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

   1 #include "builtin.h"
   2 #include "cache.h"
   3 #include "attr.h"
   4 #include "object.h"
   5 #include "blob.h"
   6 #include "commit.h"
   7 #include "tag.h"
   8 #include "tree.h"
   9 #include "delta.h"
  10 #include "pack.h"
  11 #include "pack-revindex.h"
  12 #include "csum-file.h"
  13 #include "tree-walk.h"
  14 #include "diff.h"
  15 #include "revision.h"
  16 #include "list-objects.h"
  17 #include "pack-objects.h"
  18 #include "progress.h"
  19 #include "refs.h"
  20 #include "streaming.h"
  21 #include "thread-utils.h"
  22 #include "pack-bitmap.h"
  23
  24 static const char *pack_usage[] = {
  25         N_("git pack-objects --stdout [options...] [< ref-list | < object-list]"),
  26         N_("git pack-objects [options...] base-name [< ref-list | < object-list]"),
  27         NULL
  28 };
  29
  30 /*
  31  * Objects we are going to pack are collected in the `to_pack` structure.
  32  * It contains an array (dynamically expanded) of the object data, and a map
  33  * that can resolve SHA1s to their position in the array.
  34  */
  35 static struct packing_data to_pack;
  36
  37 static struct pack_idx_entry **written_list;
  38 static uint32_t nr_result, nr_written;
  39
  40 static int non_empty;
  41 static int reuse_delta = 1, reuse_object = 1;
  42 static int keep_unreachable, unpack_unreachable, include_tag;
  43 static unsigned long unpack_unreachable_expiration;
  44 static int local;
  45 static int incremental;
  46 static int ignore_packed_keep;
  47 static int allow_ofs_delta;
  48 static struct pack_idx_option pack_idx_opts;
  49 static const char *base_name;
  50 static int progress = 1;
  51 static int window = 10;
  52 static unsigned long pack_size_limit;
  53 static int depth = 50;
  54 static int delta_search_threads;
  55 static int pack_to_stdout;
  56 static int num_preferred_base;
  57 static struct progress *progress_state;
  58 static int pack_compression_level = Z_DEFAULT_COMPRESSION;
  59 static int pack_compression_seen;
  60
  61 static struct packed_git *reuse_packfile;
  62 static uint32_t reuse_packfile_objects;
  63 static off_t reuse_packfile_offset;
  64
  65 static int use_bitmap_index = 1;
  66
  67 static unsigned long delta_cache_size = 0;
  68 static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
  69 static unsigned long cache_max_small_delta_size = 1000;
  70
  71 static unsigned long window_memory_limit = 0;
  72
  73 /*
  74  * stats
  75  */
  76 static uint32_t written, written_delta;
  77 static uint32_t reused, reused_delta;
  78
  79 static void *get_delta(struct object_entry *entry)
  80 {
  81         unsigned long size, base_size, delta_size;
  82         void *buf, *base_buf, *delta_buf;
  83         enum object_type type;
  84
  85         buf = read_sha1_file(entry->idx.sha1, &type, &size);
  86         if (!buf)
  87                 die("unable to read %s", sha1_to_hex(entry->idx.sha1));
  88         base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
  89         if (!base_buf)
  90                 die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
  91         delta_buf = diff_delta(base_buf, base_size,
  92                                buf, size, &delta_size, 0);
  93         if (!delta_buf || delta_size != entry->delta_size)
  94                 die("delta size changed");
  95         free(buf);
  96         free(base_buf);
  97         return delta_buf;
  98 }
  99
 100 static unsigned long do_compress(void **pptr, unsigned long size)
 101 {
 102         git_zstream stream;
 103         void *in, *out;
 104         unsigned long maxsize;
 105
 106         memset(&stream, 0, sizeof(stream));
 107         git_deflate_init(&stream, pack_compression_level);
 108         maxsize = git_deflate_bound(&stream, size);
 109
 110         in = *pptr;
 111         out = xmalloc(maxsize);
 112         *pptr = out;
 113
 114         stream.next_in = in;
 115         stream.avail_in = size;
 116         stream.next_out = out;
 117         stream.avail_out = maxsize;
 118         while (git_deflate(&stream, Z_FINISH) == Z_OK)
 119                 ; /* nothing */
 120         git_deflate_end(&stream);
 121
 122         free(in);
 123         return stream.total_out;
 124 }
 125
 126 static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
 127                                            const unsigned char *sha1)
 128 {
 129         git_zstream stream;
 130         unsigned char ibuf[1024 * 16];
 131         unsigned char obuf[1024 * 16];
 132         unsigned long olen = 0;
 133
 134         memset(&stream, 0, sizeof(stream));
 135         git_deflate_init(&stream, pack_compression_level);
 136
 137         for (;;) {
 138                 ssize_t readlen;
 139                 int zret = Z_OK;
 140                 readlen = read_istream(st, ibuf, sizeof(ibuf));
 141                 if (readlen == -1)
 142                         die(_("unable to read %s"), sha1_to_hex(sha1));
 143
 144                 stream.next_in = ibuf;
 145                 stream.avail_in = readlen;
 146                 while ((stream.avail_in || readlen == 0) &&
 147                        (zret == Z_OK || zret == Z_BUF_ERROR)) {
 148                         stream.next_out = obuf;
 149                         stream.avail_out = sizeof(obuf);
 150                         zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 151                         sha1write(f, obuf, stream.next_out - obuf);
 152                         olen += stream.next_out - obuf;
 153                 }
 154                 if (stream.avail_in)
 155                         die(_("deflate error (%d)"), zret);
 156                 if (readlen == 0) {
 157                         if (zret != Z_STREAM_END)
 158                                 die(_("deflate error (%d)"), zret);
 159                         break;
 160                 }
 161         }
 162         git_deflate_end(&stream);
 163         return olen;
 164 }
 165
 166 /*
 167  * we are going to reuse the existing object data as is.  make
 168  * sure it is not corrupt.
 169  */
 170 static int check_pack_inflate(struct packed_git *p,
 171                 struct pack_window **w_curs,
 172                 off_t offset,
 173                 off_t len,
 174                 unsigned long expect)
 175 {
 176         git_zstream stream;
 177         unsigned char fakebuf[4096], *in;
 178         int st;
 179
 180         memset(&stream, 0, sizeof(stream));
 181         git_inflate_init(&stream);
 182         do {
 183                 in = use_pack(p, w_curs, offset, &stream.avail_in);
 184                 stream.next_in = in;
 185                 stream.next_out = fakebuf;
 186                 stream.avail_out = sizeof(fakebuf);
 187                 st = git_inflate(&stream, Z_FINISH);
 188                 offset += stream.next_in - in;
 189         } while (st == Z_OK || st == Z_BUF_ERROR);
 190         git_inflate_end(&stream);
 191         return (st == Z_STREAM_END &&
 192                 stream.total_out == expect &&
 193                 stream.total_in == len) ? 0 : -1;
 194 }
 195
 196 static void copy_pack_data(struct sha1file *f,
 197                 struct packed_git *p,
 198                 struct pack_window **w_curs,
 199                 off_t offset,
 200                 off_t len)
 201 {
 202         unsigned char *in;
 203         unsigned long avail;
 204
 205         while (len) {
 206                 in = use_pack(p, w_curs, offset, &avail);
 207                 if (avail > len)
 208                         avail = (unsigned long)len;
 209                 sha1write(f, in, avail);
 210                 offset += avail;
 211                 len -= avail;
 212         }
 213 }
 214
 215 /* Return 0 if we will bust the pack-size limit */
 216 static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
 217                                            unsigned long limit, int usable_delta)
 218 {
 219         unsigned long size, datalen;
 220         unsigned char header[10], dheader[10];
 221         unsigned hdrlen;
 222         enum object_type type;
 223         void *buf;
 224         struct git_istream *st = NULL;
 225
 226         if (!usable_delta) {
 227                 if (entry->type == OBJ_BLOB &&
 228                     entry->size > big_file_threshold &&
 229                     (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
 230                         buf = NULL;
 231                 else {
 232                         buf = read_sha1_file(entry->idx.sha1, &type, &size);
 233                         if (!buf)
 234                                 die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
 235                 }
 236                 /*
 237                  * make sure no cached delta data remains from a
 238                  * previous attempt before a pack split occurred.
 239                  */
 240                 free(entry->delta_data);
 241                 entry->delta_data = NULL;
 242                 entry->z_delta_size = 0;
 243         } else if (entry->delta_data) {
 244                 size = entry->delta_size;
 245                 buf = entry->delta_data;
 246                 entry->delta_data = NULL;
 247                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 248                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 249         } else {
 250                 buf = get_delta(entry);
 251                 size = entry->delta_size;
 252                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 253                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 254         }
 255
 256         if (st) /* large blob case, just assume we don't compress well */
 257                 datalen = size;
 258         else if (entry->z_delta_size)
 259                 datalen = entry->z_delta_size;
 260         else
 261                 datalen = do_compress(&buf, size);
 262
 263         /*
 264          * The object header is a byte of 'type' followed by zero or
 265          * more bytes of length.
 266          */
 267         hdrlen = encode_in_pack_object_header(type, size, header);
 268
 269         if (type == OBJ_OFS_DELTA) {
 270                 /*
 271                  * Deltas with relative base contain an additional
 272                  * encoding of the relative offset for the delta
 273                  * base from this object's position in the pack.
 274                  */
 275                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 276                 unsigned pos = sizeof(dheader) - 1;
 277                 dheader[pos] = ofs & 127;
 278                 while (ofs >>= 7)
 279                         dheader[--pos] = 128 | (--ofs & 127);
 280                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 281                         if (st)
 282                                 close_istream(st);
 283                         free(buf);
 284                         return 0;
 285                 }
 286                 sha1write(f, header, hdrlen);
 287                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 288                 hdrlen += sizeof(dheader) - pos;
 289         } else if (type == OBJ_REF_DELTA) {
 290                 /*
 291                  * Deltas with a base reference contain
 292                  * an additional 20 bytes for the base sha1.
 293                  */
 294                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 295                         if (st)
 296                                 close_istream(st);
 297                         free(buf);
 298                         return 0;
 299                 }
 300                 sha1write(f, header, hdrlen);
 301                 sha1write(f, entry->delta->idx.sha1, 20);
 302                 hdrlen += 20;
 303         } else {
 304                 if (limit && hdrlen + datalen + 20 >= limit) {
 305                         if (st)
 306                                 close_istream(st);
 307                         free(buf);
 308                         return 0;
 309                 }
 310                 sha1write(f, header, hdrlen);
 311         }
 312         if (st) {
 313                 datalen = write_large_blob_data(st, f, entry->idx.sha1);
 314                 close_istream(st);
 315         } else {
 316                 sha1write(f, buf, datalen);
 317                 free(buf);
 318         }
 319
 320         return hdrlen + datalen;
 321 }
 322
 323 /* Return 0 if we will bust the pack-size limit */
 324 static unsigned long write_reuse_object(struct sha1file *f, struct object_entry *entry,
 325                                         unsigned long limit, int usable_delta)
 326 {
 327         struct packed_git *p = entry->in_pack;
 328         struct pack_window *w_curs = NULL;
 329         struct revindex_entry *revidx;
 330         off_t offset;
 331         enum object_type type = entry->type;
 332         unsigned long datalen;
 333         unsigned char header[10], dheader[10];
 334         unsigned hdrlen;
 335
 336         if (entry->delta)
 337                 type = (allow_ofs_delta && entry->delta->idx.offset) ?
