builtin / pack-objects.con commit regex: use regexec_buf() (b7d36ff)
   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#include "reachable.h"
  24#include "sha1-array.h"
  25#include "argv-array.h"
  26
  27static const char *pack_usage[] = {
  28        N_("git pack-objects --stdout [<options>...] [< <ref-list> | < <object-list>]"),
  29        N_("git pack-objects [<options>...] <base-name> [< <ref-list> | < <object-list>]"),
  30        NULL
  31};
  32
  33/*
  34 * Objects we are going to pack are collected in the `to_pack` structure.
  35 * It contains an array (dynamically expanded) of the object data, and a map
  36 * that can resolve SHA1s to their position in the array.
  37 */
  38static struct packing_data to_pack;
  39
  40static struct pack_idx_entry **written_list;
  41static uint32_t nr_result, nr_written;
  42
  43static int non_empty;
  44static int reuse_delta = 1, reuse_object = 1;
  45static int keep_unreachable, unpack_unreachable, include_tag;
  46static unsigned long unpack_unreachable_expiration;
  47static int local;
  48static int incremental;
  49static int ignore_packed_keep;
  50static int allow_ofs_delta;
  51static struct pack_idx_option pack_idx_opts;
  52static const char *base_name;
  53static int progress = 1;
  54static int window = 10;
  55static unsigned long pack_size_limit;
  56static int depth = 50;
  57static int delta_search_threads;
  58static int pack_to_stdout;
  59static int num_preferred_base;
  60static struct progress *progress_state;
  61static int pack_compression_level = Z_DEFAULT_COMPRESSION;
  62static int pack_compression_seen;
  63
  64static struct packed_git *reuse_packfile;
  65static uint32_t reuse_packfile_objects;
  66static off_t reuse_packfile_offset;
  67
  68static int use_bitmap_index = 1;
  69static int write_bitmap_index;
  70static uint16_t write_bitmap_options;
  71
  72static unsigned long delta_cache_size = 0;
  73static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
  74static unsigned long cache_max_small_delta_size = 1000;
  75
  76static unsigned long window_memory_limit = 0;
  77
  78/*
  79 * stats
  80 */
  81static uint32_t written, written_delta;
  82static uint32_t reused, reused_delta;
  83
  84/*
  85 * Indexed commits
  86 */
  87static struct commit **indexed_commits;
  88static unsigned int indexed_commits_nr;
  89static unsigned int indexed_commits_alloc;
  90
  91static void index_commit_for_bitmap(struct commit *commit)
  92{
  93        if (indexed_commits_nr >= indexed_commits_alloc) {
  94                indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
  95                REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
  96        }
  97
  98        indexed_commits[indexed_commits_nr++] = commit;
  99}
 100
 101static void *get_delta(struct object_entry *entry)
 102{
 103        unsigned long size, base_size, delta_size;
 104        void *buf, *base_buf, *delta_buf;
 105        enum object_type type;
 106
 107        buf = read_sha1_file(entry->idx.sha1, &type, &size);
 108        if (!buf)
 109                die("unable to read %s", sha1_to_hex(entry->idx.sha1));
 110        base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
 111        if (!base_buf)
 112                die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
 113        delta_buf = diff_delta(base_buf, base_size,
 114                               buf, size, &delta_size, 0);
 115        if (!delta_buf || delta_size != entry->delta_size)
 116                die("delta size changed");
 117        free(buf);
 118        free(base_buf);
 119        return delta_buf;
 120}
 121
 122static unsigned long do_compress(void **pptr, unsigned long size)
 123{
 124        git_zstream stream;
 125        void *in, *out;
 126        unsigned long maxsize;
 127
 128        git_deflate_init(&stream, pack_compression_level);
 129        maxsize = git_deflate_bound(&stream, size);
 130
 131        in = *pptr;
 132        out = xmalloc(maxsize);
 133        *pptr = out;
 134
 135        stream.next_in = in;
 136        stream.avail_in = size;
 137        stream.next_out = out;
 138        stream.avail_out = maxsize;
 139        while (git_deflate(&stream, Z_FINISH) == Z_OK)
 140                ; /* nothing */
 141        git_deflate_end(&stream);
 142
 143        free(in);
 144        return stream.total_out;
 145}
 146
 147static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
 148                                           const unsigned char *sha1)
 149{
 150        git_zstream stream;
 151        unsigned char ibuf[1024 * 16];
 152        unsigned char obuf[1024 * 16];
 153        unsigned long olen = 0;
 154
 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 struct object_id *oid, int flag,
 544                       void *cb_data)
 545{
 546        unsigned char peeled[20];
 547        struct object_entry *entry = packlist_find(&to_pack, oid->hash, 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;
 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        ALLOC_ARRAY(wo, to_pack.nr_objects);
 661        for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
 662                if (objects[i].tagged)
 663                        break;
 664                add_to_write_order(wo, &wo_end, &objects[i]);
 665        }
 666        last_untagged = i;
 667
 668        /*
 669         * Then fill all the tagged tips.
 670         */
 671        for (; i < to_pack.nr_objects; i++) {
 672                if (objects[i].tagged)
 673                        add_to_write_order(wo, &wo_end, &objects[i]);
 674        }
 675
 676        /*
 677         * And then all remaining commits and tags.
 678         */
 679        for (i = last_untagged; i < to_pack.nr_objects; i++) {
 680                if (objects[i].type != OBJ_COMMIT &&
 681                    objects[i].type != OBJ_TAG)
 682                        continue;
 683                add_to_write_order(wo, &wo_end, &objects[i]);
 684        }
 685
 686        /*
 687         * And then all the trees.
 688         */
 689        for (i = last_untagged; i < to_pack.nr_objects; i++) {
 690                if (objects[i].type != OBJ_TREE)
 691                        continue;
 692                add_to_write_order(wo, &wo_end, &objects[i]);
 693        }
 694
 695        /*
 696         * Finally all the rest in really tight order
 697         */
 698        for (i = last_untagged; i < to_pack.nr_objects; i++) {
 699                if (!objects[i].filled)
 700                        add_family_to_write_order(wo, &wo_end, &objects[i]);
 701        }
 702
 703        if (wo_end != to_pack.nr_objects)
 704                die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
 705
 706        return wo;
 707}
 708
 709static off_t write_reused_pack(struct sha1file *f)
 710{
 711        unsigned char buffer[8192];
 712        off_t to_write, total;
 713        int fd;
 714
 715        if (!is_pack_valid(reuse_packfile))
 716                die("packfile is invalid: %s", reuse_packfile->pack_name);
 717
 718        fd = git_open_noatime(reuse_packfile->pack_name);
 719        if (fd < 0)
 720                die_errno("unable to open packfile for reuse: %s",
 721                          reuse_packfile->pack_name);
 722
 723        if (lseek(fd, sizeof(struct pack_header), SEEK_SET) == -1)
 724                die_errno("unable to seek in reused packfile");
 725
 726        if (reuse_packfile_offset < 0)
 727                reuse_packfile_offset = reuse_packfile->pack_size - 20;
 728
 729        total = to_write = reuse_packfile_offset - sizeof(struct pack_header);
 730
 731        while (to_write) {
 732                int read_pack = xread(fd, buffer, sizeof(buffer));
 733
 734                if (read_pack <= 0)
 735                        die_errno("unable to read from reused packfile");
 736
 737                if (read_pack > to_write)
 738                        read_pack = to_write;
 739
 740                sha1write(f, buffer, read_pack);
 741                to_write -= read_pack;
 742
 743                /*
 744                 * We don't know the actual number of objects written,
 745                 * only how many bytes written, how many bytes total, and
 746                 * how many objects total. So we can fake it by pretending all
 747                 * objects we are writing are the same size. This gives us a
 748                 * smooth progress meter, and at the end it matches the true
 749                 * answer.
 750                 */
 751                written = reuse_packfile_objects *
 752                                (((double)(total - to_write)) / total);
 753                display_progress(progress_state, written);
 754        }
 755
 756        close(fd);
 757        written = reuse_packfile_objects;
 758        display_progress(progress_state, written);
 759        return reuse_packfile_offset - sizeof(struct pack_header);
 760}
 761
 762static const char no_split_warning[] = N_(
 763"disabling bitmap writing, packs are split due to pack.packSizeLimit"
 764);
 765
 766static void write_pack_file(void)
 767{
 768        uint32_t i = 0, j;
 769        struct sha1file *f;
 770        off_t offset;
 771        uint32_t nr_remaining = nr_result;
 772        time_t last_mtime = 0;
 773        struct object_entry **write_order;
 774
 775        if (progress > pack_to_stdout)
 776                progress_state = start_progress(_("Writing objects"), nr_result);
 777        ALLOC_ARRAY(written_list, to_pack.nr_objects);
 778        write_order = compute_write_order();
 779
 780        do {
 781                unsigned char sha1[20];
 782                char *pack_tmp_name = NULL;
 783
 784                if (pack_to_stdout)
 785                        f = sha1fd_throughput(1, "<stdout>", progress_state);
 786                else
 787                        f = create_tmp_packfile(&pack_tmp_name);
 788
 789                offset = write_pack_header(f, nr_remaining);
 790
 791                if (reuse_packfile) {
 792                        off_t packfile_size;
 793                        assert(pack_to_stdout);
 794
 795                        packfile_size = write_reused_pack(f);
 796                        offset += packfile_size;
 797                }
 798
 799                nr_written = 0;
 800                for (; i < to_pack.nr_objects; i++) {
 801                        struct object_entry *e = write_order[i];
 802                        if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 803                                break;
 804                        display_progress(progress_state, written);
 805                }
 806
 807                /*
 808                 * Did we write the wrong # entries in the header?
