builtin / pack-objects.con commit pack-objects: match prune logic for discarding objects (abcb865)
   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
  26static const char *pack_usage[] = {
  27        N_("git pack-objects --stdout [options...] [< ref-list | < object-list]"),
  28        N_("git pack-objects [options...] base-name [< ref-list | < object-list]"),
  29        NULL
  30};
  31
  32/*
  33 * Objects we are going to pack are collected in the `to_pack` structure.
  34 * It contains an array (dynamically expanded) of the object data, and a map
  35 * that can resolve SHA1s to their position in the array.
  36 */
  37static struct packing_data to_pack;
  38
  39static struct pack_idx_entry **written_list;
  40static uint32_t nr_result, nr_written;
  41
  42static int non_empty;
  43static int reuse_delta = 1, reuse_object = 1;
  44static int keep_unreachable, unpack_unreachable, include_tag;
  45static unsigned long unpack_unreachable_expiration;
  46static int local;
  47static int incremental;
  48static int ignore_packed_keep;
  49static int allow_ofs_delta;
  50static struct pack_idx_option pack_idx_opts;
  51static const char *base_name;
  52static int progress = 1;
  53static int window = 10;
  54static unsigned long pack_size_limit;
  55static int depth = 50;
  56static int delta_search_threads;
  57static int pack_to_stdout;
  58static int num_preferred_base;
  59static struct progress *progress_state;
  60static int pack_compression_level = Z_DEFAULT_COMPRESSION;
  61static int pack_compression_seen;
  62
  63static struct packed_git *reuse_packfile;
  64static uint32_t reuse_packfile_objects;
  65static off_t reuse_packfile_offset;
  66
  67static int use_bitmap_index = 1;
  68static int write_bitmap_index;
  69static uint16_t write_bitmap_options;
  70
  71static unsigned long delta_cache_size = 0;
  72static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
  73static unsigned long cache_max_small_delta_size = 1000;
  74
  75static unsigned long window_memory_limit = 0;
  76
  77/*
  78 * stats
  79 */
  80static uint32_t written, written_delta;
  81static uint32_t reused, reused_delta;
  82
  83/*
  84 * Indexed commits
  85 */
  86static struct commit **indexed_commits;
  87static unsigned int indexed_commits_nr;
  88static unsigned int indexed_commits_alloc;
  89
  90static void index_commit_for_bitmap(struct commit *commit)
  91{
  92        if (indexed_commits_nr >= indexed_commits_alloc) {
  93                indexed_commits_alloc = (indexed_commits_alloc + 32) * 2;
  94                REALLOC_ARRAY(indexed_commits, indexed_commits_alloc);
  95        }
  96
  97        indexed_commits[indexed_commits_nr++] = commit;
  98}
  99
 100static void *get_delta(struct object_entry *entry)
 101{
 102        unsigned long size, base_size, delta_size;
 103        void *buf, *base_buf, *delta_buf;
 104        enum object_type type;
 105
 106        buf = read_sha1_file(entry->idx.sha1, &type, &size);
 107        if (!buf)
 108                die("unable to read %s", sha1_to_hex(entry->idx.sha1));
 109        base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
 110        if (!base_buf)
 111                die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
 112        delta_buf = diff_delta(base_buf, base_size,
 113                               buf, size, &delta_size, 0);
 114        if (!delta_buf || delta_size != entry->delta_size)
 115                die("delta size changed");
 116        free(buf);
 117        free(base_buf);
 118        return delta_buf;
 119}
 120
 121static unsigned long do_compress(void **pptr, unsigned long size)
 122{
 123        git_zstream stream;
 124        void *in, *out;
 125        unsigned long maxsize;
 126
 127        memset(&stream, 0, sizeof(stream));
 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        memset(&stream, 0, sizeof(stream));
 156        git_deflate_init(&stream, pack_compression_level);
 157
 158        for (;;) {
 159                ssize_t readlen;
 160                int zret = Z_OK;
 161                readlen = read_istream(st, ibuf, sizeof(ibuf));
 162                if (readlen == -1)
 163                        die(_("unable to read %s"), sha1_to_hex(sha1));
 164
 165                stream.next_in = ibuf;
 166                stream.avail_in = readlen;
 167                while ((stream.avail_in || readlen == 0) &&
 168                       (zret == Z_OK || zret == Z_BUF_ERROR)) {
 169                        stream.next_out = obuf;
 170                        stream.avail_out = sizeof(obuf);
 171                        zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 172                        sha1write(f, obuf, stream.next_out - obuf);
 173                        olen += stream.next_out - obuf;
 174                }
 175                if (stream.avail_in)
 176                        die(_("deflate error (%d)"), zret);
 177                if (readlen == 0) {
 178                        if (zret != Z_STREAM_END)
 179                                die(_("deflate error (%d)"), zret);
 180                        break;
 181                }
 182        }
 183        git_deflate_end(&stream);
 184        return olen;
 185}
 186
 187/*
 188 * we are going to reuse the existing object data as is.  make
 189 * sure it is not corrupt.
 190 */
 191static int check_pack_inflate(struct packed_git *p,
 192                struct pack_window **w_curs,
 193                off_t offset,
 194                off_t len,
 195                unsigned long expect)
 196{
 197        git_zstream stream;
 198        unsigned char fakebuf[4096], *in;
 199        int st;
 200
 201        memset(&stream, 0, sizeof(stream));
 202        git_inflate_init(&stream);
 203        do {
 204                in = use_pack(p, w_curs, offset, &stream.avail_in);
 205                stream.next_in = in;
 206                stream.next_out = fakebuf;
 207                stream.avail_out = sizeof(fakebuf);
 208                st = git_inflate(&stream, Z_FINISH);
 209                offset += stream.next_in - in;
 210        } while (st == Z_OK || st == Z_BUF_ERROR);
 211        git_inflate_end(&stream);
 212        return (st == Z_STREAM_END &&
 213                stream.total_out == expect &&
 214                stream.total_in == len) ? 0 : -1;
 215}
 216
 217static void copy_pack_data(struct sha1file *f,
 218                struct packed_git *p,
 219                struct pack_window **w_curs,
 220                off_t offset,
 221                off_t len)
 222{
 223        unsigned char *in;
 224        unsigned long avail;
 225
 226        while (len) {
 227                in = use_pack(p, w_curs, offset, &avail);
 228                if (avail > len)
 229                        avail = (unsigned long)len;
 230                sha1write(f, in, avail);
 231                offset += avail;
 232                len -= avail;
 233        }
 234}
 235
 236/* Return 0 if we will bust the pack-size limit */
 237static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
 238                                           unsigned long limit, int usable_delta)
 239{
 240        unsigned long size, datalen;
 241        unsigned char header[10], dheader[10];
 242        unsigned hdrlen;
 243        enum object_type type;
 244        void *buf;
 245        struct git_istream *st = NULL;
 246
 247        if (!usable_delta) {
 248                if (entry->type == OBJ_BLOB &&
 249                    entry->size > big_file_threshold &&
 250                    (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
 251                        buf = NULL;
 252                else {
 253                        buf = read_sha1_file(entry->idx.sha1, &type, &size);
 254                        if (!buf)
 255                                die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
 256                }
 257                /*
 258                 * make sure no cached delta data remains from a
 259                 * previous attempt before a pack split occurred.
 260                 */
 261                free(entry->delta_data);
 262                entry->delta_data = NULL;
 263                entry->z_delta_size = 0;
 264        } else if (entry->delta_data) {
 265                size = entry->delta_size;
 266                buf = entry->delta_data;
 267                entry->delta_data = NULL;
 268                type = (allow_ofs_delta && entry->delta->idx.offset) ?
 269                        OBJ_OFS_DELTA : OBJ_REF_DELTA;
 270        } else {
 271                buf = get_delta(entry);
 272                size = entry->delta_size;
 273                type = (allow_ofs_delta && entry->delta->idx.offset) ?
 274                        OBJ_OFS_DELTA : OBJ_REF_DELTA;
 275        }
 276
 277        if (st) /* large blob case, just assume we don't compress well */
 278                datalen = size;
 279        else if (entry->z_delta_size)
 280                datalen = entry->z_delta_size;
 281        else
 282                datalen = do_compress(&buf, size);
 283
 284        /*
 285         * The object header is a byte of 'type' followed by zero or
 286         * more bytes of length.
 287         */
 288        hdrlen = encode_in_pack_object_header(type, size, header);
 289
 290        if (type == OBJ_OFS_DELTA) {
 291                /*
 292                 * Deltas with relative base contain an additional
 293                 * encoding of the relative offset for the delta
 294                 * base from this object's position in the pack.
 295                 */
 296                off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 297                unsigned pos = sizeof(dheader) - 1;
 298                dheader[pos] = ofs & 127;
 299                while (ofs >>= 7)
 300                        dheader[--pos] = 128 | (--ofs & 127);
 301                if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 302                        if (st)
 303                                close_istream(st);
 304                        free(buf);
 305                        return 0;
 306                }
 307                sha1write(f, header, hdrlen);
 308                sha1write(f, dheader + pos, sizeof(dheader) - pos);
 309                hdrlen += sizeof(dheader) - pos;
 310        } else if (type == OBJ_REF_DELTA) {
 311                /*
 312                 * Deltas with a base reference contain
 313                 * an additional 20 bytes for the base sha1.
 314                 */
 315                if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 316                        if (st)
 317                                close_istream(st);
 318                        free(buf);
 319                        return 0;
 320                }
 321                sha1write(f, header, hdrlen);
 322                sha1write(f, entry->delta->idx.sha1, 20);
 323                hdrlen += 20;
 324        } else {
 325                if (limit && hdrlen + datalen + 20 >= limit) {
 326                        if (st)
 327                                close_istream(st);
 328                        free(buf);
 329                        return 0;
 330                }
 331                sha1write(f, header, hdrlen);
 332        }
 333        if (st) {
 334                datalen = write_large_blob_data(st, f, entry->idx.sha1);
 335                close_istream(st);
 336        } else {
 337                sha1write(f, buf, datalen);
 338                free(buf);
 339        }
 340
 341        return hdrlen + datalen;
 342}
 343
 344/* Return 0 if we will bust the pack-size limit */
 345static unsigned long write_reuse_object(struct sha1file *f, struct object_entry *entry,
 346                                        unsigned long limit, int usable_delta)
 347{
 348        struct packed_git *p = entry->in_pack;
 349        struct pack_window *w_curs = NULL;
 350        struct revindex_entry *revidx;
 351        off_t offset;
 352        enum object_type type = entry->type;
 353        unsigned long datalen;
 354        unsigned char header[10], dheader[10];
 355        unsigned hdrlen;
 356
 357        if (entry->delta)
 358                type = (allow_ofs_delta && entry->delta->idx.offset) ?