 338                         OBJ_OFS_DELTA : OBJ_REF_DELTA;
 339         hdrlen = encode_in_pack_object_header(type, entry->size, header);
 340
 341         offset = entry->in_pack_offset;
 342         revidx = find_pack_revindex(p, offset);
 343         datalen = revidx[1].offset - offset;
 344         if (!pack_to_stdout && p->index_version > 1 &&
 345             check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 346                 error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
 347                 unuse_pack(&w_curs);
 348                 return write_no_reuse_object(f, entry, limit, usable_delta);
 349         }
 350
 351         offset += entry->in_pack_header_size;
 352         datalen -= entry->in_pack_header_size;
 353
 354         if (!pack_to_stdout && p->index_version == 1 &&
 355             check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 356                 error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
 357                 unuse_pack(&w_curs);
 358                 return write_no_reuse_object(f, entry, limit, usable_delta);
 359         }
 360
 361         if (type == OBJ_OFS_DELTA) {
 362                 off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 363                 unsigned pos = sizeof(dheader) - 1;
 364                 dheader[pos] = ofs & 127;
 365                 while (ofs >>= 7)
 366                         dheader[--pos] = 128 | (--ofs & 127);
 367                 if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 368                         unuse_pack(&w_curs);
 369                         return 0;
 370                 }
 371                 sha1write(f, header, hdrlen);
 372                 sha1write(f, dheader + pos, sizeof(dheader) - pos);
 373                 hdrlen += sizeof(dheader) - pos;
 374                 reused_delta++;
 375         } else if (type == OBJ_REF_DELTA) {
 376                 if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 377                         unuse_pack(&w_curs);
 378                         return 0;
 379                 }
 380                 sha1write(f, header, hdrlen);
 381                 sha1write(f, entry->delta->idx.sha1, 20);
 382                 hdrlen += 20;
 383                 reused_delta++;
 384         } else {
 385                 if (limit && hdrlen + datalen + 20 >= limit) {
 386                         unuse_pack(&w_curs);
 387                         return 0;
 388                 }
 389                 sha1write(f, header, hdrlen);
 390         }
 391         copy_pack_data(f, p, &w_curs, offset, datalen);
 392         unuse_pack(&w_curs);
 393         reused++;
 394         return hdrlen + datalen;
 395 }
 396
 397 /* Return 0 if we will bust the pack-size limit */
 398 static unsigned long write_object(struct sha1file *f,
 399                                   struct object_entry *entry,
 400                                   off_t write_offset)
 401 {
 402         unsigned long limit, len;
 403         int usable_delta, to_reuse;
 404
 405         if (!pack_to_stdout)
 406                 crc32_begin(f);
 407
 408         /* apply size limit if limited packsize and not first object */
 409         if (!pack_size_limit || !nr_written)
 410                 limit = 0;
 411         else if (pack_size_limit <= write_offset)
 412                 /*
 413                  * the earlier object did not fit the limit; avoid
 414                  * mistaking this with unlimited (i.e. limit = 0).
 415                  */
 416                 limit = 1;
 417         else
 418                 limit = pack_size_limit - write_offset;
 419
 420         if (!entry->delta)
 421                 usable_delta = 0;       /* no delta */
 422         else if (!pack_size_limit)
 423                usable_delta = 1;        /* unlimited packfile */
 424         else if (entry->delta->idx.offset == (off_t)-1)
 425                 usable_delta = 0;       /* base was written to another pack */
 426         else if (entry->delta->idx.offset)
 427                 usable_delta = 1;       /* base already exists in this pack */
 428         else
 429                 usable_delta = 0;       /* base could end up in another pack */
 430
 431         if (!reuse_object)
 432                 to_reuse = 0;   /* explicit */
 433         else if (!entry->in_pack)
 434                 to_reuse = 0;   /* can't reuse what we don't have */
 435         else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
 436                                 /* check_object() decided it for us ... */
 437                 to_reuse = usable_delta;
 438                                 /* ... but pack split may override that */
 439         else if (entry->type != entry->in_pack_type)
 440                 to_reuse = 0;   /* pack has delta which is unusable */
 441         else if (entry->delta)
 442                 to_reuse = 0;   /* we want to pack afresh */
 443         else
 444                 to_reuse = 1;   /* we have it in-pack undeltified,
 445                                  * and we do not need to deltify it.
 446                                  */
 447
 448         if (!to_reuse)
 449                 len = write_no_reuse_object(f, entry, limit, usable_delta);
 450         else
 451                 len = write_reuse_object(f, entry, limit, usable_delta);
 452         if (!len)
 453                 return 0;
 454
 455         if (usable_delta)
 456                 written_delta++;
 457         written++;
 458         if (!pack_to_stdout)
 459                 entry->idx.crc32 = crc32_end(f);
 460         return len;
 461 }
 462
 463 enum write_one_status {
 464         WRITE_ONE_SKIP = -1, /* already written */
 465         WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 466         WRITE_ONE_WRITTEN = 1, /* normal */
 467         WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 468 };
 469
 470 static enum write_one_status write_one(struct sha1file *f,
 471                                        struct object_entry *e,
 472                                        off_t *offset)
 473 {
 474         unsigned long size;
 475         int recursing;
 476
 477         /*
 478          * we set offset to 1 (which is an impossible value) to mark
 479          * the fact that this object is involved in "write its base
 480          * first before writing a deltified object" recursion.
 481          */
 482         recursing = (e->idx.offset == 1);
 483         if (recursing) {
 484                 warning("recursive delta detected for object %s",
 485                         sha1_to_hex(e->idx.sha1));
 486                 return WRITE_ONE_RECURSIVE;
 487         } else if (e->idx.offset || e->preferred_base) {
 488                 /* offset is non zero if object is written already. */
 489                 return WRITE_ONE_SKIP;
 490         }
 491
 492         /* if we are deltified, write out base object first. */
 493         if (e->delta) {
 494                 e->idx.offset = 1; /* now recurse */
 495                 switch (write_one(f, e->delta, offset)) {
 496                 case WRITE_ONE_RECURSIVE:
 497                         /* we cannot depend on this one */
 498                         e->delta = NULL;
 499                         break;
 500                 default:
 501                         break;
 502                 case WRITE_ONE_BREAK:
 503                         e->idx.offset = recursing;
 504                         return WRITE_ONE_BREAK;
 505                 }
 506         }
 507
 508         e->idx.offset = *offset;
 509         size = write_object(f, e, *offset);
 510         if (!size) {
 511                 e->idx.offset = recursing;
 512                 return WRITE_ONE_BREAK;
 513         }
 514         written_list[nr_written++] = &e->idx;
 515
 516         /* make sure off_t is sufficiently large not to wrap */
 517         if (signed_add_overflows(*offset, size))
 518                 die("pack too large for current definition of off_t");
 519         *offset += size;
 520         return WRITE_ONE_WRITTEN;
 521 }
 522
 523 static int mark_tagged(const char *path, const unsigned char *sha1, int flag,
 524                        void *cb_data)
 525 {
 526         unsigned char peeled[20];
 527         struct object_entry *entry = packlist_find(&to_pack, sha1, NULL);
 528
 529         if (entry)
 530                 entry->tagged = 1;
 531         if (!peel_ref(path, peeled)) {
 532                 entry = packlist_find(&to_pack, peeled, NULL);
 533                 if (entry)
 534                         entry->tagged = 1;
 535         }
 536         return 0;
 537 }
 538
 539 static inline void add_to_write_order(struct object_entry **wo,
 540                                unsigned int *endp,
 541                                struct object_entry *e)
 542 {
 543         if (e->filled)
 544                 return;
 545         wo[(*endp)++] = e;
 546         e->filled = 1;
 547 }
 548
 549 static void add_descendants_to_write_order(struct object_entry **wo,
 550                                            unsigned int *endp,
 551                                            struct object_entry *e)
 552 {
 553         int add_to_order = 1;
 554         while (e) {
 555                 if (add_to_order) {
 556                         struct object_entry *s;
 557                         /* add this node... */
 558                         add_to_write_order(wo, endp, e);
 559                         /* all its siblings... */
 560                         for (s = e->delta_sibling; s; s = s->delta_sibling) {
 561                                 add_to_write_order(wo, endp, s);
 562                         }
 563                 }
 564                 /* drop down a level to add left subtree nodes if possible */
 565                 if (e->delta_child) {
 566                         add_to_order = 1;
 567                         e = e->delta_child;
 568                 } else {
 569                         add_to_order = 0;
 570                         /* our sibling might have some children, it is next */
 571                         if (e->delta_sibling) {
 572                                 e = e->delta_sibling;
 573                                 continue;
 574                         }
 575                         /* go back to our parent node */
 576                         e = e->delta;
 577                         while (e && !e->delta_sibling) {
 578                                 /* we're on the right side of a subtree, keep
 579                                  * going up until we can go right again */
 580                                 e = e->delta;
 581                         }
 582                         if (!e) {
 583                                 /* done- we hit our original root node */
 584                                 return;
 585                         }
 586                         /* pass it off to sibling at this level */
 587                         e = e->delta_sibling;
 588                 }
 589         };
 590 }
 591
 592 static void add_family_to_write_order(struct object_entry **wo,
 593                                       unsigned int *endp,
 594                                       struct object_entry *e)
 595 {
 596         struct object_entry *root;
 597
 598         for (root = e; root->delta; root = root->delta)
 599                 ; /* nothing */
 600         add_descendants_to_write_order(wo, endp, root);
 601 }
 602
 603 static struct object_entry **compute_write_order(void)
 604 {
 605         unsigned int i, wo_end, last_untagged;
 606
 607         struct object_entry **wo = xmalloc(to_pack.nr_objects * sizeof(*wo));
 608         struct object_entry *objects = to_pack.objects;
 609
 610         for (i = 0; i < to_pack.nr_objects; i++) {
 611                 objects[i].tagged = 0;
 612                 objects[i].filled = 0;
 613                 objects[i].delta_child = NULL;
 614                 objects[i].delta_sibling = NULL;
 615         }
 616
 617         /*
 618          * Fully connect delta_child/delta_sibling network.
 619          * Make sure delta_sibling is sorted in the original
 620          * recency order.
 621          */
 622         for (i = to_pack.nr_objects; i > 0;) {
 623                 struct object_entry *e = &objects[--i];
 624                 if (!e->delta)
 625                         continue;
 626                 /* Mark me as the first child */
 627                 e->delta_sibling = e->delta->delta_child;
 628                 e->delta->delta_child = e;
 629         }
 630
 631         /*
 632          * Mark objects that are at the tip of tags.
 633          */
 634         for_each_tag_ref(mark_tagged, NULL);
 635
 636         /*
 637          * Give the objects in the original recency order until
 638          * we see a tagged tip.
 639          */
 640         for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
 641                 if (objects[i].tagged)
 642                         break;
 643                 add_to_write_order(wo, &wo_end, &objects[i]);
 644         }
 645         last_untagged = i;
 646
 647         /*
 648          * Then fill all the tagged tips.
 649          */
 650         for (; i < to_pack.nr_objects; i++) {
 651                 if (objects[i].tagged)
 652                         add_to_write_order(wo, &wo_end, &objects[i]);
 653         }
 654
 655         /*
 656          * And then all remaining commits and tags.