 809                 * If so, rewrite it like in fast-import
 810                 */
 811                if (pack_to_stdout) {
 812                        sha1close(f, sha1, CSUM_CLOSE);
 813                } else if (nr_written == nr_remaining) {
 814                        sha1close(f, sha1, CSUM_FSYNC);
 815                } else {
 816                        int fd = sha1close(f, sha1, 0);
 817                        fixup_pack_header_footer(fd, sha1, pack_tmp_name,
 818                                                 nr_written, sha1, offset);
 819                        close(fd);
 820                        if (write_bitmap_index) {
 821                                warning(_(no_split_warning));
 822                                write_bitmap_index = 0;
 823                        }
 824                }
 825
 826                if (!pack_to_stdout) {
 827                        struct stat st;
 828                        struct strbuf tmpname = STRBUF_INIT;
 829
 830                        /*
 831                         * Packs are runtime accessed in their mtime
 832                         * order since newer packs are more likely to contain
 833                         * younger objects.  So if we are creating multiple
 834                         * packs then we should modify the mtime of later ones
 835                         * to preserve this property.
 836                         */
 837                        if (stat(pack_tmp_name, &st) < 0) {
 838                                warning("failed to stat %s: %s",
 839                                        pack_tmp_name, strerror(errno));
 840                        } else if (!last_mtime) {
 841                                last_mtime = st.st_mtime;
 842                        } else {
 843                                struct utimbuf utb;
 844                                utb.actime = st.st_atime;
 845                                utb.modtime = --last_mtime;
 846                                if (utime(pack_tmp_name, &utb) < 0)
 847                                        warning("failed utime() on %s: %s",
 848                                                pack_tmp_name, strerror(errno));
 849                        }
 850
 851                        strbuf_addf(&tmpname, "%s-", base_name);
 852
 853                        if (write_bitmap_index) {
 854                                bitmap_writer_set_checksum(sha1);
 855                                bitmap_writer_build_type_index(written_list, nr_written);
 856                        }
 857
 858                        finish_tmp_packfile(&tmpname, pack_tmp_name,
 859                                            written_list, nr_written,
 860                                            &pack_idx_opts, sha1);
 861
 862                        if (write_bitmap_index) {
 863                                strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
 864
 865                                stop_progress(&progress_state);
 866
 867                                bitmap_writer_show_progress(progress);
 868                                bitmap_writer_reuse_bitmaps(&to_pack);
 869                                bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
 870                                bitmap_writer_build(&to_pack);
 871                                bitmap_writer_finish(written_list, nr_written,
 872                                                     tmpname.buf, write_bitmap_options);
 873                                write_bitmap_index = 0;
 874                        }
 875
 876                        strbuf_release(&tmpname);
 877                        free(pack_tmp_name);
 878                        puts(sha1_to_hex(sha1));
 879                }
 880
 881                /* mark written objects as written to previous pack */
 882                for (j = 0; j < nr_written; j++) {
 883                        written_list[j]->offset = (off_t)-1;
 884                }
 885                nr_remaining -= nr_written;
 886        } while (nr_remaining && i < to_pack.nr_objects);
 887
 888        free(written_list);
 889        free(write_order);
 890        stop_progress(&progress_state);
 891        if (written != nr_result)
 892                die("wrote %"PRIu32" objects while expecting %"PRIu32,
 893                        written, nr_result);
 894}
 895
 896static void setup_delta_attr_check(struct git_attr_check *check)
 897{
 898        static struct git_attr *attr_delta;
 899
 900        if (!attr_delta)
 901                attr_delta = git_attr("delta");
 902
 903        check[0].attr = attr_delta;
 904}
 905
 906static int no_try_delta(const char *path)
 907{
 908        struct git_attr_check check[1];
 909
 910        setup_delta_attr_check(check);
 911        if (git_check_attr(path, ARRAY_SIZE(check), check))
 912                return 0;
 913        if (ATTR_FALSE(check->value))
 914                return 1;
 915        return 0;
 916}
 917
 918/*
 919 * When adding an object, check whether we have already added it
 920 * to our packing list. If so, we can skip. However, if we are
 921 * being asked to excludei t, but the previous mention was to include
 922 * it, make sure to adjust its flags and tweak our numbers accordingly.
 923 *
 924 * As an optimization, we pass out the index position where we would have
 925 * found the item, since that saves us from having to look it up again a
 926 * few lines later when we want to add the new entry.
 927 */
 928static int have_duplicate_entry(const unsigned char *sha1,
 929                                int exclude,
 930                                uint32_t *index_pos)
 931{
 932        struct object_entry *entry;
 933
 934        entry = packlist_find(&to_pack, sha1, index_pos);
 935        if (!entry)
 936                return 0;
 937
 938        if (exclude) {
 939                if (!entry->preferred_base)
 940                        nr_result--;
 941                entry->preferred_base = 1;
 942        }
 943
 944        return 1;
 945}
 946
 947/*
 948 * Check whether we want the object in the pack (e.g., we do not want
 949 * objects found in non-local stores if the "--local" option was used).
 950 *
 951 * As a side effect of this check, we will find the packed version of this
 952 * object, if any. We therefore pass out the pack information to avoid having
 953 * to look it up again later.
 954 */
 955static int want_object_in_pack(const unsigned char *sha1,
 956                               int exclude,
 957                               struct packed_git **found_pack,
 958                               off_t *found_offset)
 959{
 960        struct packed_git *p;
 961
 962        if (!exclude && local && has_loose_object_nonlocal(sha1))
 963                return 0;
 964
 965        *found_pack = NULL;
 966        *found_offset = 0;
 967
 968        for (p = packed_git; p; p = p->next) {
 969                off_t offset = find_pack_entry_one(sha1, p);
 970                if (offset) {
 971                        if (!*found_pack) {
 972                                if (!is_pack_valid(p))
 973                                        continue;
 974                                *found_offset = offset;
 975                                *found_pack = p;
 976                        }
 977                        if (exclude)
 978                                return 1;
 979                        if (incremental)
 980                                return 0;
 981                        if (local && !p->pack_local)
 982                                return 0;
 983                        if (ignore_packed_keep && p->pack_local && p->pack_keep)
 984                                return 0;
 985                }
 986        }
 987
 988        return 1;
 989}
 990
 991static void create_object_entry(const unsigned char *sha1,
 992                                enum object_type type,
 993                                uint32_t hash,
 994                                int exclude,
 995                                int no_try_delta,
 996                                uint32_t index_pos,
 997                                struct packed_git *found_pack,
 998                                off_t found_offset)
 999{
1000        struct object_entry *entry;
1001
1002        entry = packlist_alloc(&to_pack, sha1, index_pos);
1003        entry->hash = hash;
1004        if (type)
1005                entry->type = type;
1006        if (exclude)
1007                entry->preferred_base = 1;
1008        else
1009                nr_result++;
1010        if (found_pack) {
1011                entry->in_pack = found_pack;
1012                entry->in_pack_offset = found_offset;
1013        }
1014
1015        entry->no_try_delta = no_try_delta;
1016}
1017
1018static const char no_closure_warning[] = N_(
1019"disabling bitmap writing, as some objects are not being packed"
1020);
1021
1022static int add_object_entry(const unsigned char *sha1, enum object_type type,
1023                            const char *name, int exclude)
1024{
1025        struct packed_git *found_pack;
1026        off_t found_offset;
1027        uint32_t index_pos;
1028
1029        if (have_duplicate_entry(sha1, exclude, &index_pos))
1030                return 0;
1031
1032        if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1033                /* The pack is missing an object, so it will not have closure */
1034                if (write_bitmap_index) {
1035                        warning(_(no_closure_warning));
1036                        write_bitmap_index = 0;
1037                }
1038                return 0;
1039        }
1040
1041        create_object_entry(sha1, type, pack_name_hash(name),
1042                            exclude, name && no_try_delta(name),
1043                            index_pos, found_pack, found_offset);
1044
1045        display_progress(progress_state, nr_result);
1046        return 1;
1047}
1048
1049static int add_object_entry_from_bitmap(const unsigned char *sha1,
1050                                        enum object_type type,
1051                                        int flags, uint32_t name_hash,
1052                                        struct packed_git *pack, off_t offset)
1053{
1054        uint32_t index_pos;
1055
1056        if (have_duplicate_entry(sha1, 0, &index_pos))
1057                return 0;
1058
1059        create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1060
1061        display_progress(progress_state, nr_result);
1062        return 1;
1063}
1064
1065struct pbase_tree_cache {
1066        unsigned char sha1[20];
1067        int ref;
1068        int temporary;
1069        void *tree_data;
1070        unsigned long tree_size;
1071};
1072
1073static struct pbase_tree_cache *(pbase_tree_cache[256]);
1074static int pbase_tree_cache_ix(const unsigned char *sha1)
1075{
1076        return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1077}
1078static int pbase_tree_cache_ix_incr(int ix)
1079{
1080        return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1081}
1082
1083static struct pbase_tree {
1084        struct pbase_tree *next;
1085        /* This is a phony "cache" entry; we are not
1086         * going to evict it or find it through _get()
1087         * mechanism -- this is for the toplevel node that
1088         * would almost always change with any commit.
