builtin / pack-objects.con commit ewah: support platforms that require aligned reads (a201c20)
   1#include "builtin.h"
   2#include "cache.h"
   3#include "attr.h"
   4#include "object.h"
   5#include "blob.h"
   6#include "commit.h"
   7#include "tag.h"
   8#include "tree.h"
   9#include "delta.h"
  10#include "pack.h"
  11#include "pack-revindex.h"
  12#include "csum-file.h"
  13#include "tree-walk.h"
  14#include "diff.h"
  15#include "revision.h"
  16#include "list-objects.h"
  17#include "pack-objects.h"
  18#include "progress.h"
  19#include "refs.h"
  20#include "streaming.h"
  21#include "thread-utils.h"
  22#include "pack-bitmap.h"
  23
  24static const char *pack_usage[] = {
  25        N_("git pack-objects --stdout [options...] [< ref-list | < object-list]"),
  26        N_("git pack-objects [options...] base-name [< ref-list | < object-list]"),
  27        NULL
  28};
  29
  30/*
  31 * Objects we are going to pack are collected in the `to_pack` structure.
  32 * It contains an array (dynamically expanded) of the object data, and a map
  33 * that can resolve SHA1s to their position in the array.
  34 */
  35static struct packing_data to_pack;
  36
  37static struct pack_idx_entry **written_list;
  38static uint32_t nr_result, nr_written;
  39
  40static int non_empty;
  41static int reuse_delta = 1, reuse_object = 1;
  42static int keep_unreachable, unpack_unreachable, include_tag;
  43static unsigned long unpack_unreachable_expiration;
  44static int local;
  45static int incremental;
  46static int ignore_packed_keep;
  47static int allow_ofs_delta;
  48static struct pack_idx_option pack_idx_opts;
  49static const char *base_name;
  50static int progress = 1;
  51static int window = 10;
  52static unsigned long pack_size_limit;
  53static int depth = 50;
  54static int delta_search_threads;
  55static int pack_to_stdout;
  56static int num_preferred_base;
  57static struct progress *progress_state;
  58static int pack_compression_level = Z_DEFAULT_COMPRESSION;
  59static int pack_compression_seen;
  60
  61static struct packed_git *reuse_packfile;
  62static uint32_t reuse_packfile_objects;
  63static off_t reuse_packfile_offset;
  64
  65static int use_bitmap_index = 1;
  66static int write_bitmap_index;
  67static uint16_t write_bitmap_options;
  68
  69static unsigned long delta_cache_size = 0;
  70static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
  71static unsigned long cache_max_small_delta_size = 1000;
  72
  73static unsigned long window_memory_limit = 0;
  74
  75/*
  76 * stats
  77 */
  78static uint32_t written, written_delta;
  79static uint32_t reused, reused_delta;
  80
  81/*
  82 * Indexed commits
  83 */
  84static struct commit **indexed_commits;
  85static unsigned int indexed_commits_nr;
  86static unsigned int indexed_commits_alloc;
  87
  88static void index_commit_for_bitmap(struct commit *commit)
  89{
  90        if (indexed_commits_nr >= indexed_commits_alloc) {
  91                indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
  92                indexed_commits = xrealloc(indexed_commits,
  93                        indexed_commits_alloc * sizeof(struct commit *));
  94        }
  95
  96        indexed_commits[indexed_commits_nr++] = commit;
  97}
  98
  99static void *get_delta(struct object_entry *entry)
 100{
 101        unsigned long size, base_size, delta_size;
 102        void *buf, *base_buf, *delta_buf;
 103        enum object_type type;
 104
 105        buf = read_sha1_file(entry->idx.sha1, &type, &size);
 106        if (!buf)
 107                die("unable to read %s", sha1_to_hex(entry->idx.sha1));
 108        base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
 109        if (!base_buf)
 110                die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
 111        delta_buf = diff_delta(base_buf, base_size,
 112                               buf, size, &delta_size, 0);
 113        if (!delta_buf || delta_size != entry->delta_size)
 114                die("delta size changed");
 115        free(buf);
 116        free(base_buf);
 117        return delta_buf;
 118}
 119
 120static unsigned long do_compress(void **pptr, unsigned long size)
 121{
 122        git_zstream stream;
 123        void *in, *out;
 124        unsigned long maxsize;
 125
 126        memset(&stream, 0, sizeof(stream));
 127        git_deflate_init(&stream, pack_compression_level);
 128        maxsize = git_deflate_bound(&stream, size);
 129
 130        in = *pptr;
 131        out = xmalloc(maxsize);
 132        *pptr = out;
 133
 134        stream.next_in = in;
 135        stream.avail_in = size;
 136        stream.next_out = out;
 137        stream.avail_out = maxsize;
 138        while (git_deflate(&stream, Z_FINISH) == Z_OK)
 139                ; /* nothing */
 140        git_deflate_end(&stream);
 141
 142        free(in);
 143        return stream.total_out;
 144}
 145
 146static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
 147                                           const unsigned char *sha1)
 148{
 149        git_zstream stream;
 150        unsigned char ibuf[1024 * 16];
 151        unsigned char obuf[1024 * 16];
 152        unsigned long olen = 0;
 153
 154        memset(&stream, 0, sizeof(stream));
 155        git_deflate_init(&stream, pack_compression_level);
 156
 157        for (;;) {
 158                ssize_t readlen;
 159                int zret = Z_OK;
 160                readlen = read_istream(st, ibuf, sizeof(ibuf));
 161                if (readlen == -1)
 162                        die(_("unable to read %s"), sha1_to_hex(sha1));
 163
 164                stream.next_in = ibuf;
 165                stream.avail_in = readlen;
 166                while ((stream.avail_in || readlen == 0) &&
 167                       (zret == Z_OK || zret == Z_BUF_ERROR)) {
 168                        stream.next_out = obuf;
 169                        stream.avail_out = sizeof(obuf);
 170                        zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 171                        sha1write(f, obuf, stream.next_out - obuf);
 172                        olen += stream.next_out - obuf;
 173                }
 174                if (stream.avail_in)
 175                        die(_("deflate error (%d)"), zret);
 176                if (readlen == 0) {
 177                        if (zret != Z_STREAM_END)
 178                                die(_("deflate error (%d)"), zret);
 179                        break;
 180                }
 181        }
 182        git_deflate_end(&stream);
 183        return olen;
 184}
 185
 186/*
 187 * we are going to reuse the existing object data as is.  make
 188 * sure it is not corrupt.
 189 */
 190static int check_pack_inflate(struct packed_git *p,
 191                struct pack_window **w_curs,
 192                off_t offset,
 193                off_t len,
 194                unsigned long expect)
 195{
 196        git_zstream stream;
 197        unsigned char fakebuf[4096], *in;
 198        int st;
 199
 200        memset(&stream, 0, sizeof(stream));
 201        git_inflate_init(&stream);
 202        do {
 203                in = use_pack(p, w_curs, offset, &stream.avail_in);
 204                stream.next_in = in;
 205                stream.next_out = fakebuf;
 206                stream.avail_out = sizeof(fakebuf);
 207                st = git_inflate(&stream, Z_FINISH);
 208                offset += stream.next_in - in;
 209        } while (st == Z_OK || st == Z_BUF_ERROR);
 210        git_inflate_end(&stream);
 211        return (st == Z_STREAM_END &&
 212                stream.total_out == expect &&
 213                stream.total_in == len) ? 0 : -1;
 214}
 215
 216static void copy_pack_data(struct sha1file *f,
 217                struct packed_git *p,
 218                struct pack_window **w_curs,
 219                off_t offset,
 220                off_t len)
 221{
 222        unsigned char *in;
 223        unsigned long avail;
 224
 225        while (len) {
 226                in = use_pack(p, w_curs, offset, &avail);
 227                if (avail > len)
 228                        avail = (unsigned long)len;
 229                sha1write(f, in, avail);
 230                offset += avail;
 231                len -= avail;
 232        }
 233}
 234
 235/* Return 0 if we will bust the pack-size limit */
 236static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
 237                                           unsigned long limit, int usable_delta)
 238{
 239        unsigned long size, datalen;
 240        unsigned char header[10], dheader[10];
 241        unsigned hdrlen;
 242        enum object_type type;
 243        void *buf;
 244        struct git_istream *st = NULL;
 245
 246        if (!usable_delta) {
 247                if (entry->type == OBJ_BLOB &&
 248                    entry->size > big_file_threshold &&
 249                    (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
 250                        buf = NULL;
 251                else {
 252                        buf = read_sha1_file(entry->idx.sha1, &type, &size);
 253                        if (!buf)
 254                                die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
 255                }
 256                /*
 257                 * make sure no cached delta data remains from a
 258                 * previous attempt before a pack split occurred.
 259                 */
 260                free(entry->delta_data);
 261                entry->delta_data = NULL;
 262                entry->z_delta_size = 0;
 263        } else if (entry->delta_data) {
 264                size = entry->delta_size;
 265                buf = entry->delta_data;
 266                entry->delta_data = NULL;
 267                type = (allow_ofs_delta && entry->delta->idx.offset) ?
 268                        OBJ_OFS_DELTA : OBJ_REF_DELTA;
 269        } else {
 270                buf = get_delta(entry);
 271                size = entry->delta_size;
 272                type = (allow_ofs_delta && entry->delta->idx.offset) ?
 273                        OBJ_OFS_DELTA : OBJ_REF_DELTA;
 274        }
 275
 276        if (st) /* large blob case, just assume we don't compress well */
 277                datalen = size;
 278        else if (entry->z_delta_size)
 279                datalen = entry->z_delta_size;
 280        else
 281                datalen = do_compress(&buf, size);
 282
 283        /*
 284         * The object header is a byte of 'type' followed by zero or
 285         * more bytes of length.
 286         */
 287        hdrlen = encode_in_pack_object_header(type, size, header);
 288
 289        if (type == OBJ_OFS_DELTA) {
 290                /*
 291                 * Deltas with relative base contain an additional
 292                 * encoding of the relative offset for the delta
 293                 * base from this object's position in the pack.
 294                 */
 295                off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 296                unsigned pos = sizeof(dheader) - 1;
 297                dheader[pos] = ofs & 127;
 298                while (ofs >>= 7)
 299                        dheader[--pos] = 128 | (--ofs & 127);
 300                if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 301                        if (st)
 302                                close_istream(st);
 303                        free(buf);
 304                        return 0;
 305                }
 306                sha1write(f, header, hdrlen);
 307                sha1write(f, dheader + pos, sizeof(dheader) - pos);
 308                hdrlen += sizeof(dheader) - pos;
 309        } else if (type == OBJ_REF_DELTA) {
 310                /*
 311                 * Deltas with a base reference contain
 312                 * an additional 20 bytes for the base sha1.
