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;
  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 = 256 * 1024 * 1024;
  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                         hashclose(f, oid.hash, CSUM_CLOSE);
 843                 } else if (nr_written == nr_remaining) {
 844                         hashclose(f, oid.hash, CSUM_FSYNC);
 845                 } else {
 846                         int fd = hashclose(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         if (have_duplicate_entry(oid, exclude, &index_pos))
1100                 return 0;
1101
1102         if (!want_object_in_pack(oid, exclude, &found_pack, &found_offset)) {
1103                 /* The pack is missing an object, so it will not have closure */
1104                 if (write_bitmap_index) {
1105                         warning(_(no_closure_warning));
1106                         write_bitmap_index = 0;
1107                 }
1108                 return 0;
1109         }
1110
1111         create_object_entry(oid, type, pack_name_hash(name),
1112                             exclude, name && no_try_delta(name),
1113                             index_pos, found_pack, found_offset);
1114
1115         display_progress(progress_state, nr_result);
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         if (have_duplicate_entry(oid, 0, &index_pos))
1127                 return 0;
1128
1129         if (!want_object_in_pack(oid, 0, &pack, &offset))
1130                 return 0;
1131
1132         create_object_entry(oid, type, name_hash, 0, 0, index_pos, pack, offset);
1133
1134         display_progress(progress_state, nr_result);
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(&entry->idx.oid, &entry->size);
1525         /*
1526          * The error condition is checked in prepare_pack().  This is
1527          * to permit a missing preferred base object to be ignored
1528          * as a preferred base.  Doing so can result in a larger
1529          * pack file, but the transfer will still take place.
1530          */
1531 }
1532
1533 static int pack_offset_sort(const void *_a, const void *_b)
1534 {
1535         const struct object_entry *a = *(struct object_entry **)_a;
1536         const struct object_entry *b = *(struct object_entry **)_b;
1537
1538         /* avoid filesystem trashing with loose objects */
1539         if (!a->in_pack && !b->in_pack)
1540                 return oidcmp(&a->idx.oid, &b->idx.oid);
1541
1542         if (a->in_pack < b->in_pack)
1543                 return -1;
1544         if (a->in_pack > b->in_pack)
1545                 return 1;
1546         return a->in_pack_offset < b->in_pack_offset ? -1 :
1547                         (a->in_pack_offset > b->in_pack_offset);
1548 }
1549
1550 /*
1551  * Drop an on-disk delta we were planning to reuse. Naively, this would
1552  * just involve blanking out the "delta" field, but we have to deal
1553  * with some extra book-keeping:
1554  *
1555  *   1. Removing ourselves from the delta_sibling linked list.
1556  *
1557  *   2. Updating our size/type to the non-delta representation. These were
1558  *      either not recorded initially (size) or overwritten with the delta type
1559  *      (type) when check_object() decided to reuse the delta.
1560  *
1561  *   3. Resetting our delta depth, as we are now a base object.
1562  */
1563 static void drop_reused_delta(struct object_entry *entry)
1564 {
1565         struct object_entry **p = &entry->delta->delta_child;
1566         struct object_info oi = OBJECT_INFO_INIT;
1567
1568         while (*p) {
1569                 if (*p == entry)
1570                         *p = (*p)->delta_sibling;
1571                 else
1572                         p = &(*p)->delta_sibling;
1573         }
1574         entry->delta = NULL;
1575         entry->depth = 0;
1576
1577         oi.sizep = &entry->size;
1578         oi.typep = &entry->type;
1579         if (packed_object_info(entry->in_pack, entry->in_pack_offset, &oi) < 0) {
1580                 /*
1581                  * We failed to get the info from this pack for some reason;
1582                  * fall back to sha1_object_info, which may find another copy.
1583                  * And if that fails, the error will be recorded in entry->type
1584                  * and dealt with in prepare_pack().
1585                  */
1586                 entry->type = oid_object_info(&entry->idx.oid, &entry->size);
1587         }
1588 }
1589
1590 /*
1591  * Follow the chain of deltas from this entry onward, throwing away any links
1592  * that cause us to hit a cycle (as determined by the DFS state flags in
1593  * the entries).
1594  *
1595  * We also detect too-long reused chains that would violate our --depth
1596  * limit.
1597  */
1598 static void break_delta_chains(struct object_entry *entry)
1599 {
1600         /*
1601          * The actual depth of each object we will write is stored as an int,
1602          * as it cannot exceed our int "depth" limit. But before we break
1603          * changes based no that limit, we may potentially go as deep as the
1604          * number of objects, which is elsewhere bounded to a uint32_t.
1605          */
1606         uint32_t total_depth;
1607         struct object_entry *cur, *next;
1608
1609         for (cur = entry, total_depth = 0;
1610              cur;
1611              cur = cur->delta, total_depth++) {
1612                 if (cur->dfs_state == DFS_DONE) {
1613                         /*
1614                          * We've already seen this object and know it isn't
1615                          * part of a cycle. We do need to append its depth
1616                          * to our count.
1617                          */
1618                         total_depth += cur->depth;
1619                         break;
1620                 }
1621
1622                 /*
1623                  * We break cycles before looping, so an ACTIVE state (or any
1624                  * other cruft which made its way into the state variable)
1625                  * is a bug.
1626                  */
1627                 if (cur->dfs_state != DFS_NONE)
1628                         die("BUG: confusing delta dfs state in first pass: %d",
1629                             cur->dfs_state);
1630
1631                 /*
1632                  * Now we know this is the first time we've seen the object. If
1633                  * it's not a delta, we're done traversing, but we'll mark it
1634                  * done to save time on future traversals.
1635                  */
1636                 if (!cur->delta) {
1637                         cur->dfs_state = DFS_DONE;
1638                         break;
1639                 }
1640
1641                 /*
1642                  * Mark ourselves as active and see if the next step causes
1643                  * us to cycle to another active object. It's important to do
1644                  * this _before_ we loop, because it impacts where we make the
1645                  * cut, and thus how our total_depth counter works.
1646                  * E.g., We may see a partial loop like:
1647                  *
1648                  *   A -> B -> C -> D -> B
1649                  *
1650                  * Cutting B->C breaks the cycle. But now the depth of A is
1651                  * only 1, and our total_depth counter is at 3. The size of the
1652                  * error is always one less than the size of the cycle we
1653                  * broke. Commits C and D were "lost" from A's chain.
1654                  *
1655                  * If we instead cut D->B, then the depth of A is correct at 3.
1656                  * We keep all commits in the chain that we examined.
1657                  */
1658                 cur->dfs_state = DFS_ACTIVE;
1659                 if (cur->delta->dfs_state == DFS_ACTIVE) {
1660                         drop_reused_delta(cur);
1661                         cur->dfs_state = DFS_DONE;
1662                         break;
1663                 }
1664         }
1665
1666         /*
1667          * And now that we've gone all the way to the bottom of the chain, we
1668          * need to clear the active flags and set the depth fields as
1669          * appropriate. Unlike the loop above, which can quit when it drops a
1670          * delta, we need to keep going to look for more depth cuts. So we need
1671          * an extra "next" pointer to keep going after we reset cur->delta.
1672          */
1673         for (cur = entry; cur; cur = next) {
1674                 next = cur->delta;
1675
1676                 /*
1677                  * We should have a chain of zero or more ACTIVE states down to
1678                  * a final DONE. We can quit after the DONE, because either it
1679                  * has no bases, or we've already handled them in a previous
1680                  * call.