1089         */
1090        struct pbase_tree_cache pcache;
1091} *pbase_tree;
1092
1093static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1094{
1095        struct pbase_tree_cache *ent, *nent;
1096        void *data;
1097        unsigned long size;
1098        enum object_type type;
1099        int neigh;
1100        int my_ix = pbase_tree_cache_ix(sha1);
1101        int available_ix = -1;
1102
1103        /* pbase-tree-cache acts as a limited hashtable.
1104         * your object will be found at your index or within a few
1105         * slots after that slot if it is cached.
1106         */
1107        for (neigh = 0; neigh < 8; neigh++) {
1108                ent = pbase_tree_cache[my_ix];
1109                if (ent && !hashcmp(ent->sha1, sha1)) {
1110                        ent->ref++;
1111                        return ent;
1112                }
1113                else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1114                         ((0 <= available_ix) &&
1115                          (!ent && pbase_tree_cache[available_ix])))
1116                        available_ix = my_ix;
1117                if (!ent)
1118                        break;
1119                my_ix = pbase_tree_cache_ix_incr(my_ix);
1120        }
1121
1122        /* Did not find one.  Either we got a bogus request or
1123         * we need to read and perhaps cache.
1124         */
1125        data = read_sha1_file(sha1, &type, &size);
1126        if (!data)
1127                return NULL;
1128        if (type != OBJ_TREE) {
1129                free(data);
1130                return NULL;
1131        }
1132
1133        /* We need to either cache or return a throwaway copy */
1134
1135        if (available_ix < 0)
1136                ent = NULL;
1137        else {
1138                ent = pbase_tree_cache[available_ix];
1139                my_ix = available_ix;
1140        }
1141
1142        if (!ent) {
1143                nent = xmalloc(sizeof(*nent));
1144                nent->temporary = (available_ix < 0);
1145        }
1146        else {
1147                /* evict and reuse */
1148                free(ent->tree_data);
1149                nent = ent;
1150        }
1151        hashcpy(nent->sha1, sha1);
1152        nent->tree_data = data;
1153        nent->tree_size = size;
1154        nent->ref = 1;
1155        if (!nent->temporary)
1156                pbase_tree_cache[my_ix] = nent;
1157        return nent;
1158}
1159
1160static void pbase_tree_put(struct pbase_tree_cache *cache)
1161{
1162        if (!cache->temporary) {
1163                cache->ref--;
1164                return;
1165        }
1166        free(cache->tree_data);
1167        free(cache);
1168}
1169
1170static int name_cmp_len(const char *name)
1171{
1172        int i;
1173        for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1174                ;
1175        return i;
1176}
1177
1178static void add_pbase_object(struct tree_desc *tree,
1179                             const char *name,
1180                             int cmplen,
1181                             const char *fullname)
1182{
1183        struct name_entry entry;
1184        int cmp;
1185
1186        while (tree_entry(tree,&entry)) {
1187                if (S_ISGITLINK(entry.mode))
1188                        continue;
1189                cmp = tree_entry_len(&entry) != cmplen ? 1 :
1190                      memcmp(name, entry.path, cmplen);
1191                if (cmp > 0)
1192                        continue;
1193                if (cmp < 0)
1194                        return;
1195                if (name[cmplen] != '/') {
1196                        add_object_entry(entry.sha1,
1197                                         object_type(entry.mode),
1198                                         fullname, 1);
1199                        return;
1200                }
1201                if (S_ISDIR(entry.mode)) {
1202                        struct tree_desc sub;
1203                        struct pbase_tree_cache *tree;
1204                        const char *down = name+cmplen+1;
1205                        int downlen = name_cmp_len(down);
1206
1207                        tree = pbase_tree_get(entry.sha1);
1208                        if (!tree)
1209                                return;
1210                        init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1211
1212                        add_pbase_object(&sub, down, downlen, fullname);
1213                        pbase_tree_put(tree);
1214                }
1215        }
1216}
1217
1218static unsigned *done_pbase_paths;
1219static int done_pbase_paths_num;
1220static int done_pbase_paths_alloc;
1221static int done_pbase_path_pos(unsigned hash)
1222{
1223        int lo = 0;
1224        int hi = done_pbase_paths_num;
1225        while (lo < hi) {
1226                int mi = (hi + lo) / 2;
1227                if (done_pbase_paths[mi] == hash)
1228                        return mi;
1229                if (done_pbase_paths[mi] < hash)
1230                        hi = mi;
1231                else
1232                        lo = mi + 1;
1233        }
1234        return -lo-1;
1235}
1236
1237static int check_pbase_path(unsigned hash)
1238{
1239        int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1240        if (0 <= pos)
1241                return 1;
1242        pos = -pos - 1;
1243        ALLOC_GROW(done_pbase_paths,
1244                   done_pbase_paths_num + 1,
1245                   done_pbase_paths_alloc);
1246        done_pbase_paths_num++;
1247        if (pos < done_pbase_paths_num)
1248                memmove(done_pbase_paths + pos + 1,
1249                        done_pbase_paths + pos,
1250                        (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1251        done_pbase_paths[pos] = hash;
1252        return 0;
1253}
1254
1255static void add_preferred_base_object(const char *name)
1256{
1257        struct pbase_tree *it;
1258        int cmplen;
1259        unsigned hash = pack_name_hash(name);
1260
1261        if (!num_preferred_base || check_pbase_path(hash))
1262                return;
1263
1264        cmplen = name_cmp_len(name);
1265        for (it = pbase_tree; it; it = it->next) {
1266                if (cmplen == 0) {
1267                        add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1268                }
1269                else {
1270                        struct tree_desc tree;
1271                        init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1272                        add_pbase_object(&tree, name, cmplen, name);
1273                }
1274        }
1275}
1276
1277static void add_preferred_base(unsigned char *sha1)
1278{
1279        struct pbase_tree *it;
1280        void *data;
1281        unsigned long size;
1282        unsigned char tree_sha1[20];
1283
1284        if (window <= num_preferred_base++)
1285                return;
1286
1287        data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1288        if (!data)
1289                return;
1290
1291        for (it = pbase_tree; it; it = it->next) {
1292                if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1293                        free(data);
1294                        return;
1295                }
1296        }
1297
1298        it = xcalloc(1, sizeof(*it));
1299        it->next = pbase_tree;
1300        pbase_tree = it;
1301
1302        hashcpy(it->pcache.sha1, tree_sha1);
1303        it->pcache.tree_data = data;
1304        it->pcache.tree_size = size;
1305}
1306
1307static void cleanup_preferred_base(void)
1308{
1309        struct pbase_tree *it;
1310        unsigned i;
1311
1312        it = pbase_tree;
1313        pbase_tree = NULL;
1314        while (it) {
1315                struct pbase_tree *this = it;
1316                it = this->next;
1317                free(this->pcache.tree_data);
1318                free(this);
1319        }
1320
1321        for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1322                if (!pbase_tree_cache[i])
1323                        continue;
1324                free(pbase_tree_cache[i]->tree_data);
1325                free(pbase_tree_cache[i]);
1326                pbase_tree_cache[i] = NULL;
1327        }
1328
1329        free(done_pbase_paths);
1330        done_pbase_paths = NULL;
1331        done_pbase_paths_num = done_pbase_paths_alloc = 0;
1332}
1333
1334static void check_object(struct object_entry *entry)
1335{
1336        if (entry->in_pack) {
1337                struct packed_git *p = entry->in_pack;
1338                struct pack_window *w_curs = NULL;
1339                const unsigned char *base_ref = NULL;
1340                struct object_entry *base_entry;
1341                unsigned long used, used_0;
1342                unsigned long avail;
1343                off_t ofs;
1344                unsigned char *buf, c;
1345
1346                buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1347
1348                /*
1349                 * We want in_pack_type even if we do not reuse delta
1350                 * since non-delta representations could still be reused.
1351                 */
1352                used = unpack_object_header_buffer(buf, avail,
1353                                                   &entry->in_pack_type,
1354                                                   &entry->size);
1355                if (used == 0)
1356                        goto give_up;
1357
1358                /*
1359                 * Determine if this is a delta and if so whether we can
1360                 * reuse it or not.  Otherwise let's find out as cheaply as
1361                 * possible what the actual type and size for this object is.
1362                 */
1363                switch (entry->in_pack_type) {
1364                default:
1365                        /* Not a delta hence we've already got all we need. */
1366                        entry->type = entry->in_pack_type;
1367                        entry->in_pack_header_size = used;
1368                        if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1369                                goto give_up;
1370                        unuse_pack(&w_curs);
1371                        return;
1372                case OBJ_REF_DELTA:
1373                        if (reuse_delta && !entry->preferred_base)
1374                                base_ref = use_pack(p, &w_curs,
1375                                                entry->in_pack_offset + used, NULL);
1376                        entry->in_pack_header_size = used + 20;
1377                        break;
1378                case OBJ_OFS_DELTA:
1379                        buf = use_pack(p, &w_curs,
1380                                       entry->in_pack_offset + used, NULL);
1381                        used_0 = 0;
1382                        c = buf[used_0++];
1383                        ofs = c & 127;
1384                        while (c & 128) {
1385                                ofs += 1;
1386                                if (!ofs || MSB(ofs, 7)) {
1387                                        error("delta base offset overflow in pack for %s",
1388                                              sha1_to_hex(entry->idx.sha1));
1389                                        goto give_up;
1390                                }
1391                                c = buf[used_0++];
1392                                ofs = (ofs << 7) + (c & 127);
1393                        }
1394                        ofs = entry->in_pack_offset - ofs;
1395                        if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1396                                error("delta base offset out of bound for %s",
1397                                      sha1_to_hex(entry->idx.sha1));
1398                                goto give_up;
1399                        }
1400                        if (reuse_delta && !entry->preferred_base) {
1401                                struct revindex_entry *revidx;
1402                                revidx = find_pack_revindex(p, ofs);
1403                                if (!revidx)
1404                                        goto give_up;
1405                                base_ref = nth_packed_object_sha1(p, revidx->nr);
1406                        }
1407                        entry->in_pack_header_size = used + used_0;
1408                        break;
1409                }
1410
1411                if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1412                        /*
1413                         * If base_ref was set above that means we wish to
1414                         * reuse delta data, and we even found that base
1415                         * in the list of objects we want to pack. Goodie!
