1;
2674         else if (a->offset > b->offset)
2675                 return 1;
2676         else
2677                 return oidcmp(&a->object->oid, &b->object->oid);
2678 }
2679
2680 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2681 {
2682         struct packed_git *p;
2683         struct in_pack in_pack;
2684         uint32_t i;
2685
2686         memset(&in_pack, 0, sizeof(in_pack));
2687
2688         for (p = get_packed_git(the_repository); p; p = p->next) {
2689                 struct object_id oid;
2690                 struct object *o;
2691
2692                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
2693                         continue;
2694                 if (open_pack_index(p))
2695                         die("cannot open pack index");
2696
2697                 ALLOC_GROW(in_pack.array,
2698                            in_pack.nr + p->num_objects,
2699                            in_pack.alloc);
2700
2701                 for (i = 0; i < p->num_objects; i++) {
2702                         nth_packed_object_oid(&oid, p, i);
2703                         o = lookup_unknown_object(oid.hash);
2704                         if (!(o->flags & OBJECT_ADDED))
2705                                 mark_in_pack_object(o, p, &in_pack);
2706                         o->flags |= OBJECT_ADDED;
2707                 }
2708         }
2709
2710         if (in_pack.nr) {
2711                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2712                 for (i = 0; i < in_pack.nr; i++) {
2713                         struct object *o = in_pack.array[i].object;
2714                         add_object_entry(&o->oid, o->type, "", 0);
2715                 }
2716         }
2717         free(in_pack.array);
2718 }
2719
2720 static int add_loose_object(const struct object_id *oid, const char *path,
2721                             void *data)
2722 {
2723         enum object_type type = oid_object_info(the_repository, oid, NULL);
2724
2725         if (type < 0) {
2726                 warning("loose object at %s could not be examined", path);
2727                 return 0;
2728         }
2729
2730         add_object_entry(oid, type, "", 0);
2731         return 0;
2732 }
2733
2734 /*
2735  * We actually don't even have to worry about reachability here.
2736  * add_object_entry will weed out duplicates, so we just add every
2737  * loose object we find.
2738  */
2739 static void add_unreachable_loose_objects(void)
2740 {
2741         for_each_loose_file_in_objdir(get_object_directory(),
2742                                       add_loose_object,
2743                                       NULL, NULL, NULL);
2744 }
2745
2746 static int has_sha1_pack_kept_or_nonlocal(const struct object_id *oid)
2747 {
2748         static struct packed_git *last_found = (void *)1;
2749         struct packed_git *p;
2750
2751         p = (last_found != (void *)1) ? last_found :
2752                                         get_packed_git(the_repository);
2753
2754         while (p) {
2755                 if ((!p->pack_local || p->pack_keep ||
2756                                 p->pack_keep_in_core) &&
2757                         find_pack_entry_one(oid->hash, p)) {
2758                         last_found = p;
2759                         return 1;
2760                 }
2761                 if (p == last_found)
2762                         p = get_packed_git(the_repository);
2763                 else
2764                         p = p->next;
2765                 if (p == last_found)
2766                         p = p->next;
2767         }
2768         return 0;
2769 }
2770
2771 /*
2772  * Store a list of sha1s that are should not be discarded
2773  * because they are either written too recently, or are
2774  * reachable from another object that was.
2775  *
2776  * This is filled by get_object_list.
2777  */
2778 static struct oid_array recent_objects;
2779
2780 static int loosened_object_can_be_discarded(const struct object_id *oid,
2781                                             timestamp_t mtime)
2782 {
2783         if (!unpack_unreachable_expiration)
2784                 return 0;
2785         if (mtime > unpack_unreachable_expiration)
2786                 return 0;
2787         if (oid_array_lookup(&recent_objects, oid) >= 0)
2788                 return 0;
2789         return 1;
2790 }
2791
2792 static void loosen_unused_packed_objects(struct rev_info *revs)
2793 {
2794         struct packed_git *p;
2795         uint32_t i;
2796         struct object_id oid;
2797
2798         for (p = get_packed_git(the_repository); p; p = p->next) {
2799                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
2800                         continue;
2801
2802                 if (open_pack_index(p))
2803                         die("cannot open pack index");
2804
2805                 for (i = 0; i < p->num_objects; i++) {
2806                         nth_packed_object_oid(&oid, p, i);
2807                         if (!packlist_find(&to_pack, oid.hash, NULL) &&
2808                             !has_sha1_pack_kept_or_nonlocal(&oid) &&
2809                             !loosened_object_can_be_discarded(&oid, p->mtime))
2810                                 if (force_object_loose(&oid, p->mtime))
2811                                         die("unable to force loose object");
2812                 }
2813         }
2814 }
2815
2816 /*
2817  * This tracks any options which pack-reuse code expects to be on, or which a
2818  * reader of the pack might not understand, and which would therefore prevent
2819  * blind reuse of what we have on disk.
