builtin / pack-objects.con commit pack-objects: split add_object_entry (ce2bc42)
   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
  23static const char *pack_usage[] = {
  24        N_("git pack-objects --stdout [options...] [< ref-list | < object-list]"),
  25        N_("git pack-objects [options...] base-name [< ref-list | < object-list]"),
  26        NULL
  27};
  28
  29/*
  30 * Objects we are going to pack are collected in the `to_pack` structure.
  31 * It contains an array (dynamically expanded) of the object data, and a map
  32 * that can resolve SHA1s to their position in the array.
  33 */
  34static struct packing_data to_pack;
  35
  36static struct pack_idx_entry **written_list;
  37static uint32_t nr_result, nr_written;
  38
  39static int non_empty;
  40static int reuse_delta = 1, reuse_object = 1;
  41static int keep_unreachable, unpack_unreachable, include_tag;
  42static unsigned long unpack_unreachable_expiration;
  43static int local;
  44static int incremental;
  45static int ignore_packed_keep;
  46static int allow_ofs_delta;
  47static struct pack_idx_option pack_idx_opts;
  48static const char *base_name;
  49static int progress = 1;
  50static int window = 10;
  51static unsigned long pack_size_limit;
  52static int depth = 50;
  53static int delta_search_threads;
  54static int pack_to_stdout;
  55static int num_preferred_base;
  56static struct progress *progress_state;
  57static int pack_compression_level = Z_DEFAULT_COMPRESSION;
  58static int pack_compression_seen;
  59
  60static unsigned long delta_cache_size = 0;
  61static unsigned long max_delta_cache_size = 256 * 1024 * 1024;
  62static unsigned long cache_max_small_delta_size = 1000;
  63
  64static unsigned long window_memory_limit = 0;
  65
  66/*
  67 * stats
  68 */
  69static uint32_t written, written_delta;
  70static uint32_t reused, reused_delta;
  71
  72static void *get_delta(struct object_entry *entry)
  73{
  74        unsigned long size, base_size, delta_size;
  75        void *buf, *base_buf, *delta_buf;
  76        enum object_type type;
  77
  78        buf = read_sha1_file(entry->idx.sha1, &type, &size);
  79        if (!buf)
  80                die("unable to read %s", sha1_to_hex(entry->idx.sha1));
  81        base_buf = read_sha1_file(entry->delta->idx.sha1, &type, &base_size);
  82        if (!base_buf)
  83                die("unable to read %s", sha1_to_hex(entry->delta->idx.sha1));
  84        delta_buf = diff_delta(base_buf, base_size,
  85                               buf, size, &delta_size, 0);
  86        if (!delta_buf || delta_size != entry->delta_size)
  87                die("delta size changed");
  88        free(buf);
  89        free(base_buf);
  90        return delta_buf;
  91}
  92
  93static unsigned long do_compress(void **pptr, unsigned long size)
  94{
  95        git_zstream stream;
  96        void *in, *out;
  97        unsigned long maxsize;
  98
  99        memset(&stream, 0, sizeof(stream));
 100        git_deflate_init(&stream, pack_compression_level);
 101        maxsize = git_deflate_bound(&stream, size);
 102
 103        in = *pptr;
 104        out = xmalloc(maxsize);
 105        *pptr = out;
 106
 107        stream.next_in = in;
 108        stream.avail_in = size;
 109        stream.next_out = out;
 110        stream.avail_out = maxsize;
 111        while (git_deflate(&stream, Z_FINISH) == Z_OK)
 112                ; /* nothing */
 113        git_deflate_end(&stream);
 114
 115        free(in);
 116        return stream.total_out;
 117}
 118
 119static unsigned long write_large_blob_data(struct git_istream *st, struct sha1file *f,
 120                                           const unsigned char *sha1)
 121{
 122        git_zstream stream;
 123        unsigned char ibuf[1024 * 16];
 124        unsigned char obuf[1024 * 16];
 125        unsigned long olen = 0;
 126
 127        memset(&stream, 0, sizeof(stream));
 128        git_deflate_init(&stream, pack_compression_level);
 129
 130        for (;;) {
 131                ssize_t readlen;
 132                int zret = Z_OK;
 133                readlen = read_istream(st, ibuf, sizeof(ibuf));
 134                if (readlen == -1)
 135                        die(_("unable to read %s"), sha1_to_hex(sha1));
 136
 137                stream.next_in = ibuf;
 138                stream.avail_in = readlen;
 139                while ((stream.avail_in || readlen == 0) &&
 140                       (zret == Z_OK || zret == Z_BUF_ERROR)) {
 141                        stream.next_out = obuf;
 142                        stream.avail_out = sizeof(obuf);
 143                        zret = git_deflate(&stream, readlen ? 0 : Z_FINISH);
 144                        sha1write(f, obuf, stream.next_out - obuf);
 145                        olen += stream.next_out - obuf;
 146                }
 147                if (stream.avail_in)
 148                        die(_("deflate error (%d)"), zret);
 149                if (readlen == 0) {
 150                        if (zret != Z_STREAM_END)
 151                                die(_("deflate error (%d)"), zret);
 152                        break;
 153                }
 154        }
 155        git_deflate_end(&stream);
 156        return olen;
 157}
 158
 159/*
 160 * we are going to reuse the existing object data as is.  make
 161 * sure it is not corrupt.