1681                  */
1682                 if (cur->dfs_state == DFS_DONE)
1683                         break;
1684                 else if (cur->dfs_state != DFS_ACTIVE)
1685                         die("BUG: confusing delta dfs state in second pass: %d",
1686                             cur->dfs_state);
1687
1688                 /*
1689                  * If the total_depth is more than depth, then we need to snip
1690                  * the chain into two or more smaller chains that don't exceed
1691                  * the maximum depth. Most of the resulting chains will contain
1692                  * (depth + 1) entries (i.e., depth deltas plus one base), and
1693                  * the last chain (i.e., the one containing entry) will contain
1694                  * whatever entries are left over, namely
1695                  * (total_depth % (depth + 1)) of them.
1696                  *
1697                  * Since we are iterating towards decreasing depth, we need to
1698                  * decrement total_depth as we go, and we need to write to the
1699                  * entry what its final depth will be after all of the
1700                  * snipping. Since we're snipping into chains of length (depth
1701                  * + 1) entries, the final depth of an entry will be its
1702                  * original depth modulo (depth + 1). Any time we encounter an
1703                  * entry whose final depth is supposed to be zero, we snip it
1704                  * from its delta base, thereby making it so.
1705                  */
1706                 cur->depth = (total_depth--) % (depth + 1);
1707                 if (!cur->depth)
1708                         drop_reused_delta(cur);
1709
1710                 cur->dfs_state = DFS_DONE;
1711         }
1712 }
1713
1714 static void get_object_details(void)
1715 {
1716         uint32_t i;
1717         struct object_entry **sorted_by_offset;
1718
1719         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1720         for (i = 0; i < to_pack.nr_objects; i++)
1721                 sorted_by_offset[i] = to_pack.objects + i;
1722         QSORT(sorted_by_offset, to_pack.nr_objects, pack_offset_sort);
1723
1724         for (i = 0; i < to_pack.nr_objects; i++) {
1725                 struct object_entry *entry = sorted_by_offset[i];
1726                 check_object(entry);
1727                 if (big_file_threshold < entry->size)
1728                         entry->no_try_delta = 1;
1729         }
1730
1731         /*
1732          * This must happen in a second pass, since we rely on the delta
1733          * information for the whole list being completed.
1734          */
1735         for (i = 0; i < to_pack.nr_objects; i++)
1736                 break_delta_chains(&to_pack.objects[i]);
1737
1738         free(sorted_by_offset);
1739 }
1740
1741 /*
1742  * We search for deltas in a list sorted by type, by filename hash, and then
1743  * by size, so that we see progressively smaller and smaller files.
1744  * That's because we prefer deltas to be from the bigger file
1745  * to the smaller -- deletes are potentially cheaper, but perhaps
1746  * more importantly, the bigger file is likely the more recent
1747  * one.  The deepest deltas are therefore the oldest objects which are
1748  * less susceptible to be accessed often.
1749  */
1750 static int type_size_sort(const void *_a, const void *_b)
1751 {
1752         const struct object_entry *a = *(struct object_entry **)_a;
1753         const struct object_entry *b = *(struct object_entry **)_b;
1754
1755         if (a->type > b->type)
1756                 return -1;
1757         if (a->type < b->type)
1758                 return 1;
1759         if (a->hash > b->hash)
1760                 return -1;
1761         if (a->hash < b->hash)
1762                 return 1;
1763         if (a->preferred_base > b->preferred_base)
1764                 return -1;
1765         if (a->preferred_base < b->preferred_base)
1766                 return 1;
1767         if (a->size > b->size)
1768                 return -1;
1769         if (a->size < b->size)
1770                 return 1;
1771         return a < b ? -1 : (a > b);  /* newest first */
1772 }
1773
1774 struct unpacked {
1775         struct object_entry *entry;
1776         void *data;
1777         struct delta_index *index;
1778         unsigned depth;
1779 };
1780
1781 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1782                            unsigned long delta_size)
1783 {
1784         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1785                 return 0;
1786
1787         if (delta_size < cache_max_small_delta_size)
1788                 return 1;
1789
1790         /* cache delta, if objects are large enough compared to delta size */
1791         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1792                 return 1;
1793
1794         return 0;
1795 }
1796
1797 #ifndef NO_PTHREADS
1798
1799 static pthread_mutex_t read_mutex;
1800 #define read_lock()             pthread_mutex_lock(&read_mutex)
1801 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1802
1803 static pthread_mutex_t cache_mutex;
1804 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1805 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1806
1807 static pthread_mutex_t progress_mutex;
1808 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1809 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1810
1811 #else
1812
1813 #define read_lock()             (void)0
1814 #define read_unlock()           (void)0
1815 #define cache_lock()            (void)0
1816 #define cache_unlock()          (void)0
1817 #define progress_lock()         (void)0
1818 #define progress_unlock()       (void)0
1819
1820 #endif
1821
1822 static int try_delta(struct unpacked *trg, struct unpacked *src,
1823                      unsigned max_depth, unsigned long *mem_usage)
1824 {
1825         struct object_entry *trg_entry = trg->entry;
1826         struct object_entry *src_entry = src->entry;
1827         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1828         unsigned ref_depth;
1829         enum object_type type;
1830         void *delta_buf;
1831
1832         /* Don't bother doing diffs between different types */
1833         if (trg_entry->type != src_entry->type)
1834                 return -1;
1835
1836         /*
1837          * We do not bother to try a delta that we discarded on an
1838          * earlier try, but only when reusing delta data.  Note that
1839          * src_entry that is marked as the preferred_base should always
1840          * be considered, as even if we produce a suboptimal delta against
1841          * it, we will still save the transfer cost, as we already know
1842          * the other side has it and we won't send src_entry at all.
1843          */
1844         if (reuse_delta && trg_entry->in_pack &&
1845             trg_entry->in_pack == src_entry->in_pack &&
1846             !src_entry->preferred_base &&
1847             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1848             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1849                 return 0;
1850
1851         /* Let's not bust the allowed depth. */
1852         if (src->depth >= max_depth)
1853                 return 0;
1854
1855         /* Now some size filtering heuristics. */
1856         trg_size = trg_entry->size;
1857         if (!trg_entry->delta) {
1858                 max_size = trg_size/2 - 20;
1859                 ref_depth = 1;
1860         } else {
1861                 max_size = trg_entry->delta_size;
1862                 ref_depth = trg->depth;
1863         }
1864         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1865                                                 (max_depth - ref_depth + 1);
1866         if (max_size == 0)
1867                 return 0;
1868         src_size = src_entry->size;
1869         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1870         if (sizediff >= max_size)
1871                 return 0;
1872         if (trg_size < src_size / 32)
1873                 return 0;
1874
1875         /* Load data if not already done */
1876         if (!trg->data) {
1877                 read_lock();
1878                 trg->data = read_object_file(&trg_entry->idx.oid, &type, &sz);
1879                 read_unlock();
1880                 if (!trg->data)
1881                         die("object %s cannot be read",
1882                             oid_to_hex(&trg_entry->idx.oid));
1883                 if (sz != trg_size)
1884                         die("object %s inconsistent object length (%lu vs %lu)",
1885                             oid_to_hex(&trg_entry->idx.oid), sz,
1886                             trg_size);
1887                 *mem_usage += sz;
1888         }
1889         if (!src->data) {
1890                 read_lock();
1891                 src->data = read_object_file(&src_entry->idx.oid, &type, &sz);
1892                 read_unlock();
1893                 if (!src->data) {
1894                         if (src_entry->preferred_base) {
1895                                 static int warned = 0;
1896                                 if (!warned++)
1897                                         warning("object %s cannot be read",
1898                                                 oid_to_hex(&src_entry->idx.oid));
1899                                 /*
1900                                  * Those objects are not included in the
1901                                  * resulting pack.  Be resilient and ignore
1902                                  * them if they can't be read, in case the
1903                                  * pack could be created nevertheless.