 359                        OBJ_OFS_DELTA : OBJ_REF_DELTA;
 360        hdrlen = encode_in_pack_object_header(type, entry->size, header);
 361
 362        offset = entry->in_pack_offset;
 363        revidx = find_pack_revindex(p, offset);
 364        datalen = revidx[1].offset - offset;
 365        if (!pack_to_stdout && p->index_version > 1 &&
 366            check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 367                error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
 368                unuse_pack(&w_curs);
 369                return write_no_reuse_object(f, entry, limit, usable_delta);
 370        }
 371
 372        offset += entry->in_pack_header_size;
 373        datalen -= entry->in_pack_header_size;
 374
 375        if (!pack_to_stdout && p->index_version == 1 &&
 376            check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 377                error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
 378                unuse_pack(&w_curs);
 379                return write_no_reuse_object(f, entry, limit, usable_delta);
 380        }
 381
 382        if (type == OBJ_OFS_DELTA) {
 383                off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 384                unsigned pos = sizeof(dheader) - 1;
 385                dheader[pos] = ofs & 127;
 386                while (ofs >>= 7)
 387                        dheader[--pos] = 128 | (--ofs & 127);
 388                if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 389                        unuse_pack(&w_curs);
 390                        return 0;
 391                }
 392                sha1write(f, header, hdrlen);
 393                sha1write(f, dheader + pos, sizeof(dheader) - pos);
 394                hdrlen += sizeof(dheader) - pos;
 395                reused_delta++;
 396        } else if (type == OBJ_REF_DELTA) {
 397                if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 398                        unuse_pack(&w_curs);
 399                        return 0;
 400                }
 401                sha1write(f, header, hdrlen);
 402                sha1write(f, entry->delta->idx.sha1, 20);
 403                hdrlen += 20;
 404                reused_delta++;
 405        } else {
 406                if (limit && hdrlen + datalen + 20 >= limit) {
 407                        unuse_pack(&w_curs);
 408                        return 0;
 409                }
 410                sha1write(f, header, hdrlen);
 411        }
 412        copy_pack_data(f, p, &w_curs, offset, datalen);
 413        unuse_pack(&w_curs);
 414        reused++;
 415        return hdrlen + datalen;
 416}
 417
 418/* Return 0 if we will bust the pack-size limit */
 419static unsigned long write_object(struct sha1file *f,
 420                                  struct object_entry *entry,
 421                                  off_t write_offset)
 422{
 423        unsigned long limit, len;
 424        int usable_delta, to_reuse;
 425
 426        if (!pack_to_stdout)
 427                crc32_begin(f);
 428
 429        /* apply size limit if limited packsize and not first object */
 430        if (!pack_size_limit || !nr_written)
 431                limit = 0;
 432        else if (pack_size_limit <= write_offset)
 433                /*
 434                 * the earlier object did not fit the limit; avoid
 435                 * mistaking this with unlimited (i.e. limit = 0).
 436                 */
 437                limit = 1;
 438        else
 439                limit = pack_size_limit - write_offset;
 440
 441        if (!entry->delta)
 442                usable_delta = 0;       /* no delta */
 443        else if (!pack_size_limit)
 444               usable_delta = 1;        /* unlimited packfile */
 445        else if (entry->delta->idx.offset == (off_t)-1)
 446                usable_delta = 0;       /* base was written to another pack */
 447        else if (entry->delta->idx.offset)
 448                usable_delta = 1;       /* base already exists in this pack */
 449        else
 450                usable_delta = 0;       /* base could end up in another pack */
 451
 452        if (!reuse_object)
 453                to_reuse = 0;   /* explicit */
 454        else if (!entry->in_pack)
 455                to_reuse = 0;   /* can't reuse what we don't have */
 456        else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
 457                                /* check_object() decided it for us ... */
 458                to_reuse = usable_delta;
 459                                /* ... but pack split may override that */
 460        else if (entry->type != entry->in_pack_type)
 461                to_reuse = 0;   /* pack has delta which is unusable */
 462        else if (entry->delta)
 463                to_reuse = 0;   /* we want to pack afresh */
 464        else
 465                to_reuse = 1;   /* we have it in-pack undeltified,
 466                                 * and we do not need to deltify it.
 467                                 */
 468
 469        if (!to_reuse)
 470                len = write_no_reuse_object(f, entry, limit, usable_delta);
 471        else
 472                len = write_reuse_object(f, entry, limit, usable_delta);
 473        if (!len)
 474                return 0;
 475
 476        if (usable_delta)
 477                written_delta++;
 478        written++;
 479        if (!pack_to_stdout)
 480                entry->idx.crc32 = crc32_end(f);
 481        return len;
 482}
 483
 484enum write_one_status {
 485        WRITE_ONE_SKIP = -1, /* already written */
 486        WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 487        WRITE_ONE_WRITTEN = 1, /* normal */
 488        WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 489};
 490
 491static enum write_one_status write_one(struct sha1file *f,
 492                                       struct object_entry *e,
 493                                       off_t *offset)
 494{
 495        unsigned long size;
 496        int recursing;
 497
 498        /*
 499         * we set offset to 1 (which is an impossible value) to mark
 500         * the fact that this object is involved in "write its base
 501         * first before writing a deltified object" recursion.
 502         */
 503        recursing = (e->idx.offset == 1);
 504        if (recursing) {
 505                warning("recursive delta detected for object %s",
 506                        sha1_to_hex(e->idx.sha1));
 507                return WRITE_ONE_RECURSIVE;
 508        } else if (e->idx.offset || e->preferred_base) {
 509                /* offset is non zero if object is written already. */
 510                return WRITE_ONE_SKIP;
 511        }
 512
 513        /* if we are deltified, write out base object first. */
 514        if (e->delta) {
 515                e->idx.offset = 1; /* now recurse */
 516                switch (write_one(f, e->delta, offset)) {
 517                case WRITE_ONE_RECURSIVE:
 518                        /* we cannot depend on this one */
 519                        e->delta = NULL;
 520                        break;
 521                default:
 522                        break;
 523                case WRITE_ONE_BREAK:
 524                        e->idx.offset = recursing;
 525                        return WRITE_ONE_BREAK;
 526                }
 527        }
 528
 529        e->idx.offset = *offset;
 530        size = write_object(f, e, *offset);
 531        if (!size) {
 532                e->idx.offset = recursing;
 533                return WRITE_ONE_BREAK;
 534        }
 535        written_list[nr_written++] = &e->idx;
 536
 537        /* make sure off_t is sufficiently large not to wrap */
 538        if (signed_add_overflows(*offset, size))
 539                die("pack too large for current definition of off_t");
 540        *offset += size;
 541        return WRITE_ONE_WRITTEN;
 542}
 543
 544static int mark_tagged(const char *path, const unsigned char *sha1, int flag,
 545                       void *cb_data)
 546{
 547        unsigned char peeled[20];
 548        struct object_entry *entry = packlist_find(&to_pack, sha1, NULL);
 549
 550        if (entry)
 551                entry->tagged = 1;
 552        if (!peel_ref(path, peeled)) {
 553                entry = packlist_find(&to_pack, peeled, NULL);
 554                if (entry)
 555                        entry->tagged = 1;
 556        }
 557        return 0;
 558}
 559
 560static inline void add_to_write_order(struct object_entry **wo,
 561                               unsigned int *endp,
 562                               struct object_entry *e)
 563{
 564        if (e->filled)
 565                return;
 566        wo[(*endp)++] = e;
 567        e->filled = 1;
 568}
 569
 570static void add_descendants_to_write_order(struct object_entry **wo,
 571                                           unsigned int *endp,
 572                                           struct object_entry *e)
 573{
 574        int add_to_order = 1;
 575        while (e) {
 576                if (add_to_order) {
 577                        struct object_entry *s;
 578                        /* add this node... */
 579                        add_to_write_order(wo, endp, e);
 580                        /* all its siblings... */
 581                        for (s = e->delta_sibling; s; s = s->delta_sibling) {
 582                                add_to_write_order(wo, endp, s);
 583                        }
 584                }
 585                /* drop down a level to add left subtree nodes if possible */
 586                if (e->delta_child) {
 587                        add_to_order = 1;
 588                        e = e->delta_child;
 589                } else {
 590                        add_to_order = 0;
 591                        /* our sibling might have some children, it is next */
 592                        if (e->delta_sibling) {
 593                                e = e->delta_sibling;
 594                                continue;
 595                        }
 596                        /* go back to our parent node */
 597                        e = e->delta;
 598                        while (e && !e->delta_sibling) {
 599                                /* we're on the right side of a subtree, keep
 600                                 * going up until we can go right again */
 601                                e = e->delta;
 602                        }
 603                        if (!e) {
 604                                /* done- we hit our original root node */
 605                                return;
 606                        }
 607                        /* pass it off to sibling at this level */
 608                        e = e->delta_sibling;
 609                }
 610        };
 611}
 612
 613static void add_family_to_write_order(struct object_entry **wo,
 614                                      unsigned int *endp,
 615                                      struct object_entry *e)
 616{
 617        struct object_entry *root;
 618
 619        for (root = e; root->delta; root = root->delta)
 620                ; /* nothing */
 621        add_descendants_to_write_order(wo, endp, root);
 622}
 623
 624static struct object_entry **compute_write_order(void)
 625{
 626        unsigned int i, wo_end, last_untagged;
 627
 628        struct object_entry **wo = xmalloc(to_pack.nr_objects * sizeof(*wo));
 629        struct object_entry *objects = to_pack.objects;
 630
 631        for (i = 0; i < to_pack.nr_objects; i++) {
 632                objects[i].tagged = 0;
 633                objects[i].filled = 0;
 634                objects[i].delta_child = NULL;
 635                objects[i].delta_sibling = NULL;
 636        }
 637
 638        /*
 639         * Fully connect delta_child/delta_sibling network.
 640         * Make sure delta_sibling is sorted in the original
 641         * recency order.
 642         */
 643        for (i = to_pack.nr_objects; i > 0;) {
 644                struct object_entry *e = &objects[--i];
 645                if (!e->delta)
 646                        continue;
 647                /* Mark me as the first child */
 648                e->delta_sibling = e->delta->delta_child;
 649                e->delta->delta_child = e;
 650        }
 651
 652        /*
 653         * Mark objects that are at the tip of tags.
 654         */
 655        for_each_tag_ref(mark_tagged, NULL);
 656
 657        /*
 658         * Give the objects in the original recency order until
 659         * we see a tagged tip.
 660         */
 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 void write_pack_file(void)
 763{
 764        uint32_t i = 0, j;
 765        struct sha1file *f;
 766        off_t offset;
 767        uint32_t nr_remaining = nr_result;
 768        time_t last_mtime = 0;
 769        struct object_entry **write_order;
 770
 771        if (progress > pack_to_stdout)
 772                progress_state = start_progress(_("Writing objects"), nr_result);
 773        written_list = xmalloc(to_pack.nr_objects * sizeof(*written_list));
 774        write_order = compute_write_order();
 775
 776        do {
 777                unsigned char sha1[20];
 778                char *pack_tmp_name = NULL;
 779
 780                if (pack_to_stdout)
 781                        f = sha1fd_throughput(1, "<stdout>", progress_state);
 782                else
 783                        f = create_tmp_packfile(&pack_tmp_name);
 784
 785                offset = write_pack_header(f, nr_remaining);
 786
 787                if (reuse_packfile) {
 788                        off_t packfile_size;
 789                        assert(pack_to_stdout);
 790
 791                        packfile_size = write_reused_pack(f);
 792                        offset += packfile_size;
 793                }
 794
 795                nr_written = 0;
 796                for (; i < to_pack.nr_objects; i++) {
 797                        struct object_entry *e = write_order[i];
 798                        if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 799                                break;
 800                        display_progress(progress_state, written);
 801                }
 802
 803                /*
 804                 * Did we write the wrong # entries in the header?