 657          */
 658         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 659                 if (objects[i].type != OBJ_COMMIT &&
 660                     objects[i].type != OBJ_TAG)
 661                         continue;
 662                 add_to_write_order(wo, &wo_end, &objects[i]);
 663         }
 664
 665         /*
 666          * And then all the trees.
 667          */
 668         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 669                 if (objects[i].type != OBJ_TREE)
 670                         continue;
 671                 add_to_write_order(wo, &wo_end, &objects[i]);
 672         }
 673
 674         /*
 675          * Finally all the rest in really tight order
 676          */
 677         for (i = last_untagged; i < to_pack.nr_objects; i++) {
 678                 if (!objects[i].filled)
 679                         add_family_to_write_order(wo, &wo_end, &objects[i]);
 680         }
 681
 682         if (wo_end != to_pack.nr_objects)
 683                 die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
 684
 685         return wo;
 686 }
 687
 688 static off_t write_reused_pack(struct sha1file *f)
 689 {
 690         unsigned char buffer[8192];
 691         off_t to_write;
 692         int fd;
 693
 694         if (!is_pack_valid(reuse_packfile))
 695                 die("packfile is invalid: %s", reuse_packfile->pack_name);
 696
 697         fd = git_open_noatime(reuse_packfile->pack_name);
 698         if (fd < 0)
 699                 die_errno("unable to open packfile for reuse: %s",
 700                           reuse_packfile->pack_name);
 701
 702         if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
 703                 die_errno("unable to seek in reused packfile");
 704
 705         if (reuse_packfile_offset < 0)
 706                 reuse_packfile_offset = reuse_packfile->pack_size - 20;
 707
 708         to_write = reuse_packfile_offset - sizeof(struct pack_header);
 709
 710         while (to_write) {
 711                 int read_pack = xread(fd, buffer, sizeof(buffer));
 712
 713                 if (read_pack <= 0)
 714                         die_errno("unable to read from reused packfile");
 715
 716                 if (read_pack > to_write)
 717                         read_pack = to_write;
 718
 719                 sha1write(f, buffer, read_pack);
 720                 to_write -= read_pack;
 721         }
 722
 723         close(fd);
 724         written += reuse_packfile_objects;
 725         return reuse_packfile_offset - sizeof(struct pack_header);
 726 }
 727
 728 static void write_pack_file(void)
 729 {
 730         uint32_t i = 0, j;
 731         struct sha1file *f;
 732         off_t offset;
 733         uint32_t nr_remaining = nr_result;
 734         time_t last_mtime = 0;
 735         struct object_entry **write_order;
 736
 737         if (progress > pack_to_stdout)
 738                 progress_state = start_progress("Writing objects", nr_result);
 739         written_list = xmalloc(to_pack.nr_objects * sizeof(*written_list));
 740         write_order = compute_write_order();
 741
 742         do {
 743                 unsigned char sha1[20];
 744                 char *pack_tmp_name = NULL;
 745
 746                 if (pack_to_stdout)
 747                         f = sha1fd_throughput(1, "<stdout>", progress_state);
 748                 else
 749                         f = create_tmp_packfile(&pack_tmp_name);
 750
 751                 offset = write_pack_header(f, nr_remaining);
 752                 if (!offset)
 753                         die_errno("unable to write pack header");
 754
 755                 if (reuse_packfile) {
 756                         off_t packfile_size;
 757                         assert(pack_to_stdout);
 758
 759                         packfile_size = write_reused_pack(f);
 760                         offset += packfile_size;
 761                 }
 762
 763                 nr_written = 0;
 764                 for (; i < to_pack.nr_objects; i++) {
 765                         struct object_entry *e = write_order[i];
 766                         if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 767                                 break;
 768                         display_progress(progress_state, written);
 769                 }
 770
 771                 /*
 772                  * Did we write the wrong # entries in the header?
 773                  * If so, rewrite it like in fast-import
 774                  */
 775                 if (pack_to_stdout) {
 776                         sha1close(f, sha1, CSUM_CLOSE);
 777                 } else if (nr_written == nr_remaining) {
 778                         sha1close(f, sha1, CSUM_FSYNC);
 779                 } else {
 780                         int fd = sha1close(f, sha1, 0);
 781                         fixup_pack_header_footer(fd, sha1, pack_tmp_name,
 782                                                  nr_written, sha1, offset);
 783                         close(fd);
 784                 }
 785
 786                 if (!pack_to_stdout) {
 787                         struct stat st;
 788                         char tmpname[PATH_MAX];
 789
 790                         /*
 791                          * Packs are runtime accessed in their mtime
 792                          * order since newer packs are more likely to contain
 793                          * younger objects.  So if we are creating multiple
 794                          * packs then we should modify the mtime of later ones
 795                          * to preserve this property.
 796                          */
 797                         if (stat(pack_tmp_name, &st) < 0) {
 798                                 warning("failed to stat %s: %s",
 799                                         pack_tmp_name, strerror(errno));
 800                         } else if (!last_mtime) {
 801                                 last_mtime = st.st_mtime;
 802                         } else {
 803                                 struct utimbuf utb;
 804                                 utb.actime = st.st_atime;
 805                                 utb.modtime = --last_mtime;
 806                                 if (utime(pack_tmp_name, &utb) < 0)
 807                                         warning("failed utime() on %s: %s",
 808                                                 tmpname, strerror(errno));
 809                         }
 810
 811                         /* Enough space for "-<sha-1>.pack"? */
 812                         if (sizeof(tmpname) <= strlen(base_name) + 50)
 813                                 die("pack base name '%s' too long", base_name);
 814                         snprintf(tmpname, sizeof(tmpname), "%s-", base_name);
 815                         finish_tmp_packfile(tmpname, pack_tmp_name,
 816                                             written_list, nr_written,
 817                                             &pack_idx_opts, sha1);
 818                         free(pack_tmp_name);
 819                         puts(sha1_to_hex(sha1));
 820                 }
 821
 822                 /* mark written objects as written to previous pack */
 823                 for (j = 0; j < nr_written; j++) {
 824                         written_list[j]->offset = (off_t)-1;
 825                 }
 826                 nr_remaining -= nr_written;
 827         } while (nr_remaining && i < to_pack.nr_objects);
 828
 829         free(written_list);
 830         free(write_order);
 831         stop_progress(&progress_state);
 832         if (written != nr_result)
 833                 die("wrote %"PRIu32" objects while expecting %"PRIu32,
 834                         written, nr_result);
 835 }
 836
 837 static void setup_delta_attr_check(struct git_attr_check *check)
 838 {
 839         static struct git_attr *attr_delta;
 840
 841         if (!attr_delta)
 842                 attr_delta = git_attr("delta");
 843
 844         check[0].attr = attr_delta;
 845 }
 846
 847 static int no_try_delta(const char *path)
 848 {
 849         struct git_attr_check check[1];
 850
 851         setup_delta_attr_check(check);
 852         if (git_check_attr(path, ARRAY_SIZE(check), check))
 853                 return 0;
 854         if (ATTR_FALSE(check->value))
 855                 return 1;
 856         return 0;
 857 }
 858
 859 /*
 860  * When adding an object, check whether we have already added it
 861  * to our packing list. If so, we can skip. However, if we are
 862  * being asked to excludei t, but the previous mention was to include
 863  * it, make sure to adjust its flags and tweak our numbers accordingly.
 864  *
 865  * As an optimization, we pass out the index position where we would have
 866  * found the item, since that saves us from having to look it up again a
 867  * few lines later when we want to add the new entry.
 868  */
 869 static int have_duplicate_entry(const unsigned char *sha1,
 870                                 int exclude,
 871                                 uint32_t *index_pos)
 872 {
 873         struct object_entry *entry;
 874
 875         entry = packlist_find(&to_pack, sha1, index_pos);
 876         if (!entry)
 877                 return 0;
 878
 879         if (exclude) {
 880                 if (!entry->preferred_base)
 881                         nr_result--;
 882                 entry->preferred_base = 1;
 883         }
 884
 885         return 1;
 886 }
 887
 888 /*
 889  * Check whether we want the object in the pack (e.g., we do not want
 890  * objects found in non-local stores if the "--local" option was used).
 891  *
 892  * As a side effect of this check, we will find the packed version of this
 893  * object, if any. We therefore pass out the pack information to avoid having
 894  * to look it up again later.
 895  */
 896 static int want_object_in_pack(const unsigned char *sha1,
 897                                int exclude,
 898                                struct packed_git **found_pack,
 899                                off_t *found_offset)
 900 {
 901         struct packed_git *p;
 902
 903         if (!exclude && local && has_loose_object_nonlocal(sha1))
 904                 return 0;
 905
 906         *found_pack = NULL;
 907         *found_offset = 0;
 908
 909         for (p = packed_git; p; p = p->next) {
 910                 off_t offset = find_pack_entry_one(sha1, p);
 911                 if (offset) {
 912                         if (!*found_pack) {
 913                                 if (!is_pack_valid(p)) {
 914                                         warning("packfile %s cannot be accessed", p->pack_name);
 915                                         continue;
 916                                 }
 917                                 *found_offset = offset;
 918                                 *found_pack = p;
 919                         }
 920                         if (exclude)
 921                                 return 1;
 922                         if (incremental)
 923                                 return 0;
 924                         if (local && !p->pack_local)
 925                                 return 0;
 926                         if (ignore_packed_keep && p->pack_local && p->pack_keep)
 927                                 return 0;
 928                 }
 929         }
 930
 931         return 1;
 932 }
 933
 934 static void create_object_entry(const unsigned char *sha1,
 935                                 enum object_type type,
 936                                 uint32_t hash,
 937                                 int exclude,
 938                                 int no_try_delta,
 939                                 uint32_t index_pos,
 940                                 struct packed_git *found_pack,
 941                                 off_t found_offset)
 942 {
 943         struct object_entry *entry;
 944
 945         entry = packlist_alloc(&to_pack, sha1, index_pos);
 946         entry->hash = hash;
 947         if (type)
 948                 entry->type = type;
 949         if (exclude)
 950                 entry->preferred_base = 1;
 951         else
 952                 nr_result++;
 953         if (found_pack) {
 954                 entry->in_pack = found_pack;
 955                 entry->in_pack_offset = found_offset;
 956         }
 957
 958         entry->no_try_delta = no_try_delta;
 959 }
 960
 961 static int add_object_entry(const unsigned char *sha1, enum object_type type,
 962                             const char *name, int exclude)
 963 {
 964         struct packed_git *found_pack;
 965         off_t found_offset;
 966         uint32_t index_pos;
 967
 968         if (have_duplicate_entry(sha1, exclude, &index_pos))
 969                 return 0;
 970
 971         if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset))
 972                 return 0;
 973
 974         create_object_entry(sha1, type, pack_name_hash(name),
 975                             exclude, name && no_try_delta(name),
 976                             index_pos, found_pack, found_offset);
 977
 978         display_progress(progress_state, to_pack.nr_objects);
 979         return 1;
 980 }
 981
 982 static int add_object_entry_from_bitmap(const unsigned char *sha1,
 983                                         enum object_type type,
 984                                         int flags, uint32_t name_hash,
 985                                         struct packed_git *pack, off_t offset)
 986 {
 987         uint32_t index_pos;
 988
 989         if (have_duplicate_entry(sha1, 0, &index_pos))
 990                 return 0;
 991
 992         create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
 993
 994         display_progress(progress_state, to_pack.nr_objects);
 995         return 1;
 996 }
 997
 998 struct pbase_tree_cache {
 999         unsigned char sha1[20];
1000         int ref;
1001         int temporary;
1002         void *tree_data;
1003         unsigned long tree_size;
1004 };
1005
1006 static struct pbase_tree_cache *(pbase_tree_cache[256]);
1007 static int pbase_tree_cache_ix(const unsigned char *sha1)
1008 {
1009         return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1010 }
1011 static int pbase_tree_cache_ix_incr(int ix)
1012 {
1013         return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1014 }
1015
1016 static struct pbase_tree {
1017         struct pbase_tree *next;
1018         /* This is a phony "cache" entry; we are not
1019          * going to evict it nor find it through _get()
1020          * mechanism -- this is for the toplevel node that
1021          * would almost always change with any commit.