1904                                  */
1905                                 return 0;
1906                         }
1907                         die("object %s cannot be read",
1908                             oid_to_hex(&src_entry->idx.oid));
1909                 }
1910                 if (sz != src_size)
1911                         die("object %s inconsistent object length (%lu vs %lu)",
1912                             oid_to_hex(&src_entry->idx.oid), sz,
1913                             src_size);
1914                 *mem_usage += sz;
1915         }
1916         if (!src->index) {
1917                 src->index = create_delta_index(src->data, src_size);
1918                 if (!src->index) {
1919                         static int warned = 0;
1920                         if (!warned++)
1921                                 warning("suboptimal pack - out of memory");
1922                         return 0;
1923                 }
1924                 *mem_usage += sizeof_delta_index(src->index);
1925         }
1926
1927         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1928         if (!delta_buf)
1929                 return 0;
1930
1931         if (trg_entry->delta) {
1932                 /* Prefer only shallower same-sized deltas. */
1933                 if (delta_size == trg_entry->delta_size &&
1934                     src->depth + 1 >= trg->depth) {
1935                         free(delta_buf);
1936                         return 0;
1937                 }
1938         }
1939
1940         /*
1941          * Handle memory allocation outside of the cache
1942          * accounting lock.  Compiler will optimize the strangeness
1943          * away when NO_PTHREADS is defined.
1944          */
1945         free(trg_entry->delta_data);
1946         cache_lock();
1947         if (trg_entry->delta_data) {
1948                 delta_cache_size -= trg_entry->delta_size;
1949                 trg_entry->delta_data = NULL;
1950         }
1951         if (delta_cacheable(src_size, trg_size, delta_size)) {
1952                 delta_cache_size += delta_size;
1953                 cache_unlock();
1954                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1955         } else {
1956                 cache_unlock();
1957                 free(delta_buf);
1958         }
1959
1960         trg_entry->delta = src_entry;
1961         trg_entry->delta_size = delta_size;
1962         trg->depth = src->depth + 1;
1963
1964         return 1;
1965 }
1966
1967 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1968 {
1969         struct object_entry *child = me->delta_child;
1970         unsigned int m = n;
1971         while (child) {
1972                 unsigned int c = check_delta_limit(child, n + 1);
1973                 if (m < c)
1974                         m = c;
1975                 child = child->delta_sibling;
1976         }
1977         return m;
1978 }
1979
1980 static unsigned long free_unpacked(struct unpacked *n)
1981 {
1982         unsigned long freed_mem = sizeof_delta_index(n->index);
1983         free_delta_index(n->index);
1984         n->index = NULL;
1985         if (n->data) {
1986                 freed_mem += n->entry->size;
1987                 FREE_AND_NULL(n->data);
1988         }
1989         n->entry = NULL;
1990         n->depth = 0;
1991         return freed_mem;
1992 }
1993
1994 static void find_deltas(struct object_entry **list, unsigned *list_size,
1995                         int window, int depth, unsigned *processed)
1996 {
1997         uint32_t i, idx = 0, count = 0;
1998         struct unpacked *array;
1999         unsigned long mem_usage = 0;
2000
2001         array = xcalloc(window, sizeof(struct unpacked));
2002
2003         for (;;) {
2004                 struct object_entry *entry;
2005                 struct unpacked *n = array + idx;
2006                 int j, max_depth, best_base = -1;
2007
2008                 progress_lock();
2009                 if (!*list_size) {
2010                         progress_unlock();
2011                         break;
2012                 }
2013                 entry = *list++;
2014                 (*list_size)--;
2015                 if (!entry->preferred_base) {
2016                         (*processed)++;
2017                         display_progress(progress_state, *processed);
2018                 }
2019                 progress_unlock();
2020
2021                 mem_usage -= free_unpacked(n);
2022                 n->entry = entry;
2023
2024                 while (window_memory_limit &&
2025                        mem_usage > window_memory_limit &&
2026                        count > 1) {
2027                         uint32_t tail = (idx + window - count) % window;
2028                         mem_usage -= free_unpacked(array + tail);
2029                         count--;
2030                 }
2031
2032                 /* We do not compute delta to *create* objects we are not
2033                  * going to pack.
2034                  */
2035                 if (entry->preferred_base)
2036                         goto next;
2037
2038                 /*
2039                  * If the current object is at pack edge, take the depth the
2040                  * objects that depend on the current object into account
2041                  * otherwise they would become too deep.
2042                  */
2043                 max_depth = depth;
2044                 if (entry->delta_child) {
2045                         max_depth -= check_delta_limit(entry, 0);
2046                         if (max_depth <= 0)
2047                                 goto next;
2048                 }
2049
2050                 j = window;
2051                 while (--j > 0) {
2052                         int ret;
2053                         uint32_t other_idx = idx + j;
2054                         struct unpacked *m;
2055                         if (other_idx >= window)
2056                                 other_idx -= window;
2057                         m = array + other_idx;
2058                         if (!m->entry)
2059                                 break;
2060                         ret = try_delta(n, m, max_depth, &mem_usage);
2061                         if (ret < 0)
2062                                 break;
2063                         else if (ret > 0)
2064                                 best_base = other_idx;
2065                 }
2066
2067                 /*
2068                  * If we decided to cache the delta data, then it is best
2069                  * to compress it right away.  First because we have to do
2070                  * it anyway, and doing it here while we're threaded will
2071                  * save a lot of time in the non threaded write phase,
2072                  * as well as allow for caching more deltas within
2073                  * the same cache size limit.
2074                  * ...
2075                  * But only if not writing to stdout, since in that case
2076                  * the network is most likely throttling writes anyway,
2077                  * and therefore it is best to go to the write phase ASAP
2078                  * instead, as we can afford spending more time compressing
2079                  * between writes at that moment.
2080                  */
2081                 if (entry->delta_data && !pack_to_stdout) {
2082                         entry->z_delta_size = do_compress(&entry->delta_data,
2083                                                           entry->delta_size);
2084                         cache_lock();
2085                         delta_cache_size -= entry->delta_size;
2086                         delta_cache_size += entry->z_delta_size;
2087                         cache_unlock();
2088                 }
2089
2090                 /* if we made n a delta, and if n is already at max
2091                  * depth, leaving it in the window is pointless.  we
2092                  * should evict it first.
2093                  */
2094                 if (entry->delta && max_depth <= n->depth)
2095                         continue;
2096
2097                 /*
2098                  * Move the best delta base up in the window, after the
2099                  * currently deltified object, to keep it longer.  It will
2100                  * be the first base object to be attempted next.
2101                  */
2102                 if (entry->delta) {
2103                         struct unpacked swap = array[best_base];
2104                         int dist = (window + idx - best_base) % window;
2105                         int dst = best_base;
2106                         while (dist--) {
2107                                 int src = (dst + 1) % window;
2108                                 array[dst] = array[src];
2109                                 dst = src;
2110                         }
2111                         array[dst] = swap;
2112                 }
2113
2114                 next:
2115                 idx++;
2116                 if (count + 1 < window)
2117                         count++;
2118                 if (idx >= window)
2119                         idx = 0;
2120         }
2121
2122         for (i = 0; i < window; ++i) {
2123                 free_delta_index(array[i].index);
2124                 free(array[i].data);
2125         }
2126         free(array);
2127 }
2128
2129 #ifndef NO_PTHREADS
2130
2131 static void try_to_free_from_threads(size_t size)
2132 {
2133         read_lock();
2134         release_pack_memory(size);
2135         read_unlock();
2136 }
2137
2138 static try_to_free_t old_try_to_free_routine;
2139
2140 /*
2141  * The main thread waits on the condition that (at least) one of the workers
2142  * has stopped working (which is indicated in the .working member of
2143  * struct thread_params).