 162 */
 163static int check_pack_inflate(struct packed_git *p,
 164                struct pack_window **w_curs,
 165                off_t offset,
 166                off_t len,
 167                unsigned long expect)
 168{
 169        git_zstream stream;
 170        unsigned char fakebuf[4096], *in;
 171        int st;
 172
 173        memset(&stream, 0, sizeof(stream));
 174        git_inflate_init(&stream);
 175        do {
 176                in = use_pack(p, w_curs, offset, &stream.avail_in);
 177                stream.next_in = in;
 178                stream.next_out = fakebuf;
 179                stream.avail_out = sizeof(fakebuf);
 180                st = git_inflate(&stream, Z_FINISH);
 181                offset += stream.next_in - in;
 182        } while (st == Z_OK || st == Z_BUF_ERROR);
 183        git_inflate_end(&stream);
 184        return (st == Z_STREAM_END &&
 185                stream.total_out == expect &&
 186                stream.total_in == len) ? 0 : -1;
 187}
 188
 189static void copy_pack_data(struct sha1file *f,
 190                struct packed_git *p,
 191                struct pack_window **w_curs,
 192                off_t offset,
 193                off_t len)
 194{
 195        unsigned char *in;
 196        unsigned long avail;
 197
 198        while (len) {
 199                in = use_pack(p, w_curs, offset, &avail);
 200                if (avail > len)
 201                        avail = (unsigned long)len;
 202                sha1write(f, in, avail);
 203                offset += avail;
 204                len -= avail;
 205        }
 206}
 207
 208/* Return 0 if we will bust the pack-size limit */
 209static unsigned long write_no_reuse_object(struct sha1file *f, struct object_entry *entry,
 210                                           unsigned long limit, int usable_delta)
 211{
 212        unsigned long size, datalen;
 213        unsigned char header[10], dheader[10];
 214        unsigned hdrlen;
 215        enum object_type type;
 216        void *buf;
 217        struct git_istream *st = NULL;
 218
 219        if (!usable_delta) {
 220                if (entry->type == OBJ_BLOB &&
 221                    entry->size > big_file_threshold &&
 222                    (st = open_istream(entry->idx.sha1, &type, &size, NULL)) != NULL)
 223                        buf = NULL;
 224                else {
 225                        buf = read_sha1_file(entry->idx.sha1, &type, &size);
 226                        if (!buf)
 227                                die(_("unable to read %s"), sha1_to_hex(entry->idx.sha1));
 228                }
 229                /*
 230                 * make sure no cached delta data remains from a
 231                 * previous attempt before a pack split occurred.
 232                 */
 233                free(entry->delta_data);
 234                entry->delta_data = NULL;
 235                entry->z_delta_size = 0;
 236        } else if (entry->delta_data) {
 237                size = entry->delta_size;
 238                buf = entry->delta_data;
 239                entry->delta_data = NULL;
 240                type = (allow_ofs_delta && entry->delta->idx.offset) ?
 241                        OBJ_OFS_DELTA : OBJ_REF_DELTA;
 242        } else {
 243                buf = get_delta(entry);
 244                size = entry->delta_size;
 245                type = (allow_ofs_delta && entry->delta->idx.offset) ?
 246                        OBJ_OFS_DELTA : OBJ_REF_DELTA;
 247        }
 248
 249        if (st) /* large blob case, just assume we don't compress well */
 250                datalen = size;
 251        else if (entry->z_delta_size)
 252                datalen = entry->z_delta_size;
 253        else
 254                datalen = do_compress(&buf, size);
 255
 256        /*
 257         * The object header is a byte of 'type' followed by zero or
 258         * more bytes of length.
 259         */
 260        hdrlen = encode_in_pack_object_header(type, size, header);
 261
 262        if (type == OBJ_OFS_DELTA) {
 263                /*
 264                 * Deltas with relative base contain an additional
 265                 * encoding of the relative offset for the delta
 266                 * base from this object's position in the pack.
 267                 */
 268                off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 269                unsigned pos = sizeof(dheader) - 1;
 270                dheader[pos] = ofs & 127;
 271                while (ofs >>= 7)
 272                        dheader[--pos] = 128 | (--ofs & 127);
 273                if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 274                        if (st)
 275                                close_istream(st);
 276                        free(buf);
 277                        return 0;
 278                }
 279                sha1write(f, header, hdrlen);
 280                sha1write(f, dheader + pos, sizeof(dheader) - pos);
 281                hdrlen += sizeof(dheader) - pos;
 282        } else if (type == OBJ_REF_DELTA) {
 283                /*
 284                 * Deltas with a base reference contain
 285                 * an additional 20 bytes for the base sha1.
 286                 */
 287                if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 288                        if (st)
 289                                close_istream(st);
 290                        free(buf);
 291                        return 0;
 292                }
 293                sha1write(f, header, hdrlen);
 294                sha1write(f, entry->delta->idx.sha1, 20);
 295                hdrlen += 20;
 296        } else {
 297                if (limit && hdrlen + datalen + 20 >= limit) {
 298                        if (st)
 299                                close_istream(st);
 300                        free(buf);
 301                        return 0;
 302                }
 303                sha1write(f, header, hdrlen);
 304        }
 305        if (st) {
 306                datalen = write_large_blob_data(st, f, entry->idx.sha1);
 307                close_istream(st);
 308        } else {
 309                sha1write(f, buf, datalen);
 310                free(buf);
 311        }
 312
 313        return hdrlen + datalen;
 314}
 315
 316/* Return 0 if we will bust the pack-size limit */
 317static unsigned long write_reuse_object(struct sha1file *f, struct object_entry *entry,
 318                                        unsigned long limit, int usable_delta)
 319{
 320        struct packed_git *p = entry->in_pack;
 321        struct pack_window *w_curs = NULL;
 322        struct revindex_entry *revidx;
 323        off_t offset;
 324        enum object_type type = entry->type;
 325        unsigned long datalen;
 326        unsigned char header[10], dheader[10];
 327        unsigned hdrlen;
 328
 329        if (entry->delta)
 330                type = (allow_ofs_delta && entry->delta->idx.offset) ?