1022          */
1023         struct pbase_tree_cache pcache;
1024 } *pbase_tree;
1025
1026 static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1027 {
1028         struct pbase_tree_cache *ent, *nent;
1029         void *data;
1030         unsigned long size;
1031         enum object_type type;
1032         int neigh;
1033         int my_ix = pbase_tree_cache_ix(sha1);
1034         int available_ix = -1;
1035
1036         /* pbase-tree-cache acts as a limited hashtable.
1037          * your object will be found at your index or within a few
1038          * slots after that slot if it is cached.
1039          */
1040         for (neigh = 0; neigh < 8; neigh++) {
1041                 ent = pbase_tree_cache[my_ix];
1042                 if (ent && !hashcmp(ent->sha1, sha1)) {
1043                         ent->ref++;
1044                         return ent;
1045                 }
1046                 else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1047                          ((0 <= available_ix) &&
1048                           (!ent && pbase_tree_cache[available_ix])))
1049                         available_ix = my_ix;
1050                 if (!ent)
1051                         break;
1052                 my_ix = pbase_tree_cache_ix_incr(my_ix);
1053         }
1054
1055         /* Did not find one.  Either we got a bogus request or
1056          * we need to read and perhaps cache.
1057          */
1058         data = read_sha1_file(sha1, &type, &size);
1059         if (!data)
1060                 return NULL;
1061         if (type != OBJ_TREE) {
1062                 free(data);
1063                 return NULL;
1064         }
1065
1066         /* We need to either cache or return a throwaway copy */
1067
1068         if (available_ix < 0)
1069                 ent = NULL;
1070         else {
1071                 ent = pbase_tree_cache[available_ix];
1072                 my_ix = available_ix;
1073         }
1074
1075         if (!ent) {
1076                 nent = xmalloc(sizeof(*nent));
1077                 nent->temporary = (available_ix < 0);
1078         }
1079         else {
1080                 /* evict and reuse */
1081                 free(ent->tree_data);
1082                 nent = ent;
1083         }
1084         hashcpy(nent->sha1, sha1);
1085         nent->tree_data = data;
1086         nent->tree_size = size;
1087         nent->ref = 1;
1088         if (!nent->temporary)
1089                 pbase_tree_cache[my_ix] = nent;
1090         return nent;
1091 }
1092
1093 static void pbase_tree_put(struct pbase_tree_cache *cache)
1094 {
1095         if (!cache->temporary) {
1096                 cache->ref--;
1097                 return;
1098         }
1099         free(cache->tree_data);
1100         free(cache);
1101 }
1102
1103 static int name_cmp_len(const char *name)
1104 {
1105         int i;
1106         for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1107                 ;
1108         return i;
1109 }
1110
1111 static void add_pbase_object(struct tree_desc *tree,
1112                              const char *name,
1113                              int cmplen,
1114                              const char *fullname)
1115 {
1116         struct name_entry entry;
1117         int cmp;
1118
1119         while (tree_entry(tree,&entry)) {
1120                 if (S_ISGITLINK(entry.mode))
1121                         continue;
1122                 cmp = tree_entry_len(&entry) != cmplen ? 1 :
1123                       memcmp(name, entry.path, cmplen);
1124                 if (cmp > 0)
1125                         continue;
1126                 if (cmp < 0)
1127                         return;
1128                 if (name[cmplen] != '/') {
1129                         add_object_entry(entry.sha1,
1130                                          object_type(entry.mode),
1131                                          fullname, 1);
1132                         return;
1133                 }
1134                 if (S_ISDIR(entry.mode)) {
1135                         struct tree_desc sub;
1136                         struct pbase_tree_cache *tree;
1137                         const char *down = name+cmplen+1;
1138                         int downlen = name_cmp_len(down);
1139
1140                         tree = pbase_tree_get(entry.sha1);
1141                         if (!tree)
1142                                 return;
1143                         init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1144
1145                         add_pbase_object(&sub, down, downlen, fullname);
1146                         pbase_tree_put(tree);
1147                 }
1148         }
1149 }
1150
1151 static unsigned *done_pbase_paths;
1152 static int done_pbase_paths_num;
1153 static int done_pbase_paths_alloc;
1154 static int done_pbase_path_pos(unsigned hash)
1155 {
1156         int lo = 0;
1157         int hi = done_pbase_paths_num;
1158         while (lo < hi) {
1159                 int mi = (hi + lo) / 2;
1160                 if (done_pbase_paths[mi] == hash)
1161                         return mi;
1162                 if (done_pbase_paths[mi] < hash)
1163                         hi = mi;
1164                 else
1165                         lo = mi + 1;
1166         }
1167         return -lo-1;
1168 }
1169
1170 static int check_pbase_path(unsigned hash)
1171 {
1172         int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1173         if (0 <= pos)
1174                 return 1;
1175         pos = -pos - 1;
1176         if (done_pbase_paths_alloc <= done_pbase_paths_num) {
1177                 done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
1178                 done_pbase_paths = xrealloc(done_pbase_paths,
1179                                             done_pbase_paths_alloc *
1180                                             sizeof(unsigned));
1181         }
1182         done_pbase_paths_num++;
1183         if (pos < done_pbase_paths_num)
1184                 memmove(done_pbase_paths + pos + 1,
1185                         done_pbase_paths + pos,
1186                         (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1187         done_pbase_paths[pos] = hash;
1188         return 0;
1189 }
1190
1191 static void add_preferred_base_object(const char *name)
1192 {
1193         struct pbase_tree *it;
1194         int cmplen;
1195         unsigned hash = pack_name_hash(name);
1196
1197         if (!num_preferred_base || check_pbase_path(hash))
1198                 return;
1199
1200         cmplen = name_cmp_len(name);
1201         for (it = pbase_tree; it; it = it->next) {
1202                 if (cmplen == 0) {
1203                         add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1204                 }
1205                 else {
1206                         struct tree_desc tree;
1207                         init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1208                         add_pbase_object(&tree, name, cmplen, name);
1209                 }
1210         }
1211 }
1212
1213 static void add_preferred_base(unsigned char *sha1)
1214 {
1215         struct pbase_tree *it;
1216         void *data;
1217         unsigned long size;
1218         unsigned char tree_sha1[20];
1219
1220         if (window <= num_preferred_base++)
1221                 return;
1222
1223         data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1224         if (!data)
1225                 return;
1226
1227         for (it = pbase_tree; it; it = it->next) {
1228                 if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1229                         free(data);
1230                         return;
1231                 }
1232         }
1233
1234         it = xcalloc(1, sizeof(*it));
1235         it->next = pbase_tree;
1236         pbase_tree = it;
1237
1238         hashcpy(it->pcache.sha1, tree_sha1);
1239         it->pcache.tree_data = data;
1240         it->pcache.tree_size = size;
1241 }
1242
1243 static void cleanup_preferred_base(void)
1244 {
1245         struct pbase_tree *it;
1246         unsigned i;
1247
1248         it = pbase_tree;
1249         pbase_tree = NULL;
1250         while (it) {
1251                 struct pbase_tree *this = it;
1252                 it = this->next;
1253                 free(this->pcache.tree_data);
1254                 free(this);
1255         }
1256
1257         for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1258                 if (!pbase_tree_cache[i])
1259                         continue;
1260                 free(pbase_tree_cache[i]->tree_data);
1261                 free(pbase_tree_cache[i]);
1262                 pbase_tree_cache[i] = NULL;
1263         }
1264
1265         free(done_pbase_paths);
1266         done_pbase_paths = NULL;
1267         done_pbase_paths_num = done_pbase_paths_alloc = 0;
1268 }
1269
1270 static void check_object(struct object_entry *entry)
1271 {
1272         if (entry->in_pack) {
1273                 struct packed_git *p = entry->in_pack;
1274                 struct pack_window *w_curs = NULL;
1275                 const unsigned char *base_ref = NULL;
1276                 struct object_entry *base_entry;
1277                 unsigned long used, used_0;
1278                 unsigned long avail;
1279                 off_t ofs;
1280                 unsigned char *buf, c;
1281
1282                 buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1283
1284                 /*
1285                  * We want in_pack_type even if we do not reuse delta
1286                  * since non-delta representations could still be reused.
1287                  */
1288                 used = unpack_object_header_buffer(buf, avail,
1289                                                    &entry->in_pack_type,
1290                                                    &entry->size);
1291                 if (used == 0)
1292                         goto give_up;
1293
1294                 /*
1295                  * Determine if this is a delta and if so whether we can
1296                  * reuse it or not.  Otherwise let's find out as cheaply as
1297                  * possible what the actual type and size for this object is.
1298                  */
1299                 switch (entry->in_pack_type) {
1300                 default:
1301                         /* Not a delta hence we've already got all we need. */
1302                         entry->type = entry->in_pack_type;
1303                         entry->in_pack_header_size = used;
1304                         if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1305                                 goto give_up;
1306                         unuse_pack(&w_curs);
1307                         return;
1308                 case OBJ_REF_DELTA:
1309                         if (reuse_delta && !entry->preferred_base)
1310                                 base_ref = use_pack(p, &w_curs,
1311                                                 entry->in_pack_offset + used, NULL);
1312                         entry->in_pack_header_size = used + 20;
1313                         break;
1314                 case OBJ_OFS_DELTA:
1315                         buf = use_pack(p, &w_curs,
1316                                        entry->in_pack_offset + used, NULL);
1317                         used_0 = 0;
1318                         c = buf[used_0++];
1319                         ofs = c & 127;
1320                         while (c & 128) {
1321                                 ofs += 1;
1322                                 if (!ofs || MSB(ofs, 7)) {
1323                                         error("delta base offset overflow in pack for %s",
1324                                               sha1_to_hex(entry->idx.sha1));
1325                                         goto give_up;
1326                                 }
1327                                 c = buf[used_0++];
1328                                 ofs = (ofs << 7) + (c & 127);
1329                         }
1330                         ofs = entry->in_pack_offset - ofs;
1331                         if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1332                                 error("delta base offset out of bound for %s",
1333                                       sha1_to_hex(entry->idx.sha1));
1334                                 goto give_up;
1335                         }
1336                         if (reuse_delta && !entry->preferred_base) {
1337                                 struct revindex_entry *revidx;
1338                                 revidx = find_pack_revindex(p, ofs);
1339                                 if (!revidx)
1340                                         goto give_up;
1341                                 base_ref = nth_packed_object_sha1(p, revidx->nr);
1342                         }
1343                         entry->in_pack_header_size = used + used_0;
1344                         break;
1345                 }
1346
1347                 if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1348                         /*
1349                          * If base_ref was set above that means we wish to
1350                          * reuse delta data, and we even found that base
1351                          * in the list of objects we want to pack. Goodie!