1416                         *
1417                         * Depth value does not matter - find_deltas() will
1418                         * never consider reused delta as the base object to
1419                         * deltify other objects against, in order to avoid
1420                         * circular deltas.
1421                         */
1422                        entry->type = entry->in_pack_type;
1423                        entry->delta = base_entry;
1424                        entry->delta_size = entry->size;
1425                        entry->delta_sibling = base_entry->delta_child;
1426                        base_entry->delta_child = entry;
1427                        unuse_pack(&w_curs);
1428                        return;
1429                }
1430
1431                if (entry->type) {
1432                        /*
1433                         * This must be a delta and we already know what the
1434                         * final object type is.  Let's extract the actual
1435                         * object size from the delta header.
1436                         */
1437                        entry->size = get_size_from_delta(p, &w_curs,
1438                                        entry->in_pack_offset + entry->in_pack_header_size);
1439                        if (entry->size == 0)
1440                                goto give_up;
1441                        unuse_pack(&w_curs);
1442                        return;
1443                }
1444
1445                /*
1446                 * No choice but to fall back to the recursive delta walk
1447                 * with sha1_object_info() to find about the object type
1448                 * at this point...
1449                 */
1450                give_up:
1451                unuse_pack(&w_curs);
1452        }
1453
1454        entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1455        /*
1456         * The error condition is checked in prepare_pack().  This is
1457         * to permit a missing preferred base object to be ignored
1458         * as a preferred base.  Doing so can result in a larger
1459         * pack file, but the transfer will still take place.
1460         */
1461}
1462
1463static int pack_offset_sort(const void *_a, const void *_b)
1464{
1465        const struct object_entry *a = *(struct object_entry **)_a;
1466        const struct object_entry *b = *(struct object_entry **)_b;
1467
1468        /* avoid filesystem trashing with loose objects */
1469        if (!a->in_pack && !b->in_pack)
1470                return hashcmp(a->idx.sha1, b->idx.sha1);
1471
1472        if (a->in_pack < b->in_pack)
1473                return -1;
1474        if (a->in_pack > b->in_pack)
1475                return 1;
1476        return a->in_pack_offset < b->in_pack_offset ? -1 :
1477                        (a->in_pack_offset > b->in_pack_offset);
1478}
1479
1480static void get_object_details(void)
1481{
1482        uint32_t i;
1483        struct object_entry **sorted_by_offset;
1484
1485        sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1486        for (i = 0; i < to_pack.nr_objects; i++)
1487                sorted_by_offset[i] = to_pack.objects + i;
1488        qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1489
1490        for (i = 0; i < to_pack.nr_objects; i++) {
1491                struct object_entry *entry = sorted_by_offset[i];
1492                check_object(entry);
1493                if (big_file_threshold < entry->size)
1494                        entry->no_try_delta = 1;
1495        }
1496
1497        free(sorted_by_offset);
1498}
1499
1500/*
1501 * We search for deltas in a list sorted by type, by filename hash, and then
1502 * by size, so that we see progressively smaller and smaller files.
1503 * That's because we prefer deltas to be from the bigger file
1504 * to the smaller -- deletes are potentially cheaper, but perhaps
1505 * more importantly, the bigger file is likely the more recent
1506 * one.  The deepest deltas are therefore the oldest objects which are
1507 * less susceptible to be accessed often.
1508 */
1509static int type_size_sort(const void *_a, const void *_b)
1510{
1511        const struct object_entry *a = *(struct object_entry **)_a;
1512        const struct object_entry *b = *(struct object_entry **)_b;
1513
1514        if (a->type > b->type)
1515                return -1;
1516        if (a->type < b->type)
1517                return 1;
1518        if (a->hash > b->hash)
1519                return -1;
1520        if (a->hash < b->hash)
1521                return 1;
1522        if (a->preferred_base > b->preferred_base)
1523                return -1;
1524        if (a->preferred_base < b->preferred_base)
1525                return 1;
1526        if (a->size > b->size)
1527                return -1;
1528        if (a->size < b->size)
1529                return 1;
1530        return a < b ? -1 : (a > b);  /* newest first */
1531}
1532
1533struct unpacked {
1534        struct object_entry *entry;
1535        void *data;
1536        struct delta_index *index;
1537        unsigned depth;
1538};
1539
1540static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1541                           unsigned long delta_size)
1542{
1543        if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1544                return 0;
1545
1546        if (delta_size < cache_max_small_delta_size)
1547                return 1;
1548
1549        /* cache delta, if objects are large enough compared to delta size */
1550        if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1551                return 1;
1552
1553        return 0;
1554}
1555
1556#ifndef NO_PTHREADS
1557
1558static pthread_mutex_t read_mutex;
1559#define read_lock()             pthread_mutex_lock(&read_mutex)
1560#define read_unlock()           pthread_mutex_unlock(&read_mutex)
1561
1562static pthread_mutex_t cache_mutex;
1563#define cache_lock()            pthread_mutex_lock(&cache_mutex)
1564#define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1565
1566static pthread_mutex_t progress_mutex;
1567#define progress_lock()         pthread_mutex_lock(&progress_mutex)
1568#define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1569
1570#else
1571
1572#define read_lock()             (void)0
1573#define read_unlock()           (void)0
1574#define cache_lock()            (void)0
1575#define cache_unlock()          (void)0
1576#define progress_lock()         (void)0
1577#define progress_unlock()       (void)0
1578
1579#endif
1580
1581static int try_delta(struct unpacked *trg, struct unpacked *src,
1582                     unsigned max_depth, unsigned long *mem_usage)
1583{
1584        struct object_entry *trg_entry = trg->entry;
1585        struct object_entry *src_entry = src->entry;
1586        unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1587        unsigned ref_depth;
1588        enum object_type type;
1589        void *delta_buf;
1590
1591        /* Don't bother doing diffs between different types */
1592        if (trg_entry->type != src_entry->type)
1593                return -1;
1594
1595        /*
1596         * We do not bother to try a delta that we discarded on an
1597         * earlier try, but only when reusing delta data.  Note that
1598         * src_entry that is marked as the preferred_base should always
1599         * be considered, as even if we produce a suboptimal delta against
1600         * it, we will still save the transfer cost, as we already know
1601         * the other side has it and we won't send src_entry at all.
1602         */
1603        if (reuse_delta && trg_entry->in_pack &&
1604            trg_entry->in_pack == src_entry->in_pack &&
1605            !src_entry->preferred_base &&
1606            trg_entry->in_pack_type != OBJ_REF_DELTA &&
1607            trg_entry->in_pack_type != OBJ_OFS_DELTA)
1608                return 0;
1609
1610        /* Let's not bust the allowed depth. */
1611        if (src->depth >= max_depth)
1612                return 0;
1613
1614        /* Now some size filtering heuristics. */
1615        trg_size = trg_entry->size;
1616        if (!trg_entry->delta) {
1617                max_size = trg_size/2 - 20;
1618                ref_depth = 1;
1619        } else {
1620                max_size = trg_entry->delta_size;
1621                ref_depth = trg->depth;
1622        }
1623        max_size = (uint64_t)max_size * (max_depth - src->depth) /
1624                                                (max_depth - ref_depth + 1);
1625        if (max_size == 0)
1626                return 0;
1627        src_size = src_entry->size;
1628        sizediff = src_size < trg_size ? trg_size - src_size : 0;
1629        if (sizediff >= max_size)
1630                return 0;
1631        if (trg_size < src_size / 32)
1632                return 0;
1633
1634        /* Load data if not already done */
1635        if (!trg->data) {
1636                read_lock();
1637                trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1638                read_unlock();
1639                if (!trg->data)
1640                        die("object %s cannot be read",
1641                            sha1_to_hex(trg_entry->idx.sha1));
1642                if (sz != trg_size)
1643                        die("object %s inconsistent object length (%lu vs %lu)",
1644                            sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1645                *mem_usage += sz;
1646        }
1647        if (!src->data) {
1648                read_lock();
1649                src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1650                read_unlock();
1651                if (!src->data) {
1652                        if (src_entry->preferred_base) {
1653                                static int warned = 0;
1654                                if (!warned++)
1655                                        warning("object %s cannot be read",
1656                                                sha1_to_hex(src_entry->idx.sha1));
1657                                /*
1658                                 * Those objects are not included in the
1659                                 * resulting pack.  Be resilient and ignore
1660                                 * them if they can't be read, in case the
1661                                 * pack could be created nevertheless.