2144  * When a work thread has completed its work, it sets .working to 0 and
2145  * signals the main thread and waits on the condition that .data_ready
2146  * becomes 1.
2147  */
2148
2149 struct thread_params {
2150         pthread_t thread;
2151         struct object_entry **list;
2152         unsigned list_size;
2153         unsigned remaining;
2154         int window;
2155         int depth;
2156         int working;
2157         int data_ready;
2158         pthread_mutex_t mutex;
2159         pthread_cond_t cond;
2160         unsigned *processed;
2161 };
2162
2163 static pthread_cond_t progress_cond;
2164
2165 /*
2166  * Mutex and conditional variable can't be statically-initialized on Windows.
2167  */
2168 static void init_threaded_search(void)
2169 {
2170         init_recursive_mutex(&read_mutex);
2171         pthread_mutex_init(&cache_mutex, NULL);
2172         pthread_mutex_init(&progress_mutex, NULL);
2173         pthread_cond_init(&progress_cond, NULL);
2174         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
2175 }
2176
2177 static void cleanup_threaded_search(void)
2178 {
2179         set_try_to_free_routine(old_try_to_free_routine);
2180         pthread_cond_destroy(&progress_cond);
2181         pthread_mutex_destroy(&read_mutex);
2182         pthread_mutex_destroy(&cache_mutex);
2183         pthread_mutex_destroy(&progress_mutex);
2184 }
2185
2186 static void *threaded_find_deltas(void *arg)
2187 {
2188         struct thread_params *me = arg;
2189
2190         progress_lock();
2191         while (me->remaining) {
2192                 progress_unlock();
2193
2194                 find_deltas(me->list, &me->remaining,
2195                             me->window, me->depth, me->processed);
2196
2197                 progress_lock();
2198                 me->working = 0;
2199                 pthread_cond_signal(&progress_cond);
2200                 progress_unlock();
2201
2202                 /*
2203                  * We must not set ->data_ready before we wait on the
2204                  * condition because the main thread may have set it to 1
2205                  * before we get here. In order to be sure that new
2206                  * work is available if we see 1 in ->data_ready, it
2207                  * was initialized to 0 before this thread was spawned
2208                  * and we reset it to 0 right away.
2209                  */
2210                 pthread_mutex_lock(&me->mutex);
2211                 while (!me->data_ready)
2212                         pthread_cond_wait(&me->cond, &me->mutex);
2213                 me->data_ready = 0;
2214                 pthread_mutex_unlock(&me->mutex);
2215
2216                 progress_lock();
2217         }
2218         progress_unlock();
2219         /* leave ->working 1 so that this doesn't get more work assigned */
2220         return NULL;
2221 }
2222
2223 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
2224                            int window, int depth, unsigned *processed)
2225 {
2226         struct thread_params *p;
2227         int i, ret, active_threads = 0;
2228
2229         init_threaded_search();
2230
2231         if (delta_search_threads <= 1) {
2232                 find_deltas(list, &list_size, window, depth, processed);
2233                 cleanup_threaded_search();
2234                 return;
2235         }
2236         if (progress > pack_to_stdout)
2237                 fprintf(stderr, "Delta compression using up to %d threads.\n",
2238                                 delta_search_threads);
2239         p = xcalloc(delta_search_threads, sizeof(*p));
2240
2241         /* Partition the work amongst work threads. */
2242         for (i = 0; i < delta_search_threads; i++) {
2243                 unsigned sub_size = list_size / (delta_search_threads - i);
2244
2245                 /* don't use too small segments or no deltas will be found */
2246                 if (sub_size < 2*window && i+1 < delta_search_threads)
2247                         sub_size = 0;
2248
2249                 p[i].window = window;
2250                 p[i].depth = depth;
2251                 p[i].processed = processed;
2252                 p[i].working = 1;
2253                 p[i].data_ready = 0;
2254
2255                 /* try to split chunks on "path" boundaries */
2256                 while (sub_size && sub_size < list_size &&
2257                        list[sub_size]->hash &&
2258                        list[sub_size]->hash == list[sub_size-1]->hash)
2259                         sub_size++;
2260
2261                 p[i].list = list;
2262                 p[i].list_size = sub_size;
2263                 p[i].remaining = sub_size;
2264
2265                 list += sub_size;
2266                 list_size -= sub_size;
2267         }
2268
2269         /* Start work threads. */
2270         for (i = 0; i < delta_search_threads; i++) {
2271                 if (!p[i].list_size)
2272                         continue;
2273                 pthread_mutex_init(&p[i].mutex, NULL);
2274                 pthread_cond_init(&p[i].cond, NULL);
2275                 ret = pthread_create(&p[i].thread, NULL,
2276                                      threaded_find_deltas, &p[i]);
2277                 if (ret)
2278                         die("unable to create thread: %s", strerror(ret));
2279                 active_threads++;
2280         }
2281
2282         /*
2283          * Now let's wait for work completion.  Each time a thread is done
2284          * with its work, we steal half of the remaining work from the
2285          * thread with the largest number of unprocessed objects and give
2286          * it to that newly idle thread.  This ensure good load balancing
2287          * until the remaining object list segments are simply too short
2288          * to be worth splitting anymore.
2289          */
2290         while (active_threads) {
2291                 struct thread_params *target = NULL;
2292                 struct thread_params *victim = NULL;
2293                 unsigned sub_size = 0;
2294
2295                 progress_lock();
2296                 for (;;) {
2297                         for (i = 0; !target && i < delta_search_threads; i++)
2298                                 if (!p[i].working)
2299                                         target = &p[i];
2300                         if (target)
2301                                 break;
2302                         pthread_cond_wait(&progress_cond, &progress_mutex);
2303                 }
2304
2305                 for (i = 0; i < delta_search_threads; i++)
2306                         if (p[i].remaining > 2*window &&
2307                             (!victim || victim->remaining < p[i].remaining))
2308                                 victim = &p[i];
2309                 if (victim) {
2310                         sub_size = victim->remaining / 2;
2311                         list = victim->list + victim->list_size - sub_size;
2312                         while (sub_size && list[0]->hash &&
2313                                list[0]->hash == list[-1]->hash) {
2314                                 list++;
2315                                 sub_size--;
2316                         }
2317                         if (!sub_size) {
2318                                 /*
2319                                  * It is possible for some "paths" to have
2320                                  * so many objects that no hash boundary
2321                                  * might be found.  Let's just steal the
2322                                  * exact half in that case.
2323                                  */
2324                                 sub_size = victim->remaining / 2;
2325                                 list -= sub_size;
2326                         }
2327                         target->list = list;
2328                         victim->list_size -= sub_size;
2329                         victim->remaining -= sub_size;
2330                 }
2331                 target->list_size = sub_size;
2332                 target->remaining = sub_size;
2333                 target->working = 1;
2334                 progress_unlock();
2335
2336                 pthread_mutex_lock(&target->mutex);
2337                 target->data_ready = 1;
2338                 pthread_cond_signal(&target->cond);
2339                 pthread_mutex_unlock(&target->mutex);
2340
2341                 if (!sub_size) {
2342                         pthread_join(target->thread, NULL);
2343                         pthread_cond_destroy(&target->cond);
2344                         pthread_mutex_destroy(&target->mutex);
2345                         active_threads--;
2346                 }
2347         }
2348         cleanup_threaded_search();
2349         free(p);
2350 }
2351
2352 #else
2353 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2354 #endif
2355
2356 static void add_tag_chain(const struct object_id *oid)
2357 {
2358         struct tag *tag;
2359
2360         /*
2361          * We catch duplicates already in add_object_entry(), but we'd
2362          * prefer to do this extra check to avoid having to parse the
2363          * tag at all if we already know that it's being packed (e.g., if
2364          * it was included via bitmaps, we would not have parsed it
2365          * previously).