 331                        OBJ_OFS_DELTA : OBJ_REF_DELTA;
 332        hdrlen = encode_in_pack_object_header(type, entry->size, header);
 333
 334        offset = entry->in_pack_offset;
 335        revidx = find_pack_revindex(p, offset);
 336        datalen = revidx[1].offset - offset;
 337        if (!pack_to_stdout && p->index_version > 1 &&
 338            check_pack_crc(p, &w_curs, offset, datalen, revidx->nr)) {
 339                error("bad packed object CRC for %s", sha1_to_hex(entry->idx.sha1));
 340                unuse_pack(&w_curs);
 341                return write_no_reuse_object(f, entry, limit, usable_delta);
 342        }
 343
 344        offset += entry->in_pack_header_size;
 345        datalen -= entry->in_pack_header_size;
 346
 347        if (!pack_to_stdout && p->index_version == 1 &&
 348            check_pack_inflate(p, &w_curs, offset, datalen, entry->size)) {
 349                error("corrupt packed object for %s", sha1_to_hex(entry->idx.sha1));
 350                unuse_pack(&w_curs);
 351                return write_no_reuse_object(f, entry, limit, usable_delta);
 352        }
 353
 354        if (type == OBJ_OFS_DELTA) {
 355                off_t ofs = entry->idx.offset - entry->delta->idx.offset;
 356                unsigned pos = sizeof(dheader) - 1;
 357                dheader[pos] = ofs & 127;
 358                while (ofs >>= 7)
 359                        dheader[--pos] = 128 | (--ofs & 127);
 360                if (limit && hdrlen + sizeof(dheader) - pos + datalen + 20 >= limit) {
 361                        unuse_pack(&w_curs);
 362                        return 0;
 363                }
 364                sha1write(f, header, hdrlen);
 365                sha1write(f, dheader + pos, sizeof(dheader) - pos);
 366                hdrlen += sizeof(dheader) - pos;
 367                reused_delta++;
 368        } else if (type == OBJ_REF_DELTA) {
 369                if (limit && hdrlen + 20 + datalen + 20 >= limit) {
 370                        unuse_pack(&w_curs);
 371                        return 0;
 372                }
 373                sha1write(f, header, hdrlen);
 374                sha1write(f, entry->delta->idx.sha1, 20);
 375                hdrlen += 20;
 376                reused_delta++;
 377        } else {
 378                if (limit && hdrlen + datalen + 20 >= limit) {
 379                        unuse_pack(&w_curs);
 380                        return 0;
 381                }
 382                sha1write(f, header, hdrlen);
 383        }
 384        copy_pack_data(f, p, &w_curs, offset, datalen);
 385        unuse_pack(&w_curs);
 386        reused++;
 387        return hdrlen + datalen;
 388}
 389
 390/* Return 0 if we will bust the pack-size limit */
 391static unsigned long write_object(struct sha1file *f,
 392                                  struct object_entry *entry,
 393                                  off_t write_offset)
 394{
 395        unsigned long limit, len;
 396        int usable_delta, to_reuse;
 397
 398        if (!pack_to_stdout)
 399                crc32_begin(f);
 400
 401        /* apply size limit if limited packsize and not first object */
 402        if (!pack_size_limit || !nr_written)
 403                limit = 0;
 404        else if (pack_size_limit <= write_offset)
 405                /*
 406                 * the earlier object did not fit the limit; avoid
 407                 * mistaking this with unlimited (i.e. limit = 0).
 408                 */
 409                limit = 1;
 410        else
 411                limit = pack_size_limit - write_offset;
 412
 413        if (!entry->delta)
 414                usable_delta = 0;       /* no delta */
 415        else if (!pack_size_limit)
 416               usable_delta = 1;        /* unlimited packfile */
 417        else if (entry->delta->idx.offset == (off_t)-1)
 418                usable_delta = 0;       /* base was written to another pack */
 419        else if (entry->delta->idx.offset)
 420                usable_delta = 1;       /* base already exists in this pack */
 421        else
 422                usable_delta = 0;       /* base could end up in another pack */
 423
 424        if (!reuse_object)
 425                to_reuse = 0;   /* explicit */
 426        else if (!entry->in_pack)
 427                to_reuse = 0;   /* can't reuse what we don't have */
 428        else if (entry->type == OBJ_REF_DELTA || entry->type == OBJ_OFS_DELTA)
 429                                /* check_object() decided it for us ... */
 430                to_reuse = usable_delta;
 431                                /* ... but pack split may override that */
 432        else if (entry->type != entry->in_pack_type)
 433                to_reuse = 0;   /* pack has delta which is unusable */
 434        else if (entry->delta)
 435                to_reuse = 0;   /* we want to pack afresh */
 436        else
 437                to_reuse = 1;   /* we have it in-pack undeltified,
 438                                 * and we do not need to deltify it.
 439                                 */
 440
 441        if (!to_reuse)
 442                len = write_no_reuse_object(f, entry, limit, usable_delta);
 443        else
 444                len = write_reuse_object(f, entry, limit, usable_delta);
 445        if (!len)
 446                return 0;
 447
 448        if (usable_delta)
 449                written_delta++;
 450        written++;
 451        if (!pack_to_stdout)
 452                entry->idx.crc32 = crc32_end(f);
 453        return len;
 454}
 455
 456enum write_one_status {
 457        WRITE_ONE_SKIP = -1, /* already written */
 458        WRITE_ONE_BREAK = 0, /* writing this will bust the limit; not written */
 459        WRITE_ONE_WRITTEN = 1, /* normal */
 460        WRITE_ONE_RECURSIVE = 2 /* already scheduled to be written */
 461};
 462
 463static enum write_one_status write_one(struct sha1file *f,
 464                                       struct object_entry *e,
 465                                       off_t *offset)
 466{
 467        unsigned long size;
 468        int recursing;
 469
 470        /*
 471         * we set offset to 1 (which is an impossible value) to mark
 472         * the fact that this object is involved in "write its base
 473         * first before writing a deltified object" recursion.