1662                                 */
1663                                return 0;
1664                        }
1665                        die("object %s cannot be read",
1666                            sha1_to_hex(src_entry->idx.sha1));
1667                }
1668                if (sz != src_size)
1669                        die("object %s inconsistent object length (%lu vs %lu)",
1670                            sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1671                *mem_usage += sz;
1672        }
1673        if (!src->index) {
1674                src->index = create_delta_index(src->data, src_size);
1675                if (!src->index) {
1676                        static int warned = 0;
1677                        if (!warned++)
1678                                warning("suboptimal pack - out of memory");
1679                        return 0;
1680                }
1681                *mem_usage += sizeof_delta_index(src->index);
1682        }
1683
1684        delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1685        if (!delta_buf)
1686                return 0;
1687
1688        if (trg_entry->delta) {
1689                /* Prefer only shallower same-sized deltas. */
1690                if (delta_size == trg_entry->delta_size &&
1691                    src->depth + 1 >= trg->depth) {
1692                        free(delta_buf);
1693                        return 0;
1694                }
1695        }
1696
1697        /*
1698         * Handle memory allocation outside of the cache
1699         * accounting lock.  Compiler will optimize the strangeness
1700         * away when NO_PTHREADS is defined.
1701         */
1702        free(trg_entry->delta_data);
1703        cache_lock();
1704        if (trg_entry->delta_data) {
1705                delta_cache_size -= trg_entry->delta_size;
1706                trg_entry->delta_data = NULL;
1707        }
1708        if (delta_cacheable(src_size, trg_size, delta_size)) {
1709                delta_cache_size += delta_size;
1710                cache_unlock();
1711                trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1712        } else {
1713                cache_unlock();
1714                free(delta_buf);
1715        }
1716
1717        trg_entry->delta = src_entry;
1718        trg_entry->delta_size = delta_size;
1719        trg->depth = src->depth + 1;
1720
1721        return 1;
1722}
1723
1724static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1725{
1726        struct object_entry *child = me->delta_child;
1727        unsigned int m = n;
1728        while (child) {
1729                unsigned int c = check_delta_limit(child, n + 1);
1730                if (m < c)
1731                        m = c;
1732                child = child->delta_sibling;
1733        }
1734        return m;
1735}
1736
1737static unsigned long free_unpacked(struct unpacked *n)
1738{
1739        unsigned long freed_mem = sizeof_delta_index(n->index);
1740        free_delta_index(n->index);
1741        n->index = NULL;
1742        if (n->data) {
1743                freed_mem += n->entry->size;
1744                free(n->data);
1745                n->data = NULL;
1746        }
1747        n->entry = NULL;
1748        n->depth = 0;
1749        return freed_mem;
1750}
1751
1752static void find_deltas(struct object_entry **list, unsigned *list_size,
1753                        int window, int depth, unsigned *processed)
1754{
1755        uint32_t i, idx = 0, count = 0;
1756        struct unpacked *array;
1757        unsigned long mem_usage = 0;
1758
1759        array = xcalloc(window, sizeof(struct unpacked));
1760
1761        for (;;) {
1762                struct object_entry *entry;
1763                struct unpacked *n = array + idx;
1764                int j, max_depth, best_base = -1;
1765
1766                progress_lock();
1767                if (!*list_size) {
1768                        progress_unlock();
1769                        break;
1770                }
1771                entry = *list++;
1772                (*list_size)--;
1773                if (!entry->preferred_base) {
1774                        (*processed)++;
1775                        display_progress(progress_state, *processed);
1776                }
1777                progress_unlock();
1778
1779                mem_usage -= free_unpacked(n);
1780                n->entry = entry;
1781
1782                while (window_memory_limit &&
1783                       mem_usage > window_memory_limit &&
1784                       count > 1) {
1785                        uint32_t tail = (idx + window - count) % window;
1786                        mem_usage -= free_unpacked(array + tail);
1787                        count--;
1788                }
1789
1790                /* We do not compute delta to *create* objects we are not
1791                 * going to pack.
1792                 */
1793                if (entry->preferred_base)
1794                        goto next;
1795
1796                /*
1797                 * If the current object is at pack edge, take the depth the
1798                 * objects that depend on the current object into account
1799                 * otherwise they would become too deep.
1800                 */
1801                max_depth = depth;
1802                if (entry->delta_child) {
1803                        max_depth -= check_delta_limit(entry, 0);
1804                        if (max_depth <= 0)
1805                                goto next;
1806                }
1807
1808                j = window;
1809                while (--j > 0) {
1810                        int ret;
1811                        uint32_t other_idx = idx + j;
1812                        struct unpacked *m;
1813                        if (other_idx >= window)
1814                                other_idx -= window;
1815                        m = array + other_idx;
1816                        if (!m->entry)
1817                                break;
1818                        ret = try_delta(n, m, max_depth, &mem_usage);
1819                        if (ret < 0)
1820                                break;
1821                        else if (ret > 0)
1822                                best_base = other_idx;
1823                }
1824
1825                /*
1826                 * If we decided to cache the delta data, then it is best
1827                 * to compress it right away.  First because we have to do
1828                 * it anyway, and doing it here while we're threaded will
1829                 * save a lot of time in the non threaded write phase,
1830                 * as well as allow for caching more deltas within
1831                 * the same cache size limit.
1832                 * ...
1833                 * But only if not writing to stdout, since in that case
1834                 * the network is most likely throttling writes anyway,
1835                 * and therefore it is best to go to the write phase ASAP
1836                 * instead, as we can afford spending more time compressing
1837                 * between writes at that moment.
1838                 */
1839                if (entry->delta_data && !pack_to_stdout) {
1840                        entry->z_delta_size = do_compress(&entry->delta_data,
1841                                                          entry->delta_size);
1842                        cache_lock();
1843                        delta_cache_size -= entry->delta_size;
1844                        delta_cache_size += entry->z_delta_size;
1845                        cache_unlock();
1846                }
1847
1848                /* if we made n a delta, and if n is already at max
1849                 * depth, leaving it in the window is pointless.  we
1850                 * should evict it first.
1851                 */
1852                if (entry->delta && max_depth <= n->depth)
1853                        continue;
1854
1855                /*
1856                 * Move the best delta base up in the window, after the
1857                 * currently deltified object, to keep it longer.  It will
1858                 * be the first base object to be attempted next.
1859                 */
1860                if (entry->delta) {
1861                        struct unpacked swap = array[best_base];
1862                        int dist = (window + idx - best_base) % window;
1863                        int dst = best_base;
1864                        while (dist--) {
1865                                int src = (dst + 1) % window;
1866                                array[dst] = array[src];
1867                                dst = src;
1868                        }
1869                        array[dst] = swap;
1870                }
1871
1872                next:
1873                idx++;
1874                if (count + 1 < window)
1875                        count++;
1876                if (idx >= window)
1877                        idx = 0;
1878        }
1879
1880        for (i = 0; i < window; ++i) {
1881                free_delta_index(array[i].index);
1882                free(array[i].data);
1883        }
1884        free(array);
1885}
1886
1887#ifndef NO_PTHREADS
1888
1889static void try_to_free_from_threads(size_t size)
1890{
1891        read_lock();
1892        release_pack_memory(size);
1893        read_unlock();
1894}
1895
1896static try_to_free_t old_try_to_free_routine;
1897
1898/*
1899 * The main thread waits on the condition that (at least) one of the workers
1900 * has stopped working (which is indicated in the .working member of
1901 * struct thread_params).
1902 * When a work thread has completed its work, it sets .working to 0 and
1903 * signals the main thread and waits on the condition that .data_ready
1904 * becomes 1.
1905 */
1906
1907struct thread_params {
1908        pthread_t thread;
1909        struct object_entry **list;
1910        unsigned list_size;
1911        unsigned remaining;
1912        int window;
1913        int depth;
1914        int working;
1915        int data_ready;
1916        pthread_mutex_t mutex;
1917        pthread_cond_t cond;
1918        unsigned *processed;
1919};
1920
1921static pthread_cond_t progress_cond;
1922
1923/*
1924 * Mutex and conditional variable can't be statically-initialized on Windows.
1925 */
1926static void init_threaded_search(void)
1927{
1928        init_recursive_mutex(&read_mutex);
1929        pthread_mutex_init(&cache_mutex, NULL);
1930        pthread_mutex_init(&progress_mutex, NULL);
1931        pthread_cond_init(&progress_cond, NULL);
1932        old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1933}
1934
1935static void cleanup_threaded_search(void)
1936{
1937        set_try_to_free_routine(old_try_to_free_routine);
1938        pthread_cond_destroy(&progress_cond);
1939        pthread_mutex_destroy(&read_mutex);
1940        pthread_mutex_destroy(&cache_mutex);
1941        pthread_mutex_destroy(&progress_mutex);
1942}
1943
1944static void *threaded_find_deltas(void *arg)
1945{
1946        struct thread_params *me = arg;
1947
1948        while (me->remaining) {
1949                find_deltas(me->list, &me->remaining,
1950                            me->window, me->depth, me->processed);
1951
1952                progress_lock();
1953                me->working = 0;
1954                pthread_cond_signal(&progress_cond);
1955                progress_unlock();
1956
1957                /*
1958                 * We must not set ->data_ready before we wait on the
1959                 * condition because the main thread may have set it to 1
1960                 * before we get here. In order to be sure that new
1961                 * work is available if we see 1 in ->data_ready, it
1962                 * was initialized to 0 before this thread was spawned
1963                 * and we reset it to 0 right away.