 313                 */
 314                if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 315                        if (st)
 316                                close_istream(st);
 317                        free(buf);
 318                        return 0;
 319                }
 320                sha1write(f, header, hdrlen);
 321                sha1write(f, entry->delta->idx.sha1, 20);
 322                hdrlen += 20;
 323        } else {
 324                if (limit && hdrlen + datalen + 20 >= limit) {
 325                        if (st)
 326                                close_istream(st);
 327                        free(buf);
 328                        return 0;
 329                }
 330                sha1write(f, header, hdrlen);
 331        }
 332        if (st) {
 333                datalen = write_large_blob_data(st, f, entry->idx.sha1);
 334                close_istream(st);
 335        } else {
 336                sha1write(f, buf, datalen);
 337                free(buf);
 338        }
 339
 340        return hdrlen + datalen;
 341}
 342
 343/* Return 0 if we will bust the pack-size limit */
 344static unsigned long write_reuse_object(struct sha1file *f, struct object_entry *entry,
 345                                        unsigned long limit, int usable_delta)
 346{
 347        struct packed_git *p = entry->in_pack;
 348        struct pack_window *w_curs = NULL;
 349        struct revindex_entry *revidx;
 350        off_t offset;
 351        enum object_type type = entry->type;
 352        unsigned long datalen;
 353        unsigned char header[10], dheader[10];
 354        unsigned hdrlen;
 355
 356        if (entry->delta)
 357                type = (allow_ofs_delta && entry->delta->idx.offset) ?
 358                        OBJ_OFS_DELTA : OBJ_REF_DELTA;
 359        hdrlen = encode_in_pack_object_header(type, entry->size, header);
 360
 361        offset = entry->in_pack_offset;
 362        revidx = find_pack_revindex(p, offset);
 363        datalen = revidx[1].offset - offset;
 364        if (!pack_to_stdout && p->index_version > 1 &&
 365            check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 366                error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
 367                unuse_pack(&w_curs);
 368                return write_no_reuse_object(f, entry, limit, usable_delta);
 369        }
 370
 371        offset += entry->in_pack_header_size;
 372        datalen -= entry->in_pack_header_size;
 373
 374        if (!pack_to_stdout && p->index_version == 1 &&
 375            check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 376                error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
 377                unuse_pack(&w_curs);
 378                return write_no_reuse_object(f, entry, limit, usable_delta);
 379        }
 380
 381        if (type == OBJ_OFS_DELTA) {
 382                off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 383                unsigned pos = sizeof(dheader) - 1;
 384                dheader[pos] = ofs & 127;
 385                while (ofs >>= 7)
 386                        dheader[--pos] = 128 | (--ofs & 127);
 387                if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 388                        unuse_pack(&w_curs);
 389                        return 0;
 390                }
 391                sha1write(f, header, hdrlen);
 392                sha1write(f, dheader + pos, sizeof(dheader) - pos);
 393                hdrlen += sizeof(dheader) - pos;
 394                reused_delta++;
 395        } else if (type == OBJ_REF_DELTA) {
 396                if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 397                        unuse_pack(&w_curs);
 398                        return 0;
 399                }
 400                sha1write(f, header, hdrlen);
 401                sha1write(f, entry->delta->idx.sha1, 20);
 402                hdrlen += 20;
 403                reused_delta++;
 404        } else {
 405                if (limit && hdrlen + datalen + 20 >= limit) {
 406                        unuse_pack(&w_curs);
 407                        return 0;
 408                }
 409                sha1write(f, header, hdrlen);
 410        }
 411        copy_pack_data(f, p, &w_curs, offset, datalen);
 412        unuse_pack(&w_curs);
 413        reused++;
 414        return hdrlen + datalen;
 415}
 416
 417/* Return 0 if we will bust the pack-size limit */
 418static unsigned long write_object(struct sha1file *f,
 419                                  struct object_entry *entry,
 420                                  off_t write_offset)
 421{
 422        unsigned long limit, len;
 423        int usable_delta, to_reuse;
 424
 425        if (!pack_to_stdout)
 426                crc32_begin(f);
 427
 428        /* apply size limit if limited packsize and not first object */
 429        if (!pack_size_limit || !nr_written)
 430                limit = 0;
 431        else if (pack_size_limit <= write_offset)
 432                /*
 433                 * the earlier object did not fit the limit; avoid
 434                 * mistaking this with unlimited (i.e. limit = 0).
 435                 */
 436                limit = 1;
 437        else
 438                limit = pack_size_limit - write_offset;
 439
 440        if (!entry->delta)
 441                usable_delta = 0;       /* no delta */
 442        else if (!pack_size_limit)
 443               usable_delta = 1;        /* unlimited packfile */
 444        else if (entry->delta->idx.offset == (off_t)-1)
 445                usable_delta = 0;       /* base was written to another pack */
 446        else if (entry->delta->idx.offset)
 447                usable_delta = 1;       /* base already exists in this pack */
 448        else
 449                usable_delta = 0;       /* base could end up in another pack */
 450
 451        if (!reuse_object)
 452                to_reuse = 0;   /* explicit */
 453        else if (!entry->in_pack)
 454                to_reuse = 0;   /* can't reuse what we don't have */
 455        else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
 456                                /* check_object() decided it for us ... */
 457                to_reuse = usable_delta;
 458                                /* ... but pack split may override that */
 459        else if (entry->type != entry->in_pack_type)
 460                to_reuse = 0;   /* pack has delta which is unusable */
 461        else if (entry->delta)
 462                to_reuse = 0;   /* we want to pack afresh */
 463        else
 464                to_reuse = 1;   /* we have it in-pack undeltified,
 465                                 * and we do not need to deltify it.
 466                                 */
 467
 468        if (!to_reuse)
 469                len = write_no_reuse_object(f, entry, limit, usable_delta);
 470        else
 471                len = write_reuse_object(f, entry, limit, usable_delta);
 472        if (!len)
 473                return 0;
 474
 475        if (usable_delta)
 476                written_delta++;
 477        written++;
 478        if (!pack_to_stdout)
 479                entry->idx.crc32 = crc32_end(f);
 480        return len;
 481}
 482
 483enum write_one_status {
 484        WRITE_ONE_SKIP = -1, /* already written */
 485        WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 486        WRITE_ONE_WRITTEN = 1, /* normal */
 487        WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 488};
 489
 490static enum write_one_status write_one(struct sha1file *f,
 491                                       struct object_entry *e,
 492                                       off_t *offset)
 493{
 494        unsigned long size;
 495        int recursing;
 496
 497        /*
 498         * we set offset to 1 (which is an impossible value) to mark
 499         * the fact that this object is involved in "write its base
 500         * first before writing a deltified object" recursion.
 501         */
 502        recursing = (e->idx.offset == 1);
 503        if (recursing) {
 504                warning("recursive delta detected for object %s",
 505                        sha1_to_hex(e->idx.sha1));
 506                return WRITE_ONE_RECURSIVE;
 507        } else if (e->idx.offset || e->preferred_base) {
 508                /* offset is non zero if object is written already. */
 509                return WRITE_ONE_SKIP;
 510        }
 511
 512        /* if we are deltified, write out base object first. */
 513        if (e->delta) {
 514                e->idx.offset = 1; /* now recurse */
 515                switch (write_one(f, e->delta, offset)) {
 516                case WRITE_ONE_RECURSIVE:
 517                        /* we cannot depend on this one */
 518                        e->delta = NULL;
 519                        break;
 520                default:
 521                        break;
 522                case WRITE_ONE_BREAK:
 523                        e->idx.offset = recursing;
 524                        return WRITE_ONE_BREAK;
 525                }
 526        }
 527
 528        e->idx.offset = *offset;
 529        size = write_object(f, e, *offset);
 530        if (!size) {
 531                e->idx.offset = recursing;
 532                return WRITE_ONE_BREAK;
 533        }
 534        written_list[nr_written++] = &e->idx;
 535
 536        /* make sure off_t is sufficiently large not to wrap */
 537        if (signed_add_overflows(*offset, size))
 538                die("pack too large for current definition of off_t");
 539        *offset += size;
 540        return WRITE_ONE_WRITTEN;
 541}
 542
 543static int mark_tagged(const char *path, const unsigned char *sha1, int flag,
 544                       void *cb_data)
 545{
 546        unsigned char peeled[20];
 547        struct object_entry *entry = packlist_find(&to_pack, sha1, NULL);
 548
 549        if (entry)
 550                entry->tagged = 1;
 551        if (!peel_ref(path, peeled)) {
 552                entry = packlist_find(&to_pack, peeled, NULL);
 553                if (entry)
 554                        entry->tagged = 1;
 555        }
 556        return 0;
 557}
 558
 559static inline void add_to_write_order(struct object_entry **wo,
 560                               unsigned int *endp,
 561                               struct object_entry *e)
 562{
 563        if (e->filled)
 564                return;
 565        wo[(*endp)++] = e;
 566        e->filled = 1;
 567}
 568
 569static void add_descendants_to_write_order(struct object_entry **wo,
 570                                           unsigned int *endp,
 571                                           struct object_entry *e)
 572{
 573        int add_to_order = 1;
 574        while (e) {
 575                if (add_to_order) {
 576                        struct object_entry *s;
 577                        /* add this node... */
 578                        add_to_write_order(wo, endp, e);
 579                        /* all its siblings... */
 580                        for (s = e->delta_sibling; s; s = s->delta_sibling) {
 581                                add_to_write_order(wo, endp, s);
 582                        }
 583                }
 584                /* drop down a level to add left subtree nodes if possible */
 585                if (e->delta_child) {
 586                        add_to_order = 1;
 587                        e = e->delta_child;
 588                } else {
 589                        add_to_order = 0;
 590                        /* our sibling might have some children, it is next */
 591                        if (e->delta_sibling) {
 592                                e = e->delta_sibling;
 593                                continue;
 594                        }
 595                        /* go back to our parent node */
 596                        e = e->delta;
 597                        while (e && !e->delta_sibling) {
 598                                /* we're on the right side of a subtree, keep
 599                                 * going up until we can go right again */
 600                                e = e->delta;
 601                        }
 602                        if (!e) {
 603                                /* done- we hit our original root node */
 604                                return;
 605                        }
 606                        /* pass it off to sibling at this level */
 607                        e = e->delta_sibling;
 608                }
 609        };
 610}
 611
 612static void add_family_to_write_order(struct object_entry **wo,
 613                                      unsigned int *endp,
 614                                      struct object_entry *e)
 615{
 616        struct object_entry *root;
 617
 618        for (root = e; root->delta; root = root->delta)
 619                ; /* nothing */
 620        add_descendants_to_write_order(wo, endp, root);
 621}
 622
 623static struct object_entry **compute_write_order(void)
 624{
 625        unsigned int i, wo_end, last_untagged;
 626
 627        struct object_entry **wo = xmalloc(to_pack.nr_objects * sizeof(*wo));
 628        struct object_entry *objects = to_pack.objects;
 629
 630        for (i = 0; i < to_pack.nr_objects; i++) {
 631                objects[i].tagged = 0;
 632                objects[i].filled = 0;
 633                objects[i].delta_child = NULL;
 634                objects[i].delta_sibling = NULL;
 635        }
 636
 637        /*
 638         * Fully connect delta_child/delta_sibling network.
 639         * Make sure delta_sibling is sorted in the original
 640         * recency order.
 641         */
 642        for (i = to_pack.nr_objects; i > 0;) {
 643                struct object_entry *e = &objects[--i];
 644                if (!e->delta)
 645                        continue;
 646                /* Mark me as the first child */
 647                e->delta_sibling = e->delta->delta_child;
 648                e->delta->delta_child = e;
 649        }
 650
 651        /*
 652         * Mark objects that are at the tip of tags.
 653         */
 654        for_each_tag_ref(mark_tagged, NULL);
 655
 656        /*
 657         * Give the objects in the original recency order until
 658         * we see a tagged tip.