 474         */
 475        recursing = (e->idx.offset == 1);
 476        if (recursing) {
 477                warning("recursive delta detected for object %s",
 478                        sha1_to_hex(e->idx.sha1));
 479                return WRITE_ONE_RECURSIVE;
 480        } else if (e->idx.offset || e->preferred_base) {
 481                /* offset is non zero if object is written already. */
 482                return WRITE_ONE_SKIP;
 483        }
 484
 485        /* if we are deltified, write out base object first. */
 486        if (e->delta) {
 487                e->idx.offset = 1; /* now recurse */
 488                switch (write_one(f, e->delta, offset)) {
 489                case WRITE_ONE_RECURSIVE:
 490                        /* we cannot depend on this one */
 491                        e->delta = NULL;
 492                        break;
 493                default:
 494                        break;
 495                case WRITE_ONE_BREAK:
 496                        e->idx.offset = recursing;
 497                        return WRITE_ONE_BREAK;
 498                }
 499        }
 500
 501        e->idx.offset = *offset;
 502        size = write_object(f, e, *offset);
 503        if (!size) {
 504                e->idx.offset = recursing;
 505                return WRITE_ONE_BREAK;
 506        }
 507        written_list[nr_written++] = &e->idx;
 508
 509        /* make sure off_t is sufficiently large not to wrap */
 510        if (signed_add_overflows(*offset, size))
 511                die("pack too large for current definition of off_t");
 512        *offset += size;
 513        return WRITE_ONE_WRITTEN;
 514}
 515
 516static int mark_tagged(const char *path, const unsigned char *sha1, int flag,
 517                       void *cb_data)
 518{
 519        unsigned char peeled[20];
 520        struct object_entry *entry = packlist_find(&to_pack, sha1, NULL);
 521
 522        if (entry)
 523                entry->tagged = 1;
 524        if (!peel_ref(path, peeled)) {
 525                entry = packlist_find(&to_pack, peeled, NULL);
 526                if (entry)
 527                        entry->tagged = 1;
 528        }
 529        return 0;
 530}
 531
 532static inline void add_to_write_order(struct object_entry **wo,
 533                               unsigned int *endp,
 534                               struct object_entry *e)
 535{
 536        if (e->filled)
 537                return;
 538        wo[(*endp)++] = e;
 539        e->filled = 1;
 540}
 541
 542static void add_descendants_to_write_order(struct object_entry **wo,
 543                                           unsigned int *endp,
 544                                           struct object_entry *e)
 545{
 546        int add_to_order = 1;
 547        while (e) {
 548                if (add_to_order) {
 549                        struct object_entry *s;
 550                        /* add this node... */
 551                        add_to_write_order(wo, endp, e);
 552                        /* all its siblings... */
 553                        for (s = e->delta_sibling; s; s = s->delta_sibling) {
 554                                add_to_write_order(wo, endp, s);
 555                        }
 556                }
 557                /* drop down a level to add left subtree nodes if possible */
 558                if (e->delta_child) {
 559                        add_to_order = 1;
 560                        e = e->delta_child;
 561                } else {
 562                        add_to_order = 0;
 563                        /* our sibling might have some children, it is next */
 564                        if (e->delta_sibling) {
 565                                e = e->delta_sibling;
 566                                continue;
 567                        }
 568                        /* go back to our parent node */
 569                        e = e->delta;
 570                        while (e && !e->delta_sibling) {
 571                                /* we're on the right side of a subtree, keep
 572                                 * going up until we can go right again */
 573                                e = e->delta;
 574                        }
 575                        if (!e) {
 576                                /* done- we hit our original root node */
 577                                return;
 578                        }
 579                        /* pass it off to sibling at this level */
 580                        e = e->delta_sibling;
 581                }
 582        };
 583}
 584
 585static void add_family_to_write_order(struct object_entry **wo,
 586                                      unsigned int *endp,
 587                                      struct object_entry *e)
 588{
 589        struct object_entry *root;
 590
 591        for (root = e; root->delta; root = root->delta)
 592                ; /* nothing */
 593        add_descendants_to_write_order(wo, endp, root);
 594}
 595
 596static struct object_entry **compute_write_order(void)
 597{
 598        unsigned int i, wo_end, last_untagged;
 599
 600        struct object_entry **wo = xmalloc(to_pack.nr_objects * sizeof(*wo));
 601        struct object_entry *objects = to_pack.objects;
 602
 603        for (i = 0; i < to_pack.nr_objects; i++) {
 604                objects[i].tagged = 0;
 605                objects[i].filled = 0;
 606                objects[i].delta_child = NULL;
 607                objects[i].delta_sibling = NULL;
 608        }
 609
 610        /*
 611         * Fully connect delta_child/delta_sibling network.
 612         * Make sure delta_sibling is sorted in the original
 613         * recency order.
 614         */
 615        for (i = to_pack.nr_objects; i > 0;) {
 616                struct object_entry *e = &objects[--i];
 617                if (!e->delta)
 618                        continue;
 619                /* Mark me as the first child */
 620                e->delta_sibling = e->delta->delta_child;
 621                e->delta->delta_child = e;
 622        }
 623
 624        /*
 625         * Mark objects that are at the tip of tags.
 626         */
 627        for_each_tag_ref(mark_tagged, NULL);
 628
 629        /*
 630         * Give the objects in the original recency order until
 631         * we see a tagged tip.
 632         */
 633        for (i = wo_end = 0; i < to_pack.nr_objects; i++) {
 634                if (objects[i].tagged)
 635                        break;
 636                add_to_write_order(wo, &wo_end, &objects[i]);
 637        }
 638        last_untagged = i;
 639
 640        /*
 641         * Then fill all the tagged tips.
 642         */
 643        for (; i < to_pack.nr_objects; i++) {
 644                if (objects[i].tagged)
 645                        add_to_write_order(wo, &wo_end, &objects[i]);
 646        }
 647
 648        /*
 649         * And then all remaining commits and tags.
 650         */
 651        for (i = last_untagged; i < to_pack.nr_objects; i++) {
 652                if (objects[i].type != OBJ_COMMIT &&
 653                    objects[i].type != OBJ_TAG)
 654                        continue;
 655                add_to_write_order(wo, &wo_end, &objects[i]);
 656        }
 657
 658        /*
 659         * And then all the trees.
 660         */
 661        for (i = last_untagged; i < to_pack.nr_objects; i++) {
 662                if (objects[i].type != OBJ_TREE)
 663                        continue;
 664                add_to_write_order(wo, &wo_end, &objects[i]);
 665        }
 666
 667        /*
 668         * Finally all the rest in really tight order
 669         */
 670        for (i = last_untagged; i < to_pack.nr_objects; i++) {
 671                if (!objects[i].filled)
 672                        add_family_to_write_order(wo, &wo_end, &objects[i]);
 673        }
 674
 675        if (wo_end != to_pack.nr_objects)
 676                die("ordered %u objects, expected %"PRIu32, wo_end, to_pack.nr_objects);
 677
 678        return wo;
 679}
 680
 681static void write_pack_file(void)
 682{
 683        uint32_t i = 0, j;
 684        struct sha1file *f;
 685        off_t offset;
 686        uint32_t nr_remaining = nr_result;
 687        time_t last_mtime = 0;
 688        struct object_entry **write_order;
 689
 690        if (progress > pack_to_stdout)
 691                progress_state = start_progress("Writing objects", nr_result);
 692        written_list = xmalloc(to_pack.nr_objects * sizeof(*written_list));
 693        write_order = compute_write_order();
 694
 695        do {
 696                unsigned char sha1[20];
 697                char *pack_tmp_name = NULL;
 698
 699                if (pack_to_stdout)
 700                        f = sha1fd_throughput(1, "<stdout>", progress_state);
 701                else
 702                        f = create_tmp_packfile(&pack_tmp_name);
 703
 704                offset = write_pack_header(f, nr_remaining);
 705                if (!offset)
 706                        die_errno("unable to write pack header");
 707                nr_written = 0;
 708                for (; i < to_pack.nr_objects; i++) {
 709                        struct object_entry *e = write_order[i];
 710                        if (write_one(f, e, &offset) == WRITE_ONE_BREAK)
 711                                break;
 712                        display_progress(progress_state, written);
 713                }
 714
 715                /*
 716                 * Did we write the wrong # entries in the header?