2366          */
2367         if (packlist_find(&to_pack, oid->hash, NULL))
2368                 return;
2369
2370         tag = lookup_tag(oid);
2371         while (1) {
2372                 if (!tag || parse_tag(tag) || !tag->tagged)
2373                         die("unable to pack objects reachable from tag %s",
2374                             oid_to_hex(oid));
2375
2376                 add_object_entry(&tag->object.oid, OBJ_TAG, NULL, 0);
2377
2378                 if (tag->tagged->type != OBJ_TAG)
2379                         return;
2380
2381                 tag = (struct tag *)tag->tagged;
2382         }
2383 }
2384
2385 static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2386 {
2387         struct object_id peeled;
2388
2389         if (starts_with(path, "refs/tags/") && /* is a tag? */
2390             !peel_ref(path, &peeled)    && /* peelable? */
2391             packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2392                 add_tag_chain(oid);
2393         return 0;
2394 }
2395
2396 static void prepare_pack(int window, int depth)
2397 {
2398         struct object_entry **delta_list;
2399         uint32_t i, nr_deltas;
2400         unsigned n;
2401
2402         get_object_details();
2403
2404         /*
2405          * If we're locally repacking then we need to be doubly careful
2406          * from now on in order to make sure no stealth corruption gets
2407          * propagated to the new pack.  Clients receiving streamed packs
2408          * should validate everything they get anyway so no need to incur
2409          * the additional cost here in that case.
2410          */
2411         if (!pack_to_stdout)
2412                 do_check_packed_object_crc = 1;
2413
2414         if (!to_pack.nr_objects || !window || !depth)
2415                 return;
2416
2417         ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2418         nr_deltas = n = 0;
2419
2420         for (i = 0; i < to_pack.nr_objects; i++) {
2421                 struct object_entry *entry = to_pack.objects + i;
2422
2423                 if (entry->delta)
2424                         /* This happens if we decided to reuse existing
2425                          * delta from a pack.  "reuse_delta &&" is implied.
2426                          */
2427                         continue;
2428
2429                 if (entry->size < 50)
2430                         continue;
2431
2432                 if (entry->no_try_delta)
2433                         continue;
2434
2435                 if (!entry->preferred_base) {
2436                         nr_deltas++;
2437                         if (entry->type < 0)
2438                                 die("unable to get type of object %s",
2439                                     oid_to_hex(&entry->idx.oid));
2440                 } else {
2441                         if (entry->type < 0) {
2442                                 /*
2443                                  * This object is not found, but we
2444                                  * don't have to include it anyway.
2445                                  */
2446                                 continue;
2447                         }
2448                 }
2449
2450                 delta_list[n++] = entry;
2451         }
2452
2453         if (nr_deltas && n > 1) {
2454                 unsigned nr_done = 0;
2455                 if (progress)
2456                         progress_state = start_progress(_("Compressing objects"),
2457                                                         nr_deltas);
2458                 QSORT(delta_list, n, type_size_sort);
2459                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2460                 stop_progress(&progress_state);
2461                 if (nr_done != nr_deltas)
2462                         die("inconsistency with delta count");
2463         }
2464         free(delta_list);
2465 }
2466
2467 static int git_pack_config(const char *k, const char *v, void *cb)
2468 {
2469         if (!strcmp(k, "pack.window")) {
2470                 window = git_config_int(k, v);
2471                 return 0;
2472         }
2473         if (!strcmp(k, "pack.windowmemory")) {
2474                 window_memory_limit = git_config_ulong(k, v);
2475                 return 0;
2476         }
2477         if (!strcmp(k, "pack.depth")) {
2478                 depth = git_config_int(k, v);
2479                 return 0;
2480         }
2481         if (!strcmp(k, "pack.deltacachesize")) {
2482                 max_delta_cache_size = git_config_int(k, v);
2483                 return 0;
2484         }
2485         if (!strcmp(k, "pack.deltacachelimit")) {
2486                 cache_max_small_delta_size = git_config_int(k, v);
2487                 return 0;
2488         }
2489         if (!strcmp(k, "pack.writebitmaphashcache")) {
2490                 if (git_config_bool(k, v))
2491                         write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2492                 else
2493                         write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2494         }
2495         if (!strcmp(k, "pack.usebitmaps")) {
2496                 use_bitmap_index_default = git_config_bool(k, v);
2497                 return 0;
2498         }
2499         if (!strcmp(k, "pack.threads")) {
2500                 delta_search_threads = git_config_int(k, v);
2501                 if (delta_search_threads < 0)
2502                         die("invalid number of threads specified (%d)",
2503                             delta_search_threads);
2504 #ifdef NO_PTHREADS
2505                 if (delta_search_threads != 1) {
2506                         warning("no threads support, ignoring %s", k);
2507                         delta_search_threads = 0;
2508                 }
2509 #endif
2510                 return 0;
2511         }
2512         if (!strcmp(k, "pack.indexversion")) {
2513                 pack_idx_opts.version = git_config_int(k, v);
2514                 if (pack_idx_opts.version > 2)
2515                         die("bad pack.indexversion=%"PRIu32,
2516                             pack_idx_opts.version);
2517                 return 0;
2518         }
2519         return git_default_config(k, v, cb);
2520 }
2521
2522 static void read_object_list_from_stdin(void)
2523 {
2524         char line[GIT_MAX_HEXSZ + 1 + PATH_MAX + 2];
2525         struct object_id oid;
2526         const char *p;
2527
2528         for (;;) {
2529                 if (!fgets(line, sizeof(line), stdin)) {
2530                         if (feof(stdin))
2531                                 break;
2532                         if (!ferror(stdin))
2533                                 die("fgets returned NULL, not EOF, not error!");
2534                         if (errno != EINTR)
2535                                 die_errno("fgets");
2536                         clearerr(stdin);
2537                         continue;
2538                 }
2539                 if (line[0] == '-') {
2540                         if (get_oid_hex(line+1, &oid))
2541                                 die("expected edge object ID, got garbage:\n %s",
2542                                     line);
2543                         add_preferred_base(&oid);
2544                         continue;
2545                 }
2546                 if (parse_oid_hex(line, &oid, &p))
2547                         die("expected object ID, got garbage:\n %s", line);
2548
2549                 add_preferred_base_object(p + 1);
2550                 add_object_entry(&oid, 0, p + 1, 0);
2551         }
2552 }
2553
2554 /* Remember to update object flag allocation in object.h */
2555 #define OBJECT_ADDED (1u<<20)
2556
2557 static void show_commit(struct commit *commit, void *data)
2558 {
2559         add_object_entry(&commit->object.oid, OBJ_COMMIT, NULL, 0);
2560         commit->object.flags |= OBJECT_ADDED;
2561
2562         if (write_bitmap_index)
2563                 index_commit_for_bitmap(commit);
2564 }
2565
2566 static void show_object(struct object *obj, const char *name, void *data)
2567 {
2568         add_preferred_base_object(name);
2569         add_object_entry(&obj->oid, obj->type, name, 0);
2570         obj->flags |= OBJECT_ADDED;
2571 }
2572
2573 static void show_object__ma_allow_any(struct object *obj, const char *name, void *data)
2574 {
2575         assert(arg_missing_action == MA_ALLOW_ANY);
2576
2577         /*
2578          * Quietly ignore ALL missing objects.  This avoids problems with
2579          * staging them now and getting an odd error later.
2580          */
2581         if (!has_object_file(&obj->oid))
2582                 return;
2583
2584         show_object(obj, name, data);
2585 }
2586
2587 static void show_object__ma_allow_promisor(struct object *obj, const char *name, void *data)
2588 {
2589         assert(arg_missing_action == MA_ALLOW_PROMISOR);
2590
2591         /*
2592          * Quietly ignore EXPECTED missing objects.  This avoids problems with
2593          * staging them now and getting an odd error later.