 805                 * If so, rewrite it like in fast-import
 806                 */
 807                if (pack_to_stdout) {
 808                        sha1close(f, sha1, CSUM_CLOSE);
 809                } else if (nr_written == nr_remaining) {
 810                        sha1close(f, sha1, CSUM_FSYNC);
 811                } else {
 812                        int fd = sha1close(f, sha1, 0);
 813                        fixup_pack_header_footer(fd, sha1, pack_tmp_name,
 814                                                 nr_written, sha1, offset);
 815                        close(fd);
 816                }
 817
 818                if (!pack_to_stdout) {
 819                        struct stat st;
 820                        struct strbuf tmpname = STRBUF_INIT;
 821
 822                        /*
 823                         * Packs are runtime accessed in their mtime
 824                         * order since newer packs are more likely to contain
 825                         * younger objects.  So if we are creating multiple
 826                         * packs then we should modify the mtime of later ones
 827                         * to preserve this property.
 828                         */
 829                        if (stat(pack_tmp_name, &st) < 0) {
 830                                warning("failed to stat %s: %s",
 831                                        pack_tmp_name, strerror(errno));
 832                        } else if (!last_mtime) {
 833                                last_mtime = st.st_mtime;
 834                        } else {
 835                                struct utimbuf utb;
 836                                utb.actime = st.st_atime;
 837                                utb.modtime = --last_mtime;
 838                                if (utime(pack_tmp_name, &utb) < 0)
 839                                        warning("failed utime() on %s: %s",
 840                                                pack_tmp_name, strerror(errno));
 841                        }
 842
 843                        strbuf_addf(&tmpname, "%s-", base_name);
 844
 845                        if (write_bitmap_index) {
 846                                bitmap_writer_set_checksum(sha1);
 847                                bitmap_writer_build_type_index(written_list, nr_written);
 848                        }
 849
 850                        finish_tmp_packfile(&tmpname, pack_tmp_name,
 851                                            written_list, nr_written,
 852                                            &pack_idx_opts, sha1);
 853
 854                        if (write_bitmap_index) {
 855                                strbuf_addf(&tmpname, "%s.bitmap", sha1_to_hex(sha1));
 856
 857                                stop_progress(&progress_state);
 858
 859                                bitmap_writer_show_progress(progress);
 860                                bitmap_writer_reuse_bitmaps(&to_pack);
 861                                bitmap_writer_select_commits(indexed_commits, indexed_commits_nr, -1);
 862                                bitmap_writer_build(&to_pack);
 863                                bitmap_writer_finish(written_list, nr_written,
 864                                                     tmpname.buf, write_bitmap_options);
 865                                write_bitmap_index = 0;
 866                        }
 867
 868                        strbuf_release(&tmpname);
 869                        free(pack_tmp_name);
 870                        puts(sha1_to_hex(sha1));
 871                }
 872
 873                /* mark written objects as written to previous pack */
 874                for (j = 0; j < nr_written; j++) {
 875                        written_list[j]->offset = (off_t)-1;
 876                }
 877                nr_remaining -= nr_written;
 878        } while (nr_remaining && i < to_pack.nr_objects);
 879
 880        free(written_list);
 881        free(write_order);
 882        stop_progress(&progress_state);
 883        if (written != nr_result)
 884                die("wrote %"PRIu32" objects while expecting %"PRIu32,
 885                        written, nr_result);
 886}
 887
 888static void setup_delta_attr_check(struct git_attr_check *check)
 889{
 890        static struct git_attr *attr_delta;
 891
 892        if (!attr_delta)
 893                attr_delta = git_attr("delta");
 894
 895        check[0].attr = attr_delta;
 896}
 897
 898static int no_try_delta(const char *path)
 899{
 900        struct git_attr_check check[1];
 901
 902        setup_delta_attr_check(check);
 903        if (git_check_attr(path, ARRAY_SIZE(check), check))
 904                return 0;
 905        if (ATTR_FALSE(check->value))
 906                return 1;
 907        return 0;
 908}
 909
 910/*
 911 * When adding an object, check whether we have already added it
 912 * to our packing list. If so, we can skip. However, if we are
 913 * being asked to excludei t, but the previous mention was to include
 914 * it, make sure to adjust its flags and tweak our numbers accordingly.
 915 *
 916 * As an optimization, we pass out the index position where we would have
 917 * found the item, since that saves us from having to look it up again a
 918 * few lines later when we want to add the new entry.
 919 */
 920static int have_duplicate_entry(const unsigned char *sha1,
 921                                int exclude,
 922                                uint32_t *index_pos)
 923{
 924        struct object_entry *entry;
 925
 926        entry = packlist_find(&to_pack, sha1, index_pos);
 927        if (!entry)
 928                return 0;
 929
 930        if (exclude) {
 931                if (!entry->preferred_base)
 932                        nr_result--;
 933                entry->preferred_base = 1;
 934        }
 935
 936        return 1;
 937}
 938
 939/*
 940 * Check whether we want the object in the pack (e.g., we do not want
 941 * objects found in non-local stores if the "--local" option was used).
 942 *
 943 * As a side effect of this check, we will find the packed version of this
 944 * object, if any. We therefore pass out the pack information to avoid having
 945 * to look it up again later.
 946 */
 947static int want_object_in_pack(const unsigned char *sha1,
 948                               int exclude,
 949                               struct packed_git **found_pack,
 950                               off_t *found_offset)
 951{
 952        struct packed_git *p;
 953
 954        if (!exclude && local && has_loose_object_nonlocal(sha1))
 955                return 0;
 956
 957        *found_pack = NULL;
 958        *found_offset = 0;
 959
 960        for (p = packed_git; p; p = p->next) {
 961                off_t offset = find_pack_entry_one(sha1, p);
 962                if (offset) {
 963                        if (!*found_pack) {
 964                                if (!is_pack_valid(p)) {
 965                                        warning("packfile %s cannot be accessed", p->pack_name);
 966                                        continue;
 967                                }
 968                                *found_offset = offset;
 969                                *found_pack = p;
 970                        }
 971                        if (exclude)
 972                                return 1;
 973                        if (incremental)
 974                                return 0;
 975                        if (local && !p->pack_local)
 976                                return 0;
 977                        if (ignore_packed_keep && p->pack_local && p->pack_keep)
 978                                return 0;
 979                }
 980        }
 981
 982        return 1;
 983}
 984
 985static void create_object_entry(const unsigned char *sha1,
 986                                enum object_type type,
 987                                uint32_t hash,
 988                                int exclude,
 989                                int no_try_delta,
 990                                uint32_t index_pos,
 991                                struct packed_git *found_pack,
 992                                off_t found_offset)
 993{
 994        struct object_entry *entry;
 995
 996        entry = packlist_alloc(&to_pack, sha1, index_pos);
 997        entry->hash = hash;
 998        if (type)
 999                entry->type = type;
1000        if (exclude)
1001                entry->preferred_base = 1;
1002        else
1003                nr_result++;
1004        if (found_pack) {
1005                entry->in_pack = found_pack;
1006                entry->in_pack_offset = found_offset;
1007        }
1008
1009        entry->no_try_delta = no_try_delta;
1010}
1011
1012static const char no_closure_warning[] = N_(
1013"disabling bitmap writing, as some objects are not being packed"
1014);
1015
1016static int add_object_entry(const unsigned char *sha1, enum object_type type,
1017                            const char *name, int exclude)
1018{
1019        struct packed_git *found_pack;
1020        off_t found_offset;
1021        uint32_t index_pos;
1022
1023        if (have_duplicate_entry(sha1, exclude, &index_pos))
1024                return 0;
1025
1026        if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset)) {
1027                /* The pack is missing an object, so it will not have closure */
1028                if (write_bitmap_index) {
1029                        warning(_(no_closure_warning));
1030                        write_bitmap_index = 0;
1031                }
1032                return 0;
1033        }
1034
1035        create_object_entry(sha1, type, pack_name_hash(name),
1036                            exclude, name && no_try_delta(name),
1037                            index_pos, found_pack, found_offset);
1038
1039        display_progress(progress_state, nr_result);
1040        return 1;
1041}
1042
1043static int add_object_entry_from_bitmap(const unsigned char *sha1,
1044                                        enum object_type type,
1045                                        int flags, uint32_t name_hash,
1046                                        struct packed_git *pack, off_t offset)
1047{
1048        uint32_t index_pos;
1049
1050        if (have_duplicate_entry(sha1, 0, &index_pos))
1051                return 0;
1052
1053        create_object_entry(sha1, type, name_hash, 0, 0, index_pos, pack, offset);
1054
1055        display_progress(progress_state, nr_result);
1056        return 1;
1057}
1058
1059struct pbase_tree_cache {
1060        unsigned char sha1[20];
1061        int ref;
1062        int temporary;
1063        void *tree_data;
1064        unsigned long tree_size;
1065};
1066
1067static struct pbase_tree_cache *(pbase_tree_cache[256]);
1068static int pbase_tree_cache_ix(const unsigned char *sha1)
1069{
1070        return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
1071}
1072static int pbase_tree_cache_ix_incr(int ix)
1073{
1074        return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
1075}
1076
1077static struct pbase_tree {
1078        struct pbase_tree *next;
1079        /* This is a phony "cache" entry; we are not
1080         * going to evict it or find it through _get()
1081         * mechanism -- this is for the toplevel node that
1082         * would almost always change with any commit.
1083         */
1084        struct pbase_tree_cache pcache;
1085} *pbase_tree;
1086
1087static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
1088{
1089        struct pbase_tree_cache *ent, *nent;
1090        void *data;
1091        unsigned long size;
1092        enum object_type type;
1093        int neigh;
1094        int my_ix = pbase_tree_cache_ix(sha1);
1095        int available_ix = -1;
1096
1097        /* pbase-tree-cache acts as a limited hashtable.
1098         * your object will be found at your index or within a few
1099         * slots after that slot if it is cached.
1100         */
1101        for (neigh = 0; neigh < 8; neigh++) {
1102                ent = pbase_tree_cache[my_ix];
1103                if (ent && !hashcmp(ent->sha1, sha1)) {
1104                        ent->ref++;
1105                        return ent;
1106                }
1107                else if (((available_ix < 0) && (!ent || !ent->ref)) ||
1108                         ((0 <= available_ix) &&
1109                          (!ent && pbase_tree_cache[available_ix])))
1110                        available_ix = my_ix;
1111                if (!ent)
1112                        break;
1113                my_ix = pbase_tree_cache_ix_incr(my_ix);
1114        }
1115
1116        /* Did not find one.  Either we got a bogus request or
1117         * we need to read and perhaps cache.