 717                 * If so, rewrite it like in fast-import
 718                 */
 719                if (pack_to_stdout) {
 720                        sha1close(f, sha1, CSUM_CLOSE);
 721                } else if (nr_written == nr_remaining) {
 722                        sha1close(f, sha1, CSUM_FSYNC);
 723                } else {
 724                        int fd = sha1close(f, sha1, 0);
 725                        fixup_pack_header_footer(fd, sha1, pack_tmp_name,
 726                                                 nr_written, sha1, offset);
 727                        close(fd);
 728                }
 729
 730                if (!pack_to_stdout) {
 731                        struct stat st;
 732                        char tmpname[PATH_MAX];
 733
 734                        /*
 735                         * Packs are runtime accessed in their mtime
 736                         * order since newer packs are more likely to contain
 737                         * younger objects.  So if we are creating multiple
 738                         * packs then we should modify the mtime of later ones
 739                         * to preserve this property.
 740                         */
 741                        if (stat(pack_tmp_name, &st) < 0) {
 742                                warning("failed to stat %s: %s",
 743                                        pack_tmp_name, strerror(errno));
 744                        } else if (!last_mtime) {
 745                                last_mtime = st.st_mtime;
 746                        } else {
 747                                struct utimbuf utb;
 748                                utb.actime = st.st_atime;
 749                                utb.modtime = --last_mtime;
 750                                if (utime(pack_tmp_name, &utb) < 0)
 751                                        warning("failed utime() on %s: %s",
 752                                                tmpname, strerror(errno));
 753                        }
 754
 755                        /* Enough space for "-<sha-1>.pack"? */
 756                        if (sizeof(tmpname) <= strlen(base_name) + 50)
 757                                die("pack base name '%s' too long", base_name);
 758                        snprintf(tmpname, sizeof(tmpname), "%s-", base_name);
 759                        finish_tmp_packfile(tmpname, pack_tmp_name,
 760                                            written_list, nr_written,
 761                                            &pack_idx_opts, sha1);
 762                        free(pack_tmp_name);
 763                        puts(sha1_to_hex(sha1));
 764                }
 765
 766                /* mark written objects as written to previous pack */
 767                for (j = 0; j < nr_written; j++) {
 768                        written_list[j]->offset = (off_t)-1;
 769                }
 770                nr_remaining -= nr_written;
 771        } while (nr_remaining && i < to_pack.nr_objects);
 772
 773        free(written_list);
 774        free(write_order);
 775        stop_progress(&progress_state);
 776        if (written != nr_result)
 777                die("wrote %"PRIu32" objects while expecting %"PRIu32,
 778                        written, nr_result);
 779}
 780
 781static void setup_delta_attr_check(struct git_attr_check *check)
 782{
 783        static struct git_attr *attr_delta;
 784
 785        if (!attr_delta)
 786                attr_delta = git_attr("delta");
 787
 788        check[0].attr = attr_delta;
 789}
 790
 791static int no_try_delta(const char *path)
 792{
 793        struct git_attr_check check[1];
 794
 795        setup_delta_attr_check(check);
 796        if (git_check_attr(path, ARRAY_SIZE(check), check))
 797                return 0;
 798        if (ATTR_FALSE(check->value))
 799                return 1;
 800        return 0;
 801}
 802
 803/*
 804 * When adding an object, check whether we have already added it
 805 * to our packing list. If so, we can skip. However, if we are
 806 * being asked to excludei t, but the previous mention was to include
 807 * it, make sure to adjust its flags and tweak our numbers accordingly.
 808 *
 809 * As an optimization, we pass out the index position where we would have
 810 * found the item, since that saves us from having to look it up again a
 811 * few lines later when we want to add the new entry.
 812 */
 813static int have_duplicate_entry(const unsigned char *sha1,
 814                                int exclude,
 815                                uint32_t *index_pos)
 816{
 817        struct object_entry *entry;
 818
 819        entry = packlist_find(&to_pack, sha1, index_pos);
 820        if (!entry)
 821                return 0;
 822
 823        if (exclude) {
 824                if (!entry->preferred_base)
 825                        nr_result--;
 826                entry->preferred_base = 1;
 827        }
 828
 829        return 1;
 830}
 831
 832/*
 833 * Check whether we want the object in the pack (e.g., we do not want
 834 * objects found in non-local stores if the "--local" option was used).
 835 *
 836 * As a side effect of this check, we will find the packed version of this
 837 * object, if any. We therefore pass out the pack information to avoid having
 838 * to look it up again later.