2594          */
2595         if (!has_object_file(&obj->oid) && is_promisor_object(&obj->oid))
2596                 return;
2597
2598         show_object(obj, name, data);
2599 }
2600
2601 static int option_parse_missing_action(const struct option *opt,
2602                                        const char *arg, int unset)
2603 {
2604         assert(arg);
2605         assert(!unset);
2606
2607         if (!strcmp(arg, "error")) {
2608                 arg_missing_action = MA_ERROR;
2609                 fn_show_object = show_object;
2610                 return 0;
2611         }
2612
2613         if (!strcmp(arg, "allow-any")) {
2614                 arg_missing_action = MA_ALLOW_ANY;
2615                 fetch_if_missing = 0;
2616                 fn_show_object = show_object__ma_allow_any;
2617                 return 0;
2618         }
2619
2620         if (!strcmp(arg, "allow-promisor")) {
2621                 arg_missing_action = MA_ALLOW_PROMISOR;
2622                 fetch_if_missing = 0;
2623                 fn_show_object = show_object__ma_allow_promisor;
2624                 return 0;
2625         }
2626
2627         die(_("invalid value for --missing"));
2628         return 0;
2629 }
2630
2631 static void show_edge(struct commit *commit)
2632 {
2633         add_preferred_base(&commit->object.oid);
2634 }
2635
2636 struct in_pack_object {
2637         off_t offset;
2638         struct object *object;
2639 };
2640
2641 struct in_pack {
2642         unsigned int alloc;
2643         unsigned int nr;
2644         struct in_pack_object *array;
2645 };
2646
2647 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2648 {
2649         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2650         in_pack->array[in_pack->nr].object = object;
2651         in_pack->nr++;
2652 }
2653
2654 /*
2655  * Compare the objects in the offset order, in order to emulate the
2656  * "git rev-list --objects" output that produced the pack originally.
2657  */
2658 static int ofscmp(const void *a_, const void *b_)
2659 {
2660         struct in_pack_object *a = (struct in_pack_object *)a_;
2661         struct in_pack_object *b = (struct in_pack_object *)b_;
2662
2663         if (a->offset < b->offset)
2664                 return -1;
2665         else if (a->offset > b->offset)
2666                 return 1;
2667         else
2668                 return oidcmp(&a->object->oid, &b->object->oid);
2669 }
2670
2671 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2672 {
2673         struct packed_git *p;
2674         struct in_pack in_pack;
2675         uint32_t i;
2676
2677         memset(&in_pack, 0, sizeof(in_pack));
2678
2679         for (p = get_packed_git(the_repository); p; p = p->next) {
2680                 struct object_id oid;
2681                 struct object *o;
2682
2683                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
2684                         continue;
2685                 if (open_pack_index(p))
2686                         die("cannot open pack index");
2687
2688                 ALLOC_GROW(in_pack.array,
2689                            in_pack.nr + p->num_objects,
2690                            in_pack.alloc);
2691
2692                 for (i = 0; i < p->num_objects; i++) {
2693                         nth_packed_object_oid(&oid, p, i);
2694                         o = lookup_unknown_object(oid.hash);
2695                         if (!(o->flags & OBJECT_ADDED))
2696                                 mark_in_pack_object(o, p, &in_pack);
2697                         o->flags |= OBJECT_ADDED;
2698                 }
2699         }
2700
2701         if (in_pack.nr) {
2702                 QSORT(in_pack.array, in_pack.nr, ofscmp);
2703                 for (i = 0; i < in_pack.nr; i++) {
2704                         struct object *o = in_pack.array[i].object;
2705                         add_object_entry(&o->oid, o->type, "", 0);
2706                 }
2707         }
2708         free(in_pack.array);
2709 }
2710
2711 static int add_loose_object(const struct object_id *oid, const char *path,
2712                             void *data)
2713 {
2714         enum object_type type = oid_object_info(oid, NULL);
2715
2716         if (type < 0) {
2717                 warning("loose object at %s could not be examined", path);
2718                 return 0;
2719         }
2720
2721         add_object_entry(oid, type, "", 0);
2722         return 0;
2723 }
2724
2725 /*
2726  * We actually don't even have to worry about reachability here.
2727  * add_object_entry will weed out duplicates, so we just add every
2728  * loose object we find.
2729  */
2730 static void add_unreachable_loose_objects(void)
2731 {
2732         for_each_loose_file_in_objdir(get_object_directory(),
2733                                       add_loose_object,
2734                                       NULL, NULL, NULL);
2735 }
2736
2737 static int has_sha1_pack_kept_or_nonlocal(const struct object_id *oid)
2738 {
2739         static struct packed_git *last_found = (void *)1;
2740         struct packed_git *p;
2741
2742         p = (last_found != (void *)1) ? last_found :
2743                                         get_packed_git(the_repository);
2744
2745         while (p) {
2746                 if ((!p->pack_local || p->pack_keep ||
2747                                 p->pack_keep_in_core) &&
2748                         find_pack_entry_one(oid->hash, p)) {
2749                         last_found = p;
2750                         return 1;
2751                 }
2752                 if (p == last_found)
2753                         p = get_packed_git(the_repository);
2754                 else
2755                         p = p->next;
2756                 if (p == last_found)
2757                         p = p->next;
2758         }
2759         return 0;
2760 }
2761
2762 /*
2763  * Store a list of sha1s that are should not be discarded
2764  * because they are either written too recently, or are
2765  * reachable from another object that was.
2766  *
2767  * This is filled by get_object_list.
2768  */
2769 static struct oid_array recent_objects;
2770
2771 static int loosened_object_can_be_discarded(const struct object_id *oid,
2772                                             timestamp_t mtime)
2773 {
2774         if (!unpack_unreachable_expiration)
2775                 return 0;
2776         if (mtime > unpack_unreachable_expiration)
2777                 return 0;
2778         if (oid_array_lookup(&recent_objects, oid) >= 0)
2779                 return 0;
2780         return 1;
2781 }
2782
2783 static void loosen_unused_packed_objects(struct rev_info *revs)
2784 {
2785         struct packed_git *p;
2786         uint32_t i;
2787         struct object_id oid;
2788
2789         for (p = get_packed_git(the_repository); p; p = p->next) {
2790                 if (!p->pack_local || p->pack_keep || p->pack_keep_in_core)
2791                         continue;
2792
2793                 if (open_pack_index(p))
2794                         die("cannot open pack index");
2795
2796                 for (i = 0; i < p->num_objects; i++) {
2797                         nth_packed_object_oid(&oid, p, i);
2798                         if (!packlist_find(&to_pack, oid.hash, NULL) &&
2799                             !has_sha1_pack_kept_or_nonlocal(&oid) &&
2800                             !loosened_object_can_be_discarded(&oid, p->mtime))
2801                                 if (force_object_loose(&oid, p->mtime))
2802                                         die("unable to force loose object");
2803                 }
2804         }
2805 }
2806
2807 /*
2808  * This tracks any options which pack-reuse code expects to be on, or which a
2809  * reader of the pack might not understand, and which would therefore prevent
2810  * blind reuse of what we have on disk.