 659         */
 660        for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
 661                if (objects[i].tagged)
 662                        break;
 663                add_to_write_order(wo, &wo_end, &objects[i]);
 664        }
 665        last_untagged = i;
 666
 667        /*
 668         * Then fill all the tagged tips.
 669         */
 670        for (; i < to_pack.nr_objects; i++) {
 671                if (objects[i].tagged)
 672                        add_to_write_order(wo, &wo_end, &objects[i]);
 673        }
 674
 675        /*
 676         * And then all remaining commits and tags.
 677         */
 678        for (i = last_untagged; i < to_pack.nr_objects; i++) {
 679                if (objects[i].type != OBJ_COMMIT &&
 680                    objects[i].type != OBJ_TAG)
 681                        continue;
 682                add_to_write_order(wo, &wo_end, &objects[i]);
 683        }
 684
 685        /*
 686         * And then all the trees.
 687         */
 688        for (i = last_untagged; i < to_pack.nr_objects; i++) {
 689                if (objects[i].type != OBJ_TREE)
 690                        continue;
 691                add_to_write_order(wo, &wo_end, &objects[i]);
 692        }
 693
 694        /*
 695         * Finally all the rest in really tight order
 696         */
 697        for (i = last_untagged; i < to_pack.nr_objects; i++) {
 698                if (!objects[i].filled)
 699                        add_family_to_write_order(wo, &wo_end, &objects[i]);
 700        }
 701
 702        if (wo_end != to_pack.nr_objects)
 703                die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
 704
 705        return wo;
 706}
 707
 708static off_t write_reused_pack(struct sha1file *f)
 709{
 710        unsigned char buffer[8192];
 711        off_t to_write;
 712        int fd;
 713
 714        if (!is_pack_valid(reuse_packfile))
 715                die("packfile is invalid: %s", reuse_packfile->pack_name);
 716
 717        fd = git_open_noatime(reuse_packfile->pack_name);
 718        if (fd < 0)
 719                die_errno("unable to open packfile for reuse: %s",
 720                          reuse_packfile->pack_name);
 721
 722        if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
 723                die_errno("unable to seek in reused packfile");
 724
 725        if (reuse_packfile_offset < 0)
 726                reuse_packfile_offset = reuse_packfile->pack_size - 20;
 727
 728        to_write = reuse_packfile_offset - sizeof(struct pack_header);
 729
 730        while (to_write) {
 731                int read_pack = xread(fd, buffer, sizeof(buffer));
 732
 733                if (read_pack <= 0)
 734                        die_errno("unable to read from reused packfile");
 735
 736                if (read_pack > to_write)
 737                        read_pack = to_write;
 738
 739                sha1write(f, buffer, read_pack);
 740                to_write -= read_pack;
 741        }
 742
 743        close(fd);
 744        written += reuse_packfile_objects;
 745        return reuse_packfile_offset - sizeof(struct pack_header);
 746}
 747
 748static void write_pack_file(void)
 749{
 750        uint32_t i = 0, j;
 751        struct sha1file *f;
 752        off_t offset;
 753        uint32_t nr_remaining = nr_result;
 754        time_t last_mtime = 0;
 755        struct object_entry **write_order;
 756
 757        if (progress > pack_to_stdout)
 758                progress_state = start_progress("Writing objects", nr_result);
 759        written_list = xmalloc(to_pack.nr_objects * sizeof(*written_list));
 760        write_order = compute_write_order();
 761
 762        do {
 763                unsigned char sha1[20];
 764                char *pack_tmp_name = NULL;
 765
 766                if (pack_to_stdout)
 767                        f = sha1fd_throughput(1, "<stdout>", progress_state);
 768                else
 769                        f = create_tmp_packfile(&pack_tmp_name);
 770
 771                offset = write_pack_header(f, nr_remaining);
 772                if (!offset)
 773                        die_errno("unable to write pack header");
 774
 775                if (reuse_packfile) {
 776                        off_t packfile_size;
 777                        assert(pack_to_stdout);
 778
 779                        packfile_size = write_reused_pack(f);
 780                        offset += packfile_size;
 781                }
 782
 783                nr_written = 0;
 784                for (; i < to_pack.nr_objects; i++) {
 785                        struct object_entry *e = write_order[i];
 786                        if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 787                                break;
 788                        display_progress(progress_state, written);
 789                }
 790
 791                /*
 792                 * Did we write the wrong # entries in the header?
 793                 * If so, rewrite it like in fast-import
 794                 */
 795                if (pack_to_stdout) {
 796                        sha1close(f, sha1, CSUM_CLOSE);
 797                } else if (nr_written == nr_remaining) {
 798                        sha1close(f, sha1, CSUM_FSYNC);
 799                } else {
 800                        int fd = sha1close(f, sha1, 0);
 801                        fixup_pack_header_footer(fd, sha1, pack_tmp_name,
 802                                                 nr_written, sha1, offset);
 803                        close(fd);
 804                }
 805
 806                if (!pack_to_stdout) {
 807                        struct stat st;
 808                        char tmpname[PATH_MAX];
 809
 810                        /*
 811                         * Packs are runtime accessed in their mtime
 812                         * order since newer packs are more likely to contain
 813                         * younger objects.  So if we are creating multiple
 814                         * packs then we should modify the mtime of later ones
 815                         * to preserve this property.
 816                         */
 817                        if (stat(pack_tmp_name, &st) < 0) {
 818                                warning("failed to stat %s: %s",
 819                                        pack_tmp_name, strerror(errno));
 820                        } else if (!last_mtime) {
 821                                last_mtime = st.st_mtime;
 822                        } else {
 823                                struct utimbuf utb;
 824                                utb.actime = st.st_atime;
 825                                utb.modtime = --last_mtime;
 826                                if (utime(pack_tmp_name, &utb) < 0)
 827                                        warning("failed utime() on %s: %s",
 828                                                tmpname, strerror(errno));
 829                        }
 830
 831                        /* Enough space for "-<sha-1>.pack"? */
 832                        if (sizeof(tmpname) <= strlen(base_name) + 50)
 833                                die("pack base name '%s' too long", base_name);
 834                        snprintf(tmpname, sizeof(tmpname), "%s-", base_name);
 835
 836                        if (write_bitmap_index) {
 837                                bitmap_writer_set_checksum(sha1);
 838                                bitmap_writer_build_type_index(written_list, nr_written);
 839                        }
 840
 841                        finish_tmp_packfile(tmpname, pack_tmp_name,
 842                                            written_list, nr_written,
 843                                            &pack_idx_opts, sha1);
 844
 845                        if (write_bitmap_index) {
 846                                char *end_of_name_prefix = strrchr(tmpname, 0);
 847                                sprintf(end_of_name_prefix, "%s.bitmap", sha1_to_hex(sha1));
 848
 849                                stop_progress(&progress_state);
 850
 851                                bitmap_writer_show_progress(progress);
 852                                bitmap_writer_reuse_bitmaps(&to_pack);
 853                                bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
 854                                bitmap_writer_build(&to_pack);
 855                                bitmap_writer_finish(written_list, nr_written,
 856                                                     tmpname, write_bitmap_options);
 857                                write_bitmap_index = 0;
 858                        }
 859
 860                        free(pack_tmp_name);
 861                        puts(sha1_to_hex(sha1));
 862                }
 863
 864                /* mark written objects as written to previous pack */
 865                for (j = 0; j < nr_written; j++) {
 866                        written_list[j]->offset = (off_t)-1;
 867                }
 868                nr_remaining -= nr_written;
 869        } while (nr_remaining && i < to_pack.nr_objects);
 870
 871        free(written_list);
 872        free(write_order);
 873        stop_progress(&progress_state);
 874        if (written != nr_result)
 875                die("wrote %"PRIu32" objects while expecting %"PRIu32,
 876                        written, nr_result);
 877}
 878
 879static void setup_delta_attr_check(struct git_attr_check *check)
 880{
 881        static struct git_attr *attr_delta;
 882
 883        if (!attr_delta)
 884                attr_delta = git_attr("delta");
 885
 886        check[0].attr = attr_delta;
 887}
 888
 889static int no_try_delta(const char *path)
 890{
 891        struct git_attr_check check[1];
 892
 893        setup_delta_attr_check(check);
 894        if (git_check_attr(path, ARRAY_SIZE(check), check))
 895                return 0;
 896        if (ATTR_FALSE(check->value))
 897                return 1;
 898        return 0;
 899}
 900
 901/*
 902 * When adding an object, check whether we have already added it
 903 * to our packing list. If so, we can skip. However, if we are
 904 * being asked to excludei t, but the previous mention was to include
 905 * it, make sure to adjust its flags and tweak our numbers accordingly.
 906 *
 907 * As an optimization, we pass out the index position where we would have
 908 * found the item, since that saves us from having to look it up again a
 909 * few lines later when we want to add the new entry.
 910 */
 911static int have_duplicate_entry(const unsigned char *sha1,
 912                                int exclude,
 913                                uint32_t *index_pos)
 914{
 915        struct object_entry *entry;
 916
 917        entry = packlist_find(&to_pack, sha1, index_pos);
 918        if (!entry)
 919                return 0;
 920
 921        if (exclude) {
 922                if (!entry->preferred_base)
 923                        nr_result--;
 924                entry->preferred_base = 1;
 925        }
 926
 927        return 1;
 928}
 929
 930/*
 931 * Check whether we want the object in the pack (e.g., we do not want
 932 * objects found in non-local stores if the "--local" option was used).
 933 *
 934 * As a side effect of this check, we will find the packed version of this
 935 * object, if any. We therefore pass out the pack information to avoid having
 936 * to look it up again later.