2820  */
2821 static int pack_options_allow_reuse(void)
2822 {
2823         return pack_to_stdout &&
2824                allow_ofs_delta &&
2825                !ignore_packed_keep_on_disk &&
2826                !ignore_packed_keep_in_core &&
2827                (!local || !have_non_local_packs) &&
2828                !incremental;
2829 }
2830
2831 static int get_object_list_from_bitmap(struct rev_info *revs)
2832 {
2833         if (prepare_bitmap_walk(revs) < 0)
2834                 return -1;
2835
2836         if (pack_options_allow_reuse() &&
2837             !reuse_partial_packfile_from_bitmap(
2838                         &reuse_packfile,
2839                         &reuse_packfile_objects,
2840                         &reuse_packfile_offset)) {
2841                 assert(reuse_packfile_objects);
2842                 nr_result += reuse_packfile_objects;
2843                 display_progress(progress_state, nr_result);
2844         }
2845
2846         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2847         return 0;
2848 }
2849
2850 static void record_recent_object(struct object *obj,
2851                                  const char *name,
2852                                  void *data)
2853 {
2854         oid_array_append(&recent_objects, &obj->oid);
2855 }
2856
2857 static void record_recent_commit(struct commit *commit, void *data)
2858 {
2859         oid_array_append(&recent_objects, &commit->object.oid);
2860 }
2861
2862 static void get_object_list(int ac, const char **av)
2863 {
2864         struct rev_info revs;
2865         char line[1000];
2866         int flags = 0;
2867
2868         init_revisions(&revs, NULL);
2869         save_commit_buffer = 0;
2870         setup_revisions(ac, av, &revs, NULL);
2871
2872         /* make sure shallows are read */
2873         is_repository_shallow();
2874
2875         while (fgets(line, sizeof(line), stdin) != NULL) {
2876                 int len = strlen(line);
2877                 if (len && line[len - 1] == '\n')
2878                         line[--len] = 0;
2879                 if (!len)
2880                         break;
2881                 if (*line == '-') {
2882                         if (!strcmp(line, "--not")) {
2883                                 flags ^= UNINTERESTING;
2884                                 write_bitmap_index = 0;
2885                                 continue;
2886                         }
2887                         if (starts_with(line, "--shallow ")) {
2888                                 struct object_id oid;
2889                                 if (get_oid_hex(line + 10, &oid))
2890                                         die("not an SHA-1 '%s'", line + 10);
2891                                 register_shallow(&oid);
2892                                 use_bitmap_index = 0;
2893                                 continue;
2894                         }
2895                         die("not a rev '%s'", line);
2896                 }
2897                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2898                         die("bad revision '%s'", line);
2899         }
2900
2901         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2902                 return;
2903
2904         if (prepare_revision_walk(&revs))
2905                 die("revision walk setup failed");
2906         mark_edges_uninteresting(&revs, show_edge);
2907
2908         if (!fn_show_object)
2909                 fn_show_object = show_object;
2910         traverse_commit_list_filtered(&filter_options, &revs,
2911                                       show_commit, fn_show_object, NULL,
2912                                       NULL);
2913
2914         if (unpack_unreachable_expiration) {
2915                 revs.ignore_missing_links = 1;
2916                 if (add_unseen_recent_objects_to_traversal(&revs,
2917                                 unpack_unreachable_expiration))
2918                         die("unable to add recent objects");
2919                 if (prepare_revision_walk(&revs))
2920                         die("revision walk setup failed");
2921                 traverse_commit_list(&revs, record_recent_commit,
2922                                      record_recent_object, NULL);
2923         }
2924
2925         if (keep_unreachable)
2926                 add_objects_in_unpacked_packs(&revs);
2927         if (pack_loose_unreachable)
2928                 add_unreachable_loose_objects();
2929         if (unpack_unreachable)
2930                 loosen_unused_packed_objects(&revs);
2931
2932         oid_array_clear(&recent_objects);
2933 }
2934
2935 static void add_extra_kept_packs(const struct string_list *names)
2936 {
2937         struct packed_git *p;
2938
2939         if (!names->nr)
2940                 return;
2941
2942         for (p = get_packed_git(the_repository); p; p = p->next) {
