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