2811  */
2812 static int pack_options_allow_reuse(void)
2813 {
2814         return pack_to_stdout &&
2815                allow_ofs_delta &&
2816                !ignore_packed_keep_on_disk &&
2817                !ignore_packed_keep_in_core &&
2818                (!local || !have_non_local_packs) &&
2819                !incremental;
2820 }
2821
2822 static int get_object_list_from_bitmap(struct rev_info *revs)
2823 {
2824         if (prepare_bitmap_walk(revs) < 0)
2825                 return -1;
2826
2827         if (pack_options_allow_reuse() &&
2828             !reuse_partial_packfile_from_bitmap(
2829                         &reuse_packfile,
2830                         &reuse_packfile_objects,
2831                         &reuse_packfile_offset)) {
2832                 assert(reuse_packfile_objects);
2833                 nr_result += reuse_packfile_objects;
2834                 display_progress(progress_state, nr_result);
2835         }
2836
2837         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2838         return 0;
2839 }
2840
2841 static void record_recent_object(struct object *obj,
2842                                  const char *name,
2843                                  void *data)
2844 {
2845         oid_array_append(&recent_objects, &obj->oid);
2846 }
2847
2848 static void record_recent_commit(struct commit *commit, void *data)
2849 {
2850         oid_array_append(&recent_objects, &commit->object.oid);
2851 }
2852
2853 static void get_object_list(int ac, const char **av)
2854 {
2855         struct rev_info revs;
2856         char line[1000];
2857         int flags = 0;
2858
2859         init_revisions(&revs, NULL);
2860         save_commit_buffer = 0;
2861         setup_revisions(ac, av, &revs, NULL);
2862
2863         /* make sure shallows are read */
2864         is_repository_shallow();
2865
2866         while (fgets(line, sizeof(line), stdin) != NULL) {
2867                 int len = strlen(line);
2868                 if (len && line[len - 1] == '\n')
2869                         line[--len] = 0;
2870                 if (!len)
2871                         break;
2872                 if (*line == '-') {
2873                         if (!strcmp(line, "--not")) {
2874                                 flags ^= UNINTERESTING;
2875                                 write_bitmap_index = 0;
2876                                 continue;
2877                         }
2878                         if (starts_with(line, "--shallow ")) {
2879                                 struct object_id oid;
2880                                 if (get_oid_hex(line + 10, &oid))
2881                                         die("not an SHA-1 '%s'", line + 10);
2882                                 register_shallow(&oid);
2883                                 use_bitmap_index = 0;
2884                                 continue;
2885                         }
2886                         die("not a rev '%s'", line);
2887                 }
2888                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2889                         die("bad revision '%s'", line);
2890         }
2891
2892         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2893                 return;
2894
2895         if (prepare_revision_walk(&revs))
2896                 die("revision walk setup failed");
2897         mark_edges_uninteresting(&revs, show_edge);
2898
2899         if (!fn_show_object)
2900                 fn_show_object = show_object;
2901         traverse_commit_list_filtered(&filter_options, &revs,
2902                                       show_commit, fn_show_object, NULL,
2903                                       NULL);
2904
2905         if (unpack_unreachable_expiration) {
2906                 revs.ignore_missing_links = 1;
2907                 if (add_unseen_recent_objects_to_traversal(&revs,
2908                                 unpack_unreachable_expiration))
2909                         die("unable to add recent objects");
2910                 if (prepare_revision_walk(&revs))
2911                         die("revision walk setup failed");
2912                 traverse_commit_list(&revs, record_recent_commit,
2913                                      record_recent_object, NULL);
2914         }
2915
2916         if (keep_unreachable)
2917                 add_objects_in_unpacked_packs(&revs);
2918         if (pack_loose_unreachable)
2919                 add_unreachable_loose_objects();
2920         if (unpack_unreachable)
2921                 loosen_unused_packed_objects(&revs);
2922
2923         oid_array_clear(&recent_objects);
2924 }
2925
2926 static void add_extra_kept_packs(const struct string_list *names)
2927 {
2928         struct packed_git *p;
2929
2930         if (!names->nr)
2931                 return;
2932
2933         for (p = get_packed_git(the_repository); p; p = p->next) {
2934                 const char *name = basename(p->pack_name);
2935                 int i;
2936
2937                 if (!p->pack_local)
2938                         continue;
2939
2940                 for (i = 0; i < names->nr; i++)
2941                         if (!fspathcmp(name, names->items[i].string))
2942                                 break;
2943
2944                 if (i < names->nr) {
2945                         p->pack_keep_in_core = 1;
2946                         ignore_packed_keep_in_core = 1;
2947                         continue;
2948                 }
2949         }
2950 }
2951
2952 static int option_parse_index_version(const struct option *opt,
2953                                       const char *arg, int unset)
2954 {
2955         char *c;
2956         const char *val = arg;
2957         pack_idx_opts.version = strtoul(val, &c, 10);
2958         if (pack_idx_opts.version > 2)
2959                 die(_("unsupported index version %s"), val);
2960         if (*c == ',' && c[1])
2961                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2962         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2963                 die(_("bad index version '%s'"), val);
2964         return 0;
2965 }
2966
2967 static int option_parse_unpack_unreachable(const struct option *opt,
2968                                            const char *arg, int unset)
2969 {
2970         if (unset) {
2971                 unpack_unreachable = 0;
2972                 unpack_unreachable_expiration = 0;
2973         }
2974         else {
2975                 unpack_unreachable = 1;
2976                 if (arg)
2977                         unpack_unreachable_expiration = approxidate(arg);
2978         }
2979         return 0;
2980 }
2981
2982 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2983 {
2984         int use_internal_rev_list = 0;
2985         int thin = 0;
2986         int shallow = 0;
2987         int all_progress_implied = 0;
2988         struct argv_array rp = ARGV_ARRAY_INIT;
2989         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2990         int rev_list_index = 0;
2991         struct string_list keep_pack_list = STRING_LIST_INIT_NODUP;
2992         struct option pack_objects_options[] = {
2993                 OPT_SET_INT('q', "quiet", &progress,
2994                             N_("do not show progress meter"), 0),
2995                 OPT_SET_INT(0, "progress", &progress,
2996                             N_("show progress meter"), 1),
2997                 OPT_SET_INT(0, "all-progress", &progress,
2998                             N_("show progress meter during object writing phase"), 2),
2999                 OPT_BOOL(0, "all-progress-implied",
3000                          &all_progress_implied,
3001                          N_("similar to --all-progress when progress meter is shown")),
3002                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
3003                   N_("write the pack index file in the specified idx format version"),
3004                   0, option_parse_index_version },
3005                 OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
3006                               N_("maximum size of each output pack file")),
3007                 OPT_BOOL(0, "local", &local,
3008                          N_("ignore borrowed objects from alternate object store")),
3009                 OPT_BOOL(0, "incremental", &incremental,
3010                          N_("ignore packed objects")),
3011                 OPT_INTEGER(0, "window", &window,
3012                             N_("limit pack window by objects")),
3013                 OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
3014                               N_("limit pack window by memory in addition to object limit")),
3015                 OPT_INTEGER(0, "depth", &depth,
3016                             N_("maximum length of delta chain allowed in the resulting pack")),
3017                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
3018                          N_("reuse existing deltas")),
3019                 OPT_BOOL(0, "reuse-object", &reuse_object,
3020                          N_("reuse existing objects")),
3021                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
3022                          N_("use OFS_DELTA objects")),
3023                 OPT_INTEGER(0, "threads", &delta_search_threads,
3024                             N_("use threads when searching for best delta matches")),
3025                 OPT_BOOL(0, "non-empty", &non_empty,
3026                          N_("do not create an empty pack output")),
3027                 OPT_BOOL(0, "revs", &use_internal_rev_list,
3028                          N_("read revision arguments from standard input")),
3029                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
3030                   N_("limit the objects to those that are not yet packed"),