 839 */
 840static int want_object_in_pack(const unsigned char *sha1,
 841                               int exclude,
 842                               struct packed_git **found_pack,
 843                               off_t *found_offset)
 844{
 845        struct packed_git *p;
 846
 847        if (!exclude && local && has_loose_object_nonlocal(sha1))
 848                return 0;
 849
 850        *found_pack = NULL;
 851        *found_offset = 0;
 852
 853        for (p = packed_git; p; p = p->next) {
 854                off_t offset = find_pack_entry_one(sha1, p);
 855                if (offset) {
 856                        if (!*found_pack) {
 857                                if (!is_pack_valid(p)) {
 858                                        warning("packfile %s cannot be accessed", p->pack_name);
 859                                        continue;
 860                                }
 861                                *found_offset = offset;
 862                                *found_pack = p;
 863                        }
 864                        if (exclude)
 865                                return 1;
 866                        if (incremental)
 867                                return 0;
 868                        if (local && !p->pack_local)
 869                                return 0;
 870                        if (ignore_packed_keep && p->pack_local && p->pack_keep)
 871                                return 0;
 872                }
 873        }
 874
 875        return 1;
 876}
 877
 878static void create_object_entry(const unsigned char *sha1,
 879                                enum object_type type,
 880                                uint32_t hash,
 881                                int exclude,
 882                                int no_try_delta,
 883                                uint32_t index_pos,
 884                                struct packed_git *found_pack,
 885                                off_t found_offset)
 886{
 887        struct object_entry *entry;
 888
 889        entry = packlist_alloc(&to_pack, sha1, index_pos);
 890        entry->hash = hash;
 891        if (type)
 892                entry->type = type;
 893        if (exclude)
 894                entry->preferred_base = 1;
 895        else
 896                nr_result++;
 897        if (found_pack) {
 898                entry->in_pack = found_pack;
 899                entry->in_pack_offset = found_offset;
 900        }
 901
 902        entry->no_try_delta = no_try_delta;
 903}
 904
 905static int add_object_entry(const unsigned char *sha1, enum object_type type,
 906                            const char *name, int exclude)
 907{
 908        struct packed_git *found_pack;
 909        off_t found_offset;
 910        uint32_t index_pos;
 911
 912        if (have_duplicate_entry(sha1, exclude, &index_pos))
 913                return 0;
 914
 915        if (!want_object_in_pack(sha1, exclude, &found_pack, &found_offset))
 916                return 0;
 917
 918        create_object_entry(sha1, type, pack_name_hash(name),
 919                            exclude, name && no_try_delta(name),
 920                            index_pos, found_pack, found_offset);
 921
 922        display_progress(progress_state, to_pack.nr_objects);
 923        return 1;
 924}
 925
 926struct pbase_tree_cache {
 927        unsigned char sha1[20];
 928        int ref;
 929        int temporary;
 930        void *tree_data;
 931        unsigned long tree_size;
 932};
 933
 934static struct pbase_tree_cache *(pbase_tree_cache[256]);
 935static int pbase_tree_cache_ix(const unsigned char *sha1)
 936{
 937        return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
 938}
 939static int pbase_tree_cache_ix_incr(int ix)
 940{
 941        return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
 942}
 943
 944static struct pbase_tree {
 945        struct pbase_tree *next;
 946        /* This is a phony "cache" entry; we are not
 947         * going to evict it nor find it through _get()
 948         * mechanism -- this is for the toplevel node that
 949         * would almost always change with any commit.
 950         */
 951        struct pbase_tree_cache pcache;
 952} *pbase_tree;
 953
 954static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
 955{
 956        struct pbase_tree_cache *ent, *nent;
 957        void *data;
 958        unsigned long size;
 959        enum object_type type;
 960        int neigh;
 961        int my_ix = pbase_tree_cache_ix(sha1);
 962        int available_ix = -1;
 963
 964        /* pbase-tree-cache acts as a limited hashtable.
 965         * your object will be found at your index or within a few
 966         * slots after that slot if it is cached.
 967         */
 968        for (neigh = 0; neigh < 8; neigh++) {
 969                ent = pbase_tree_cache[my_ix];
 970                if (ent && !hashcmp(ent->sha1, sha1)) {
 971                        ent->ref++;
 972                        return ent;
 973                }
 974                else if (((available_ix < 0) && (!ent || !ent->ref)) ||
 975                         ((0 <= available_ix) &&
 976                          (!ent && pbase_tree_cache[available_ix])))
 977                        available_ix = my_ix;
 978                if (!ent)
 979                        break;
 980                my_ix = pbase_tree_cache_ix_incr(my_ix);
 981        }
 982
 983        /* Did not find one.  Either we got a bogus request or
 984         * we need to read and perhaps cache.
 985         */
 986        data = read_sha1_file(sha1, &type, &size);
 987        if (!data)
 988                return NULL;
 989        if (type != OBJ_TREE) {
 990                free(data);
 991                return NULL;
 992        }
 993
 994        /* We need to either cache or return a throwaway copy */
 995
 996        if (available_ix < 0)
 997                ent = NULL;
 998        else {
 999                ent = pbase_tree_cache[available_ix];
1000                my_ix = available_ix;
1001        }
1002
1003        if (!ent) {
1004                nent = xmalloc(sizeof(*nent));
1005                nent->temporary = (available_ix < 0);
1006        }
1007        else {
1008                /* evict and reuse */
1009                free(ent->tree_data);
1010                nent = ent;
1011        }
1012        hashcpy(nent->sha1, sha1);
1013        nent->tree_data = data;
1014        nent->tree_size = size;
1015        nent->ref = 1;
1016        if (!nent->temporary)
1017                pbase_tree_cache[my_ix] = nent;
1018        return nent;
1019}
1020
1021static void pbase_tree_put(struct pbase_tree_cache *cache)
1022{
1023        if (!cache->temporary) {
1024                cache->ref--;
1025                return;
1026        }
1027        free(cache->tree_data);
1028        free(cache);
1029}
1030
1031static int name_cmp_len(const char *name)
1032{
1033        int i;
1034        for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
1035                ;
1036        return i;
1037}
1038
1039static void add_pbase_object(struct tree_desc *tree,
1040                             const char *name,
1041                             int cmplen,
1042                             const char *fullname)
1043{
1044        struct name_entry entry;
1045        int cmp;
1046
1047        while (tree_entry(tree,&entry)) {
1048                if (S_ISGITLINK(entry.mode))
1049                        continue;
1050                cmp = tree_entry_len(&entry) != cmplen ? 1 :
1051                      memcmp(name, entry.path, cmplen);
1052                if (cmp > 0)
1053                        continue;
1054                if (cmp < 0)
1055                        return;
1056                if (name[cmplen] != '/') {
1057                        add_object_entry(entry.sha1,
1058                                         object_type(entry.mode),
1059                                         fullname, 1);
1060                        return;
1061                }
1062                if (S_ISDIR(entry.mode)) {
1063                        struct tree_desc sub;
1064                        struct pbase_tree_cache *tree;
1065                        const char *down = name+cmplen+1;
1066                        int downlen = name_cmp_len(down);
1067
1068                        tree = pbase_tree_get(entry.sha1);
1069                        if (!tree)
1070                                return;
1071                        init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1072
1073                        add_pbase_object(&sub, down, downlen, fullname);