 937 */
 938static int want_object_in_pack(const unsigned char *sha1,
 939                               int exclude,
 940                               struct packed_git **found_pack,
 941                               off_t *found_offset)
 942{
 943        struct packed_git *p;
 944
 945        if (!exclude && local && has_loose_object_nonlocal(sha1))
 946                return 0;
 947
 948        *found_pack = NULL;
 949        *found_offset = 0;
 950
 951        for (p = packed_git; p; p = p->next) {
 952                off_t offset = find_pack_entry_one(sha1, p);
 953                if (offset) {
 954                        if (!*found_pack) {
 955                                if (!is_pack_valid(p)) {
 956                                        warning("packfile %s cannot be accessed", p->pack_name);
 957                                        continue;
 958                                }
 959                                *found_offset = offset;
 960                                *found_pack = p;
 961                        }
 962                        if (exclude)
 963                                return 1;
 964                        if (incremental)
 965                                return 0;
 966                        if (local && !p->pack_local)
 967                                return 0;
 968                        if (ignore_packed_keep && p->pack_local && p->pack_keep)
 969                                return 0;
 970                }
 971        }
 972
 973        return 1;
 974}
 975
 976static void create_object_entry(const unsigned char *sha1,
 977                                enum object_type type,
 978                                uint32_t hash,
 979                                int exclude,
 980                                int no_try_delta,
 981                                uint32_t index_pos,
 982                                struct packed_git *found_pack,
 983                                off_t found_offset)
 984{
 985        struct object_entry *entry;
 986
 987        entry = packlist_alloc(&to_pack, sha1, index_pos);
 988        entry->hash = hash;
 989        if (type)
 990                entry->type = type;
 991        if (exclude)
 992                entry->preferred_base = 1;
 993        else
 994                nr_result++;
 995        if (found_pack) {
 996                entry->in_pack = found_pack;
 997                entry->in_pack_offset = found_offset;
 998        }
 999
1000        entry->no_try_delta = no_try_delta;
1001}
1002
1003static int add_object_entry(const unsigned char *sha1, enum object_type type,
1004                            const char *name, int exclude)
1005{
1006        struct packed_git *found_pack;
1007        off_t found_offset;
1008        uint32_t index_pos;
1009
1010        if (have_duplicate_entry(sha1, exclude, &index_pos))
1011                return 0;
1012
1013        if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset))
1014                return 0;
1015
1016        create_object_entry(sha1, type, pack_name_hash(name),
1017                            exclude, name && no_try_delta(name),
1018                            index_pos, found_pack, found_offset);
1019
1020        display_progress(progress_state, to_pack.nr_objects);
1021        return 1;
1022}
1023
1024static int add_object_entry_from_bitmap(const unsigned char *sha1,
1025                                        enum object_type type,
1026                                        int flags, uint32_t name_hash,
1027                                        struct packed_git *pack, off_t offset)
1028{
1029        uint32_t index_pos;
1030
1031        if (have_duplicate_entry(sha1, 0, &index_pos))
1032                return 0;
1033
1034        create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1035
1036        display_progress(progress_state, to_pack.nr_objects);
1037        return 1;
1038}
1039
1040struct pbase_tree_cache {
1041        unsigned char sha1[20];
1042        int ref;
1043        int temporary;
1044        void *tree_data;
1045        unsigned long tree_size;
1046};
1047
1048static struct pbase_tree_cache *(pbase_tree_cache[256]);
1049static int pbase_tree_cache_ix(const unsigned char *sha1)
1050{
1051        return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1052}
1053static int pbase_tree_cache_ix_incr(int ix)
1054{
1055        return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1056}
1057
1058static struct pbase_tree {
1059        struct pbase_tree *next;
1060        /* This is a phony "cache" entry; we are not
1061         * going to evict it nor find it through _get()
1062         * mechanism -- this is for the toplevel node that
1063         * would almost always change with any commit.
1064         */
1065        struct pbase_tree_cache pcache;
1066} *pbase_tree;
1067
1068static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1069{
1070        struct pbase_tree_cache *ent, *nent;
1071        void *data;
1072        unsigned long size;
1073        enum object_type type;
1074        int neigh;
1075        int my_ix = pbase_tree_cache_ix(sha1);
1076        int available_ix = -1;
1077
1078        /* pbase-tree-cache acts as a limited hashtable.
1079         * your object will be found at your index or within a few
1080         * slots after that slot if it is cached.
1081         */
1082        for (neigh = 0; neigh < 8; neigh++) {
1083                ent = pbase_tree_cache[my_ix];
1084                if (ent && !hashcmp(ent->sha1, sha1)) {
1085                        ent->ref++;
1086                        return ent;
1087                }
1088                else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1089                         ((0 <= available_ix) &&
1090                          (!ent && pbase_tree_cache[available_ix])))
1091                        available_ix = my_ix;
1092                if (!ent)
1093                        break;
1094                my_ix = pbase_tree_cache_ix_incr(my_ix);
1095        }
1096
1097        /* Did not find one.  Either we got a bogus request or
1098         * we need to read and perhaps cache.
1099         */
1100        data = read_sha1_file(sha1, &type, &size);
1101        if (!data)
1102                return NULL;
1103        if (type != OBJ_TREE) {
1104                free(data);
1105                return NULL;
1106        }
1107
1108        /* We need to either cache or return a throwaway copy */
1109
1110        if (available_ix < 0)
1111                ent = NULL;
1112        else {
1113                ent = pbase_tree_cache[available_ix];
1114                my_ix = available_ix;
1115        }
1116
1117        if (!ent) {
1118                nent = xmalloc(sizeof(*nent));
1119                nent->temporary = (available_ix < 0);
1120        }
1121        else {
1122                /* evict and reuse */
1123                free(ent->tree_data);
1124                nent = ent;
1125        }
1126        hashcpy(nent->sha1, sha1);
1127        nent->tree_data = data;
1128        nent->tree_size = size;
1129        nent->ref = 1;
1130        if (!nent->temporary)
1131                pbase_tree_cache[my_ix] = nent;
1132        return nent;
1133}
1134
1135static void pbase_tree_put(struct pbase_tree_cache *cache)
1136{
1137        if (!cache->temporary) {
1138                cache->ref--;
1139                return;
1140        }
1141        free(cache->tree_data);
1142        free(cache);
1143}
1144
1145static int name_cmp_len(const char *name)
1146{
1147        int i;
1148        for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1149                ;
1150        return i;
1151}
1152
1153static void add_pbase_object(struct tree_desc *tree,
1154                             const char *name,
1155                             int cmplen,
1156                             const char *fullname)
1157{
1158        struct name_entry entry;
1159        int cmp;
1160
1161        while (tree_entry(tree,&entry)) {
1162                if (S_ISGITLINK(entry.mode))
1163                        continue;
1164                cmp = tree_entry_len(&entry) != cmplen ? 1 :
1165                      memcmp(name, entry.path, cmplen);
1166                if (cmp > 0)
1167                        continue;
1168                if (cmp < 0)
1169                        return;
1170                if (name[cmplen] != '/') {
1171                        add_object_entry(entry.sha1,
1172                                         object_type(entry.mode),
1173                                         fullname, 1);
1174                        return;
1175                }
1176                if (S_ISDIR(entry.mode)) {
1177                        struct tree_desc sub;
1178                        struct pbase_tree_cache *tree;
1179                        const char *down = name+cmplen+1;
1180                        int downlen = name_cmp_len(down);
1181
1182                        tree = pbase_tree_get(entry.sha1);
1183                        if (!tree)
1184                                return;
1185                        init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1186
1187                        add_pbase_object(&sub, down, downlen, fullname);
1188                        pbase_tree_put(tree);
1189                }
1190        }
1191}
1192
1193static unsigned *done_pbase_paths;
1194static int done_pbase_paths_num;
1195static int done_pbase_paths_alloc;
1196static int done_pbase_path_pos(unsigned hash)
1197{
1198        int lo = 0;
1199        int hi = done_pbase_paths_num;
1200        while (lo < hi) {
1201                int mi = (hi + lo) / 2;
1202                if (done_pbase_paths[mi] == hash)
1203                        return mi;
1204                if (done_pbase_paths[mi] < hash)
1205                        hi = mi;
1206                else
1207                        lo = mi + 1;
1208        }
1209        return -lo-1;
1210}
1211
1212static int check_pbase_path(unsigned hash)
1213{
1214        int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1215        if (0 <= pos)
1216                return 1;
1217        pos = -pos - 1;
1218        if (done_pbase_paths_alloc <= done_pbase_paths_num) {
1219                done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
1220                done_pbase_paths = xrealloc(done_pbase_paths,
1221                                            done_pbase_paths_alloc *
1222                                            sizeof(unsigned));
1223        }
1224        done_pbase_paths_num++;
1225        if (pos < done_pbase_paths_num)
1226                memmove(done_pbase_paths + pos + 1,
1227                        done_pbase_paths + pos,
1228                        (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1229        done_pbase_paths[pos] = hash;
1230        return 0;
1231}
1232
1233static void add_preferred_base_object(const char *name)
1234{
1235        struct pbase_tree *it;
1236        int cmplen;
1237        unsigned hash = pack_name_hash(name);
1238
1239        if (!num_preferred_base || check_pbase_path(hash))
1240                return;
1241
1242        cmplen = name_cmp_len(name);
1243        for (it = pbase_tree; it; it = it->next) {
1244                if (cmplen == 0) {
1245                        add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1246                }
1247                else {
1248                        struct tree_desc tree;
1249                        init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1250                        add_pbase_object(&tree, name, cmplen, name);
1251                }
1252        }
1253}
1254
1255static void add_preferred_base(unsigned char *sha1)
1256{
1257        struct pbase_tree *it;
1258        void *data;
1259        unsigned long size;
1260        unsigned char tree_sha1[20];
1261
1262        if (window <= num_preferred_base++)
1263                return;
1264
1265        data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1266        if (!data)
1267                return;
1268
1269        for (it = pbase_tree; it; it = it->next) {
1270                if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1271                        free(data);
1272                        return;
1273                }
1274        }
1275
1276        it = xcalloc(1, sizeof(*it));
1277        it->next = pbase_tree;
1278        pbase_tree = it;
1279
1280        hashcpy(it->pcache.sha1, tree_sha1);
1281        it->pcache.tree_data = data;
1282        it->pcache.tree_size = size;
1283}
1284
1285static void cleanup_preferred_base(void)
1286{
1287        struct pbase_tree *it;
1288        unsigned i;
1289
1290        it = pbase_tree;
1291        pbase_tree = NULL;
1292        while (it) {
1293                struct pbase_tree *this = it;
1294                it = this->next;
1295                free(this->pcache.tree_data);
1296                free(this);
1297        }
1298
1299        for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1300                if (!pbase_tree_cache[i])
1301                        continue;
1302                free(pbase_tree_cache[i]->tree_data);
1303                free(pbase_tree_cache[i]);
1304                pbase_tree_cache[i] = NULL;
1305        }
1306
1307        free(done_pbase_paths);
1308        done_pbase_paths = NULL;
1309        done_pbase_paths_num = done_pbase_paths_alloc = 0;
1310}
1311
1312static void check_object(struct object_entry *entry)
1313{
1314        if (entry->in_pack) {
1315                struct packed_git *p = entry->in_pack;
1316                struct pack_window *w_curs = NULL;
1317                const unsigned char *base_ref = NULL;
1318                struct object_entry *base_entry;
1319                unsigned long used, used_0;
1320                unsigned long avail;
1321                off_t ofs;
1322                unsigned char *buf, c;
1323
1324                buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1325
1326                /*
1327                 * We want in_pack_type even if we do not reuse delta
1328                 * since non-delta representations could still be reused.
1329                 */
1330                used = unpack_object_header_buffer(buf, avail,
1331                                                   &entry->in_pack_type,
1332                                                   &entry->size);
1333                if (used == 0)
1334                        goto give_up;
1335
1336                /*
1337                 * Determine if this is a delta and if so whether we can
1338                 * reuse it or not.  Otherwise let's find out as cheaply as
1339                 * possible what the actual type and size for this object is.
1340                 */
1341                switch (entry->in_pack_type) {
1342                default:
1343                        /* Not a delta hence we've already got all we need. */
1344                        entry->type = entry->in_pack_type;
1345                        entry->in_pack_header_size = used;
1346                        if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1347                                goto give_up;
1348                        unuse_pack(&w_curs);
1349                        return;
1350                case OBJ_REF_DELTA:
1351                        if (reuse_delta && !entry->preferred_base)
1352                                base_ref = use_pack(p, &w_curs,
1353                                                entry->in_pack_offset + used, NULL);
1354                        entry->in_pack_header_size = used + 20;
1355                        break;
1356                case OBJ_OFS_DELTA:
1357                        buf = use_pack(p, &w_curs,
1358                                       entry->in_pack_offset + used, NULL);
1359                        used_0 = 0;
1360                        c = buf[used_0++];
1361                        ofs = c & 127;
1362                        while (c & 128) {
1363                                ofs += 1;
1364                                if (!ofs || MSB(ofs, 7)) {
1365                                        error("delta base offset overflow in pack for %s",
1366                                              sha1_to_hex(entry->idx.sha1));
1367                                        goto give_up;
1368                                }
1369                                c = buf[used_0++];
1370                                ofs = (ofs << 7) + (c & 127);
1371                        }
1372                        ofs = entry->in_pack_offset - ofs;
1373                        if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1374                                error("delta base offset out of bound for %s",
1375                                      sha1_to_hex(entry->idx.sha1));
1376                                goto give_up;
1377                        }
1378                        if (reuse_delta && !entry->preferred_base) {
1379                                struct revindex_entry *revidx;
1380                                revidx = find_pack_revindex(p, ofs);
1381                                if (!revidx)
1382                                        goto give_up;
1383                                base_ref = nth_packed_object_sha1(p, revidx->nr);
1384                        }
1385                        entry->in_pack_header_size = used + used_0;
1386                        break;
1387                }
1388
1389                if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1390                        /*
1391                         * If base_ref was set above that means we wish to
1392                         * reuse delta data, and we even found that base
1393                         * in the list of objects we want to pack. Goodie!