1352                          *
1353                          * Depth value does not matter - find_deltas() will
1354                          * never consider reused delta as the base object to
1355                          * deltify other objects against, in order to avoid
1356                          * circular deltas.
1357                          */
1358                         entry->type = entry->in_pack_type;
1359                         entry->delta = base_entry;
1360                         entry->delta_size = entry->size;
1361                         entry->delta_sibling = base_entry->delta_child;
1362                         base_entry->delta_child = entry;
1363                         unuse_pack(&w_curs);
1364                         return;
1365                 }
1366
1367                 if (entry->type) {
1368                         /*
1369                          * This must be a delta and we already know what the
1370                          * final object type is.  Let's extract the actual
1371                          * object size from the delta header.
1372                          */
1373                         entry->size = get_size_from_delta(p, &w_curs,
1374                                         entry->in_pack_offset + entry->in_pack_header_size);
1375                         if (entry->size == 0)
1376                                 goto give_up;
1377                         unuse_pack(&w_curs);
1378                         return;
1379                 }
1380
1381                 /*
1382                  * No choice but to fall back to the recursive delta walk
1383                  * with sha1_object_info() to find about the object type
1384                  * at this point...
1385                  */
1386                 give_up:
1387                 unuse_pack(&w_curs);
1388         }
1389
1390         entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1391         /*
1392          * The error condition is checked in prepare_pack().  This is
1393          * to permit a missing preferred base object to be ignored
1394          * as a preferred base.  Doing so can result in a larger
1395          * pack file, but the transfer will still take place.
1396          */
1397 }
1398
1399 static int pack_offset_sort(const void *_a, const void *_b)
1400 {
1401         const struct object_entry *a = *(struct object_entry **)_a;
1402         const struct object_entry *b = *(struct object_entry **)_b;
1403
1404         /* avoid filesystem trashing with loose objects */
1405         if (!a->in_pack && !b->in_pack)
1406                 return hashcmp(a->idx.sha1, b->idx.sha1);
1407
1408         if (a->in_pack < b->in_pack)
1409                 return -1;
1410         if (a->in_pack > b->in_pack)
1411                 return 1;
1412         return a->in_pack_offset < b->in_pack_offset ? -1 :
1413                         (a->in_pack_offset > b->in_pack_offset);
1414 }
1415
1416 static void get_object_details(void)
1417 {
1418         uint32_t i;
1419         struct object_entry **sorted_by_offset;
1420
1421         sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1422         for (i = 0; i < to_pack.nr_objects; i++)
1423                 sorted_by_offset[i] = to_pack.objects + i;
1424         qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1425
1426         for (i = 0; i < to_pack.nr_objects; i++) {
1427                 struct object_entry *entry = sorted_by_offset[i];
1428                 check_object(entry);
1429                 if (big_file_threshold < entry->size)
1430                         entry->no_try_delta = 1;
1431         }
1432
1433         free(sorted_by_offset);
1434 }
1435
1436 /*
1437  * We search for deltas in a list sorted by type, by filename hash, and then
1438  * by size, so that we see progressively smaller and smaller files.
1439  * That's because we prefer deltas to be from the bigger file
1440  * to the smaller -- deletes are potentially cheaper, but perhaps
1441  * more importantly, the bigger file is likely the more recent
1442  * one.  The deepest deltas are therefore the oldest objects which are
1443  * less susceptible to be accessed often.
1444  */
1445 static int type_size_sort(const void *_a, const void *_b)
1446 {
1447         const struct object_entry *a = *(struct object_entry **)_a;
1448         const struct object_entry *b = *(struct object_entry **)_b;
1449
1450         if (a->type > b->type)
1451                 return -1;
1452         if (a->type < b->type)
1453                 return 1;
1454         if (a->hash > b->hash)
1455                 return -1;
1456         if (a->hash < b->hash)
1457                 return 1;
1458         if (a->preferred_base > b->preferred_base)
1459                 return -1;
1460         if (a->preferred_base < b->preferred_base)
1461                 return 1;
1462         if (a->size > b->size)
1463                 return -1;
1464         if (a->size < b->size)
1465                 return 1;
1466         return a < b ? -1 : (a > b);  /* newest first */
1467 }
1468
1469 struct unpacked {
1470         struct object_entry *entry;
1471         void *data;
1472         struct delta_index *index;
1473         unsigned depth;
1474 };
1475
1476 static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1477                            unsigned long delta_size)
1478 {
1479         if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1480                 return 0;
1481
1482         if (delta_size < cache_max_small_delta_size)
1483                 return 1;
1484
1485         /* cache delta, if objects are large enough compared to delta size */
1486         if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1487                 return 1;
1488
1489         return 0;
1490 }
1491
1492 #ifndef NO_PTHREADS
1493
1494 static pthread_mutex_t read_mutex;
1495 #define read_lock()             pthread_mutex_lock(&read_mutex)
1496 #define read_unlock()           pthread_mutex_unlock(&read_mutex)
1497
1498 static pthread_mutex_t cache_mutex;
1499 #define cache_lock()            pthread_mutex_lock(&cache_mutex)
1500 #define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1501
1502 static pthread_mutex_t progress_mutex;
1503 #define progress_lock()         pthread_mutex_lock(&progress_mutex)
1504 #define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1505
1506 #else
1507
1508 #define read_lock()             (void)0
1509 #define read_unlock()           (void)0
1510 #define cache_lock()            (void)0
1511 #define cache_unlock()          (void)0
1512 #define progress_lock()         (void)0
1513 #define progress_unlock()       (void)0
1514
1515 #endif
1516
1517 static int try_delta(struct unpacked *trg, struct unpacked *src,
1518                      unsigned max_depth, unsigned long *mem_usage)
1519 {
1520         struct object_entry *trg_entry = trg->entry;
1521         struct object_entry *src_entry = src->entry;
1522         unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1523         unsigned ref_depth;
1524         enum object_type type;
1525         void *delta_buf;
1526
1527         /* Don't bother doing diffs between different types */
1528         if (trg_entry->type != src_entry->type)
1529                 return -1;
1530
1531         /*
1532          * We do not bother to try a delta that we discarded on an
1533          * earlier try, but only when reusing delta data.  Note that
1534          * src_entry that is marked as the preferred_base should always
1535          * be considered, as even if we produce a suboptimal delta against
1536          * it, we will still save the transfer cost, as we already know
1537          * the other side has it and we won't send src_entry at all.
1538          */
1539         if (reuse_delta && trg_entry->in_pack &&
1540             trg_entry->in_pack == src_entry->in_pack &&
1541             !src_entry->preferred_base &&
1542             trg_entry->in_pack_type != OBJ_REF_DELTA &&
1543             trg_entry->in_pack_type != OBJ_OFS_DELTA)
1544                 return 0;
1545
1546         /* Let's not bust the allowed depth. */
1547         if (src->depth >= max_depth)
1548                 return 0;
1549
1550         /* Now some size filtering heuristics. */
1551         trg_size = trg_entry->size;
1552         if (!trg_entry->delta) {
1553                 max_size = trg_size/2 - 20;
1554                 ref_depth = 1;
1555         } else {
1556                 max_size = trg_entry->delta_size;
1557                 ref_depth = trg->depth;
1558         }
1559         max_size = (uint64_t)max_size * (max_depth - src->depth) /
1560                                                 (max_depth - ref_depth + 1);
1561         if (max_size == 0)
1562                 return 0;
1563         src_size = src_entry->size;
1564         sizediff = src_size < trg_size ? trg_size - src_size : 0;
1565         if (sizediff >= max_size)
1566                 return 0;
1567         if (trg_size < src_size / 32)
1568                 return 0;
1569
1570         /* Load data if not already done */
1571         if (!trg->data) {
1572                 read_lock();
1573                 trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1574                 read_unlock();
1575                 if (!trg->data)
1576                         die("object %s cannot be read",
1577                             sha1_to_hex(trg_entry->idx.sha1));
1578                 if (sz != trg_size)
1579                         die("object %s inconsistent object length (%lu vs %lu)",
1580                             sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1581                 *mem_usage += sz;
1582         }
1583         if (!src->data) {
1584                 read_lock();
1585                 src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1586                 read_unlock();
1587                 if (!src->data) {
1588                         if (src_entry->preferred_base) {
1589                                 static int warned = 0;
1590                                 if (!warned++)
1591                                         warning("object %s cannot be read",
1592                                                 sha1_to_hex(src_entry->idx.sha1));
1593                                 /*
1594                                  * Those objects are not included in the
1595                                  * resulting pack.  Be resilient and ignore
1596                                  * them if they can't be read, in case the
1597                                  * pack could be created nevertheless.
1598                                  */
1599                                 return 0;
1600                         }
1601                         die("object %s cannot be read",
1602                             sha1_to_hex(src_entry->idx.sha1));
1603                 }
1604                 if (sz != src_size)
1605                         die("object %s inconsistent object length (%lu vs %lu)",
1606                             sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1607                 *mem_usage += sz;
1608         }
1609         if (!src->index) {
1610                 src->index = create_delta_index(src->data, src_size);
1611                 if (!src->index) {
1612                         static int warned = 0;
1613                         if (!warned++)
1614                                 warning("suboptimal pack - out of memory");
1615                         return 0;
1616                 }
1617                 *mem_usage += sizeof_delta_index(src->index);
1618         }
1619
1620         delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1621         if (!delta_buf)
1622                 return 0;
1623
1624         if (trg_entry->delta) {
1625                 /* Prefer only shallower same-sized deltas. */
1626                 if (delta_size == trg_entry->delta_size &&
1627                     src->depth + 1 >= trg->depth) {
1628                         free(delta_buf);
1629                         return 0;
1630                 }
1631         }
1632
1633         /*
1634          * Handle memory allocation outside of the cache
1635          * accounting lock.  Compiler will optimize the strangeness
1636          * away when NO_PTHREADS is defined.