2943                 const char *name = basename(p->pack_name);
2944                 int i;
2945
2946                 if (!p->pack_local)
2947                         continue;
2948
2949                 for (i = 0; i < names->nr; i++)
2950                         if (!fspathcmp(name, names->items[i].string))
2951                                 break;
2952
2953                 if (i < names->nr) {
2954                         p->pack_keep_in_core = 1;
2955                         ignore_packed_keep_in_core = 1;
2956                         continue;
2957                 }
2958         }
2959 }
2960
2961 static int option_parse_index_version(const struct option *opt,
2962                                       const char *arg, int unset)
2963 {
2964         char *c;
2965         const char *val = arg;
2966         pack_idx_opts.version = strtoul(val, &c, 10);
2967         if (pack_idx_opts.version > 2)
2968                 die(_("unsupported index version %s"), val);
2969         if (*c == ',' && c[1])
2970                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2971         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2972                 die(_("bad index version '%s'"), val);
2973         return 0;
2974 }
2975
2976 static int option_parse_unpack_unreachable(const struct option *opt,
2977                                            const char *arg, int unset)
2978 {
2979         if (unset) {
2980                 unpack_unreachable = 0;
2981                 unpack_unreachable_expiration = 0;
2982         }
2983         else {
2984                 unpack_unreachable = 1;
2985                 if (arg)
2986                         unpack_unreachable_expiration = approxidate(arg);
2987         }
2988         return 0;
2989 }
2990
2991 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2992 {
2993         int use_internal_rev_list = 0;
2994         int thin = 0;
2995         int shallow = 0;
2996         int all_progress_implied = 0;
2997         struct argv_array rp = ARGV_ARRAY_INIT;
2998         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2999         int rev_list_index = 0;
3000         struct string_list keep_pack_list = STRING_LIST_INIT_NODUP;
3001         struct option pack_objects_options[] = {
3002                 OPT_SET_INT('q', "quiet", &progress,
3003                             N_("do not show progress meter"), 0),
3004                 OPT_SET_INT(0, "progress", &progress,
3005                             N_("show progress meter"), 1),
3006                 OPT_SET_INT(0, "all-progress", &progress,
3007                             N_("show progress meter during object writing phase"), 2),
3008                 OPT_BOOL(0, "all-progress-implied",
3009                          &all_progress_implied,
3010                          N_("similar to --all-progress when progress meter is shown")),
3011                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
3012                   N_("write the pack index file in the specified idx format version"),
3013                   0, option_parse_index_version },
3014                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
3015                               N_("maximum size of each output pack file")),
3016                 OPT_BOOL(0, "local", &local,
3017                          N_("ignore borrowed objects from alternate object store")),
3018                 OPT_BOOL(0, "incremental", &incremental,
3019                          N_("ignore packed objects")),
3020                 OPT_INTEGER(0, "window", &window,
3021                             N_("limit pack window by objects")),
3022                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
3023                               N_("limit pack window by memory in addition to object limit")),
3024                 OPT_INTEGER(0, "depth", &depth,
3025                             N_("maximum length of delta chain allowed in the resulting pack")),
3026                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
3027                          N_("reuse existing deltas")),
3028                 OPT_BOOL(0, "reuse-object", &reuse_object,
3029                          N_("reuse existing objects")),
3030                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
3031                          N_("use OFS_DELTA objects")),
3032                 OPT_INTEGER(0, "threads", &delta_search_threads,
3033                             N_("use threads when searching for best delta matches")),
3034                 OPT_BOOL(0, "non-empty", &non_empty,
3035                          N_("do not create an empty pack output")),
3036                 OPT_BOOL(0, "revs", &use_internal_rev_list,
3037                          N_("read revision arguments from standard input")),
3038                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
3039                   N_("limit the objects to those that are not yet packed"),