1118         */
1119        data = read_sha1_file(sha1, &type, &size);
1120        if (!data)
1121                return NULL;
1122        if (type != OBJ_TREE) {
1123                free(data);
1124                return NULL;
1125        }
1126
1127        /* We need to either cache or return a throwaway copy */
1128
1129        if (available_ix < 0)
1130                ent = NULL;
1131        else {
1132                ent = pbase_tree_cache[available_ix];
1133                my_ix = available_ix;
1134        }
1135
1136        if (!ent) {
1137                nent = xmalloc(sizeof(*nent));
1138                nent->temporary = (available_ix < 0);
1139        }
1140        else {
1141                /* evict and reuse */
1142                free(ent->tree_data);
1143                nent = ent;
1144        }
1145        hashcpy(nent->sha1, sha1);
1146        nent->tree_data = data;
1147        nent->tree_size = size;
1148        nent->ref = 1;
1149        if (!nent->temporary)
1150                pbase_tree_cache[my_ix] = nent;
1151        return nent;
1152}
1153
1154static void pbase_tree_put(struct pbase_tree_cache *cache)
1155{
1156        if (!cache->temporary) {
1157                cache->ref--;
1158                return;
1159        }
1160        free(cache->tree_data);
1161        free(cache);
1162}
1163
1164static int name_cmp_len(const char *name)
1165{
1166        int i;
1167        for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1168                ;
1169        return i;
1170}
1171
1172static void add_pbase_object(struct tree_desc *tree,
1173                             const char *name,
1174                             int cmplen,
1175                             const char *fullname)
1176{
1177        struct name_entry entry;
1178        int cmp;
1179
1180        while (tree_entry(tree,&entry)) {
1181                if (S_ISGITLINK(entry.mode))
1182                        continue;
1183                cmp = tree_entry_len(&entry) != cmplen ? 1 :
1184                      memcmp(name, entry.path, cmplen);
1185                if (cmp > 0)
1186                        continue;
1187                if (cmp < 0)
1188                        return;
1189                if (name[cmplen] != '/') {
1190                        add_object_entry(entry.sha1,
1191                                         object_type(entry.mode),
1192                                         fullname, 1);
1193                        return;
1194                }
1195                if (S_ISDIR(entry.mode)) {
1196                        struct tree_desc sub;
1197                        struct pbase_tree_cache *tree;
1198                        const char *down = name+cmplen+1;
1199                        int downlen = name_cmp_len(down);
1200
1201                        tree = pbase_tree_get(entry.sha1);
1202                        if (!tree)
1203                                return;
1204                        init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1205
1206                        add_pbase_object(&sub, down, downlen, fullname);
1207                        pbase_tree_put(tree);
1208                }
1209        }
1210}
1211
1212static unsigned *done_pbase_paths;
1213static int done_pbase_paths_num;
1214static int done_pbase_paths_alloc;
1215static int done_pbase_path_pos(unsigned hash)
1216{
1217        int lo = 0;
1218        int hi = done_pbase_paths_num;
1219        while (lo < hi) {
1220                int mi = (hi + lo) / 2;
1221                if (done_pbase_paths[mi] == hash)
1222                        return mi;
1223                if (done_pbase_paths[mi] < hash)
1224                        hi = mi;
1225                else
1226                        lo = mi + 1;
1227        }
1228        return -lo-1;
1229}
1230
1231static int check_pbase_path(unsigned hash)
1232{
1233        int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1234        if (0 <= pos)
1235                return 1;
1236        pos = -pos - 1;
1237        ALLOC_GROW(done_pbase_paths,
1238                   done_pbase_paths_num + 1,
1239                   done_pbase_paths_alloc);
1240        done_pbase_paths_num++;
1241        if (pos < done_pbase_paths_num)
1242                memmove(done_pbase_paths + pos + 1,
1243                        done_pbase_paths + pos,
1244                        (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1245        done_pbase_paths[pos] = hash;
1246        return 0;
1247}
1248
1249static void add_preferred_base_object(const char *name)
1250{
1251        struct pbase_tree *it;
1252        int cmplen;
1253        unsigned hash = pack_name_hash(name);
1254
1255        if (!num_preferred_base || check_pbase_path(hash))
1256                return;
1257
1258        cmplen = name_cmp_len(name);
1259        for (it = pbase_tree; it; it = it->next) {
1260                if (cmplen == 0) {
1261                        add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1262                }
1263                else {
1264                        struct tree_desc tree;
1265                        init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1266                        add_pbase_object(&tree, name, cmplen, name);
1267                }
1268        }
1269}
1270
1271static void add_preferred_base(unsigned char *sha1)
1272{
1273        struct pbase_tree *it;
1274        void *data;
1275        unsigned long size;
1276        unsigned char tree_sha1[20];
1277
1278        if (window <= num_preferred_base++)
1279                return;
1280
1281        data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1282        if (!data)
1283                return;
1284
1285        for (it = pbase_tree; it; it = it->next) {
1286                if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1287                        free(data);
1288                        return;
1289                }
1290        }
1291
1292        it = xcalloc(1, sizeof(*it));
1293        it->next = pbase_tree;
1294        pbase_tree = it;
1295
1296        hashcpy(it->pcache.sha1, tree_sha1);
1297        it->pcache.tree_data = data;
1298        it->pcache.tree_size = size;
1299}
1300
1301static void cleanup_preferred_base(void)
1302{
1303        struct pbase_tree *it;
1304        unsigned i;
1305
1306        it = pbase_tree;
1307        pbase_tree = NULL;
1308        while (it) {
1309                struct pbase_tree *this = it;
1310                it = this->next;
1311                free(this->pcache.tree_data);
1312                free(this);
1313        }
1314
1315        for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1316                if (!pbase_tree_cache[i])
1317                        continue;
1318                free(pbase_tree_cache[i]->tree_data);
1319                free(pbase_tree_cache[i]);
1320                pbase_tree_cache[i] = NULL;
1321        }
1322
1323        free(done_pbase_paths);
1324        done_pbase_paths = NULL;
1325        done_pbase_paths_num = done_pbase_paths_alloc = 0;
1326}
1327
1328static void check_object(struct object_entry *entry)
1329{
1330        if (entry->in_pack) {
1331                struct packed_git *p = entry->in_pack;
1332                struct pack_window *w_curs = NULL;
1333                const unsigned char *base_ref = NULL;
1334                struct object_entry *base_entry;
1335                unsigned long used, used_0;
1336                unsigned long avail;
1337                off_t ofs;
1338                unsigned char *buf, c;
1339
1340                buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1341
1342                /*
1343                 * We want in_pack_type even if we do not reuse delta
1344                 * since non-delta representations could still be reused.
1345                 */
1346                used = unpack_object_header_buffer(buf, avail,
1347                                                   &entry->in_pack_type,
1348                                                   &entry->size);
1349                if (used == 0)
1350                        goto give_up;
1351
1352                /*
1353                 * Determine if this is a delta and if so whether we can
1354                 * reuse it or not.  Otherwise let's find out as cheaply as
1355                 * possible what the actual type and size for this object is.
1356                 */
1357                switch (entry->in_pack_type) {
1358                default:
1359                        /* Not a delta hence we've already got all we need. */
1360                        entry->type = entry->in_pack_type;
1361                        entry->in_pack_header_size = used;
1362                        if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1363                                goto give_up;
1364                        unuse_pack(&w_curs);
1365                        return;
1366                case OBJ_REF_DELTA:
1367                        if (reuse_delta && !entry->preferred_base)
1368                                base_ref = use_pack(p, &w_curs,
1369                                                entry->in_pack_offset + used, NULL);
1370                        entry->in_pack_header_size = used + 20;
1371                        break;
1372                case OBJ_OFS_DELTA:
1373                        buf = use_pack(p, &w_curs,
1374                                       entry->in_pack_offset + used, NULL);
1375                        used_0 = 0;
1376                        c = buf[used_0++];
1377                        ofs = c & 127;
1378                        while (c & 128) {
1379                                ofs += 1;
1380                                if (!ofs || MSB(ofs, 7)) {
1381                                        error("delta base offset overflow in pack for %s",
1382                                              sha1_to_hex(entry->idx.sha1));
1383                                        goto give_up;
1384                                }
1385                                c = buf[used_0++];
1386                                ofs = (ofs << 7) + (c & 127);
1387                        }
1388                        ofs = entry->in_pack_offset - ofs;
1389                        if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1390                                error("delta base offset out of bound for %s",
1391                                      sha1_to_hex(entry->idx.sha1));
1392                                goto give_up;
1393                        }
1394                        if (reuse_delta && !entry->preferred_base) {
1395                                struct revindex_entry *revidx;
1396                                revidx = find_pack_revindex(p, ofs);
1397                                if (!revidx)
1398                                        goto give_up;
1399                                base_ref = nth_packed_object_sha1(p, revidx->nr);
1400                        }
1401                        entry->in_pack_header_size = used + used_0;
1402                        break;
1403                }
1404
1405                if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1406                        /*
1407                         * If base_ref was set above that means we wish to
1408                         * reuse delta data, and we even found that base
1409                         * in the list of objects we want to pack. Goodie!
1410                         *
1411                         * Depth value does not matter - find_deltas() will
1412                         * never consider reused delta as the base object to
1413                         * deltify other objects against, in order to avoid
1414                         * circular deltas.
1415                         */
1416                        entry->type = entry->in_pack_type;
1417                        entry->delta = base_entry;
1418                        entry->delta_size = entry->size;
1419                        entry->delta_sibling = base_entry->delta_child;
1420                        base_entry->delta_child = entry;
1421                        unuse_pack(&w_curs);
1422                        return;
1423                }
1424
1425                if (entry->type) {
1426                        /*
1427                         * This must be a delta and we already know what the
1428                         * final object type is.  Let's extract the actual
1429                         * object size from the delta header.
1430                         */
1431                        entry->size = get_size_from_delta(p, &w_curs,
1432                                        entry->in_pack_offset + entry->in_pack_header_size);
1433                        if (entry->size == 0)
1434                                goto give_up;
1435                        unuse_pack(&w_curs);
1436                        return;
1437                }
1438
1439                /*
1440                 * No choice but to fall back to the recursive delta walk
1441                 * with sha1_object_info() to find about the object type
1442                 * at this point...
1443                 */
1444                give_up:
1445                unuse_pack(&w_curs);
1446        }
1447
1448        entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1449        /*
1450         * The error condition is checked in prepare_pack().  This is
1451         * to permit a missing preferred base object to be ignored
1452         * as a preferred base.  Doing so can result in a larger
1453         * pack file, but the transfer will still take place.
1454         */
1455}
1456
1457static int pack_offset_sort(const void *_a, const void *_b)
1458{
1459        const struct object_entry *a = *(struct object_entry **)_a;
1460        const struct object_entry *b = *(struct object_entry **)_b;
1461
1462        /* avoid filesystem trashing with loose objects */
1463        if (!a->in_pack && !b->in_pack)
1464                return hashcmp(a->idx.sha1, b->idx.sha1);
1465
1466        if (a->in_pack < b->in_pack)
1467                return -1;
1468        if (a->in_pack > b->in_pack)
1469                return 1;
1470        return a->in_pack_offset < b->in_pack_offset ? -1 :
1471                        (a->in_pack_offset > b->in_pack_offset);
1472}
1473
1474static void get_object_details(void)
1475{
1476        uint32_t i;
1477        struct object_entry **sorted_by_offset;
1478
1479        sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1480        for (i = 0; i < to_pack.nr_objects; i++)
1481                sorted_by_offset[i] = to_pack.objects + i;
1482        qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1483
1484        for (i = 0; i < to_pack.nr_objects; i++) {
1485                struct object_entry *entry = sorted_by_offset[i];
1486                check_object(entry);
1487                if (big_file_threshold < entry->size)
1488                        entry->no_try_delta = 1;
1489        }
1490
1491        free(sorted_by_offset);
1492}
1493
1494/*
1495 * We search for deltas in a list sorted by type, by filename hash, and then
1496 * by size, so that we see progressively smaller and smaller files.