1637          */
1638         free(trg_entry->delta_data);
1639         cache_lock();
1640         if (trg_entry->delta_data) {
1641                 delta_cache_size -= trg_entry->delta_size;
1642                 trg_entry->delta_data = NULL;
1643         }
1644         if (delta_cacheable(src_size, trg_size, delta_size)) {
1645                 delta_cache_size += delta_size;
1646                 cache_unlock();
1647                 trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1648         } else {
1649                 cache_unlock();
1650                 free(delta_buf);
1651         }
1652
1653         trg_entry->delta = src_entry;
1654         trg_entry->delta_size = delta_size;
1655         trg->depth = src->depth + 1;
1656
1657         return 1;
1658 }
1659
1660 static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1661 {
1662         struct object_entry *child = me->delta_child;
1663         unsigned int m = n;
1664         while (child) {
1665                 unsigned int c = check_delta_limit(child, n + 1);
1666                 if (m < c)
1667                         m = c;
1668                 child = child->delta_sibling;
1669         }
1670         return m;
1671 }
1672
1673 static unsigned long free_unpacked(struct unpacked *n)
1674 {
1675         unsigned long freed_mem = sizeof_delta_index(n->index);
1676         free_delta_index(n->index);
1677         n->index = NULL;
1678         if (n->data) {
1679                 freed_mem += n->entry->size;
1680                 free(n->data);
1681                 n->data = NULL;
1682         }
1683         n->entry = NULL;
1684         n->depth = 0;
1685         return freed_mem;
1686 }
1687
1688 static void find_deltas(struct object_entry **list, unsigned *list_size,
1689                         int window, int depth, unsigned *processed)
1690 {
1691         uint32_t i, idx = 0, count = 0;
1692         struct unpacked *array;
1693         unsigned long mem_usage = 0;
1694
1695         array = xcalloc(window, sizeof(struct unpacked));
1696
1697         for (;;) {
1698                 struct object_entry *entry;
1699                 struct unpacked *n = array + idx;
1700                 int j, max_depth, best_base = -1;
1701
1702                 progress_lock();
1703                 if (!*list_size) {
1704                         progress_unlock();
1705                         break;
1706                 }
1707                 entry = *list++;
1708                 (*list_size)--;
1709                 if (!entry->preferred_base) {
1710                         (*processed)++;
1711                         display_progress(progress_state, *processed);
1712                 }
1713                 progress_unlock();
1714
1715                 mem_usage -= free_unpacked(n);
1716                 n->entry = entry;
1717
1718                 while (window_memory_limit &&
1719                        mem_usage > window_memory_limit &&
1720                        count > 1) {
1721                         uint32_t tail = (idx + window - count) % window;
1722                         mem_usage -= free_unpacked(array + tail);
1723                         count--;
1724                 }
1725
1726                 /* We do not compute delta to *create* objects we are not
1727                  * going to pack.
1728                  */
1729                 if (entry->preferred_base)
1730                         goto next;
1731
1732                 /*
1733                  * If the current object is at pack edge, take the depth the
1734                  * objects that depend on the current object into account
1735                  * otherwise they would become too deep.
1736                  */
1737                 max_depth = depth;
1738                 if (entry->delta_child) {
1739                         max_depth -= check_delta_limit(entry, 0);
1740                         if (max_depth <= 0)
1741                                 goto next;
1742                 }
1743
1744                 j = window;
1745                 while (--j > 0) {
1746                         int ret;
1747                         uint32_t other_idx = idx + j;
1748                         struct unpacked *m;
1749                         if (other_idx >= window)
1750                                 other_idx -= window;
1751                         m = array + other_idx;
1752                         if (!m->entry)
1753                                 break;
1754                         ret = try_delta(n, m, max_depth, &mem_usage);
1755                         if (ret < 0)
1756                                 break;
1757                         else if (ret > 0)
1758                                 best_base = other_idx;
1759                 }
1760
1761                 /*
1762                  * If we decided to cache the delta data, then it is best
1763                  * to compress it right away.  First because we have to do
1764                  * it anyway, and doing it here while we're threaded will
1765                  * save a lot of time in the non threaded write phase,
1766                  * as well as allow for caching more deltas within
1767                  * the same cache size limit.
1768                  * ...
1769                  * But only if not writing to stdout, since in that case
1770                  * the network is most likely throttling writes anyway,
1771                  * and therefore it is best to go to the write phase ASAP
1772                  * instead, as we can afford spending more time compressing
1773                  * between writes at that moment.
1774                  */
1775                 if (entry->delta_data && !pack_to_stdout) {
1776                         entry->z_delta_size = do_compress(&entry->delta_data,
1777                                                           entry->delta_size);
1778                         cache_lock();
1779                         delta_cache_size -= entry->delta_size;
1780                         delta_cache_size += entry->z_delta_size;
1781                         cache_unlock();
1782                 }
1783
1784                 /* if we made n a delta, and if n is already at max
1785                  * depth, leaving it in the window is pointless.  we
1786                  * should evict it first.
1787                  */
1788                 if (entry->delta && max_depth <= n->depth)
1789                         continue;
1790
1791                 /*
1792                  * Move the best delta base up in the window, after the
1793                  * currently deltified object, to keep it longer.  It will
1794                  * be the first base object to be attempted next.
1795                  */
1796                 if (entry->delta) {
1797                         struct unpacked swap = array[best_base];
1798                         int dist = (window + idx - best_base) % window;
1799                         int dst = best_base;
1800                         while (dist--) {
1801                                 int src = (dst + 1) % window;
1802                                 array[dst] = array[src];
1803                                 dst = src;
1804                         }
1805                         array[dst] = swap;
1806                 }
1807
1808                 next:
1809                 idx++;
1810                 if (count + 1 < window)
1811                         count++;
1812                 if (idx >= window)
1813                         idx = 0;
1814         }
1815
1816         for (i = 0; i < window; ++i) {
1817                 free_delta_index(array[i].index);
1818                 free(array[i].data);
1819         }
1820         free(array);
1821 }
1822
1823 #ifndef NO_PTHREADS
1824
1825 static void try_to_free_from_threads(size_t size)
1826 {
1827         read_lock();
1828         release_pack_memory(size);
1829         read_unlock();
1830 }
1831
1832 static try_to_free_t old_try_to_free_routine;
1833
1834 /*
1835  * The main thread waits on the condition that (at least) one of the workers
1836  * has stopped working (which is indicated in the .working member of
1837  * struct thread_params).
1838  * When a work thread has completed its work, it sets .working to 0 and
1839  * signals the main thread and waits on the condition that .data_ready
1840  * becomes 1.
1841  */
1842
1843 struct thread_params {
1844         pthread_t thread;
1845         struct object_entry **list;
1846         unsigned list_size;
1847         unsigned remaining;
1848         int window;
1849         int depth;
1850         int working;
1851         int data_ready;
1852         pthread_mutex_t mutex;
1853         pthread_cond_t cond;
1854         unsigned *processed;
1855 };
1856
1857 static pthread_cond_t progress_cond;
1858
1859 /*
1860  * Mutex and conditional variable can't be statically-initialized on Windows.
1861  */
1862 static void init_threaded_search(void)
1863 {
1864         init_recursive_mutex(&read_mutex);
1865         pthread_mutex_init(&cache_mutex, NULL);
1866         pthread_mutex_init(&progress_mutex, NULL);
1867         pthread_cond_init(&progress_cond, NULL);
1868         old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1869 }
1870
1871 static void cleanup_threaded_search(void)
1872 {
1873         set_try_to_free_routine(old_try_to_free_routine);
1874         pthread_cond_destroy(&progress_cond);
1875         pthread_mutex_destroy(&read_mutex);
1876         pthread_mutex_destroy(&cache_mutex);
1877         pthread_mutex_destroy(&progress_mutex);
1878 }
1879
1880 static void *threaded_find_deltas(void *arg)
1881 {
1882         struct thread_params *me = arg;
1883
1884         while (me->remaining) {
1885                 find_deltas(me->list, &me->remaining,
1886                             me->window, me->depth, me->processed);
1887
1888                 progress_lock();
1889                 me->working = 0;
1890                 pthread_cond_signal(&progress_cond);
1891                 progress_unlock();
1892
1893                 /*
1894                  * We must not set ->data_ready before we wait on the
1895                  * condition because the main thread may have set it to 1
1896                  * before we get here. In order to be sure that new
1897                  * work is available if we see 1 in ->data_ready, it
1898                  * was initialized to 0 before this thread was spawned
1899                  * and we reset it to 0 right away.
1900                  */
1901                 pthread_mutex_lock(&me->mutex);
1902                 while (!me->data_ready)
1903                         pthread_cond_wait(&me->cond, &me->mutex);
1904                 me->data_ready = 0;
1905                 pthread_mutex_unlock(&me->mutex);
1906         }
1907         /* leave ->working 1 so that this doesn't get more work assigned */
1908         return NULL;
1909 }
1910
1911 static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1912                            int window, int depth, unsigned *processed)
1913 {
1914         struct thread_params *p;
1915         int i, ret, active_threads = 0;
1916
1917         init_threaded_search();
1918
1919         if (!delta_search_threads)      /* --threads=0 means autodetect */
1920                 delta_search_threads = online_cpus();
1921         if (delta_search_threads <= 1) {
1922                 find_deltas(list, &list_size, window, depth, processed);
1923                 cleanup_threaded_search();
1924                 return;
1925         }
1926         if (progress > pack_to_stdout)
1927                 fprintf(stderr, "Delta compression using up to %d threads.\n",
1928                                 delta_search_threads);
1929         p = xcalloc(delta_search_threads, sizeof(*p));
1930
1931         /* Partition the work amongst work threads. */
1932         for (i = 0; i < delta_search_threads; i++) {
1933                 unsigned sub_size = list_size / (delta_search_threads - i);
1934
1935                 /* don't use too small segments or no deltas will be found */
1936                 if (sub_size < 2*window && i+1 < delta_search_threads)
1937                         sub_size = 0;
1938
1939                 p[i].window = window;
1940                 p[i].depth = depth;
1941                 p[i].processed = processed;
1942                 p[i].working = 1;
1943                 p[i].data_ready = 0;
1944
1945                 /* try to split chunks on "path" boundaries */
1946                 while (sub_size && sub_size < list_size &&
1947                        list[sub_size]->hash &&
1948                        list[sub_size]->hash == list[sub_size-1]->hash)
1949                         sub_size++;
1950
1951                 p[i].list = list;
1952                 p[i].list_size = sub_size;
1953                 p[i].remaining = sub_size;
1954
1955                 list += sub_size;
1956                 list_size -= sub_size;
1957         }
1958
1959         /* Start work threads. */
1960         for (i = 0; i < delta_search_threads; i++) {
1961                 if (!p[i].list_size)
1962                         continue;
1963                 pthread_mutex_init(&p[i].mutex, NULL);
1964                 pthread_cond_init(&p[i].cond, NULL);
1965                 ret = pthread_create(&p[i].thread, NULL,
1966                                      threaded_find_deltas, &p[i]);
1967                 if (ret)
1968                         die("unable to create thread: %s", strerror(ret));
1969                 active_threads++;
1970         }
1971
1972         /*
1973          * Now let's wait for work completion.  Each time a thread is done
1974          * with its work, we steal half of the remaining work from the
1975          * thread with the largest number of unprocessed objects and give
1976          * it to that newly idle thread.  This ensure good load balancing
1977          * until the remaining object list segments are simply too short
1978          * to be worth splitting anymore.
1979          */
1980         while (active_threads) {
1981                 struct thread_params *target = NULL;
1982                 struct thread_params *victim = NULL;
1983                 unsigned sub_size = 0;
1984
1985                 progress_lock();
1986                 for (;;) {
1987                         for (i = 0; !target && i < delta_search_threads; i++)
1988                                 if (!p[i].working)
1989                                         target = &p[i];
1990                         if (target)
1991                                 break;
1992                         pthread_cond_wait(&progress_cond, &progress_mutex);
1993                 }
1994
1995                 for (i = 0; i < delta_search_threads; i++)
1996                         if (p[i].remaining > 2*window &&
1997                             (!victim || victim->remaining < p[i].remaining))
1998                                 victim = &p[i];
1999                 if (victim) {
2000                         sub_size = victim->remaining / 2;
2001                         list = victim->list + victim->list_size - sub_size;
2002                         while (sub_size && list[0]->hash &&
2003                                list[0]->hash == list[-1]->hash) {
2004                                 list++;
2005                                 sub_size--;
2006                         }
2007                         if (!sub_size) {
2008                                 /*
2009                                  * It is possible for some "paths" to have
2010                                  * so many objects that no hash boundary
2011                                  * might be found.  Let's just steal the
2012                                  * exact half in that case.