1394                         *
1395                         * Depth value does not matter - find_deltas() will
1396                         * never consider reused delta as the base object to
1397                         * deltify other objects against, in order to avoid
1398                         * circular deltas.
1399                         */
1400                        entry->type = entry->in_pack_type;
1401                        entry->delta = base_entry;
1402                        entry->delta_size = entry->size;
1403                        entry->delta_sibling = base_entry->delta_child;
1404                        base_entry->delta_child = entry;
1405                        unuse_pack(&w_curs);
1406                        return;
1407                }
1408
1409                if (entry->type) {
1410                        /*
1411                         * This must be a delta and we already know what the
1412                         * final object type is.  Let's extract the actual
1413                         * object size from the delta header.
1414                         */
1415                        entry->size = get_size_from_delta(p, &w_curs,
1416                                        entry->in_pack_offset + entry->in_pack_header_size);
1417                        if (entry->size == 0)
1418                                goto give_up;
1419                        unuse_pack(&w_curs);
1420                        return;
1421                }
1422
1423                /*
1424                 * No choice but to fall back to the recursive delta walk
1425                 * with sha1_object_info() to find about the object type
1426                 * at this point...
1427                 */
1428                give_up:
1429                unuse_pack(&w_curs);
1430        }
1431
1432        entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1433        /*
1434         * The error condition is checked in prepare_pack().  This is
1435         * to permit a missing preferred base object to be ignored
1436         * as a preferred base.  Doing so can result in a larger
1437         * pack file, but the transfer will still take place.
1438         */
1439}
1440
1441static int pack_offset_sort(const void *_a, const void *_b)
1442{
1443        const struct object_entry *a = *(struct object_entry **)_a;
1444        const struct object_entry *b = *(struct object_entry **)_b;
1445
1446        /* avoid filesystem trashing with loose objects */
1447        if (!a->in_pack && !b->in_pack)
1448                return hashcmp(a->idx.sha1, b->idx.sha1);
1449
1450        if (a->in_pack < b->in_pack)
1451                return -1;
1452        if (a->in_pack > b->in_pack)
1453                return 1;
1454        return a->in_pack_offset < b->in_pack_offset ? -1 :
1455                        (a->in_pack_offset > b->in_pack_offset);
1456}
1457
1458static void get_object_details(void)
1459{
1460        uint32_t i;
1461        struct object_entry **sorted_by_offset;
1462
1463        sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1464        for (i = 0; i < to_pack.nr_objects; i++)
1465                sorted_by_offset[i] = to_pack.objects + i;
1466        qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1467
1468        for (i = 0; i < to_pack.nr_objects; i++) {
1469                struct object_entry *entry = sorted_by_offset[i];
1470                check_object(entry);
1471                if (big_file_threshold < entry->size)
1472                        entry->no_try_delta = 1;
1473        }
1474
1475        free(sorted_by_offset);
1476}
1477
1478/*
1479 * We search for deltas in a list sorted by type, by filename hash, and then
1480 * by size, so that we see progressively smaller and smaller files.
1481 * That's because we prefer deltas to be from the bigger file
1482 * to the smaller -- deletes are potentially cheaper, but perhaps
1483 * more importantly, the bigger file is likely the more recent
1484 * one.  The deepest deltas are therefore the oldest objects which are
1485 * less susceptible to be accessed often.
1486 */
1487static int type_size_sort(const void *_a, const void *_b)
1488{
1489        const struct object_entry *a = *(struct object_entry **)_a;
1490        const struct object_entry *b = *(struct object_entry **)_b;
1491
1492        if (a->type > b->type)
1493                return -1;
1494        if (a->type < b->type)
1495                return 1;
1496        if (a->hash > b->hash)
1497                return -1;
1498        if (a->hash < b->hash)
1499                return 1;
1500        if (a->preferred_base > b->preferred_base)
1501                return -1;
1502        if (a->preferred_base < b->preferred_base)
1503                return 1;
1504        if (a->size > b->size)
1505                return -1;
1506        if (a->size < b->size)
1507                return 1;
1508        return a < b ? -1 : (a > b);  /* newest first */
1509}
1510
1511struct unpacked {
1512        struct object_entry *entry;
1513        void *data;
1514        struct delta_index *index;
1515        unsigned depth;
1516};
1517
1518static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1519                           unsigned long delta_size)
1520{
1521        if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1522                return 0;
1523
1524        if (delta_size < cache_max_small_delta_size)
1525                return 1;
1526
1527        /* cache delta, if objects are large enough compared to delta size */
1528        if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1529                return 1;
1530
1531        return 0;
1532}
1533
1534#ifndef NO_PTHREADS
1535
1536static pthread_mutex_t read_mutex;
1537#define read_lock()             pthread_mutex_lock(&read_mutex)
1538#define read_unlock()           pthread_mutex_unlock(&read_mutex)
1539
1540static pthread_mutex_t cache_mutex;
1541#define cache_lock()            pthread_mutex_lock(&cache_mutex)
1542#define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1543
1544static pthread_mutex_t progress_mutex;
1545#define progress_lock()         pthread_mutex_lock(&progress_mutex)
1546#define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1547
1548#else
1549
1550#define read_lock()             (void)0
1551#define read_unlock()           (void)0
1552#define cache_lock()            (void)0
1553#define cache_unlock()          (void)0
1554#define progress_lock()         (void)0
1555#define progress_unlock()       (void)0
1556
1557#endif
1558
1559static int try_delta(struct unpacked *trg, struct unpacked *src,
1560                     unsigned max_depth, unsigned long *mem_usage)
1561{
1562        struct object_entry *trg_entry = trg->entry;
1563        struct object_entry *src_entry = src->entry;
1564        unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1565        unsigned ref_depth;
1566        enum object_type type;
1567        void *delta_buf;
1568
1569        /* Don't bother doing diffs between different types */
1570        if (trg_entry->type != src_entry->type)
1571                return -1;
1572
1573        /*
1574         * We do not bother to try a delta that we discarded on an
1575         * earlier try, but only when reusing delta data.  Note that
1576         * src_entry that is marked as the preferred_base should always
1577         * be considered, as even if we produce a suboptimal delta against
1578         * it, we will still save the transfer cost, as we already know
1579         * the other side has it and we won't send src_entry at all.
1580         */
1581        if (reuse_delta && trg_entry->in_pack &&
1582            trg_entry->in_pack == src_entry->in_pack &&
1583            !src_entry->preferred_base &&
1584            trg_entry->in_pack_type != OBJ_REF_DELTA &&
1585            trg_entry->in_pack_type != OBJ_OFS_DELTA)
1586                return 0;
1587
1588        /* Let's not bust the allowed depth. */
1589        if (src->depth >= max_depth)
1590                return 0;
1591
1592        /* Now some size filtering heuristics. */
1593        trg_size = trg_entry->size;
1594        if (!trg_entry->delta) {
1595                max_size = trg_size/2 - 20;
1596                ref_depth = 1;
1597        } else {
1598                max_size = trg_entry->delta_size;
1599                ref_depth = trg->depth;
1600        }
1601        max_size = (uint64_t)max_size * (max_depth - src->depth) /
1602                                                (max_depth - ref_depth + 1);
1603        if (max_size == 0)
1604                return 0;
1605        src_size = src_entry->size;
1606        sizediff = src_size < trg_size ? trg_size - src_size : 0;
1607        if (sizediff >= max_size)
1608                return 0;
1609        if (trg_size < src_size / 32)
1610                return 0;
1611
1612        /* Load data if not already done */
1613        if (!trg->data) {
1614                read_lock();
1615                trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1616                read_unlock();
1617                if (!trg->data)
1618                        die("object %s cannot be read",
1619                            sha1_to_hex(trg_entry->idx.sha1));
1620                if (sz != trg_size)
1621                        die("object %s inconsistent object length (%lu vs %lu)",
1622                            sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1623                *mem_usage += sz;
1624        }
1625        if (!src->data) {
1626                read_lock();
1627                src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1628                read_unlock();
1629                if (!src->data) {
1630                        if (src_entry->preferred_base) {
1631                                static int warned = 0;
1632                                if (!warned++)
1633                                        warning("object %s cannot be read",
1634                                                sha1_to_hex(src_entry->idx.sha1));
1635                                /*
1636                                 * Those objects are not included in the
1637                                 * resulting pack.  Be resilient and ignore
1638                                 * them if they can't be read, in case the
1639                                 * pack could be created nevertheless.
1640                                 */
1641                                return 0;
1642                        }
1643                        die("object %s cannot be read",
1644                            sha1_to_hex(src_entry->idx.sha1));
1645                }
1646                if (sz != src_size)
1647                        die("object %s inconsistent object length (%lu vs %lu)",
1648                            sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1649                *mem_usage += sz;
1650        }
1651        if (!src->index) {
1652                src->index = create_delta_index(src->data, src_size);
1653                if (!src->index) {
1654                        static int warned = 0;
1655                        if (!warned++)
1656                                warning("suboptimal pack - out of memory");
1657                        return 0;
1658                }
1659                *mem_usage += sizeof_delta_index(src->index);
1660        }
1661
1662        delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1663        if (!delta_buf)
1664                return 0;
1665
1666        if (trg_entry->delta) {
1667                /* Prefer only shallower same-sized deltas. */
1668                if (delta_size == trg_entry->delta_size &&
1669                    src->depth + 1 >= trg->depth) {
1670                        free(delta_buf);
1671                        return 0;
1672                }
1673        }
1674
1675        /*
1676         * Handle memory allocation outside of the cache
1677         * accounting lock.  Compiler will optimize the strangeness
1678         * away when NO_PTHREADS is defined.