1497 * That's because we prefer deltas to be from the bigger file
1498 * to the smaller -- deletes are potentially cheaper, but perhaps
1499 * more importantly, the bigger file is likely the more recent
1500 * one.  The deepest deltas are therefore the oldest objects which are
1501 * less susceptible to be accessed often.
1502 */
1503static int type_size_sort(const void *_a, const void *_b)
1504{
1505        const struct object_entry *a = *(struct object_entry **)_a;
1506        const struct object_entry *b = *(struct object_entry **)_b;
1507
1508        if (a->type > b->type)
1509                return -1;
1510        if (a->type < b->type)
1511                return 1;
1512        if (a->hash > b->hash)
1513                return -1;
1514        if (a->hash < b->hash)
1515                return 1;
1516        if (a->preferred_base > b->preferred_base)
1517                return -1;
1518        if (a->preferred_base < b->preferred_base)
1519                return 1;
1520        if (a->size > b->size)
1521                return -1;
1522        if (a->size < b->size)
1523                return 1;
1524        return a < b ? -1 : (a > b);  /* newest first */
1525}
1526
1527struct unpacked {
1528        struct object_entry *entry;
1529        void *data;
1530        struct delta_index *index;
1531        unsigned depth;
1532};
1533
1534static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1535                           unsigned long delta_size)
1536{
1537        if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1538                return 0;
1539
1540        if (delta_size < cache_max_small_delta_size)
1541                return 1;
1542
1543        /* cache delta, if objects are large enough compared to delta size */
1544        if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1545                return 1;
1546
1547        return 0;
1548}
1549
1550#ifndef NO_PTHREADS
1551
1552static pthread_mutex_t read_mutex;
1553#define read_lock()             pthread_mutex_lock(&read_mutex)
1554#define read_unlock()           pthread_mutex_unlock(&read_mutex)
1555
1556static pthread_mutex_t cache_mutex;
1557#define cache_lock()            pthread_mutex_lock(&cache_mutex)
1558#define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1559
1560static pthread_mutex_t progress_mutex;
1561#define progress_lock()         pthread_mutex_lock(&progress_mutex)
1562#define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1563
1564#else
1565
1566#define read_lock()             (void)0
1567#define read_unlock()           (void)0
1568#define cache_lock()            (void)0
1569#define cache_unlock()          (void)0
1570#define progress_lock()         (void)0
1571#define progress_unlock()       (void)0
1572
1573#endif
1574
1575static int try_delta(struct unpacked *trg, struct unpacked *src,
1576                     unsigned max_depth, unsigned long *mem_usage)
1577{
1578        struct object_entry *trg_entry = trg->entry;
1579        struct object_entry *src_entry = src->entry;
1580        unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1581        unsigned ref_depth;
1582        enum object_type type;
1583        void *delta_buf;
1584
1585        /* Don't bother doing diffs between different types */
1586        if (trg_entry->type != src_entry->type)
1587                return -1;
1588
1589        /*
1590         * We do not bother to try a delta that we discarded on an
1591         * earlier try, but only when reusing delta data.  Note that
1592         * src_entry that is marked as the preferred_base should always
1593         * be considered, as even if we produce a suboptimal delta against
1594         * it, we will still save the transfer cost, as we already know
1595         * the other side has it and we won't send src_entry at all.
1596         */
1597        if (reuse_delta && trg_entry->in_pack &&
1598            trg_entry->in_pack == src_entry->in_pack &&
1599            !src_entry->preferred_base &&
1600            trg_entry->in_pack_type != OBJ_REF_DELTA &&
1601            trg_entry->in_pack_type != OBJ_OFS_DELTA)
1602                return 0;
1603
1604        /* Let's not bust the allowed depth. */
1605        if (src->depth >= max_depth)
1606                return 0;
1607
1608        /* Now some size filtering heuristics. */
1609        trg_size = trg_entry->size;
1610        if (!trg_entry->delta) {
1611                max_size = trg_size/2 - 20;
1612                ref_depth = 1;
1613        } else {
1614                max_size = trg_entry->delta_size;
1615                ref_depth = trg->depth;
1616        }
1617        max_size = (uint64_t)max_size * (max_depth - src->depth) /
1618                                                (max_depth - ref_depth + 1);
1619        if (max_size == 0)
1620                return 0;
1621        src_size = src_entry->size;
1622        sizediff = src_size < trg_size ? trg_size - src_size : 0;
1623        if (sizediff >= max_size)
1624                return 0;
1625        if (trg_size < src_size / 32)
1626                return 0;
1627
1628        /* Load data if not already done */
1629        if (!trg->data) {
1630                read_lock();
1631                trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1632                read_unlock();
1633                if (!trg->data)
1634                        die("object %s cannot be read",
1635                            sha1_to_hex(trg_entry->idx.sha1));
1636                if (sz != trg_size)
1637                        die("object %s inconsistent object length (%lu vs %lu)",
1638                            sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1639                *mem_usage += sz;
1640        }
1641        if (!src->data) {
1642                read_lock();
1643                src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1644                read_unlock();
1645                if (!src->data) {
1646                        if (src_entry->preferred_base) {
1647                                static int warned = 0;
1648                                if (!warned++)
1649                                        warning("object %s cannot be read",
1650                                                sha1_to_hex(src_entry->idx.sha1));
1651                                /*
1652                                 * Those objects are not included in the
1653                                 * resulting pack.  Be resilient and ignore
1654                                 * them if they can't be read, in case the
1655                                 * pack could be created nevertheless.
1656                                 */
1657                                return 0;
1658                        }
1659                        die("object %s cannot be read",
1660                            sha1_to_hex(src_entry->idx.sha1));
1661                }
1662                if (sz != src_size)
1663                        die("object %s inconsistent object length (%lu vs %lu)",
1664                            sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1665                *mem_usage += sz;
1666        }
1667        if (!src->index) {
1668                src->index = create_delta_index(src->data, src_size);
1669                if (!src->index) {
1670                        static int warned = 0;
1671                        if (!warned++)
1672                                warning("suboptimal pack - out of memory");
1673                        return 0;
1674                }
1675                *mem_usage += sizeof_delta_index(src->index);
1676        }
1677
1678        delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1679        if (!delta_buf)
1680                return 0;
1681
1682        if (trg_entry->delta) {
1683                /* Prefer only shallower same-sized deltas. */
1684                if (delta_size == trg_entry->delta_size &&
1685                    src->depth + 1 >= trg->depth) {
1686                        free(delta_buf);
1687                        return 0;
1688                }
1689        }
1690
1691        /*
1692         * Handle memory allocation outside of the cache
1693         * accounting lock.  Compiler will optimize the strangeness
1694         * away when NO_PTHREADS is defined.
1695         */
1696        free(trg_entry->delta_data);
1697        cache_lock();
1698        if (trg_entry->delta_data) {
1699                delta_cache_size -= trg_entry->delta_size;
1700                trg_entry->delta_data = NULL;
1701        }
1702        if (delta_cacheable(src_size, trg_size, delta_size)) {
1703                delta_cache_size += delta_size;
1704                cache_unlock();
1705                trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1706        } else {
1707                cache_unlock();
1708                free(delta_buf);
1709        }
1710
1711        trg_entry->delta = src_entry;
1712        trg_entry->delta_size = delta_size;
1713        trg->depth = src->depth + 1;
1714
1715        return 1;
1716}
1717
1718static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1719{
1720        struct object_entry *child = me->delta_child;
1721        unsigned int m = n;
1722        while (child) {
1723                unsigned int c = check_delta_limit(child, n + 1);
1724                if (m < c)
1725                        m = c;
1726                child = child->delta_sibling;
1727        }
1728        return m;
1729}
1730
1731static unsigned long free_unpacked(struct unpacked *n)
1732{
1733        unsigned long freed_mem = sizeof_delta_index(n->index);
1734        free_delta_index(n->index);
1735        n->index = NULL;
1736        if (n->data) {
1737                freed_mem += n->entry->size;
1738                free(n->data);
1739                n->data = NULL;
1740        }
1741        n->entry = NULL;
1742        n->depth = 0;
1743        return freed_mem;
1744}
1745
1746static void find_deltas(struct object_entry **list, unsigned *list_size,
1747                        int window, int depth, unsigned *processed)
1748{
1749        uint32_t i, idx = 0, count = 0;
1750        struct unpacked *array;
1751        unsigned long mem_usage = 0;
1752
1753        array = xcalloc(window, sizeof(struct unpacked));
1754
1755        for (;;) {
1756                struct object_entry *entry;
1757                struct unpacked *n = array + idx;
1758                int j, max_depth, best_base = -1;
1759
1760                progress_lock();
1761                if (!*list_size) {
1762                        progress_unlock();
1763                        break;
1764                }
1765                entry = *list++;
1766                (*list_size)--;
1767                if (!entry->preferred_base) {
1768                        (*processed)++;
1769                        display_progress(progress_state, *processed);
1770                }
1771                progress_unlock();
1772
1773                mem_usage -= free_unpacked(n);
1774                n->entry = entry;
1775
1776                while (window_memory_limit &&
1777                       mem_usage > window_memory_limit &&
1778                       count > 1) {
1779                        uint32_t tail = (idx + window - count) % window;
1780                        mem_usage -= free_unpacked(array + tail);
1781                        count--;
1782                }
1783
1784                /* We do not compute delta to *create* objects we are not
1785                 * going to pack.
1786                 */
1787                if (entry->preferred_base)
1788                        goto next;
1789
1790                /*
1791                 * If the current object is at pack edge, take the depth the
1792                 * objects that depend on the current object into account
1793                 * otherwise they would become too deep.
1794                 */
1795                max_depth = depth;
1796                if (entry->delta_child) {
1797                        max_depth -= check_delta_limit(entry, 0);
1798                        if (max_depth <= 0)
1799                                goto next;
1800                }
1801
1802                j = window;
1803                while (--j > 0) {
1804                        int ret;
1805                        uint32_t other_idx = idx + j;
1806                        struct unpacked *m;
1807                        if (other_idx >= window)
1808                                other_idx -= window;
1809                        m = array + other_idx;
1810                        if (!m->entry)
1811                                break;
1812                        ret = try_delta(n, m, max_depth, &mem_usage);
1813                        if (ret < 0)
1814                                break;
1815                        else if (ret > 0)
1816                                best_base = other_idx;
1817                }
1818
1819                /*
1820                 * If we decided to cache the delta data, then it is best
1821                 * to compress it right away.  First because we have to do
1822                 * it anyway, and doing it here while we're threaded will
1823                 * save a lot of time in the non threaded write phase,
1824                 * as well as allow for caching more deltas within
1825                 * the same cache size limit.