1964                 */
1965                pthread_mutex_lock(&me->mutex);
1966                while (!me->data_ready)
1967                        pthread_cond_wait(&me->cond, &me->mutex);
1968                me->data_ready = 0;
1969                pthread_mutex_unlock(&me->mutex);
1970        }
1971        /* leave ->working 1 so that this doesn't get more work assigned */
1972        return NULL;
1973}
1974
1975static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1976                           int window, int depth, unsigned *processed)
1977{
1978        struct thread_params *p;
1979        int i, ret, active_threads = 0;
1980
1981        init_threaded_search();
1982
1983        if (delta_search_threads <= 1) {
1984                find_deltas(list, &list_size, window, depth, processed);
1985                cleanup_threaded_search();
1986                return;
1987        }
1988        if (progress > pack_to_stdout)
1989                fprintf(stderr, "Delta compression using up to %d threads.\n",
1990                                delta_search_threads);
1991        p = xcalloc(delta_search_threads, sizeof(*p));
1992
1993        /* Partition the work amongst work threads. */
1994        for (i = 0; i < delta_search_threads; i++) {
1995                unsigned sub_size = list_size / (delta_search_threads - i);
1996
1997                /* don't use too small segments or no deltas will be found */
1998                if (sub_size < 2*window && i+1 < delta_search_threads)
1999                        sub_size = 0;
2000
2001                p[i].window = window;
2002                p[i].depth = depth;
2003                p[i].processed = processed;
2004                p[i].working = 1;
2005                p[i].data_ready = 0;
2006
2007                /* try to split chunks on "path" boundaries */
2008                while (sub_size && sub_size < list_size &&
2009                       list[sub_size]->hash &&
2010                       list[sub_size]->hash == list[sub_size-1]->hash)
2011                        sub_size++;
2012
2013                p[i].list = list;
2014                p[i].list_size = sub_size;
2015                p[i].remaining = sub_size;
2016
2017                list += sub_size;
2018                list_size -= sub_size;
2019        }
2020
2021        /* Start work threads. */
2022        for (i = 0; i < delta_search_threads; i++) {
2023                if (!p[i].list_size)
2024                        continue;
2025                pthread_mutex_init(&p[i].mutex, NULL);
2026                pthread_cond_init(&p[i].cond, NULL);
2027                ret = pthread_create(&p[i].thread, NULL,
2028                                     threaded_find_deltas, &p[i]);
2029                if (ret)
2030                        die("unable to create thread: %s", strerror(ret));
2031                active_threads++;
2032        }
2033
2034        /*
2035         * Now let's wait for work completion.  Each time a thread is done
2036         * with its work, we steal half of the remaining work from the
2037         * thread with the largest number of unprocessed objects and give
2038         * it to that newly idle thread.  This ensure good load balancing
2039         * until the remaining object list segments are simply too short
2040         * to be worth splitting anymore.
2041         */
2042        while (active_threads) {
2043                struct thread_params *target = NULL;
2044                struct thread_params *victim = NULL;
2045                unsigned sub_size = 0;
2046
2047                progress_lock();
2048                for (;;) {
2049                        for (i = 0; !target && i < delta_search_threads; i++)
2050                                if (!p[i].working)
2051                                        target = &p[i];
2052                        if (target)
2053                                break;
2054                        pthread_cond_wait(&progress_cond, &progress_mutex);
2055                }
2056
2057                for (i = 0; i < delta_search_threads; i++)
2058                        if (p[i].remaining > 2*window &&
2059                            (!victim || victim->remaining < p[i].remaining))
2060                                victim = &p[i];
2061                if (victim) {
2062                        sub_size = victim->remaining / 2;
2063                        list = victim->list + victim->list_size - sub_size;
2064                        while (sub_size && list[0]->hash &&
2065                               list[0]->hash == list[-1]->hash) {
2066                                list++;
2067                                sub_size--;
2068                        }
2069                        if (!sub_size) {
2070                                /*
2071                                 * It is possible for some "paths" to have
2072                                 * so many objects that no hash boundary
2073                                 * might be found.  Let's just steal the
2074                                 * exact half in that case.
2075                                 */
2076                                sub_size = victim->remaining / 2;
2077                                list -= sub_size;
2078                        }
2079                        target->list = list;
2080                        victim->list_size -= sub_size;
2081                        victim->remaining -= sub_size;
2082                }
2083                target->list_size = sub_size;
2084                target->remaining = sub_size;
2085                target->working = 1;
2086                progress_unlock();
2087
2088                pthread_mutex_lock(&target->mutex);
2089                target->data_ready = 1;
2090                pthread_cond_signal(&target->cond);
2091                pthread_mutex_unlock(&target->mutex);
2092
2093                if (!sub_size) {
2094                        pthread_join(target->thread, NULL);
2095                        pthread_cond_destroy(&target->cond);
2096                        pthread_mutex_destroy(&target->mutex);
2097                        active_threads--;
2098                }
2099        }
2100        cleanup_threaded_search();
2101        free(p);
2102}
2103
2104#else
2105#define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2106#endif
2107
2108static int add_ref_tag(const char *path, const struct object_id *oid, int flag, void *cb_data)
2109{
2110        struct object_id peeled;
2111
2112        if (starts_with(path, "refs/tags/") && /* is a tag? */
2113            !peel_ref(path, peeled.hash)    && /* peelable? */
2114            packlist_find(&to_pack, peeled.hash, NULL))      /* object packed? */
2115                add_object_entry(oid->hash, OBJ_TAG, NULL, 0);
2116        return 0;
2117}
2118
2119static void prepare_pack(int window, int depth)
2120{
2121        struct object_entry **delta_list;
2122        uint32_t i, nr_deltas;
2123        unsigned n;
2124
2125        get_object_details();
2126
2127        /*
2128         * If we're locally repacking then we need to be doubly careful
2129         * from now on in order to make sure no stealth corruption gets
2130         * propagated to the new pack.  Clients receiving streamed packs
2131         * should validate everything they get anyway so no need to incur
2132         * the additional cost here in that case.
2133         */
2134        if (!pack_to_stdout)
2135                do_check_packed_object_crc = 1;
2136
2137        if (!to_pack.nr_objects || !window || !depth)
2138                return;
2139
2140        ALLOC_ARRAY(delta_list, to_pack.nr_objects);
2141        nr_deltas = n = 0;
2142
2143        for (i = 0; i < to_pack.nr_objects; i++) {
2144                struct object_entry *entry = to_pack.objects + i;
2145
2146                if (entry->delta)
2147                        /* This happens if we decided to reuse existing
2148                         * delta from a pack.  "reuse_delta &&" is implied.
2149                         */
2150                        continue;
2151
2152                if (entry->size < 50)
2153                        continue;
2154
2155                if (entry->no_try_delta)
2156                        continue;
2157
2158                if (!entry->preferred_base) {
2159                        nr_deltas++;
2160                        if (entry->type < 0)
2161                                die("unable to get type of object %s",
2162                                    sha1_to_hex(entry->idx.sha1));
2163                } else {
2164                        if (entry->type < 0) {
2165                                /*
2166                                 * This object is not found, but we
2167                                 * don't have to include it anyway.