1074                        pbase_tree_put(tree);
1075                }
1076        }
1077}
1078
1079static unsigned *done_pbase_paths;
1080static int done_pbase_paths_num;
1081static int done_pbase_paths_alloc;
1082static int done_pbase_path_pos(unsigned hash)
1083{
1084        int lo = 0;
1085        int hi = done_pbase_paths_num;
1086        while (lo < hi) {
1087                int mi = (hi + lo) / 2;
1088                if (done_pbase_paths[mi] == hash)
1089                        return mi;
1090                if (done_pbase_paths[mi] < hash)
1091                        hi = mi;
1092                else
1093                        lo = mi + 1;
1094        }
1095        return -lo-1;
1096}
1097
1098static int check_pbase_path(unsigned hash)
1099{
1100        int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1101        if (0 <= pos)
1102                return 1;
1103        pos = -pos - 1;
1104        if (done_pbase_paths_alloc <= done_pbase_paths_num) {
1105                done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
1106                done_pbase_paths = xrealloc(done_pbase_paths,
1107                                            done_pbase_paths_alloc *
1108                                            sizeof(unsigned));
1109        }
1110        done_pbase_paths_num++;
1111        if (pos < done_pbase_paths_num)
1112                memmove(done_pbase_paths + pos + 1,
1113                        done_pbase_paths + pos,
1114                        (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1115        done_pbase_paths[pos] = hash;
1116        return 0;
1117}
1118
1119static void add_preferred_base_object(const char *name)
1120{
1121        struct pbase_tree *it;
1122        int cmplen;
1123        unsigned hash = pack_name_hash(name);
1124
1125        if (!num_preferred_base || check_pbase_path(hash))
1126                return;
1127
1128        cmplen = name_cmp_len(name);
1129        for (it = pbase_tree; it; it = it->next) {
1130                if (cmplen == 0) {
1131                        add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1132                }
1133                else {
1134                        struct tree_desc tree;
1135                        init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1136                        add_pbase_object(&tree, name, cmplen, name);
1137                }
1138        }
1139}
1140
1141static void add_preferred_base(unsigned char *sha1)
1142{
1143        struct pbase_tree *it;
1144        void *data;
1145        unsigned long size;
1146        unsigned char tree_sha1[20];
1147
1148        if (window <= num_preferred_base++)
1149                return;
1150
1151        data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1152        if (!data)
1153                return;
1154
1155        for (it = pbase_tree; it; it = it->next) {
1156                if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1157                        free(data);
1158                        return;
1159                }
1160        }
1161
1162        it = xcalloc(1, sizeof(*it));
1163        it->next = pbase_tree;
1164        pbase_tree = it;
1165
1166        hashcpy(it->pcache.sha1, tree_sha1);
1167        it->pcache.tree_data = data;
1168        it->pcache.tree_size = size;
1169}
1170
1171static void cleanup_preferred_base(void)
1172{
1173        struct pbase_tree *it;
1174        unsigned i;
1175
1176        it = pbase_tree;
1177        pbase_tree = NULL;
1178        while (it) {
1179                struct pbase_tree *this = it;
1180                it = this->next;
1181                free(this->pcache.tree_data);
1182                free(this);
1183        }
1184
1185        for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1186                if (!pbase_tree_cache[i])
1187                        continue;
1188                free(pbase_tree_cache[i]->tree_data);
1189                free(pbase_tree_cache[i]);
1190                pbase_tree_cache[i] = NULL;
1191        }
1192
1193        free(done_pbase_paths);
1194        done_pbase_paths = NULL;
1195        done_pbase_paths_num = done_pbase_paths_alloc = 0;
1196}
1197
1198static void check_object(struct object_entry *entry)
1199{
1200        if (entry->in_pack) {
1201                struct packed_git *p = entry->in_pack;
1202                struct pack_window *w_curs = NULL;
1203                const unsigned char *base_ref = NULL;
1204                struct object_entry *base_entry;
1205                unsigned long used, used_0;
1206                unsigned long avail;
1207                off_t ofs;
1208                unsigned char *buf, c;
1209
1210                buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1211
1212                /*
1213                 * We want in_pack_type even if we do not reuse delta
1214                 * since non-delta representations could still be reused.
1215                 */
1216                used = unpack_object_header_buffer(buf, avail,
1217                                                   &entry->in_pack_type,
1218                                                   &entry->size);
1219                if (used == 0)
1220                        goto give_up;
1221
1222                /*
1223                 * Determine if this is a delta and if so whether we can
1224                 * reuse it or not.  Otherwise let's find out as cheaply as
1225                 * possible what the actual type and size for this object is.
1226                 */
1227                switch (entry->in_pack_type) {
1228                default:
1229                        /* Not a delta hence we've already got all we need. */
1230                        entry->type = entry->in_pack_type;
1231                        entry->in_pack_header_size = used;
1232                        if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1233                                goto give_up;
1234                        unuse_pack(&w_curs);
1235                        return;
1236                case OBJ_REF_DELTA:
1237                        if (reuse_delta && !entry->preferred_base)
1238                                base_ref = use_pack(p, &w_curs,
1239                                                entry->in_pack_offset + used, NULL);
1240                        entry->in_pack_header_size = used + 20;
1241                        break;
1242                case OBJ_OFS_DELTA:
1243                        buf = use_pack(p, &w_curs,
1244                                       entry->in_pack_offset + used, NULL);
1245                        used_0 = 0;
1246                        c = buf[used_0++];
1247                        ofs = c & 127;
1248                        while (c & 128) {
1249                                ofs += 1;
1250                                if (!ofs || MSB(ofs, 7)) {
1251                                        error("delta base offset overflow in pack for %s",
1252                                              sha1_to_hex(entry->idx.sha1));
1253                                        goto give_up;
1254                                }
1255                                c = buf[used_0++];
1256                                ofs = (ofs << 7) + (c & 127);
1257                        }
1258                        ofs = entry->in_pack_offset - ofs;
1259                        if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1260                                error("delta base offset out of bound for %s",
1261                                      sha1_to_hex(entry->idx.sha1));
1262                                goto give_up;
1263                        }
1264                        if (reuse_delta && !entry->preferred_base) {
1265                                struct revindex_entry *revidx;
1266                                revidx = find_pack_revindex(p, ofs);
1267                                if (!revidx)
1268                                        goto give_up;
1269                                base_ref = nth_packed_object_sha1(p, revidx->nr);
1270                        }
1271                        entry->in_pack_header_size = used + used_0;
1272                        break;
1273                }
1274
1275                if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1276                        /*
1277                         * If base_ref was set above that means we wish to
1278                         * reuse delta data, and we even found that base
1279                         * in the list of objects we want to pack. Goodie!