3031                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
3032                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
3033                   N_("include objects reachable from any reference"),
3034                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
3035                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
3036                   N_("include objects referred by reflog entries"),
3037                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
3038                 { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
3039                   N_("include objects referred to by the index"),
3040                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
3041                 OPT_BOOL(0, "stdout", &pack_to_stdout,
3042                          N_("output pack to stdout")),
3043                 OPT_BOOL(0, "include-tag", &include_tag,
3044                          N_("include tag objects that refer to objects to be packed")),
3045                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
3046                          N_("keep unreachable objects")),
3047                 OPT_BOOL(0, "pack-loose-unreachable", &pack_loose_unreachable,
3048                          N_("pack loose unreachable objects")),
3049                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
3050                   N_("unpack unreachable objects newer than <time>"),
3051                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
3052                 OPT_BOOL(0, "thin", &thin,
3053                          N_("create thin packs")),
3054                 OPT_BOOL(0, "shallow", &shallow,
3055                          N_("create packs suitable for shallow fetches")),
3056                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep_on_disk,
3057                          N_("ignore packs that have companion .keep file")),
3058                 OPT_STRING_LIST(0, "keep-pack", &keep_pack_list, N_("name"),
3059                                 N_("ignore this pack")),
3060                 OPT_INTEGER(0, "compression", &pack_compression_level,
3061                             N_("pack compression level")),
3062                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
3063                             N_("do not hide commits by grafts"), 0),
3064                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
3065                          N_("use a bitmap index if available to speed up counting objects")),
3066                 OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
3067                          N_("write a bitmap index together with the pack index")),
3068                 OPT_PARSE_LIST_OBJECTS_FILTER(&filter_options),
3069                 { OPTION_CALLBACK, 0, "missing", NULL, N_("action"),
3070                   N_("handling for missing objects"), PARSE_OPT_NONEG,
3071                   option_parse_missing_action },
3072                 OPT_BOOL(0, "exclude-promisor-objects", &exclude_promisor_objects,
3073                          N_("do not pack objects in promisor packfiles")),
3074                 OPT_END(),
3075         };
3076
3077         check_replace_refs = 0;
3078
3079         reset_pack_idx_option(&pack_idx_opts);
3080         git_config(git_pack_config, NULL);
3081
3082         progress = isatty(2);
3083         argc = parse_options(argc, argv, prefix, pack_objects_options,
3084                              pack_usage, 0);
3085
3086         if (argc) {
3087                 base_name = argv[0];
3088                 argc--;
3089         }
3090         if (pack_to_stdout != !base_name || argc)
3091                 usage_with_options(pack_usage, pack_objects_options);
3092
3093         argv_array_push(&rp, "pack-objects");
3094         if (thin) {
3095                 use_internal_rev_list = 1;
3096                 argv_array_push(&rp, shallow
3097                                 ? "--objects-edge-aggressive"
3098                                 : "--objects-edge");
3099         } else
3100                 argv_array_push(&rp, "--objects");
3101
3102         if (rev_list_all) {
3103                 use_internal_rev_list = 1;
3104                 argv_array_push(&rp, "--all");
3105         }
3106         if (rev_list_reflog) {
3107                 use_internal_rev_list = 1;
3108                 argv_array_push(&rp, "--reflog");
3109         }
3110         if (rev_list_index) {
3111                 use_internal_rev_list = 1;
3112                 argv_array_push(&rp, "--indexed-objects");
3113         }
3114         if (rev_list_unpacked) {
3115                 use_internal_rev_list = 1;
3116                 argv_array_push(&rp, "--unpacked");
3117         }
3118
3119         if (exclude_promisor_objects) {
3120                 use_internal_rev_list = 1;
3121                 fetch_if_missing = 0;
3122                 argv_array_push(&rp, "--exclude-promisor-objects");
3123         }
3124
3125         if (!reuse_object)
3126                 reuse_delta = 0;
3127         if (pack_compression_level == -1)
3128                 pack_compression_level = Z_DEFAULT_COMPRESSION;
3129         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
3130                 die("bad pack compression level %d", pack_compression_level);
3131
3132         if (!delta_search_threads)      /* --threads=0 means autodetect */
3133                 delta_search_threads = online_cpus();
3134
3135 #ifdef NO_PTHREADS
3136         if (delta_search_threads != 1)
3137                 warning("no threads support, ignoring --threads");
3138 #endif
3139         if (!pack_to_stdout && !pack_size_limit)
3140                 pack_size_limit = pack_size_limit_cfg;
3141         if (pack_to_stdout && pack_size_limit)
3142                 die("--max-pack-size cannot be used to build a pack for transfer.");
3143         if (pack_size_limit && pack_size_limit < 1024*1024) {
3144                 warning("minimum pack size limit is 1 MiB");
3145                 pack_size_limit = 1024*1024;
3146         }
3147
3148         if (!pack_to_stdout && thin)
3149                 die("--thin cannot be used to build an indexable pack.");
3150
3151         if (keep_unreachable && unpack_unreachable)
3152                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
3153         if (!rev_list_all || !rev_list_reflog || !rev_list_index)
3154                 unpack_unreachable_expiration = 0;
3155
3156         if (filter_options.choice) {
3157                 if (!pack_to_stdout)
3158                         die("cannot use --filter without --stdout.");
3159                 use_bitmap_index = 0;
3160         }
3161
3162         /*
3163          * "soft" reasons not to use bitmaps - for on-disk repack by default we want
3164          *
3165          * - to produce good pack (with bitmap index not-yet-packed objects are
3166          *   packed in suboptimal order).
3167          *
3168          * - to use more robust pack-generation codepath (avoiding possible
3169          *   bugs in bitmap code and possible bitmap index corruption).
3170          */
3171         if (!pack_to_stdout)
3172                 use_bitmap_index_default = 0;
3173
3174         if (use_bitmap_index < 0)
3175                 use_bitmap_index = use_bitmap_index_default;
3176
3177         /* "hard" reasons not to use bitmaps; these just won't work at all */
3178         if (!use_internal_rev_list || (!pack_to_stdout && write_bitmap_index) || is_repository_shallow())
3179                 use_bitmap_index = 0;
3180
3181         if (pack_to_stdout || !rev_list_all)
3182                 write_bitmap_index = 0;
3183
3184         if (progress && all_progress_implied)
3185                 progress = 2;
3186
3187         add_extra_kept_packs(&keep_pack_list);
3188         if (ignore_packed_keep_on_disk) {
3189                 struct packed_git *p;
3190                 for (p = get_packed_git(the_repository); p; p = p->next)
3191                         if (p->pack_local && p->pack_keep)
3192                                 break;
3193                 if (!p) /* no keep-able packs found */
3194                         ignore_packed_keep_on_disk = 0;
3195         }
3196         if (local) {
3197                 /*
3198                  * unlike ignore_packed_keep_on_disk above, we do not
3199                  * want to unset "local" based on looking at packs, as
3200                  * it also covers non-local objects
3201                  */
3202                 struct packed_git *p;
3203                 for (p = get_packed_git(the_repository); p; p = p->next) {
3204                         if (!p->pack_local) {
3205                                 have_non_local_packs = 1;
3206                                 break;
3207                         }
3208                 }
3209         }
3210
3211         if (progress)
3212                 progress_state = start_progress(_("Counting objects"), 0);
3213         if (!use_internal_rev_list)
3214                 read_object_list_from_stdin();
3215         else {
3216                 get_object_list(rp.argc, rp.argv);
3217                 argv_array_clear(&rp);
3218         }
3219         cleanup_preferred_base();
3220         if (include_tag && nr_result)
3221                 for_each_ref(add_ref_tag, NULL);
3222         stop_progress(&progress_state);
3223
3224         if (non_empty && !nr_result)
3225                 return 0;
3226         if (nr_result)
3227                 prepare_pack(window, depth);
3228         write_pack_file();
3229         if (progress)
3230                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
3231                         " reused %"PRIu32" (delta %"PRIu32")\n",
3232                         written, written_delta, reused, reused_delta);
3233         return 0;
3234 }