2168                                 */
2169                                continue;
2170                        }
2171                }
2172
2173                delta_list[n++] = entry;
2174        }
2175
2176        if (nr_deltas && n > 1) {
2177                unsigned nr_done = 0;
2178                if (progress)
2179                        progress_state = start_progress(_("Compressing objects"),
2180                                                        nr_deltas);
2181                qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2182                ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2183                stop_progress(&progress_state);
2184                if (nr_done != nr_deltas)
2185                        die("inconsistency with delta count");
2186        }
2187        free(delta_list);
2188}
2189
2190static int git_pack_config(const char *k, const char *v, void *cb)
2191{
2192        if (!strcmp(k, "pack.window")) {
2193                window = git_config_int(k, v);
2194                return 0;
2195        }
2196        if (!strcmp(k, "pack.windowmemory")) {
2197                window_memory_limit = git_config_ulong(k, v);
2198                return 0;
2199        }
2200        if (!strcmp(k, "pack.depth")) {
2201                depth = git_config_int(k, v);
2202                return 0;
2203        }
2204        if (!strcmp(k, "pack.compression")) {
2205                int level = git_config_int(k, v);
2206                if (level == -1)
2207                        level = Z_DEFAULT_COMPRESSION;
2208                else if (level < 0 || level > Z_BEST_COMPRESSION)
2209                        die("bad pack compression level %d", level);
2210                pack_compression_level = level;
2211                pack_compression_seen = 1;
2212                return 0;
2213        }
2214        if (!strcmp(k, "pack.deltacachesize")) {
2215                max_delta_cache_size = git_config_int(k, v);
2216                return 0;
2217        }
2218        if (!strcmp(k, "pack.deltacachelimit")) {
2219                cache_max_small_delta_size = git_config_int(k, v);
2220                return 0;
2221        }
2222        if (!strcmp(k, "pack.writebitmaphashcache")) {
2223                if (git_config_bool(k, v))
2224                        write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2225                else
2226                        write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2227        }
2228        if (!strcmp(k, "pack.usebitmaps")) {
2229                use_bitmap_index = git_config_bool(k, v);
2230                return 0;
2231        }
2232        if (!strcmp(k, "pack.threads")) {
2233                delta_search_threads = git_config_int(k, v);
2234                if (delta_search_threads < 0)
2235                        die("invalid number of threads specified (%d)",
2236                            delta_search_threads);
2237#ifdef NO_PTHREADS
2238                if (delta_search_threads != 1)
2239                        warning("no threads support, ignoring %s", k);
2240#endif
2241                return 0;
2242        }
2243        if (!strcmp(k, "pack.indexversion")) {
2244                pack_idx_opts.version = git_config_int(k, v);
2245                if (pack_idx_opts.version > 2)
2246                        die("bad pack.indexversion=%"PRIu32,
2247                            pack_idx_opts.version);
2248                return 0;
2249        }
2250        return git_default_config(k, v, cb);
2251}
2252
2253static void read_object_list_from_stdin(void)
2254{
2255        char line[40 + 1 + PATH_MAX + 2];
2256        unsigned char sha1[20];
2257
2258        for (;;) {
2259                if (!fgets(line, sizeof(line), stdin)) {
2260                        if (feof(stdin))
2261                                break;
2262                        if (!ferror(stdin))
2263                                die("fgets returned NULL, not EOF, not error!");
2264                        if (errno != EINTR)
2265                                die_errno("fgets");
2266                        clearerr(stdin);
2267                        continue;
2268                }
2269                if (line[0] == '-') {
2270                        if (get_sha1_hex(line+1, sha1))
2271                                die("expected edge sha1, got garbage:\n %s",
2272                                    line);
2273                        add_preferred_base(sha1);
2274                        continue;
2275                }
2276                if (get_sha1_hex(line, sha1))
2277                        die("expected sha1, got garbage:\n %s", line);
2278
2279                add_preferred_base_object(line+41);
2280                add_object_entry(sha1, 0, line+41, 0);
2281        }
2282}
2283
2284#define OBJECT_ADDED (1u<<20)
2285
2286static void show_commit(struct commit *commit, void *data)
2287{
2288        add_object_entry(commit->object.oid.hash, OBJ_COMMIT, NULL, 0);
2289        commit->object.flags |= OBJECT_ADDED;
2290
2291        if (write_bitmap_index)
2292                index_commit_for_bitmap(commit);
2293}
2294
2295static void show_object(struct object *obj, const char *name, void *data)
2296{
2297        add_preferred_base_object(name);
2298        add_object_entry(obj->oid.hash, obj->type, name, 0);
2299        obj->flags |= OBJECT_ADDED;
2300}
2301
2302static void show_edge(struct commit *commit)
2303{
2304        add_preferred_base(commit->object.oid.hash);
2305}
2306
2307struct in_pack_object {
2308        off_t offset;
2309        struct object *object;
2310};
2311
2312struct in_pack {
2313        int alloc;
2314        int nr;
2315        struct in_pack_object *array;
2316};
2317
2318static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2319{
2320        in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->oid.hash, p);
2321        in_pack->array[in_pack->nr].object = object;
2322        in_pack->nr++;
2323}
2324
2325/*
2326 * Compare the objects in the offset order, in order to emulate the
2327 * "git rev-list --objects" output that produced the pack originally.
2328 */
2329static int ofscmp(const void *a_, const void *b_)
2330{
2331        struct in_pack_object *a = (struct in_pack_object *)a_;
2332        struct in_pack_object *b = (struct in_pack_object *)b_;
2333
2334        if (a->offset < b->offset)
2335                return -1;
2336        else if (a->offset > b->offset)
2337                return 1;
2338        else
2339                return oidcmp(&a->object->oid, &b->object->oid);
2340}
2341
2342static void add_objects_in_unpacked_packs(struct rev_info *revs)
2343{
2344        struct packed_git *p;
2345        struct in_pack in_pack;
2346        uint32_t i;
2347
2348        memset(&in_pack, 0, sizeof(in_pack));
2349
2350        for (p = packed_git; p; p = p->next) {
2351                const unsigned char *sha1;
2352                struct object *o;
2353
2354                if (!p->pack_local || p->pack_keep)
2355                        continue;
2356                if (open_pack_index(p))
2357                        die("cannot open pack index");
2358
2359                ALLOC_GROW(in_pack.array,
2360                           in_pack.nr + p->num_objects,
2361                           in_pack.alloc);
2362
2363                for (i = 0; i < p->num_objects; i++) {
2364                        sha1 = nth_packed_object_sha1(p, i);
2365                        o = lookup_unknown_object(sha1);
2366                        if (!(o->flags & OBJECT_ADDED))
2367                                mark_in_pack_object(o, p, &in_pack);
2368                        o->flags |= OBJECT_ADDED;
2369                }
2370        }
2371
2372        if (in_pack.nr) {
2373                qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2374                      ofscmp);
2375                for (i = 0; i < in_pack.nr; i++) {
2376                        struct object *o = in_pack.array[i].object;
2377                        add_object_entry(o->oid.hash, o->type, "", 0);
2378                }
2379        }
2380        free(in_pack.array);
2381}
2382
2383static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2384{
2385        static struct packed_git *last_found = (void *)1;
2386        struct packed_git *p;
2387
2388        p = (last_found != (void *)1) ? last_found : packed_git;
2389
2390        while (p) {
2391                if ((!p->pack_local || p->pack_keep) &&
2392                        find_pack_entry_one(sha1, p)) {
2393                        last_found = p;
2394                        return 1;
2395                }
2396                if (p == last_found)
2397                        p = packed_git;
2398                else
2399                        p = p->next;
2400                if (p == last_found)
2401                        p = p->next;
2402        }
2403        return 0;
2404}
2405
2406/*
2407 * Store a list of sha1s that are should not be discarded
2408 * because they are either written too recently, or are
2409 * reachable from another object that was.
2410 *
2411 * This is filled by get_object_list.
2412 */
2413static struct sha1_array recent_objects;
2414
2415static int loosened_object_can_be_discarded(const unsigned char *sha1,
2416                                            unsigned long mtime)
2417{
2418        if (!unpack_unreachable_expiration)
2419                return 0;
2420        if (mtime > unpack_unreachable_expiration)
2421                return 0;
2422        if (sha1_array_lookup(&recent_objects, sha1) >= 0)
2423                return 0;
2424        return 1;
2425}
2426
2427static void loosen_unused_packed_objects(struct rev_info *revs)
2428{
2429        struct packed_git *p;
2430        uint32_t i;
2431        const unsigned char *sha1;
2432
2433        for (p = packed_git; p; p = p->next) {
2434                if (!p->pack_local || p->pack_keep)
2435                        continue;
2436
2437                if (open_pack_index(p))
2438                        die("cannot open pack index");
2439
2440                for (i = 0; i < p->num_objects; i++) {
2441                        sha1 = nth_packed_object_sha1(p, i);
2442                        if (!packlist_find(&to_pack, sha1, NULL) &&
2443                            !has_sha1_pack_kept_or_nonlocal(sha1) &&
2444                            !loosened_object_can_be_discarded(sha1, p->mtime))
2445                                if (force_object_loose(sha1, p->mtime))
2446                                        die("unable to force loose object");
2447                }
2448        }
2449}
2450
2451/*
2452 * This tracks any options which a reader of the pack might
2453 * not understand, and which would therefore prevent blind reuse
2454 * of what we have on disk.