1679         */
1680        free(trg_entry->delta_data);
1681        cache_lock();
1682        if (trg_entry->delta_data) {
1683                delta_cache_size -= trg_entry->delta_size;
1684                trg_entry->delta_data = NULL;
1685        }
1686        if (delta_cacheable(src_size, trg_size, delta_size)) {
1687                delta_cache_size += delta_size;
1688                cache_unlock();
1689                trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1690        } else {
1691                cache_unlock();
1692                free(delta_buf);
1693        }
1694
1695        trg_entry->delta = src_entry;
1696        trg_entry->delta_size = delta_size;
1697        trg->depth = src->depth + 1;
1698
1699        return 1;
1700}
1701
1702static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1703{
1704        struct object_entry *child = me->delta_child;
1705        unsigned int m = n;
1706        while (child) {
1707                unsigned int c = check_delta_limit(child, n + 1);
1708                if (m < c)
1709                        m = c;
1710                child = child->delta_sibling;
1711        }
1712        return m;
1713}
1714
1715static unsigned long free_unpacked(struct unpacked *n)
1716{
1717        unsigned long freed_mem = sizeof_delta_index(n->index);
1718        free_delta_index(n->index);
1719        n->index = NULL;
1720        if (n->data) {
1721                freed_mem += n->entry->size;
1722                free(n->data);
1723                n->data = NULL;
1724        }
1725        n->entry = NULL;
1726        n->depth = 0;
1727        return freed_mem;
1728}
1729
1730static void find_deltas(struct object_entry **list, unsigned *list_size,
1731                        int window, int depth, unsigned *processed)
1732{
1733        uint32_t i, idx = 0, count = 0;
1734        struct unpacked *array;
1735        unsigned long mem_usage = 0;
1736
1737        array = xcalloc(window, sizeof(struct unpacked));
1738
1739        for (;;) {
1740                struct object_entry *entry;
1741                struct unpacked *n = array + idx;
1742                int j, max_depth, best_base = -1;
1743
1744                progress_lock();
1745                if (!*list_size) {
1746                        progress_unlock();
1747                        break;
1748                }
1749                entry = *list++;
1750                (*list_size)--;
1751                if (!entry->preferred_base) {
1752                        (*processed)++;
1753                        display_progress(progress_state, *processed);
1754                }
1755                progress_unlock();
1756
1757                mem_usage -= free_unpacked(n);
1758                n->entry = entry;
1759
1760                while (window_memory_limit &&
1761                       mem_usage > window_memory_limit &&
1762                       count > 1) {
1763                        uint32_t tail = (idx + window - count) % window;
1764                        mem_usage -= free_unpacked(array + tail);
1765                        count--;
1766                }
1767
1768                /* We do not compute delta to *create* objects we are not
1769                 * going to pack.
1770                 */
1771                if (entry->preferred_base)
1772                        goto next;
1773
1774                /*
1775                 * If the current object is at pack edge, take the depth the
1776                 * objects that depend on the current object into account
1777                 * otherwise they would become too deep.
1778                 */
1779                max_depth = depth;
1780                if (entry->delta_child) {
1781                        max_depth -= check_delta_limit(entry, 0);
1782                        if (max_depth <= 0)
1783                                goto next;
1784                }
1785
1786                j = window;
1787                while (--j > 0) {
1788                        int ret;
1789                        uint32_t other_idx = idx + j;
1790                        struct unpacked *m;
1791                        if (other_idx >= window)
1792                                other_idx -= window;
1793                        m = array + other_idx;
1794                        if (!m->entry)
1795                                break;
1796                        ret = try_delta(n, m, max_depth, &mem_usage);
1797                        if (ret < 0)
1798                                break;
1799                        else if (ret > 0)
1800                                best_base = other_idx;
1801                }
1802
1803                /*
1804                 * If we decided to cache the delta data, then it is best
1805                 * to compress it right away.  First because we have to do
1806                 * it anyway, and doing it here while we're threaded will
1807                 * save a lot of time in the non threaded write phase,
1808                 * as well as allow for caching more deltas within
1809                 * the same cache size limit.
1810                 * ...
1811                 * But only if not writing to stdout, since in that case
1812                 * the network is most likely throttling writes anyway,
1813                 * and therefore it is best to go to the write phase ASAP
1814                 * instead, as we can afford spending more time compressing
1815                 * between writes at that moment.
1816                 */
1817                if (entry->delta_data && !pack_to_stdout) {
1818                        entry->z_delta_size = do_compress(&entry->delta_data,
1819                                                          entry->delta_size);
1820                        cache_lock();
1821                        delta_cache_size -= entry->delta_size;
1822                        delta_cache_size += entry->z_delta_size;
1823                        cache_unlock();
1824                }
1825
1826                /* if we made n a delta, and if n is already at max
1827                 * depth, leaving it in the window is pointless.  we
1828                 * should evict it first.
1829                 */
1830                if (entry->delta && max_depth <= n->depth)
1831                        continue;
1832
1833                /*
1834                 * Move the best delta base up in the window, after the
1835                 * currently deltified object, to keep it longer.  It will
1836                 * be the first base object to be attempted next.
1837                 */
1838                if (entry->delta) {
1839                        struct unpacked swap = array[best_base];
1840                        int dist = (window + idx - best_base) % window;
1841                        int dst = best_base;
1842                        while (dist--) {
1843                                int src = (dst + 1) % window;
1844                                array[dst] = array[src];
1845                                dst = src;
1846                        }
1847                        array[dst] = swap;
1848                }
1849
1850                next:
1851                idx++;
1852                if (count + 1 < window)
1853                        count++;
1854                if (idx >= window)
1855                        idx = 0;
1856        }
1857
1858        for (i = 0; i < window; ++i) {
1859                free_delta_index(array[i].index);
1860                free(array[i].data);
1861        }
1862        free(array);
1863}
1864
1865#ifndef NO_PTHREADS
1866
1867static void try_to_free_from_threads(size_t size)
1868{
1869        read_lock();
1870        release_pack_memory(size);
1871        read_unlock();
1872}
1873
1874static try_to_free_t old_try_to_free_routine;
1875
1876/*
1877 * The main thread waits on the condition that (at least) one of the workers
1878 * has stopped working (which is indicated in the .working member of
1879 * struct thread_params).
1880 * When a work thread has completed its work, it sets .working to 0 and
1881 * signals the main thread and waits on the condition that .data_ready
1882 * becomes 1.
1883 */
1884
1885struct thread_params {
1886        pthread_t thread;
1887        struct object_entry **list;
1888        unsigned list_size;
1889        unsigned remaining;
1890        int window;
1891        int depth;
1892        int working;
1893        int data_ready;
1894        pthread_mutex_t mutex;
1895        pthread_cond_t cond;
1896        unsigned *processed;
1897};
1898
1899static pthread_cond_t progress_cond;
1900
1901/*
1902 * Mutex and conditional variable can't be statically-initialized on Windows.
1903 */
1904static void init_threaded_search(void)
1905{
1906        init_recursive_mutex(&read_mutex);
1907        pthread_mutex_init(&cache_mutex, NULL);
1908        pthread_mutex_init(&progress_mutex, NULL);
1909        pthread_cond_init(&progress_cond, NULL);
1910        old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1911}
1912
1913static void cleanup_threaded_search(void)
1914{
1915        set_try_to_free_routine(old_try_to_free_routine);
1916        pthread_cond_destroy(&progress_cond);
1917        pthread_mutex_destroy(&read_mutex);
1918        pthread_mutex_destroy(&cache_mutex);
1919        pthread_mutex_destroy(&progress_mutex);
1920}
1921
1922static void *threaded_find_deltas(void *arg)
1923{
1924        struct thread_params *me = arg;
1925
1926        while (me->remaining) {
1927                find_deltas(me->list, &me->remaining,
1928                            me->window, me->depth, me->processed);
1929
1930                progress_lock();
1931                me->working = 0;
1932                pthread_cond_signal(&progress_cond);
1933                progress_unlock();
1934
1935                /*
1936                 * We must not set ->data_ready before we wait on the
1937                 * condition because the main thread may have set it to 1
1938                 * before we get here. In order to be sure that new
1939                 * work is available if we see 1 in ->data_ready, it
1940                 * was initialized to 0 before this thread was spawned
1941                 * and we reset it to 0 right away.
1942                 */
1943                pthread_mutex_lock(&me->mutex);
1944                while (!me->data_ready)
1945                        pthread_cond_wait(&me->cond, &me->mutex);
1946                me->data_ready = 0;
1947                pthread_mutex_unlock(&me->mutex);
1948        }
1949        /* leave ->working 1 so that this doesn't get more work assigned */
1950        return NULL;
1951}
1952
1953static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1954                           int window, int depth, unsigned *processed)
1955{
1956        struct thread_params *p;
1957        int i, ret, active_threads = 0;
1958
1959        init_threaded_search();
1960
1961        if (!delta_search_threads)      /* --threads=0 means autodetect */
1962                delta_search_threads = online_cpus();
1963        if (delta_search_threads <= 1) {
1964                find_deltas(list, &list_size, window, depth, processed);
1965                cleanup_threaded_search();
1966                return;
1967        }
1968        if (progress > pack_to_stdout)
1969                fprintf(stderr, "Delta compression using up to %d threads.\n",
1970                                delta_search_threads);
1971        p = xcalloc(delta_search_threads, sizeof(*p));
1972
1973        /* Partition the work amongst work threads. */
1974        for (i = 0; i < delta_search_threads; i++) {
1975                unsigned sub_size = list_size / (delta_search_threads - i);
1976
1977                /* don't use too small segments or no deltas will be found */
1978                if (sub_size < 2*window && i+1 < delta_search_threads)
1979                        sub_size = 0;
1980
1981                p[i].window = window;
1982                p[i].depth = depth;
1983                p[i].processed = processed;
1984                p[i].working = 1;
1985                p[i].data_ready = 0;
1986
1987                /* try to split chunks on "path" boundaries */
1988                while (sub_size && sub_size < list_size &&
1989                       list[sub_size]->hash &&
1990                       list[sub_size]->hash == list[sub_size-1]->hash)
1991                        sub_size++;
1992
1993                p[i].list = list;
1994                p[i].list_size = sub_size;
1995                p[i].remaining = sub_size;
1996
1997                list += sub_size;
1998                list_size -= sub_size;
1999        }
2000
2001        /* Start work threads. */
2002        for (i = 0; i < delta_search_threads; i++) {
2003                if (!p[i].list_size)
2004                        continue;
2005                pthread_mutex_init(&p[i].mutex, NULL);
2006                pthread_cond_init(&p[i].cond, NULL);
2007                ret = pthread_create(&p[i].thread, NULL,
2008                                     threaded_find_deltas, &p[i]);
2009                if (ret)
2010                        die("unable to create thread: %s", strerror(ret));
2011                active_threads++;
2012        }
2013
2014        /*
2015         * Now let's wait for work completion.  Each time a thread is done
2016         * with its work, we steal half of the remaining work from the
2017         * thread with the largest number of unprocessed objects and give
2018         * it to that newly idle thread.  This ensure good load balancing
2019         * until the remaining object list segments are simply too short
2020         * to be worth splitting anymore.
2021         */
2022        while (active_threads) {
2023                struct thread_params *target = NULL;
2024                struct thread_params *victim = NULL;
2025                unsigned sub_size = 0;
2026
2027                progress_lock();
2028                for (;;) {
2029                        for (i = 0; !target && i < delta_search_threads; i++)
2030                                if (!p[i].working)
2031                                        target = &p[i];
2032                        if (target)
2033                                break;
2034                        pthread_cond_wait(&progress_cond, &progress_mutex);
2035                }
2036
2037                for (i = 0; i < delta_search_threads; i++)
2038                        if (p[i].remaining > 2*window &&
2039                            (!victim || victim->remaining < p[i].remaining))
2040                                victim = &p[i];
2041                if (victim) {
2042                        sub_size = victim->remaining / 2;
2043                        list = victim->list + victim->list_size - sub_size;
2044                        while (sub_size && list[0]->hash &&
2045                               list[0]->hash == list[-1]->hash) {
2046                                list++;
2047                                sub_size--;
2048                        }
2049                        if (!sub_size) {
2050                                /*
2051                                 * It is possible for some "paths" to have
2052                                 * so many objects that no hash boundary
2053                                 * might be found.  Let's just steal the
2054                                 * exact half in that case.