2013                                  */
2014                                 sub_size = victim->remaining / 2;
2015                                 list -= sub_size;
2016                         }
2017                         target->list = list;
2018                         victim->list_size -= sub_size;
2019                         victim->remaining -= sub_size;
2020                 }
2021                 target->list_size = sub_size;
2022                 target->remaining = sub_size;
2023                 target->working = 1;
2024                 progress_unlock();
2025
2026                 pthread_mutex_lock(&target->mutex);
2027                 target->data_ready = 1;
2028                 pthread_cond_signal(&target->cond);
2029                 pthread_mutex_unlock(&target->mutex);
2030
2031                 if (!sub_size) {
2032                         pthread_join(target->thread, NULL);
2033                         pthread_cond_destroy(&target->cond);
2034                         pthread_mutex_destroy(&target->mutex);
2035                         active_threads--;
2036                 }
2037         }
2038         cleanup_threaded_search();
2039         free(p);
2040 }
2041
2042 #else
2043 #define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2044 #endif
2045
2046 static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
2047 {
2048         unsigned char peeled[20];
2049
2050         if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
2051             !peel_ref(path, peeled)        && /* peelable? */
2052             packlist_find(&to_pack, peeled, NULL))      /* object packed? */
2053                 add_object_entry(sha1, OBJ_TAG, NULL, 0);
2054         return 0;
2055 }
2056
2057 static void prepare_pack(int window, int depth)
2058 {
2059         struct object_entry **delta_list;
2060         uint32_t i, nr_deltas;
2061         unsigned n;
2062
2063         get_object_details();
2064
2065         /*
2066          * If we're locally repacking then we need to be doubly careful
2067          * from now on in order to make sure no stealth corruption gets
2068          * propagated to the new pack.  Clients receiving streamed packs
2069          * should validate everything they get anyway so no need to incur
2070          * the additional cost here in that case.
2071          */
2072         if (!pack_to_stdout)
2073                 do_check_packed_object_crc = 1;
2074
2075         if (!to_pack.nr_objects || !window || !depth)
2076                 return;
2077
2078         delta_list = xmalloc(to_pack.nr_objects * sizeof(*delta_list));
2079         nr_deltas = n = 0;
2080
2081         for (i = 0; i < to_pack.nr_objects; i++) {
2082                 struct object_entry *entry = to_pack.objects + i;
2083
2084                 if (entry->delta)
2085                         /* This happens if we decided to reuse existing
2086                          * delta from a pack.  "reuse_delta &&" is implied.
2087                          */
2088                         continue;
2089
2090                 if (entry->size < 50)
2091                         continue;
2092
2093                 if (entry->no_try_delta)
2094                         continue;
2095
2096                 if (!entry->preferred_base) {
2097                         nr_deltas++;
2098                         if (entry->type < 0)
2099                                 die("unable to get type of object %s",
2100                                     sha1_to_hex(entry->idx.sha1));
2101                 } else {
2102                         if (entry->type < 0) {
2103                                 /*
2104                                  * This object is not found, but we
2105                                  * don't have to include it anyway.
2106                                  */
2107                                 continue;
2108                         }
2109                 }
2110
2111                 delta_list[n++] = entry;
2112         }
2113
2114         if (nr_deltas && n > 1) {
2115                 unsigned nr_done = 0;
2116                 if (progress)
2117                         progress_state = start_progress("Compressing objects",
2118                                                         nr_deltas);
2119                 qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2120                 ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2121                 stop_progress(&progress_state);
2122                 if (nr_done != nr_deltas)
2123                         die("inconsistency with delta count");
2124         }
2125         free(delta_list);
2126 }
2127
2128 static int git_pack_config(const char *k, const char *v, void *cb)
2129 {
2130         if (!strcmp(k, "pack.window")) {
2131                 window = git_config_int(k, v);
2132                 return 0;
2133         }
2134         if (!strcmp(k, "pack.windowmemory")) {
2135                 window_memory_limit = git_config_ulong(k, v);
2136                 return 0;
2137         }
2138         if (!strcmp(k, "pack.depth")) {
2139                 depth = git_config_int(k, v);
2140                 return 0;
2141         }
2142         if (!strcmp(k, "pack.compression")) {
2143                 int level = git_config_int(k, v);
2144                 if (level == -1)
2145                         level = Z_DEFAULT_COMPRESSION;
2146                 else if (level < 0 || level > Z_BEST_COMPRESSION)
2147                         die("bad pack compression level %d", level);
2148                 pack_compression_level = level;
2149                 pack_compression_seen = 1;
2150                 return 0;
2151         }
2152         if (!strcmp(k, "pack.deltacachesize")) {
2153                 max_delta_cache_size = git_config_int(k, v);
2154                 return 0;
2155         }
2156         if (!strcmp(k, "pack.deltacachelimit")) {
2157                 cache_max_small_delta_size = git_config_int(k, v);
2158                 return 0;
2159         }
2160         if (!strcmp(k, "pack.usebitmaps")) {
2161                 use_bitmap_index = git_config_bool(k, v);
2162                 return 0;
2163         }
2164         if (!strcmp(k, "pack.threads")) {
2165                 delta_search_threads = git_config_int(k, v);
2166                 if (delta_search_threads < 0)
2167                         die("invalid number of threads specified (%d)",
2168                             delta_search_threads);
2169 #ifdef NO_PTHREADS
2170                 if (delta_search_threads != 1)
2171                         warning("no threads support, ignoring %s", k);
2172 #endif
2173                 return 0;
2174         }
2175         if (!strcmp(k, "pack.indexversion")) {
2176                 pack_idx_opts.version = git_config_int(k, v);
2177                 if (pack_idx_opts.version > 2)
2178                         die("bad pack.indexversion=%"PRIu32,
2179                             pack_idx_opts.version);
2180                 return 0;
2181         }
2182         return git_default_config(k, v, cb);
2183 }
2184
2185 static void read_object_list_from_stdin(void)
2186 {
2187         char line[40 + 1 + PATH_MAX + 2];
2188         unsigned char sha1[20];
2189
2190         for (;;) {
2191                 if (!fgets(line, sizeof(line), stdin)) {
2192                         if (feof(stdin))
2193                                 break;
2194                         if (!ferror(stdin))
2195                                 die("fgets returned NULL, not EOF, not error!");
2196                         if (errno != EINTR)
2197                                 die_errno("fgets");
2198                         clearerr(stdin);
2199                         continue;
2200                 }
2201                 if (line[0] == '-') {
2202                         if (get_sha1_hex(line+1, sha1))
2203                                 die("expected edge sha1, got garbage:\n %s",
2204                                     line);
2205                         add_preferred_base(sha1);
2206                         continue;
2207                 }
2208                 if (get_sha1_hex(line, sha1))
2209                         die("expected sha1, got garbage:\n %s", line);
2210
2211                 add_preferred_base_object(line+41);
2212                 add_object_entry(sha1, 0, line+41, 0);
2213         }
2214 }
2215
2216 #define OBJECT_ADDED (1u<<20)
2217
2218 static void show_commit(struct commit *commit, void *data)
2219 {
2220         add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
2221         commit->object.flags |= OBJECT_ADDED;
2222 }
2223
2224 static void show_object(struct object *obj,
2225                         const struct name_path *path, const char *last,
2226                         void *data)
2227 {
2228         char *name = path_name(path, last);
2229
2230         add_preferred_base_object(name);
2231         add_object_entry(obj->sha1, obj->type, name, 0);
2232         obj->flags |= OBJECT_ADDED;
2233
2234         /*
2235          * We will have generated the hash from the name,
2236          * but not saved a pointer to it - we can free it
2237          */
2238         free((char *)name);
2239 }
2240
2241 static void show_edge(struct commit *commit)
2242 {
2243         add_preferred_base(commit->object.sha1);
2244 }
2245
2246 struct in_pack_object {
2247         off_t offset;
2248         struct object *object;
2249 };
2250
2251 struct in_pack {
2252         int alloc;
2253         int nr;
2254         struct in_pack_object *array;
2255 };
2256
2257 static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2258 {
2259         in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
2260         in_pack->array[in_pack->nr].object = object;
2261         in_pack->nr++;
2262 }
2263
2264 /*
2265  * Compare the objects in the offset order, in order to emulate the
2266  * "git rev-list --objects" output that produced the pack originally.