1826                 * ...
1827                 * But only if not writing to stdout, since in that case
1828                 * the network is most likely throttling writes anyway,
1829                 * and therefore it is best to go to the write phase ASAP
1830                 * instead, as we can afford spending more time compressing
1831                 * between writes at that moment.
1832                 */
1833                if (entry->delta_data && !pack_to_stdout) {
1834                        entry->z_delta_size = do_compress(&entry->delta_data,
1835                                                          entry->delta_size);
1836                        cache_lock();
1837                        delta_cache_size -= entry->delta_size;
1838                        delta_cache_size += entry->z_delta_size;
1839                        cache_unlock();
1840                }
1841
1842                /* if we made n a delta, and if n is already at max
1843                 * depth, leaving it in the window is pointless.  we
1844                 * should evict it first.
1845                 */
1846                if (entry->delta && max_depth <= n->depth)
1847                        continue;
1848
1849                /*
1850                 * Move the best delta base up in the window, after the
1851                 * currently deltified object, to keep it longer.  It will
1852                 * be the first base object to be attempted next.
1853                 */
1854                if (entry->delta) {
1855                        struct unpacked swap = array[best_base];
1856                        int dist = (window + idx - best_base) % window;
1857                        int dst = best_base;
1858                        while (dist--) {
1859                                int src = (dst + 1) % window;
1860                                array[dst] = array[src];
1861                                dst = src;
1862                        }
1863                        array[dst] = swap;
1864                }
1865
1866                next:
1867                idx++;
1868                if (count + 1 < window)
1869                        count++;
1870                if (idx >= window)
1871                        idx = 0;
1872        }
1873
1874        for (i = 0; i < window; ++i) {
1875                free_delta_index(array[i].index);
1876                free(array[i].data);
1877        }
1878        free(array);
1879}
1880
1881#ifndef NO_PTHREADS
1882
1883static void try_to_free_from_threads(size_t size)
1884{
1885        read_lock();
1886        release_pack_memory(size);
1887        read_unlock();
1888}
1889
1890static try_to_free_t old_try_to_free_routine;
1891
1892/*
1893 * The main thread waits on the condition that (at least) one of the workers
1894 * has stopped working (which is indicated in the .working member of
1895 * struct thread_params).
1896 * When a work thread has completed its work, it sets .working to 0 and
1897 * signals the main thread and waits on the condition that .data_ready
1898 * becomes 1.
1899 */
1900
1901struct thread_params {
1902        pthread_t thread;
1903        struct object_entry **list;
1904        unsigned list_size;
1905        unsigned remaining;
1906        int window;
1907        int depth;
1908        int working;
1909        int data_ready;
1910        pthread_mutex_t mutex;
1911        pthread_cond_t cond;
1912        unsigned *processed;
1913};
1914
1915static pthread_cond_t progress_cond;
1916
1917/*
1918 * Mutex and conditional variable can't be statically-initialized on Windows.
1919 */
1920static void init_threaded_search(void)
1921{
1922        init_recursive_mutex(&read_mutex);
1923        pthread_mutex_init(&cache_mutex, NULL);
1924        pthread_mutex_init(&progress_mutex, NULL);
1925        pthread_cond_init(&progress_cond, NULL);
1926        old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1927}
1928
1929static void cleanup_threaded_search(void)
1930{
1931        set_try_to_free_routine(old_try_to_free_routine);
1932        pthread_cond_destroy(&progress_cond);
1933        pthread_mutex_destroy(&read_mutex);
1934        pthread_mutex_destroy(&cache_mutex);
1935        pthread_mutex_destroy(&progress_mutex);
1936}
1937
1938static void *threaded_find_deltas(void *arg)
1939{
1940        struct thread_params *me = arg;
1941
1942        while (me->remaining) {
1943                find_deltas(me->list, &me->remaining,
1944                            me->window, me->depth, me->processed);
1945
1946                progress_lock();
1947                me->working = 0;
1948                pthread_cond_signal(&progress_cond);
1949                progress_unlock();
1950
1951                /*
1952                 * We must not set ->data_ready before we wait on the
1953                 * condition because the main thread may have set it to 1
1954                 * before we get here. In order to be sure that new
1955                 * work is available if we see 1 in ->data_ready, it
1956                 * was initialized to 0 before this thread was spawned
1957                 * and we reset it to 0 right away.
1958                 */
1959                pthread_mutex_lock(&me->mutex);
1960                while (!me->data_ready)
1961                        pthread_cond_wait(&me->cond, &me->mutex);
1962                me->data_ready = 0;
1963                pthread_mutex_unlock(&me->mutex);
1964        }
1965        /* leave ->working 1 so that this doesn't get more work assigned */
1966        return NULL;
1967}
1968
1969static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1970                           int window, int depth, unsigned *processed)
1971{
1972        struct thread_params *p;
1973        int i, ret, active_threads = 0;
1974
1975        init_threaded_search();
1976
1977        if (!delta_search_threads)      /* --threads=0 means autodetect */
1978                delta_search_threads = online_cpus();
1979        if (delta_search_threads <= 1) {
1980                find_deltas(list, &list_size, window, depth, processed);
1981                cleanup_threaded_search();
1982                return;
1983        }
1984        if (progress > pack_to_stdout)
1985                fprintf(stderr, "Delta compression using up to %d threads.\n",
1986                                delta_search_threads);
1987        p = xcalloc(delta_search_threads, sizeof(*p));
1988
1989        /* Partition the work amongst work threads. */
1990        for (i = 0; i < delta_search_threads; i++) {
1991                unsigned sub_size = list_size / (delta_search_threads - i);
1992
1993                /* don't use too small segments or no deltas will be found */
1994                if (sub_size < 2*window && i+1 < delta_search_threads)
1995                        sub_size = 0;
1996
1997                p[i].window = window;
1998                p[i].depth = depth;
1999                p[i].processed = processed;
2000                p[i].working = 1;
2001                p[i].data_ready = 0;
2002
2003                /* try to split chunks on "path" boundaries */
2004                while (sub_size && sub_size < list_size &&
2005                       list[sub_size]->hash &&
2006                       list[sub_size]->hash == list[sub_size-1]->hash)
2007                        sub_size++;
2008
2009                p[i].list = list;
2010                p[i].list_size = sub_size;
2011                p[i].remaining = sub_size;
2012
2013                list += sub_size;
2014                list_size -= sub_size;
2015        }
2016
2017        /* Start work threads. */
2018        for (i = 0; i < delta_search_threads; i++) {
2019                if (!p[i].list_size)
2020                        continue;
2021                pthread_mutex_init(&p[i].mutex, NULL);
2022                pthread_cond_init(&p[i].cond, NULL);
2023                ret = pthread_create(&p[i].thread, NULL,
2024                                     threaded_find_deltas, &p[i]);
2025                if (ret)
2026                        die("unable to create thread: %s", strerror(ret));
2027                active_threads++;
2028        }
2029
2030        /*
2031         * Now let's wait for work completion.  Each time a thread is done
2032         * with its work, we steal half of the remaining work from the
2033         * thread with the largest number of unprocessed objects and give
2034         * it to that newly idle thread.  This ensure good load balancing
2035         * until the remaining object list segments are simply too short
2036         * to be worth splitting anymore.
2037         */
2038        while (active_threads) {
2039                struct thread_params *target = NULL;
2040                struct thread_params *victim = NULL;
2041                unsigned sub_size = 0;
2042
2043                progress_lock();
2044                for (;;) {
2045                        for (i = 0; !target && i < delta_search_threads; i++)
2046                                if (!p[i].working)
2047                                        target = &p[i];
2048                        if (target)
2049                                break;
2050                        pthread_cond_wait(&progress_cond, &progress_mutex);
2051                }
2052
2053                for (i = 0; i < delta_search_threads; i++)
2054                        if (p[i].remaining > 2*window &&
2055                            (!victim || victim->remaining < p[i].remaining))
2056                                victim = &p[i];
2057                if (victim) {
2058                        sub_size = victim->remaining / 2;
2059                        list = victim->list + victim->list_size - sub_size;
2060                        while (sub_size && list[0]->hash &&
2061                               list[0]->hash == list[-1]->hash) {
2062                                list++;
2063                                sub_size--;
2064                        }
2065                        if (!sub_size) {
2066                                /*
2067                                 * It is possible for some "paths" to have
2068                                 * so many objects that no hash boundary
2069                                 * might be found.  Let's just steal the
2070                                 * exact half in that case.
2071                                 */
2072                                sub_size = victim->remaining / 2;
2073                                list -= sub_size;
2074                        }
2075                        target->list = list;
2076                        victim->list_size -= sub_size;
2077                        victim->remaining -= sub_size;
2078                }
2079                target->list_size = sub_size;
2080                target->remaining = sub_size;
2081                target->working = 1;
2082                progress_unlock();
2083
2084                pthread_mutex_lock(&target->mutex);
2085                target->data_ready = 1;
2086                pthread_cond_signal(&target->cond);
2087                pthread_mutex_unlock(&target->mutex);
2088
2089                if (!sub_size) {
2090                        pthread_join(target->thread, NULL);
2091                        pthread_cond_destroy(&target->cond);
2092                        pthread_mutex_destroy(&target->mutex);
2093                        active_threads--;
2094                }
2095        }
2096        cleanup_threaded_search();
2097        free(p);
2098}
2099
2100#else
2101#define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
2102#endif
2103
2104static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
2105{
2106        unsigned char peeled[20];
2107
2108        if (starts_with(path, "refs/tags/") && /* is a tag? */
2109            !peel_ref(path, peeled)        && /* peelable? */
2110            packlist_find(&to_pack, peeled, NULL))      /* object packed? */
2111                add_object_entry(sha1, OBJ_TAG, NULL, 0);
2112        return 0;
2113}
2114
2115static void prepare_pack(int window, int depth)
2116{
2117        struct object_entry **delta_list;
2118        uint32_t i, nr_deltas;
2119        unsigned n;
2120
2121        get_object_details();
2122
2123        /*
2124         * If we're locally repacking then we need to be doubly careful
2125         * from now on in order to make sure no stealth corruption gets
2126         * propagated to the new pack.  Clients receiving streamed packs
2127         * should validate everything they get anyway so no need to incur
2128         * the additional cost here in that case.
2129         */
2130        if (!pack_to_stdout)
2131                do_check_packed_object_crc = 1;
2132
2133        if (!to_pack.nr_objects || !window || !depth)
2134                return;
2135
2136        delta_list = xmalloc(to_pack.nr_objects * sizeof(*delta_list));
2137        nr_deltas = n = 0;
2138
2139        for (i = 0; i < to_pack.nr_objects; i++) {
2140                struct object_entry *entry = to_pack.objects + i;
2141
2142                if (entry->delta)
2143                        /* This happens if we decided to reuse existing
2144                         * delta from a pack.  "reuse_delta &&" is implied.