2055                                 */
2056                                sub_size = victim->remaining / 2;
2057                                list -= sub_size;
2058                        }
2059                        target->list = list;
2060                        victim->list_size -= sub_size;
2061                        victim->remaining -= sub_size;
2062                }
2063                target->list_size = sub_size;
2064                target->remaining = sub_size;
2065                target->working = 1;
2066                progress_unlock();
2067
2068                pthread_mutex_lock(&target->mutex);
2069                target->data_ready = 1;
2070                pthread_cond_signal(&target->cond);
2071                pthread_mutex_unlock(&target->mutex);
2072
2073                if (!sub_size) {
2074                        pthread_join(target->thread, NULL);
2075                        pthread_cond_destroy(&target->cond);
2076                        pthread_mutex_destroy(&target->mutex);
2077                        active_threads--;
2078                }
2079        }
2080        cleanup_threaded_search();
2081        free(p);
2082}
2083
2084#else
2085#define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2086#endif
2087
2088static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
2089{
2090        unsigned char peeled[20];
2091
2092        if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
2093            !peel_ref(path, peeled)        && /* peelable? */
2094            packlist_find(&to_pack, peeled, NULL))      /* object packed? */
2095                add_object_entry(sha1, OBJ_TAG, NULL, 0);
2096        return 0;
2097}
2098
2099static void prepare_pack(int window, int depth)
2100{
2101        struct object_entry **delta_list;
2102        uint32_t i, nr_deltas;
2103        unsigned n;
2104
2105        get_object_details();
2106
2107        /*
2108         * If we're locally repacking then we need to be doubly careful
2109         * from now on in order to make sure no stealth corruption gets
2110         * propagated to the new pack.  Clients receiving streamed packs
2111         * should validate everything they get anyway so no need to incur
2112         * the additional cost here in that case.
2113         */
2114        if (!pack_to_stdout)
2115                do_check_packed_object_crc = 1;
2116
2117        if (!to_pack.nr_objects || !window || !depth)
2118                return;
2119
2120        delta_list = xmalloc(to_pack.nr_objects * sizeof(*delta_list));
2121        nr_deltas = n = 0;
2122
2123        for (i = 0; i < to_pack.nr_objects; i++) {
2124                struct object_entry *entry = to_pack.objects + i;
2125
2126                if (entry->delta)
2127                        /* This happens if we decided to reuse existing
2128                         * delta from a pack.  "reuse_delta &&" is implied.
2129                         */
2130                        continue;
2131
2132                if (entry->size < 50)
2133                        continue;
2134
2135                if (entry->no_try_delta)
2136                        continue;
2137
2138                if (!entry->preferred_base) {
2139                        nr_deltas++;
2140                        if (entry->type < 0)
2141                                die("unable to get type of object %s",
2142                                    sha1_to_hex(entry->idx.sha1));
2143                } else {
2144                        if (entry->type < 0) {
2145                                /*
2146                                 * This object is not found, but we
2147                                 * don't have to include it anyway.
2148                                 */
2149                                continue;
2150                        }
2151                }
2152
2153                delta_list[n++] = entry;
2154        }
2155
2156        if (nr_deltas && n > 1) {
2157                unsigned nr_done = 0;
2158                if (progress)
2159                        progress_state = start_progress("Compressing objects",
2160                                                        nr_deltas);
2161                qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2162                ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2163                stop_progress(&progress_state);
2164                if (nr_done != nr_deltas)
2165                        die("inconsistency with delta count");
2166        }
2167        free(delta_list);
2168}
2169
2170static int git_pack_config(const char *k, const char *v, void *cb)
2171{
2172        if (!strcmp(k, "pack.window")) {
2173                window = git_config_int(k, v);
2174                return 0;
2175        }
2176        if (!strcmp(k, "pack.windowmemory")) {
2177                window_memory_limit = git_config_ulong(k, v);
2178                return 0;
2179        }
2180        if (!strcmp(k, "pack.depth")) {
2181                depth = git_config_int(k, v);
2182                return 0;
2183        }
2184        if (!strcmp(k, "pack.compression")) {
2185                int level = git_config_int(k, v);
2186                if (level == -1)
2187                        level = Z_DEFAULT_COMPRESSION;
2188                else if (level < 0 || level > Z_BEST_COMPRESSION)
2189                        die("bad pack compression level %d", level);
2190                pack_compression_level = level;
2191                pack_compression_seen = 1;
2192                return 0;
2193        }
2194        if (!strcmp(k, "pack.deltacachesize")) {
2195                max_delta_cache_size = git_config_int(k, v);
2196                return 0;
2197        }
2198        if (!strcmp(k, "pack.deltacachelimit")) {
2199                cache_max_small_delta_size = git_config_int(k, v);
2200                return 0;
2201        }
2202        if (!strcmp(k, "pack.writebitmaps")) {
2203                write_bitmap_index = git_config_bool(k, v);
2204                return 0;
2205        }
2206        if (!strcmp(k, "pack.writebitmaphashcache")) {
2207                if (git_config_bool(k, v))
2208                        write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2209                else
2210                        write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2211        }
2212        if (!strcmp(k, "pack.usebitmaps")) {
2213                use_bitmap_index = git_config_bool(k, v);
2214                return 0;
2215        }
2216        if (!strcmp(k, "pack.threads")) {
2217                delta_search_threads = git_config_int(k, v);
2218                if (delta_search_threads < 0)
2219                        die("invalid number of threads specified (%d)",
2220                            delta_search_threads);
2221#ifdef NO_PTHREADS
2222                if (delta_search_threads != 1)
2223                        warning("no threads support, ignoring %s", k);
2224#endif
2225                return 0;
2226        }
2227        if (!strcmp(k, "pack.indexversion")) {
2228                pack_idx_opts.version = git_config_int(k, v);
2229                if (pack_idx_opts.version > 2)
2230                        die("bad pack.indexversion=%"PRIu32,
2231                            pack_idx_opts.version);
2232                return 0;
2233        }
2234        return git_default_config(k, v, cb);
2235}
2236
2237static void read_object_list_from_stdin(void)
2238{
2239        char line[40 + 1 + PATH_MAX + 2];
2240        unsigned char sha1[20];
2241
2242        for (;;) {
2243                if (!fgets(line, sizeof(line), stdin)) {
2244                        if (feof(stdin))
2245                                break;
2246                        if (!ferror(stdin))
2247                                die("fgets returned NULL, not EOF, not error!");
2248                        if (errno != EINTR)
2249                                die_errno("fgets");
2250                        clearerr(stdin);
2251                        continue;
2252                }
2253                if (line[0] == '-') {
2254                        if (get_sha1_hex(line+1, sha1))
2255                                die("expected edge sha1, got garbage:\n %s",
2256                                    line);
2257                        add_preferred_base(sha1);
2258                        continue;
2259                }
2260                if (get_sha1_hex(line, sha1))
2261                        die("expected sha1, got garbage:\n %s", line);
2262
2263                add_preferred_base_object(line+41);
2264                add_object_entry(sha1, 0, line+41, 0);
2265        }
2266}
2267
2268#define OBJECT_ADDED (1u<<20)
2269
2270static void show_commit(struct commit *commit, void *data)
2271{
2272        add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
2273        commit->object.flags |= OBJECT_ADDED;
2274
2275        if (write_bitmap_index)
2276                index_commit_for_bitmap(commit);
2277}
2278
2279static void show_object(struct object *obj,
2280                        const struct name_path *path, const char *last,
2281                        void *data)
2282{
2283        char *name = path_name(path, last);
2284
2285        add_preferred_base_object(name);
2286        add_object_entry(obj->sha1, obj->type, name, 0);
2287        obj->flags |= OBJECT_ADDED;
2288
2289        /*
2290         * We will have generated the hash from the name,
2291         * but not saved a pointer to it - we can free it
2292         */
2293        free((char *)name);
2294}
2295
2296static void show_edge(struct commit *commit)
2297{
2298        add_preferred_base(commit->object.sha1);
2299}
2300
2301struct in_pack_object {
2302        off_t offset;
2303        struct object *object;
2304};
2305
2306struct in_pack {
2307        int alloc;
2308        int nr;
2309        struct in_pack_object *array;
2310};
2311
2312static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2313{
2314        in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
2315        in_pack->array[in_pack->nr].object = object;
2316        in_pack->nr++;
2317}
2318
2319/*
2320 * Compare the objects in the offset order, in order to emulate the
2321 * "git rev-list --objects" output that produced the pack originally.
