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