3040                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
3041                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
3042                   N_("include objects reachable from any reference"),
3043                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
3044                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
3045                   N_("include objects referred by reflog entries"),
3046                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
3047                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
3048                   N_("include objects referred to by the index"),
3049                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
3050                 OPT_BOOL(0, "stdout", &pack_to_stdout,
3051                          N_("output pack to stdout")),
3052                 OPT_BOOL(0, "include-tag", &include_tag,
3053                          N_("include tag objects that refer to objects to be packed")),
3054                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
3055                          N_("keep unreachable objects")),
3056                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
3057                          N_("pack loose unreachable objects")),
3058                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
3059                   N_("unpack unreachable objects newer than <time>"),
3060                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
3061                 OPT_BOOL(0, "thin", &thin,
3062                          N_("create thin packs")),
3063                 OPT_BOOL(0, "shallow", &shallow,
3064                          N_("create packs suitable for shallow fetches")),
3065                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep_on_disk,
3066                          N_("ignore packs that have companion .keep file")),
3067                 OPT_STRING_LIST(0, "keep-pack", &keep_pack_list, N_("name"),
3068                                 N_("ignore this pack")),
3069                 OPT_INTEGER(0, "compression", &pack_compression_level,
3070                             N_("pack compression level")),
3071                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
3072                             N_("do not hide commits by grafts"), 0),
3073                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
3074                          N_("use a bitmap index if available to speed up counting objects")),
3075                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
3076                          N_("write a bitmap index together with the pack index")),
3077                 OPT_PARSE_LIST_OBJECTS_FILTER(&filter_options),
3078                 { OPTION_CALLBACK, 0, "missing", NULL, N_("action"),
3079                   N_("handling for missing objects"), PARSE_OPT_NONEG,
3080                   option_parse_missing_action },
3081                 OPT_BOOL(0, "exclude-promisor-objects", &exclude_promisor_objects,
3082                          N_("do not pack objects in promisor packfiles")),
3083                 OPT_END(),
3084         };
3085
3086         check_replace_refs = 0;
3087
3088         reset_pack_idx_option(&pack_idx_opts);
3089         git_config(git_pack_config, NULL);
3090
3091         progress = isatty(2);
3092         argc = parse_options(argc, argv, prefix, pack_objects_options,
3093                              pack_usage, 0);
3094
3095         if (argc) {
3096                 base_name = argv[0];
3097                 argc--;
3098         }
3099         if (pack_to_stdout != !base_name || argc)
3100                 usage_with_options(pack_usage, pack_objects_options);
3101
3102         argv_array_push(&rp, "pack-objects");
3103         if (thin) {
3104                 use_internal_rev_list = 1;
3105                 argv_array_push(&rp, shallow
3106                                 ? "--objects-edge-aggressive"
3107                                 : "--objects-edge");
3108         } else
3109                 argv_array_push(&rp, "--objects");
3110
3111         if (rev_list_all) {
3112                 use_internal_rev_list = 1;
3113                 argv_array_push(&rp, "--all");
3114         }
3115         if (rev_list_reflog) {
3116                 use_internal_rev_list = 1;
3117                 argv_array_push(&rp, "--reflog");
3118         }
3119         if (rev_list_index) {
3120                 use_internal_rev_list = 1;
3121                 argv_array_push(&rp, "--indexed-objects");
3122         }
3123         if (rev_list_unpacked) {
3124                 use_internal_rev_list = 1;
3125                 argv_array_push(&rp, "--unpacked");
3126         }
3127
3128         if (exclude_promisor_objects) {
3129                 use_internal_rev_list = 1;
3130                 fetch_if_missing = 0;