2322 */
2323static int ofscmp(const void *a_, const void *b_)
2324{
2325        struct in_pack_object *a = (struct in_pack_object *)a_;
2326        struct in_pack_object *b = (struct in_pack_object *)b_;
2327
2328        if (a->offset < b->offset)
2329                return -1;
2330        else if (a->offset > b->offset)
2331                return 1;
2332        else
2333                return hashcmp(a->object->sha1, b->object->sha1);
2334}
2335
2336static void add_objects_in_unpacked_packs(struct rev_info *revs)
2337{
2338        struct packed_git *p;
2339        struct in_pack in_pack;
2340        uint32_t i;
2341
2342        memset(&in_pack, 0, sizeof(in_pack));
2343
2344        for (p = packed_git; p; p = p->next) {
2345                const unsigned char *sha1;
2346                struct object *o;
2347
2348                if (!p->pack_local || p->pack_keep)
2349                        continue;
2350                if (open_pack_index(p))
2351                        die("cannot open pack index");
2352
2353                ALLOC_GROW(in_pack.array,
2354                           in_pack.nr + p->num_objects,
2355                           in_pack.alloc);
2356
2357                for (i = 0; i < p->num_objects; i++) {
2358                        sha1 = nth_packed_object_sha1(p, i);
2359                        o = lookup_unknown_object(sha1);
2360                        if (!(o->flags & OBJECT_ADDED))
2361                                mark_in_pack_object(o, p, &in_pack);
2362                        o->flags |= OBJECT_ADDED;
2363                }
2364        }
2365
2366        if (in_pack.nr) {
2367                qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2368                      ofscmp);
2369                for (i = 0; i < in_pack.nr; i++) {
2370                        struct object *o = in_pack.array[i].object;
2371                        add_object_entry(o->sha1, o->type, "", 0);
2372                }
2373        }
2374        free(in_pack.array);
2375}
2376
2377static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2378{
2379        static struct packed_git *last_found = (void *)1;
2380        struct packed_git *p;
2381
2382        p = (last_found != (void *)1) ? last_found : packed_git;
2383
2384        while (p) {
2385                if ((!p->pack_local || p->pack_keep) &&
2386                        find_pack_entry_one(sha1, p)) {
2387                        last_found = p;
2388                        return 1;
2389                }
2390                if (p == last_found)
2391                        p = packed_git;
2392                else
2393                        p = p->next;
2394                if (p == last_found)
2395                        p = p->next;
2396        }
2397        return 0;
2398}
2399
2400static void loosen_unused_packed_objects(struct rev_info *revs)
2401{
2402        struct packed_git *p;
2403        uint32_t i;
2404        const unsigned char *sha1;
2405
2406        for (p = packed_git; p; p = p->next) {
2407                if (!p->pack_local || p->pack_keep)
2408                        continue;
2409
2410                if (unpack_unreachable_expiration &&
2411                    p->mtime < unpack_unreachable_expiration)
2412                        continue;
2413
2414                if (open_pack_index(p))
2415                        die("cannot open pack index");
2416
2417                for (i = 0; i < p->num_objects; i++) {
2418                        sha1 = nth_packed_object_sha1(p, i);
2419                        if (!packlist_find(&to_pack, sha1, NULL) &&
2420                                !has_sha1_pack_kept_or_nonlocal(sha1))
2421                                if (force_object_loose(sha1, p->mtime))
2422                                        die("unable to force loose object");
2423                }
2424        }
2425}
2426
2427static int get_object_list_from_bitmap(struct rev_info *revs)
2428{
2429        if (prepare_bitmap_walk(revs) < 0)
2430                return -1;
2431
2432        if (!reuse_partial_packfile_from_bitmap(
2433                        &reuse_packfile,
2434                        &reuse_packfile_objects,
2435                        &reuse_packfile_offset)) {
2436                assert(reuse_packfile_objects);
2437                nr_result += reuse_packfile_objects;
2438
2439                if (progress) {
2440                        fprintf(stderr, "Reusing existing pack: %d, done.\n",
2441                                reuse_packfile_objects);
2442                        fflush(stderr);
2443                }
2444        }
2445
2446        traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2447        return 0;
2448}
2449
2450static void get_object_list(int ac, const char **av)
2451{
2452        struct rev_info revs;
2453        char line[1000];
2454        int flags = 0;
2455
2456        init_revisions(&revs, NULL);
2457        save_commit_buffer = 0;
2458        setup_revisions(ac, av, &revs, NULL);
2459
2460        while (fgets(line, sizeof(line), stdin) != NULL) {
2461                int len = strlen(line);
2462                if (len && line[len - 1] == '\n')
2463                        line[--len] = 0;
2464                if (!len)
2465                        break;
2466                if (*line == '-') {
2467                        if (!strcmp(line, "--not")) {
2468                                flags ^= UNINTERESTING;
2469                                write_bitmap_index = 0;
2470                                continue;
2471                        }
2472                        die("not a rev '%s'", line);
2473                }
2474                if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2475                        die("bad revision '%s'", line);
2476        }
2477
2478        if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2479                return;
2480
2481        if (prepare_revision_walk(&revs))
2482                die("revision walk setup failed");
2483        mark_edges_uninteresting(&revs, show_edge);
2484        traverse_commit_list(&revs, show_commit, show_object, NULL);
2485
2486        if (keep_unreachable)
2487                add_objects_in_unpacked_packs(&revs);
2488        if (unpack_unreachable)
2489                loosen_unused_packed_objects(&revs);
2490}
2491
2492static int option_parse_index_version(const struct option *opt,
2493                                      const char *arg, int unset)
2494{
2495        char *c;
2496        const char *val = arg;
2497        pack_idx_opts.version = strtoul(val, &c, 10);
2498        if (pack_idx_opts.version > 2)
2499                die(_("unsupported index version %s"), val);
2500        if (*c == ',' && c[1])
2501                pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2502        if (*c || pack_idx_opts.off32_limit & 0x80000000)
2503                die(_("bad index version '%s'"), val);
2504        return 0;
2505}
2506
2507static int option_parse_unpack_unreachable(const struct option *opt,
2508                                           const char *arg, int unset)
2509{
2510        if (unset) {
2511                unpack_unreachable = 0;
2512                unpack_unreachable_expiration = 0;
2513        }
2514        else {
2515                unpack_unreachable = 1;
2516                if (arg)
2517                        unpack_unreachable_expiration = approxidate(arg);
2518        }
2519        return 0;
2520}
2521
2522static int option_parse_ulong(const struct option *opt,
2523                              const char *arg, int unset)
2524{
2525        if (unset)
2526                die(_("option %s does not accept negative form"),
2527                    opt->long_name);
2528
2529        if (!git_parse_ulong(arg, opt->value))
2530                die(_("unable to parse value '%s' for option %s"),
2531                    arg, opt->long_name);
2532        return 0;
2533}
2534
2535#define OPT_ULONG(s, l, v, h) \
2536        { OPTION_CALLBACK, (s), (l), (v), "n", (h),     \
2537          PARSE_OPT_NONEG, option_parse_ulong }
2538
2539int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2540{
2541        int use_internal_rev_list = 0;
2542        int thin = 0;
2543        int all_progress_implied = 0;
2544        const char *rp_av[6];
2545        int rp_ac = 0;
2546        int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2547        struct option pack_objects_options[] = {
2548                OPT_SET_INT('q', "quiet", &progress,
2549                            N_("do not show progress meter"), 0),
2550                OPT_SET_INT(0, "progress", &progress,
2551                            N_("show progress meter"), 1),
2552                OPT_SET_INT(0, "all-progress", &progress,
2553                            N_("show progress meter during object writing phase"), 2),
2554                OPT_BOOL(0, "all-progress-implied",
2555                         &all_progress_implied,
2556                         N_("similar to --all-progress when progress meter is shown")),
2557                { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2558                  N_("write the pack index file in the specified idx format version"),
2559                  0, option_parse_index_version },
2560                OPT_ULONG(0, "max-pack-size", &pack_size_limit,
2561                          N_("maximum size of each output pack file")),
2562                OPT_BOOL(0, "local", &local,
2563                         N_("ignore borrowed objects from alternate object store")),
2564                OPT_BOOL(0, "incremental", &incremental,
2565                         N_("ignore packed objects")),
2566                OPT_INTEGER(0, "window", &window,
2567                            N_("limit pack window by objects")),
2568                OPT_ULONG(0, "window-memory", &window_memory_limit,
2569                          N_("limit pack window by memory in addition to object limit")),
2570                OPT_INTEGER(0, "depth", &depth,
2571                            N_("maximum length of delta chain allowed in the resulting pack")),
2572                OPT_BOOL(0, "reuse-delta", &reuse_delta,
2573                         N_("reuse existing deltas")),
2574                OPT_BOOL(0, "reuse-object", &reuse_object,
2575                         N_("reuse existing objects")),
2576                OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2577                         N_("use OFS_DELTA objects")),
2578                OPT_INTEGER(0, "threads", &delta_search_threads,
2579                            N_("use threads when searching for best delta matches")),
2580                OPT_BOOL(0, "non-empty", &non_empty,
2581                         N_("do not create an empty pack output")),
2582                OPT_BOOL(0, "revs", &use_internal_rev_list,
2583                         N_("read revision arguments from standard input")),
2584                { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2585                  N_("limit the objects to those that are not yet packed"),
2586                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2587                { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2588                  N_("include objects reachable from any reference"),
2589                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2590                { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2591                  N_("include objects referred by reflog entries"),
2592                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2593                OPT_BOOL(0, "stdout", &pack_to_stdout,
2594                         N_("output pack to stdout")),
2595                OPT_BOOL(0, "include-tag", &include_tag,
2596                         N_("include tag objects that refer to objects to be packed")),
2597                OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2598                         N_("keep unreachable objects")),
2599                { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2600                  N_("unpack unreachable objects newer than <time>"),
2601                  PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2602                OPT_BOOL(0, "thin", &thin,
2603                         N_("create thin packs")),
2604                OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2605                         N_("ignore packs that have companion .keep file")),
2606                OPT_INTEGER(0, "compression", &pack_compression_level,
2607                            N_("pack compression level")),
2608                OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2609                            N_("do not hide commits by grafts"), 0),
2610                OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2611                         N_("use a bitmap index if available to speed up counting objects")),
2612                OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2613                         N_("write a bitmap index together with the pack index")),
2614                OPT_END(),
2615        };
2616
2617        read_replace_refs = 0;
2618
2619        reset_pack_idx_option(&pack_idx_opts);
2620        git_config(git_pack_config, NULL);
2621        if (!pack_compression_seen && core_compression_seen)
2622                pack_compression_level = core_compression_level;
2623
2624        progress = isatty(2);
2625        argc = parse_options(argc, argv, prefix, pack_objects_options,
2626                             pack_usage, 0);
2627
2628        if (argc) {
2629                base_name = argv[0];
2630                argc--;
2631        }
2632        if (pack_to_stdout != !base_name || argc)
2633                usage_with_options(pack_usage, pack_objects_options);
2634
2635        rp_av[rp_ac++] = "pack-objects";
2636        if (thin) {
2637                use_internal_rev_list = 1;
2638                rp_av[rp_ac++] = "--objects-edge";
2639        } else
2640                rp_av[rp_ac++] = "--objects";
2641
2642        if (rev_list_all) {
2643                use_internal_rev_list = 1;
2644                rp_av[rp_ac++] = "--all";
2645        }
2646        if (rev_list_reflog) {
2647                use_internal_rev_list = 1;
2648                rp_av[rp_ac++] = "--reflog";
2649        }
2650        if (rev_list_unpacked) {
2651                use_internal_rev_list = 1;
2652                rp_av[rp_ac++] = "--unpacked";
2653        }
2654
2655        if (!reuse_object)
2656                reuse_delta = 0;
2657        if (pack_compression_level == -1)
2658                pack_compression_level = Z_DEFAULT_COMPRESSION;
2659        else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2660                die("bad pack compression level %d", pack_compression_level);
2661#ifdef NO_PTHREADS
2662        if (delta_search_threads != 1)
2663                warning("no threads support, ignoring --threads");
2664#endif
2665        if (!pack_to_stdout && !pack_size_limit)
2666                pack_size_limit = pack_size_limit_cfg;
2667        if (pack_to_stdout && pack_size_limit)
2668                die("--max-pack-size cannot be used to build a pack for transfer.");
2669        if (pack_size_limit && pack_size_limit < 1024*1024) {
2670                warning("minimum pack size limit is 1 MiB");
2671                pack_size_limit = 1024*1024;
2672        }
2673
2674        if (!pack_to_stdout && thin)
2675                die("--thin cannot be used to build an indexable pack.");
2676
2677        if (keep_unreachable && unpack_unreachable)
2678                die("--keep-unreachable and --unpack-unreachable are incompatible.");
2679
2680        if (!use_internal_rev_list || !pack_to_stdout || is_repository_shallow())
2681                use_bitmap_index = 0;
2682
2683        if (pack_to_stdout || !rev_list_all)
2684                write_bitmap_index = 0;
2685
2686        if (progress && all_progress_implied)
2687                progress = 2;
2688
2689        prepare_packed_git();
2690
2691        if (progress)
2692                progress_state = start_progress("Counting objects", 0);
2693        if (!use_internal_rev_list)
2694                read_object_list_from_stdin();
2695        else {
2696                rp_av[rp_ac] = NULL;
2697                get_object_list(rp_ac, rp_av);
2698        }
2699        cleanup_preferred_base();
2700        if (include_tag && nr_result)
2701                for_each_ref(add_ref_tag, NULL);
2702        stop_progress(&progress_state);
2703
2704        if (non_empty && !nr_result)
2705                return 0;
2706        if (nr_result)
2707                prepare_pack(window, depth);
2708        write_pack_file();
2709        if (progress)
2710                fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2711                        " reused %"PRIu32" (delta %"PRIu32")\n",
2712                        written, written_delta, reused, reused_delta);
2713        return 0;
2714}