2267  */
2268 static int ofscmp(const void *a_, const void *b_)
2269 {
2270         struct in_pack_object *a = (struct in_pack_object *)a_;
2271         struct in_pack_object *b = (struct in_pack_object *)b_;
2272
2273         if (a->offset < b->offset)
2274                 return -1;
2275         else if (a->offset > b->offset)
2276                 return 1;
2277         else
2278                 return hashcmp(a->object->sha1, b->object->sha1);
2279 }
2280
2281 static void add_objects_in_unpacked_packs(struct rev_info *revs)
2282 {
2283         struct packed_git *p;
2284         struct in_pack in_pack;
2285         uint32_t i;
2286
2287         memset(&in_pack, 0, sizeof(in_pack));
2288
2289         for (p = packed_git; p; p = p->next) {
2290                 const unsigned char *sha1;
2291                 struct object *o;
2292
2293                 if (!p->pack_local || p->pack_keep)
2294                         continue;
2295                 if (open_pack_index(p))
2296                         die("cannot open pack index");
2297
2298                 ALLOC_GROW(in_pack.array,
2299                            in_pack.nr + p->num_objects,
2300                            in_pack.alloc);
2301
2302                 for (i = 0; i < p->num_objects; i++) {
2303                         sha1 = nth_packed_object_sha1(p, i);
2304                         o = lookup_unknown_object(sha1);
2305                         if (!(o->flags & OBJECT_ADDED))
2306                                 mark_in_pack_object(o, p, &in_pack);
2307                         o->flags |= OBJECT_ADDED;
2308                 }
2309         }
2310
2311         if (in_pack.nr) {
2312                 qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2313                       ofscmp);
2314                 for (i = 0; i < in_pack.nr; i++) {
2315                         struct object *o = in_pack.array[i].object;
2316                         add_object_entry(o->sha1, o->type, "", 0);
2317                 }
2318         }
2319         free(in_pack.array);
2320 }
2321
2322 static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2323 {
2324         static struct packed_git *last_found = (void *)1;
2325         struct packed_git *p;
2326
2327         p = (last_found != (void *)1) ? last_found : packed_git;
2328
2329         while (p) {
2330                 if ((!p->pack_local || p->pack_keep) &&
2331                         find_pack_entry_one(sha1, p)) {
2332                         last_found = p;
2333                         return 1;
2334                 }
2335                 if (p == last_found)
2336                         p = packed_git;
2337                 else
2338                         p = p->next;
2339                 if (p == last_found)
2340                         p = p->next;
2341         }
2342         return 0;
2343 }
2344
2345 static void loosen_unused_packed_objects(struct rev_info *revs)
2346 {
2347         struct packed_git *p;
2348         uint32_t i;
2349         const unsigned char *sha1;
2350
2351         for (p = packed_git; p; p = p->next) {
2352                 if (!p->pack_local || p->pack_keep)
2353                         continue;
2354
2355                 if (unpack_unreachable_expiration &&
2356                     p->mtime < unpack_unreachable_expiration)
2357                         continue;
2358
2359                 if (open_pack_index(p))
2360                         die("cannot open pack index");
2361
2362                 for (i = 0; i < p->num_objects; i++) {
2363                         sha1 = nth_packed_object_sha1(p, i);
2364                         if (!packlist_find(&to_pack, sha1, NULL) &&
2365                                 !has_sha1_pack_kept_or_nonlocal(sha1))
2366                                 if (force_object_loose(sha1, p->mtime))
2367                                         die("unable to force loose object");
2368                 }
2369         }
2370 }
2371
2372 static int get_object_list_from_bitmap(struct rev_info *revs)
2373 {
2374         if (prepare_bitmap_walk(revs) < 0)
2375                 return -1;
2376
2377         if (!reuse_partial_packfile_from_bitmap(
2378                         &reuse_packfile,
2379                         &reuse_packfile_objects,
2380                         &reuse_packfile_offset)) {
2381                 assert(reuse_packfile_objects);
2382                 nr_result += reuse_packfile_objects;
2383
2384                 if (progress) {
2385                         fprintf(stderr, "Reusing existing pack: %d, done.\n",
2386                                 reuse_packfile_objects);
2387                         fflush(stderr);
2388                 }
2389         }
2390
2391         traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2392         return 0;
2393 }
2394
2395 static void get_object_list(int ac, const char **av)
2396 {
2397         struct rev_info revs;
2398         char line[1000];
2399         int flags = 0;
2400
2401         init_revisions(&revs, NULL);
2402         save_commit_buffer = 0;
2403         setup_revisions(ac, av, &revs, NULL);
2404
2405         while (fgets(line, sizeof(line), stdin) != NULL) {
2406                 int len = strlen(line);
2407                 if (len && line[len - 1] == '\n')
2408                         line[--len] = 0;
2409                 if (!len)
2410                         break;
2411                 if (*line == '-') {
2412                         if (!strcmp(line, "--not")) {
2413                                 flags ^= UNINTERESTING;
2414                                 continue;
2415                         }
2416                         die("not a rev '%s'", line);
2417                 }
2418                 if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2419                         die("bad revision '%s'", line);
2420         }
2421
2422         if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2423                 return;
2424
2425         if (prepare_revision_walk(&revs))
2426                 die("revision walk setup failed");
2427         mark_edges_uninteresting(&revs, show_edge);
2428         traverse_commit_list(&revs, show_commit, show_object, NULL);
2429
2430         if (keep_unreachable)
2431                 add_objects_in_unpacked_packs(&revs);
2432         if (unpack_unreachable)
2433                 loosen_unused_packed_objects(&revs);
2434 }
2435
2436 static int option_parse_index_version(const struct option *opt,
2437                                       const char *arg, int unset)
2438 {
2439         char *c;
2440         const char *val = arg;
2441         pack_idx_opts.version = strtoul(val, &c, 10);
2442         if (pack_idx_opts.version > 2)
2443                 die(_("unsupported index version %s"), val);
2444         if (*c == ',' && c[1])
2445                 pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2446         if (*c || pack_idx_opts.off32_limit & 0x80000000)
2447                 die(_("bad index version '%s'"), val);
2448         return 0;
2449 }
2450
2451 static int option_parse_unpack_unreachable(const struct option *opt,
2452                                            const char *arg, int unset)
2453 {
2454         if (unset) {
2455                 unpack_unreachable = 0;
2456                 unpack_unreachable_expiration = 0;
2457         }
2458         else {
2459                 unpack_unreachable = 1;
2460                 if (arg)
2461                         unpack_unreachable_expiration = approxidate(arg);
2462         }
2463         return 0;
2464 }
2465
2466 static int option_parse_ulong(const struct option *opt,
2467                               const char *arg, int unset)
2468 {
2469         if (unset)
2470                 die(_("option %s does not accept negative form"),
2471                     opt->long_name);
2472
2473         if (!git_parse_ulong(arg, opt->value))
2474                 die(_("unable to parse value '%s' for option %s"),
2475                     arg, opt->long_name);
2476         return 0;
2477 }
2478
2479 #define OPT_ULONG(s, l, v, h) \
2480         { OPTION_CALLBACK, (s), (l), (v), "n", (h),     \
2481           PARSE_OPT_NONEG, option_parse_ulong }
2482
2483 int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2484 {
2485         int use_internal_rev_list = 0;
2486         int thin = 0;
2487         int all_progress_implied = 0;
2488         const char *rp_av[6];
2489         int rp_ac = 0;
2490         int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2491         struct option pack_objects_options[] = {
2492                 OPT_SET_INT('q', "quiet", &progress,
2493                             N_("do not show progress meter"), 0),
2494                 OPT_SET_INT(0, "progress", &progress,
2495                             N_("show progress meter"), 1),
2496                 OPT_SET_INT(0, "all-progress", &progress,
2497                             N_("show progress meter during object writing phase"), 2),
2498                 OPT_BOOL(0, "all-progress-implied",
2499                          &all_progress_implied,
2500                          N_("similar to --all-progress when progress meter is shown")),
2501                 { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2502                   N_("write the pack index file in the specified idx format version"),
2503                   0, option_parse_index_version },
2504                 OPT_ULONG(0, "max-pack-size", &pack_size_limit,
2505                           N_("maximum size of each output pack file")),
2506                 OPT_BOOL(0, "local", &local,
2507                          N_("ignore borrowed objects from alternate object store")),
2508                 OPT_BOOL(0, "incremental", &incremental,
2509                          N_("ignore packed objects")),
2510                 OPT_INTEGER(0, "window", &window,
2511                             N_("limit pack window by objects")),
2512                 OPT_ULONG(0, "window-memory", &window_memory_limit,
2513                           N_("limit pack window by memory in addition to object limit")),
2514                 OPT_INTEGER(0, "depth", &depth,
2515                             N_("maximum length of delta chain allowed in the resulting pack")),
2516                 OPT_BOOL(0, "reuse-delta", &reuse_delta,
2517                          N_("reuse existing deltas")),
2518                 OPT_BOOL(0, "reuse-object", &reuse_object,
2519                          N_("reuse existing objects")),
2520                 OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2521                          N_("use OFS_DELTA objects")),
2522                 OPT_INTEGER(0, "threads", &delta_search_threads,
2523                             N_("use threads when searching for best delta matches")),
2524                 OPT_BOOL(0, "non-empty", &non_empty,
2525                          N_("do not create an empty pack output")),
2526                 OPT_BOOL(0, "revs", &use_internal_rev_list,
2527                          N_("read revision arguments from standard input")),
2528                 { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2529                   N_("limit the objects to those that are not yet packed"),
2530                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2531                 { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2532                   N_("include objects reachable from any reference"),
2533                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2534                 { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2535                   N_("include objects referred by reflog entries"),
2536                   PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2537                 OPT_BOOL(0, "stdout", &pack_to_stdout,
2538                          N_("output pack to stdout")),
2539                 OPT_BOOL(0, "include-tag", &include_tag,
2540                          N_("include tag objects that refer to objects to be packed")),
2541                 OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2542                          N_("keep unreachable objects")),
2543                 { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2544                   N_("unpack unreachable objects newer than <time>"),
2545                   PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2546                 OPT_BOOL(0, "thin", &thin,
2547                          N_("create thin packs")),
2548                 OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2549                          N_("ignore packs that have companion .keep file")),
2550                 OPT_INTEGER(0, "compression", &pack_compression_level,
2551                             N_("pack compression level")),
2552                 OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2553                             N_("do not hide commits by grafts"), 0),
2554                 OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2555                          N_("use a bitmap index if available to speed up counting objects")),
2556                 OPT_END(),
2557         };
2558
2559         read_replace_refs = 0;
2560
2561         reset_pack_idx_option(&pack_idx_opts);
2562         git_config(git_pack_config, NULL);
2563         if (!pack_compression_seen && core_compression_seen)
2564                 pack_compression_level = core_compression_level;
2565
2566         progress = isatty(2);
2567         argc = parse_options(argc, argv, prefix, pack_objects_options,
2568                              pack_usage, 0);
2569
2570         if (argc) {
2571                 base_name = argv[0];
2572                 argc--;
2573         }
2574         if (pack_to_stdout != !base_name || argc)
2575                 usage_with_options(pack_usage, pack_objects_options);
2576
2577         rp_av[rp_ac++] = "pack-objects";
2578         if (thin) {
2579                 use_internal_rev_list = 1;
2580                 rp_av[rp_ac++] = "--objects-edge";
2581         } else
2582                 rp_av[rp_ac++] = "--objects";
2583
2584         if (rev_list_all) {
2585                 use_internal_rev_list = 1;
2586                 rp_av[rp_ac++] = "--all";
2587         }
2588         if (rev_list_reflog) {
2589                 use_internal_rev_list = 1;
2590                 rp_av[rp_ac++] = "--reflog";
2591         }
2592         if (rev_list_unpacked) {
2593                 use_internal_rev_list = 1;
2594                 rp_av[rp_ac++] = "--unpacked";
2595         }
2596
2597         if (!reuse_object)
2598                 reuse_delta = 0;
2599         if (pack_compression_level == -1)
2600                 pack_compression_level = Z_DEFAULT_COMPRESSION;
2601         else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2602                 die("bad pack compression level %d", pack_compression_level);
2603 #ifdef NO_PTHREADS
2604         if (delta_search_threads != 1)
2605                 warning("no threads support, ignoring --threads");
2606 #endif
2607         if (!pack_to_stdout && !pack_size_limit)
2608                 pack_size_limit = pack_size_limit_cfg;
2609         if (pack_to_stdout && pack_size_limit)
2610                 die("--max-pack-size cannot be used to build a pack for transfer.");
2611         if (pack_size_limit && pack_size_limit < 1024*1024) {
2612                 warning("minimum pack size limit is 1 MiB");
2613                 pack_size_limit = 1024*1024;
2614         }
2615
2616         if (!pack_to_stdout && thin)
2617                 die("--thin cannot be used to build an indexable pack.");
2618
2619         if (keep_unreachable && unpack_unreachable)
2620                 die("--keep-unreachable and --unpack-unreachable are incompatible.");
2621
2622         if (!use_internal_rev_list || !pack_to_stdout || is_repository_shallow())
2623                 use_bitmap_index = 0;
2624
2625         if (progress && all_progress_implied)
2626                 progress = 2;
2627
2628         prepare_packed_git();
2629
2630         if (progress)
2631                 progress_state = start_progress("Counting objects", 0);
2632         if (!use_internal_rev_list)
2633                 read_object_list_from_stdin();
2634         else {
2635                 rp_av[rp_ac] = NULL;
2636                 get_object_list(rp_ac, rp_av);
2637         }
2638         cleanup_preferred_base();
2639         if (include_tag && nr_result)
2640                 for_each_ref(add_ref_tag, NULL);
2641         stop_progress(&progress_state);
2642
2643         if (non_empty && !nr_result)
2644                 return 0;
2645         if (nr_result)
2646                 prepare_pack(window, depth);
2647         write_pack_file();
2648         if (progress)
2649                 fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2650                         " reused %"PRIu32" (delta %"PRIu32")\n",
2651                         written, written_delta, reused, reused_delta);
2652         return 0;
2653 }