2455 */
2456static int pack_options_allow_reuse(void)
2457{
2458        return allow_ofs_delta;
2459}
2460
2461static int get_object_list_from_bitmap(struct rev_info *revs)
2462{
2463        if (prepare_bitmap_walk(revs) < 0)
2464                return -1;
2465
2466        if (pack_options_allow_reuse() &&
2467            !reuse_partial_packfile_from_bitmap(
2468                        &reuse_packfile,
2469                        &reuse_packfile_objects,
2470                        &reuse_packfile_offset)) {
2471                assert(reuse_packfile_objects);
2472                nr_result += reuse_packfile_objects;
2473                display_progress(progress_state, nr_result);
2474        }
2475
2476        traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2477        return 0;
2478}
2479
2480static void record_recent_object(struct object *obj,
2481                                 const char *name,
2482                                 void *data)
2483{
2484        sha1_array_append(&recent_objects, obj->oid.hash);
2485}
2486
2487static void record_recent_commit(struct commit *commit, void *data)
2488{
2489        sha1_array_append(&recent_objects, commit->object.oid.hash);
2490}
2491
2492static void get_object_list(int ac, const char **av)
2493{
2494        struct rev_info revs;
2495        char line[1000];
2496        int flags = 0;
2497
2498        init_revisions(&revs, NULL);
2499        save_commit_buffer = 0;
2500        setup_revisions(ac, av, &revs, NULL);
2501
2502        /* make sure shallows are read */
2503        is_repository_shallow();
2504
2505        while (fgets(line, sizeof(line), stdin) != NULL) {
2506                int len = strlen(line);
2507                if (len && line[len - 1] == '\n')
2508                        line[--len] = 0;
2509                if (!len)
2510                        break;
2511                if (*line == '-') {
2512                        if (!strcmp(line, "--not")) {
2513                                flags ^= UNINTERESTING;
2514                                write_bitmap_index = 0;
2515                                continue;
2516                        }
2517                        if (starts_with(line, "--shallow ")) {
2518                                unsigned char sha1[20];
2519                                if (get_sha1_hex(line + 10, sha1))
2520                                        die("not an SHA-1 '%s'", line + 10);
2521                                register_shallow(sha1);
2522                                use_bitmap_index = 0;
2523                                continue;
2524                        }
2525                        die("not a rev '%s'", line);
2526                }
2527                if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2528                        die("bad revision '%s'", line);
2529        }
2530
2531        if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2532                return;
2533
2534        if (prepare_revision_walk(&revs))
2535                die("revision walk setup failed");
2536        mark_edges_uninteresting(&revs, show_edge);
2537        traverse_commit_list(&revs, show_commit, show_object, NULL);
2538
2539        if (unpack_unreachable_expiration) {
2540                revs.ignore_missing_links = 1;
2541                if (add_unseen_recent_objects_to_traversal(&revs,
2542                                unpack_unreachable_expiration))
2543                        die("unable to add recent objects");
2544                if (prepare_revision_walk(&revs))
2545                        die("revision walk setup failed");
2546                traverse_commit_list(&revs, record_recent_commit,
2547                                     record_recent_object, NULL);
2548        }
2549
2550        if (keep_unreachable)
2551                add_objects_in_unpacked_packs(&revs);
2552        if (unpack_unreachable)
2553                loosen_unused_packed_objects(&revs);
2554
2555        sha1_array_clear(&recent_objects);
2556}
2557
2558static int option_parse_index_version(const struct option *opt,
2559                                      const char *arg, int unset)
2560{
2561        char *c;
2562        const char *val = arg;
2563        pack_idx_opts.version = strtoul(val, &c, 10);
2564        if (pack_idx_opts.version > 2)
2565                die(_("unsupported index version %s"), val);
2566        if (*c == ',' && c[1])
2567                pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2568        if (*c || pack_idx_opts.off32_limit & 0x80000000)
2569                die(_("bad index version '%s'"), val);
2570        return 0;
2571}
2572
2573static int option_parse_unpack_unreachable(const struct option *opt,
2574                                           const char *arg, int unset)
2575{
2576        if (unset) {
2577                unpack_unreachable = 0;
2578                unpack_unreachable_expiration = 0;
2579        }
2580        else {
2581                unpack_unreachable = 1;
2582                if (arg)
2583                        unpack_unreachable_expiration = approxidate(arg);
2584        }
2585        return 0;
2586}
2587
2588int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2589{
2590        int use_internal_rev_list = 0;
2591        int thin = 0;
2592        int shallow = 0;
2593        int all_progress_implied = 0;
2594        struct argv_array rp = ARGV_ARRAY_INIT;
2595        int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2596        int rev_list_index = 0;
2597        struct option pack_objects_options[] = {
2598                OPT_SET_INT('q', "quiet", &progress,
2599                            N_("do not show progress meter"), 0),
2600                OPT_SET_INT(0, "progress", &progress,
2601                            N_("show progress meter"), 1),
2602                OPT_SET_INT(0, "all-progress", &progress,
2603                            N_("show progress meter during object writing phase"), 2),
2604                OPT_BOOL(0, "all-progress-implied",
2605                         &all_progress_implied,
2606                         N_("similar to --all-progress when progress meter is shown")),
2607                { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2608                  N_("write the pack index file in the specified idx format version"),
2609                  0, option_parse_index_version },
2610                OPT_MAGNITUDE(0, "max-pack-size", &pack_size_limit,
2611                              N_("maximum size of each output pack file")),
2612                OPT_BOOL(0, "local", &local,
2613                         N_("ignore borrowed objects from alternate object store")),
2614                OPT_BOOL(0, "incremental", &incremental,
2615                         N_("ignore packed objects")),
2616                OPT_INTEGER(0, "window", &window,
2617                            N_("limit pack window by objects")),
2618                OPT_MAGNITUDE(0, "window-memory", &window_memory_limit,
2619                              N_("limit pack window by memory in addition to object limit")),
2620                OPT_INTEGER(0, "depth", &depth,
2621                            N_("maximum length of delta chain allowed in the resulting pack")),
2622                OPT_BOOL(0, "reuse-delta", &reuse_delta,
2623                         N_("reuse existing deltas")),
2624                OPT_BOOL(0, "reuse-object", &reuse_object,
2625                         N_("reuse existing objects")),
2626                OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2627                         N_("use OFS_DELTA objects")),
2628                OPT_INTEGER(0, "threads", &delta_search_threads,
2629                            N_("use threads when searching for best delta matches")),
2630                OPT_BOOL(0, "non-empty", &non_empty,
2631                         N_("do not create an empty pack output")),
2632                OPT_BOOL(0, "revs", &use_internal_rev_list,
2633                         N_("read revision arguments from standard input")),
2634                { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2635                  N_("limit the objects to those that are not yet packed"),
2636                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2637                { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2638                  N_("include objects reachable from any reference"),
2639                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2640                { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2641                  N_("include objects referred by reflog entries"),
2642                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2643                { OPTION_SET_INT, 0, "indexed-objects", &rev_list_index, NULL,
2644                  N_("include objects referred to by the index"),
2645                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2646                OPT_BOOL(0, "stdout", &pack_to_stdout,
2647                         N_("output pack to stdout")),
2648                OPT_BOOL(0, "include-tag", &include_tag,
2649                         N_("include tag objects that refer to objects to be packed")),
2650                OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2651                         N_("keep unreachable objects")),
2652                { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2653                  N_("unpack unreachable objects newer than <time>"),
2654                  PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2655                OPT_BOOL(0, "thin", &thin,
2656                         N_("create thin packs")),
2657                OPT_BOOL(0, "shallow", &shallow,
2658                         N_("create packs suitable for shallow fetches")),
2659                OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2660                         N_("ignore packs that have companion .keep file")),
2661                OPT_INTEGER(0, "compression", &pack_compression_level,
2662                            N_("pack compression level")),
2663                OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2664                            N_("do not hide commits by grafts"), 0),
2665                OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2666                         N_("use a bitmap index if available to speed up counting objects")),
2667                OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2668                         N_("write a bitmap index together with the pack index")),
2669                OPT_END(),
2670        };
2671
2672        check_replace_refs = 0;
2673
2674        reset_pack_idx_option(&pack_idx_opts);
2675        git_config(git_pack_config, NULL);
2676        if (!pack_compression_seen && core_compression_seen)
2677                pack_compression_level = core_compression_level;
2678
2679        progress = isatty(2);
2680        argc = parse_options(argc, argv, prefix, pack_objects_options,
2681                             pack_usage, 0);
2682
2683        if (argc) {
2684                base_name = argv[0];
2685                argc--;
2686        }
2687        if (pack_to_stdout != !base_name || argc)
2688                usage_with_options(pack_usage, pack_objects_options);
2689
2690        argv_array_push(&rp, "pack-objects");
2691        if (thin) {
2692                use_internal_rev_list = 1;
2693                argv_array_push(&rp, shallow
2694                                ? "--objects-edge-aggressive"
2695                                : "--objects-edge");
2696        } else
2697                argv_array_push(&rp, "--objects");
2698
2699        if (rev_list_all) {
2700                use_internal_rev_list = 1;
2701                argv_array_push(&rp, "--all");
2702        }
2703        if (rev_list_reflog) {
2704                use_internal_rev_list = 1;
2705                argv_array_push(&rp, "--reflog");
2706        }
2707        if (rev_list_index) {
2708                use_internal_rev_list = 1;
2709                argv_array_push(&rp, "--indexed-objects");
2710        }
2711        if (rev_list_unpacked) {
2712                use_internal_rev_list = 1;
2713                argv_array_push(&rp, "--unpacked");
2714        }
2715
2716        if (!reuse_object)
2717                reuse_delta = 0;
2718        if (pack_compression_level == -1)
2719                pack_compression_level = Z_DEFAULT_COMPRESSION;
2720        else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2721                die("bad pack compression level %d", pack_compression_level);
2722
2723        if (!delta_search_threads)      /* --threads=0 means autodetect */
2724                delta_search_threads = online_cpus();
2725
2726#ifdef NO_PTHREADS
2727        if (delta_search_threads != 1)
2728                warning("no threads support, ignoring --threads");
2729#endif
2730        if (!pack_to_stdout && !pack_size_limit)
2731                pack_size_limit = pack_size_limit_cfg;
2732        if (pack_to_stdout && pack_size_limit)
2733                die("--max-pack-size cannot be used to build a pack for transfer.");
2734        if (pack_size_limit && pack_size_limit < 1024*1024) {
2735                warning("minimum pack size limit is 1 MiB");
2736                pack_size_limit = 1024*1024;
2737        }
2738
2739        if (!pack_to_stdout && thin)
2740                die("--thin cannot be used to build an indexable pack.");
2741
2742        if (keep_unreachable && unpack_unreachable)
2743                die("--keep-unreachable and --unpack-unreachable are incompatible.");
2744        if (!rev_list_all || !rev_list_reflog || !rev_list_index)
2745                unpack_unreachable_expiration = 0;
2746
2747        if (!use_internal_rev_list || !pack_to_stdout || is_repository_shallow())
2748                use_bitmap_index = 0;
2749
2750        if (pack_to_stdout || !rev_list_all)
2751                write_bitmap_index = 0;
2752
2753        if (progress && all_progress_implied)
2754                progress = 2;
2755
2756        prepare_packed_git();
2757
2758        if (progress)
2759                progress_state = start_progress(_("Counting objects"), 0);
2760        if (!use_internal_rev_list)
2761                read_object_list_from_stdin();
2762        else {
2763                get_object_list(rp.argc, rp.argv);
2764                argv_array_clear(&rp);
2765        }
2766        cleanup_preferred_base();
2767        if (include_tag && nr_result)
2768                for_each_ref(add_ref_tag, NULL);
2769        stop_progress(&progress_state);
2770
2771        if (non_empty && !nr_result)
2772                return 0;
2773        if (nr_result)
2774                prepare_pack(window, depth);
2775        write_pack_file();
2776        if (progress)
2777                fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2778                        " reused %"PRIu32" (delta %"PRIu32")\n",
2779                        written, written_delta, reused, reused_delta);
2780        return 0;
2781}