2145                         */
2146                        continue;
2147
2148                if (entry->size < 50)
2149                        continue;
2150
2151                if (entry->no_try_delta)
2152                        continue;
2153
2154                if (!entry->preferred_base) {
2155                        nr_deltas++;
2156                        if (entry->type < 0)
2157                                die("unable to get type of object %s",
2158                                    sha1_to_hex(entry->idx.sha1));
2159                } else {
2160                        if (entry->type < 0) {
2161                                /*
2162                                 * This object is not found, but we
2163                                 * don't have to include it anyway.
2164                                 */
2165                                continue;
2166                        }
2167                }
2168
2169                delta_list[n++] = entry;
2170        }
2171
2172        if (nr_deltas && n > 1) {
2173                unsigned nr_done = 0;
2174                if (progress)
2175                        progress_state = start_progress(_("Compressing objects"),
2176                                                        nr_deltas);
2177                qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2178                ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2179                stop_progress(&progress_state);
2180                if (nr_done != nr_deltas)
2181                        die("inconsistency with delta count");
2182        }
2183        free(delta_list);
2184}
2185
2186static int git_pack_config(const char *k, const char *v, void *cb)
2187{
2188        if (!strcmp(k, "pack.window")) {
2189                window = git_config_int(k, v);
2190                return 0;
2191        }
2192        if (!strcmp(k, "pack.windowmemory")) {
2193                window_memory_limit = git_config_ulong(k, v);
2194                return 0;
2195        }
2196        if (!strcmp(k, "pack.depth")) {
2197                depth = git_config_int(k, v);
2198                return 0;
2199        }
2200        if (!strcmp(k, "pack.compression")) {
2201                int level = git_config_int(k, v);
2202                if (level == -1)
2203                        level = Z_DEFAULT_COMPRESSION;
2204                else if (level < 0 || level > Z_BEST_COMPRESSION)
2205                        die("bad pack compression level %d", level);
2206                pack_compression_level = level;
2207                pack_compression_seen = 1;
2208                return 0;
2209        }
2210        if (!strcmp(k, "pack.deltacachesize")) {
2211                max_delta_cache_size = git_config_int(k, v);
2212                return 0;
2213        }
2214        if (!strcmp(k, "pack.deltacachelimit")) {
2215                cache_max_small_delta_size = git_config_int(k, v);
2216                return 0;
2217        }
2218        if (!strcmp(k, "pack.writebitmaphashcache")) {
2219                if (git_config_bool(k, v))
2220                        write_bitmap_options |= BITMAP_OPT_HASH_CACHE;
2221                else
2222                        write_bitmap_options &= ~BITMAP_OPT_HASH_CACHE;
2223        }
2224        if (!strcmp(k, "pack.usebitmaps")) {
2225                use_bitmap_index = git_config_bool(k, v);
2226                return 0;
2227        }
2228        if (!strcmp(k, "pack.threads")) {
2229                delta_search_threads = git_config_int(k, v);
2230                if (delta_search_threads < 0)
2231                        die("invalid number of threads specified (%d)",
2232                            delta_search_threads);
2233#ifdef NO_PTHREADS
2234                if (delta_search_threads != 1)
2235                        warning("no threads support, ignoring %s", k);
2236#endif
2237                return 0;
2238        }
2239        if (!strcmp(k, "pack.indexversion")) {
2240                pack_idx_opts.version = git_config_int(k, v);
2241                if (pack_idx_opts.version > 2)
2242                        die("bad pack.indexversion=%"PRIu32,
2243                            pack_idx_opts.version);
2244                return 0;
2245        }
2246        return git_default_config(k, v, cb);
2247}
2248
2249static void read_object_list_from_stdin(void)
2250{
2251        char line[40 + 1 + PATH_MAX + 2];
2252        unsigned char sha1[20];
2253
2254        for (;;) {
2255                if (!fgets(line, sizeof(line), stdin)) {
2256                        if (feof(stdin))
2257                                break;
2258                        if (!ferror(stdin))
2259                                die("fgets returned NULL, not EOF, not error!");
2260                        if (errno != EINTR)
2261                                die_errno("fgets");
2262                        clearerr(stdin);
2263                        continue;
2264                }
2265                if (line[0] == '-') {
2266                        if (get_sha1_hex(line+1, sha1))
2267                                die("expected edge sha1, got garbage:\n %s",
2268                                    line);
2269                        add_preferred_base(sha1);
2270                        continue;
2271                }
2272                if (get_sha1_hex(line, sha1))
2273                        die("expected sha1, got garbage:\n %s", line);
2274
2275                add_preferred_base_object(line+41);
2276                add_object_entry(sha1, 0, line+41, 0);
2277        }
2278}
2279
2280#define OBJECT_ADDED (1u<<20)
2281
2282static void show_commit(struct commit *commit, void *data)
2283{
2284        add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
2285        commit->object.flags |= OBJECT_ADDED;
2286
2287        if (write_bitmap_index)
2288                index_commit_for_bitmap(commit);
2289}
2290
2291static void show_object(struct object *obj,
2292                        const struct name_path *path, const char *last,
2293                        void *data)
2294{
2295        char *name = path_name(path, last);
2296
2297        add_preferred_base_object(name);
2298        add_object_entry(obj->sha1, obj->type, name, 0);
2299        obj->flags |= OBJECT_ADDED;
2300
2301        /*
2302         * We will have generated the hash from the name,
2303         * but not saved a pointer to it - we can free it
2304         */
2305        free((char *)name);
2306}
2307
2308static void show_edge(struct commit *commit)
2309{
2310        add_preferred_base(commit->object.sha1);
2311}
2312
2313struct in_pack_object {
2314        off_t offset;
2315        struct object *object;
2316};
2317
2318struct in_pack {
2319        int alloc;
2320        int nr;
2321        struct in_pack_object *array;
2322};
2323
2324static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2325{
2326        in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
2327        in_pack->array[in_pack->nr].object = object;
2328        in_pack->nr++;
2329}
2330
2331/*
2332 * Compare the objects in the offset order, in order to emulate the
2333 * "git rev-list --objects" output that produced the pack originally.
2334 */
2335static int ofscmp(const void *a_, const void *b_)
2336{
2337        struct in_pack_object *a = (struct in_pack_object *)a_;
2338        struct in_pack_object *b = (struct in_pack_object *)b_;
2339
2340        if (a->offset < b->offset)
2341                return -1;
2342        else if (a->offset > b->offset)
2343                return 1;
2344        else
2345                return hashcmp(a->object->sha1, b->object->sha1);
2346}
2347
2348static void add_objects_in_unpacked_packs(struct rev_info *revs)
2349{
2350        struct packed_git *p;
2351        struct in_pack in_pack;
2352        uint32_t i;
2353
2354        memset(&in_pack, 0, sizeof(in_pack));
2355
2356        for (p = packed_git; p; p = p->next) {
2357                const unsigned char *sha1;
2358                struct object *o;
2359
2360                if (!p->pack_local || p->pack_keep)
2361                        continue;
2362                if (open_pack_index(p))
2363                        die("cannot open pack index");
2364
2365                ALLOC_GROW(in_pack.array,
2366                           in_pack.nr + p->num_objects,
2367                           in_pack.alloc);
2368
2369                for (i = 0; i < p->num_objects; i++) {
2370                        sha1 = nth_packed_object_sha1(p, i);
2371                        o = lookup_unknown_object(sha1);
2372                        if (!(o->flags & OBJECT_ADDED))
2373                                mark_in_pack_object(o, p, &in_pack);
2374                        o->flags |= OBJECT_ADDED;
2375                }
2376        }
2377
2378        if (in_pack.nr) {
2379                qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2380                      ofscmp);
2381                for (i = 0; i < in_pack.nr; i++) {
2382                        struct object *o = in_pack.array[i].object;
2383                        add_object_entry(o->sha1, o->type, "", 0);
2384                }
2385        }
2386        free(in_pack.array);
2387}
2388
2389static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2390{
2391        static struct packed_git *last_found = (void *)1;
2392        struct packed_git *p;
2393
2394        p = (last_found != (void *)1) ? last_found : packed_git;
2395
2396        while (p) {
2397                if ((!p->pack_local || p->pack_keep) &&
2398                        find_pack_entry_one(sha1, p)) {
2399                        last_found = p;
2400                        return 1;
2401                }
2402                if (p == last_found)
2403                        p = packed_git;
2404                else
2405                        p = p->next;
2406                if (p == last_found)
2407                        p = p->next;
2408        }
2409        return 0;
2410}
2411
2412/*
2413 * Store a list of sha1s that are should not be discarded
2414 * because they are either written too recently, or are
2415 * reachable from another object that was.
2416 *
2417 * This is filled by get_object_list.
2418 */
2419static struct sha1_array recent_objects;
2420
2421static int loosened_object_can_be_discarded(const unsigned char *sha1,
2422                                            unsigned long mtime)
2423{
2424        if (!unpack_unreachable_expiration)
2425                return 0;
2426        if (mtime > unpack_unreachable_expiration)
2427                return 0;
2428        if (sha1_array_lookup(&recent_objects, sha1) >= 0)
2429                return 0;
2430        return 1;
2431}
2432
2433static void loosen_unused_packed_objects(struct rev_info *revs)
2434{
2435        struct packed_git *p;
2436        uint32_t i;
2437        const unsigned char *sha1;
2438
2439        for (p = packed_git; p; p = p->next) {
2440                if (!p->pack_local || p->pack_keep)
2441                        continue;
2442
2443                if (open_pack_index(p))
2444                        die("cannot open pack index");
2445
2446                for (i = 0; i < p->num_objects; i++) {
2447                        sha1 = nth_packed_object_sha1(p, i);
2448                        if (!packlist_find(&to_pack, sha1, NULL) &&
2449                            !has_sha1_pack_kept_or_nonlocal(sha1) &&
2450                            !loosened_object_can_be_discarded(sha1, p->mtime))
2451                                if (force_object_loose(sha1, p->mtime))
2452                                        die("unable to force loose object");
2453                }
2454        }
2455}
2456
2457/*
2458 * This tracks any options which a reader of the pack might
2459 * not understand, and which would therefore prevent blind reuse
2460 * of what we have on disk.