3131                 argv_array_push(&rp, "--exclude-promisor-objects");
3132         }
3133
3134         if (!reuse_object)
3135                 reuse_delta = 0;
3136         if (pack_compression_level == -1)
3137                 pack_compression_level = Z_DEFAULT_COMPRESSION;
3138         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
3139                 die("bad pack compression level %d", pack_compression_level);
3140
3141         if (!delta_search_threads)      /* --threads=0 means autodetect */
3142                 delta_search_threads = online_cpus();
3143
3144 #ifdef NO_PTHREADS
3145         if (delta_search_threads != 1)
3146                 warning("no threads support, ignoring --threads");
3147 #endif
3148         if (!pack_to_stdout && !pack_size_limit)
3149                 pack_size_limit = pack_size_limit_cfg;
3150         if (pack_to_stdout && pack_size_limit)
3151                 die("--max-pack-size cannot be used to build a pack for transfer.");
3152         if (pack_size_limit && pack_size_limit < 1024*1024) {
3153                 warning("minimum pack size limit is 1 MiB");
3154                 pack_size_limit = 1024*1024;
3155         }
3156
3157         if (!pack_to_stdout && thin)
3158                 die("--thin cannot be used to build an indexable pack.");
3159
3160         if (keep_unreachable && unpack_unreachable)
3161                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
3162         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3163                 unpack_unreachable_expiration = 0;
3164
3165         if (filter_options.choice) {
3166                 if (!pack_to_stdout)
3167                         die("cannot use --filter without --stdout.");
3168                 use_bitmap_index = 0;
3169         }
3170
3171         /*
3172          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3173          *
3174          * - to produce good pack (with bitmap index not-yet-packed objects are
3175          *   packed in suboptimal order).
3176          *
3177          * - to use more robust pack-generation codepath (avoiding possible
3178          *   bugs in bitmap code and possible bitmap index corruption).
3179          */
3180         if (!pack_to_stdout)
3181                 use_bitmap_index_default = 0;
3182
3183         if (use_bitmap_index < 0)
3184                 use_bitmap_index = use_bitmap_index_default;
3185
3186         /* "hard" reasons not to use bitmaps; these just won't work at all */
3187         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
3188                 use_bitmap_index = 0;
3189
3190         if (pack_to_stdout || !rev_list_all)
3191                 write_bitmap_index = 0;
3192
3193         if (progress && all_progress_implied)
3194                 progress = 2;
3195
3196         add_extra_kept_packs(&keep_pack_list);
3197         if (ignore_packed_keep_on_disk) {
3198                 struct packed_git *p;
3199                 for (p = get_packed_git(the_repository); p; p = p->next)
3200                         if (p->pack_local && p->pack_keep)
3201                                 break;
3202                 if (!p) /* no keep-able packs found */
3203                         ignore_packed_keep_on_disk = 0;
3204         }
3205         if (local) {
3206                 /*
3207                  * unlike ignore_packed_keep_on_disk above, we do not
3208                  * want to unset "local" based on looking at packs, as
3209                  * it also covers non-local objects
3210                  */
3211                 struct packed_git *p;
3212                 for (p = get_packed_git(the_repository); p; p = p->next) {
3213                         if (!p->pack_local) {
3214                                 have_non_local_packs = 1;
3215                                 break;
3216                         }
3217                 }
3218         }
3219
3220         if (progress)
3221                 progress_state = start_progress(_("Enumerating objects"), 0);
3222         if (!use_internal_rev_list)
3223                 read_object_list_from_stdin();
3224         else {
3225                 get_object_list(rp.argc, rp.argv);
3226                 argv_array_clear(&rp);
3227         }
3228         cleanup_preferred_base();
3229         if (include_tag && nr_result)
3230                 for_each_ref(add_ref_tag, NULL);
3231         stop_progress(&progress_state);
3232
3233         if (non_empty && !nr_result)
3234                 return 0;
3235         if (nr_result)
3236                 prepare_pack(window, depth);
3237         write_pack_file();
3238         if (progress)
3239                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
3240                         " reused %"PRIu32" (delta %"PRIu32")\n",
3241                         written, written_delta, reused, reused_delta);
3242         return 0;
3243 }