2461 */
2462static int pack_options_allow_reuse(void)
2463{
2464        return allow_ofs_delta;
2465}
2466
2467static int get_object_list_from_bitmap(struct rev_info *revs)
2468{
2469        if (prepare_bitmap_walk(revs) < 0)
2470                return -1;
2471
2472        if (pack_options_allow_reuse() &&
2473            !reuse_partial_packfile_from_bitmap(
2474                        &reuse_packfile,
2475                        &reuse_packfile_objects,
2476                        &reuse_packfile_offset)) {
2477                assert(reuse_packfile_objects);
2478                nr_result += reuse_packfile_objects;
2479                display_progress(progress_state, nr_result);
2480        }
2481
2482        traverse_bitmap_commit_list(&add_object_entry_from_bitmap);
2483        return 0;
2484}
2485
2486static void record_recent_object(struct object *obj,
2487                                 const struct name_path *path,
2488                                 const char *last,
2489                                 void *data)
2490{
2491        sha1_array_append(&recent_objects, obj->sha1);
2492}
2493
2494static void record_recent_commit(struct commit *commit, void *data)
2495{
2496        sha1_array_append(&recent_objects, commit->object.sha1);
2497}
2498
2499static void get_object_list(int ac, const char **av)
2500{
2501        struct rev_info revs;
2502        char line[1000];
2503        int flags = 0;
2504
2505        init_revisions(&revs, NULL);
2506        save_commit_buffer = 0;
2507        setup_revisions(ac, av, &revs, NULL);
2508
2509        /* make sure shallows are read */
2510        is_repository_shallow();
2511
2512        while (fgets(line, sizeof(line), stdin) != NULL) {
2513                int len = strlen(line);
2514                if (len && line[len - 1] == '\n')
2515                        line[--len] = 0;
2516                if (!len)
2517                        break;
2518                if (*line == '-') {
2519                        if (!strcmp(line, "--not")) {
2520                                flags ^= UNINTERESTING;
2521                                write_bitmap_index = 0;
2522                                continue;
2523                        }
2524                        if (starts_with(line, "--shallow ")) {
2525                                unsigned char sha1[20];
2526                                if (get_sha1_hex(line + 10, sha1))
2527                                        die("not an SHA-1 '%s'", line + 10);
2528                                register_shallow(sha1);
2529                                use_bitmap_index = 0;
2530                                continue;
2531                        }
2532                        die("not a rev '%s'", line);
2533                }
2534                if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2535                        die("bad revision '%s'", line);
2536        }
2537
2538        if (use_bitmap_index && !get_object_list_from_bitmap(&revs))
2539                return;
2540
2541        if (prepare_revision_walk(&revs))
2542                die("revision walk setup failed");
2543        mark_edges_uninteresting(&revs, show_edge);
2544        traverse_commit_list(&revs, show_commit, show_object, NULL);
2545
2546        if (unpack_unreachable_expiration) {
2547                revs.ignore_missing_links = 1;
2548                if (add_unseen_recent_objects_to_traversal(&revs,
2549                                unpack_unreachable_expiration))
2550                        die("unable to add recent objects");
2551                if (prepare_revision_walk(&revs))
2552                        die("revision walk setup failed");
2553                traverse_commit_list(&revs, record_recent_commit,
2554                                     record_recent_object, NULL);
2555        }
2556
2557        if (keep_unreachable)
2558                add_objects_in_unpacked_packs(&revs);
2559        if (unpack_unreachable)
2560                loosen_unused_packed_objects(&revs);
2561
2562        sha1_array_clear(&recent_objects);
2563}
2564
2565static int option_parse_index_version(const struct option *opt,
2566                                      const char *arg, int unset)
2567{
2568        char *c;
2569        const char *val = arg;
2570        pack_idx_opts.version = strtoul(val, &c, 10);
2571        if (pack_idx_opts.version > 2)
2572                die(_("unsupported index version %s"), val);
2573        if (*c == ',' && c[1])
2574                pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2575        if (*c || pack_idx_opts.off32_limit & 0x80000000)
2576                die(_("bad index version '%s'"), val);
2577        return 0;
2578}
2579
2580static int option_parse_unpack_unreachable(const struct option *opt,
2581                                           const char *arg, int unset)
2582{
2583        if (unset) {
2584                unpack_unreachable = 0;
2585                unpack_unreachable_expiration = 0;
2586        }
2587        else {
2588                unpack_unreachable = 1;
2589                if (arg)
2590                        unpack_unreachable_expiration = approxidate(arg);
2591        }
2592        return 0;
2593}
2594
2595static int option_parse_ulong(const struct option *opt,
2596                              const char *arg, int unset)
2597{
2598        if (unset)
2599                die(_("option %s does not accept negative form"),
2600                    opt->long_name);
2601
2602        if (!git_parse_ulong(arg, opt->value))
2603                die(_("unable to parse value '%s' for option %s"),
2604                    arg, opt->long_name);
2605        return 0;
2606}
2607
2608#define OPT_ULONG(s, l, v, h) \
2609        { OPTION_CALLBACK, (s), (l), (v), "n", (h),     \
2610          PARSE_OPT_NONEG, option_parse_ulong }
2611
2612int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2613{
2614        int use_internal_rev_list = 0;
2615        int thin = 0;
2616        int all_progress_implied = 0;
2617        const char *rp_av[6];
2618        int rp_ac = 0;
2619        int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2620        struct option pack_objects_options[] = {
2621                OPT_SET_INT('q', "quiet", &progress,
2622                            N_("do not show progress meter"), 0),
2623                OPT_SET_INT(0, "progress", &progress,
2624                            N_("show progress meter"), 1),
2625                OPT_SET_INT(0, "all-progress", &progress,
2626                            N_("show progress meter during object writing phase"), 2),
2627                OPT_BOOL(0, "all-progress-implied",
2628                         &all_progress_implied,
2629                         N_("similar to --all-progress when progress meter is shown")),
2630                { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2631                  N_("write the pack index file in the specified idx format version"),
2632                  0, option_parse_index_version },
2633                OPT_ULONG(0, "max-pack-size", &pack_size_limit,
2634                          N_("maximum size of each output pack file")),
2635                OPT_BOOL(0, "local", &local,
2636                         N_("ignore borrowed objects from alternate object store")),
2637                OPT_BOOL(0, "incremental", &incremental,
2638                         N_("ignore packed objects")),
2639                OPT_INTEGER(0, "window", &window,
2640                            N_("limit pack window by objects")),
2641                OPT_ULONG(0, "window-memory", &window_memory_limit,
2642                          N_("limit pack window by memory in addition to object limit")),
2643                OPT_INTEGER(0, "depth", &depth,
2644                            N_("maximum length of delta chain allowed in the resulting pack")),
2645                OPT_BOOL(0, "reuse-delta", &reuse_delta,
2646                         N_("reuse existing deltas")),
2647                OPT_BOOL(0, "reuse-object", &reuse_object,
2648                         N_("reuse existing objects")),
2649                OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2650                         N_("use OFS_DELTA objects")),
2651                OPT_INTEGER(0, "threads", &delta_search_threads,
2652                            N_("use threads when searching for best delta matches")),
2653                OPT_BOOL(0, "non-empty", &non_empty,
2654                         N_("do not create an empty pack output")),
2655                OPT_BOOL(0, "revs", &use_internal_rev_list,
2656                         N_("read revision arguments from standard input")),
2657                { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2658                  N_("limit the objects to those that are not yet packed"),
2659                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2660                { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2661                  N_("include objects reachable from any reference"),
2662                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2663                { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2664                  N_("include objects referred by reflog entries"),
2665                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2666                OPT_BOOL(0, "stdout", &pack_to_stdout,
2667                         N_("output pack to stdout")),
2668                OPT_BOOL(0, "include-tag", &include_tag,
2669                         N_("include tag objects that refer to objects to be packed")),
2670                OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2671                         N_("keep unreachable objects")),
2672                { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2673                  N_("unpack unreachable objects newer than <time>"),
2674                  PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2675                OPT_BOOL(0, "thin", &thin,
2676                         N_("create thin packs")),
2677                OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2678                         N_("ignore packs that have companion .keep file")),
2679                OPT_INTEGER(0, "compression", &pack_compression_level,
2680                            N_("pack compression level")),
2681                OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2682                            N_("do not hide commits by grafts"), 0),
2683                OPT_BOOL(0, "use-bitmap-index", &use_bitmap_index,
2684                         N_("use a bitmap index if available to speed up counting objects")),
2685                OPT_BOOL(0, "write-bitmap-index", &write_bitmap_index,
2686                         N_("write a bitmap index together with the pack index")),
2687                OPT_END(),
2688        };
2689
2690        check_replace_refs = 0;
2691
2692        reset_pack_idx_option(&pack_idx_opts);
2693        git_config(git_pack_config, NULL);
2694        if (!pack_compression_seen && core_compression_seen)
2695                pack_compression_level = core_compression_level;
2696
2697        progress = isatty(2);
2698        argc = parse_options(argc, argv, prefix, pack_objects_options,
2699                             pack_usage, 0);
2700
2701        if (argc) {
2702                base_name = argv[0];
2703                argc--;
2704        }
2705        if (pack_to_stdout != !base_name || argc)
2706                usage_with_options(pack_usage, pack_objects_options);
2707
2708        rp_av[rp_ac++] = "pack-objects";
2709        if (thin) {
2710                use_internal_rev_list = 1;
2711                rp_av[rp_ac++] = "--objects-edge";
2712        } else
2713                rp_av[rp_ac++] = "--objects";
2714
2715        if (rev_list_all) {
2716                use_internal_rev_list = 1;
2717                rp_av[rp_ac++] = "--all";
2718        }
2719        if (rev_list_reflog) {
2720                use_internal_rev_list = 1;
2721                rp_av[rp_ac++] = "--reflog";
2722        }
2723        if (rev_list_unpacked) {
2724                use_internal_rev_list = 1;
2725                rp_av[rp_ac++] = "--unpacked";
2726        }
2727
2728        if (!reuse_object)
2729                reuse_delta = 0;
2730        if (pack_compression_level == -1)
2731                pack_compression_level = Z_DEFAULT_COMPRESSION;
2732        else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2733                die("bad pack compression level %d", pack_compression_level);
2734#ifdef NO_PTHREADS
2735        if (delta_search_threads != 1)
2736                warning("no threads support, ignoring --threads");
2737#endif
2738        if (!pack_to_stdout && !pack_size_limit)
2739                pack_size_limit = pack_size_limit_cfg;
2740        if (pack_to_stdout && pack_size_limit)
2741                die("--max-pack-size cannot be used to build a pack for transfer.");
2742        if (pack_size_limit && pack_size_limit < 1024*1024) {
2743                warning("minimum pack size limit is 1 MiB");
2744                pack_size_limit = 1024*1024;
2745        }
2746
2747        if (!pack_to_stdout && thin)
2748                die("--thin cannot be used to build an indexable pack.");
2749
2750        if (keep_unreachable && unpack_unreachable)
2751                die("--keep-unreachable and --unpack-unreachable are incompatible.");
2752
2753        if (!use_internal_rev_list || !pack_to_stdout || is_repository_shallow())
2754                use_bitmap_index = 0;
2755
2756        if (pack_to_stdout || !rev_list_all)
2757                write_bitmap_index = 0;
2758
2759        if (progress && all_progress_implied)
2760                progress = 2;
2761
2762        prepare_packed_git();
2763
2764        if (progress)
2765                progress_state = start_progress(_("Counting objects"), 0);
2766        if (!use_internal_rev_list)
2767                read_object_list_from_stdin();
2768        else {
2769                rp_av[rp_ac] = NULL;
2770                get_object_list(rp_ac, rp_av);
2771        }
2772        cleanup_preferred_base();
2773        if (include_tag && nr_result)
2774                for_each_ref(add_ref_tag, NULL);
2775        stop_progress(&progress_state);
2776
2777        if (non_empty && !nr_result)
2778                return 0;
2779        if (nr_result)
2780                prepare_pack(window, depth);
2781        write_pack_file();
2782        if (progress)
2783                fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2784                        " reused %"PRIu32" (delta %"PRIu32")\n",
2785                        written, written_delta, reused, reused_delta);
2786        return 0;
2787}