1280                         *
1281                         * Depth value does not matter - find_deltas() will
1282                         * never consider reused delta as the base object to
1283                         * deltify other objects against, in order to avoid
1284                         * circular deltas.
1285                         */
1286                        entry->type = entry->in_pack_type;
1287                        entry->delta = base_entry;
1288                        entry->delta_size = entry->size;
1289                        entry->delta_sibling = base_entry->delta_child;
1290                        base_entry->delta_child = entry;
1291                        unuse_pack(&w_curs);
1292                        return;
1293                }
1294
1295                if (entry->type) {
1296                        /*
1297                         * This must be a delta and we already know what the
1298                         * final object type is.  Let's extract the actual
1299                         * object size from the delta header.
1300                         */
1301                        entry->size = get_size_from_delta(p, &w_curs,
1302                                        entry->in_pack_offset + entry->in_pack_header_size);
1303                        if (entry->size == 0)
1304                                goto give_up;
1305                        unuse_pack(&w_curs);
1306                        return;
1307                }
1308
1309                /*
1310                 * No choice but to fall back to the recursive delta walk
1311                 * with sha1_object_info() to find about the object type
1312                 * at this point...
1313                 */
1314                give_up:
1315                unuse_pack(&w_curs);
1316        }
1317
1318        entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1319        /*
1320         * The error condition is checked in prepare_pack().  This is
1321         * to permit a missing preferred base object to be ignored
1322         * as a preferred base.  Doing so can result in a larger
1323         * pack file, but the transfer will still take place.
1324         */
1325}
1326
1327static int pack_offset_sort(const void *_a, const void *_b)
1328{
1329        const struct object_entry *a = *(struct object_entry **)_a;
1330        const struct object_entry *b = *(struct object_entry **)_b;
1331
1332        /* avoid filesystem trashing with loose objects */
1333        if (!a->in_pack && !b->in_pack)
1334                return hashcmp(a->idx.sha1, b->idx.sha1);
1335
1336        if (a->in_pack < b->in_pack)
1337                return -1;
1338        if (a->in_pack > b->in_pack)
1339                return 1;
1340        return a->in_pack_offset < b->in_pack_offset ? -1 :
1341                        (a->in_pack_offset > b->in_pack_offset);
1342}
1343
1344static void get_object_details(void)
1345{
1346        uint32_t i;
1347        struct object_entry **sorted_by_offset;
1348
1349        sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1350        for (i = 0; i < to_pack.nr_objects; i++)
1351                sorted_by_offset[i] = to_pack.objects + i;
1352        qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1353
1354        for (i = 0; i < to_pack.nr_objects; i++) {
1355                struct object_entry *entry = sorted_by_offset[i];
1356                check_object(entry);
1357                if (big_file_threshold < entry->size)
1358                        entry->no_try_delta = 1;
1359        }
1360
1361        free(sorted_by_offset);
1362}
1363
1364/*
1365 * We search for deltas in a list sorted by type, by filename hash, and then
1366 * by size, so that we see progressively smaller and smaller files.
1367 * That's because we prefer deltas to be from the bigger file
1368 * to the smaller -- deletes are potentially cheaper, but perhaps
1369 * more importantly, the bigger file is likely the more recent
1370 * one.  The deepest deltas are therefore the oldest objects which are
1371 * less susceptible to be accessed often.
1372 */
1373static int type_size_sort(const void *_a, const void *_b)
1374{
1375        const struct object_entry *a = *(struct object_entry **)_a;
1376        const struct object_entry *b = *(struct object_entry **)_b;
1377
1378        if (a->type > b->type)
1379                return -1;
1380        if (a->type < b->type)
1381                return 1;
1382        if (a->hash > b->hash)
1383                return -1;
1384        if (a->hash < b->hash)
1385                return 1;
1386        if (a->preferred_base > b->preferred_base)
1387                return -1;
1388        if (a->preferred_base < b->preferred_base)
1389                return 1;
1390        if (a->size > b->size)
1391                return -1;
1392        if (a->size < b->size)
1393                return 1;
1394        return a < b ? -1 : (a > b);  /* newest first */
1395}
1396
1397struct unpacked {
1398        struct object_entry *entry;
1399        void *data;
1400        struct delta_index *index;
1401        unsigned depth;
1402};
1403
1404static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1405                           unsigned long delta_size)
1406{
1407        if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1408                return 0;
1409
1410        if (delta_size < cache_max_small_delta_size)
1411                return 1;
1412
1413        /* cache delta, if objects are large enough compared to delta size */
1414        if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1415                return 1;
1416
1417        return 0;
1418}
1419
1420#ifndef NO_PTHREADS
1421
1422static pthread_mutex_t read_mutex;
1423#define read_lock()             pthread_mutex_lock(&read_mutex)
1424#define read_unlock()           pthread_mutex_unlock(&read_mutex)
1425
1426static pthread_mutex_t cache_mutex;
1427#define cache_lock()            pthread_mutex_lock(&cache_mutex)
1428#define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1429
1430static pthread_mutex_t progress_mutex;
1431#define progress_lock()         pthread_mutex_lock(&progress_mutex)
1432#define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1433
1434#else
1435
1436#define read_lock()             (void)0
1437#define read_unlock()           (void)0
1438#define cache_lock()            (void)0
1439#define cache_unlock()          (void)0
1440#define progress_lock()         (void)0
1441#define progress_unlock()       (void)0
1442
1443#endif
1444
1445static int try_delta(struct unpacked *trg, struct unpacked *src,
1446                     unsigned max_depth, unsigned long *mem_usage)
1447{
1448        struct object_entry *trg_entry = trg->entry;
1449        struct object_entry *src_entry = src->entry;
1450        unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1451        unsigned ref_depth;
1452        enum object_type type;
1453        void *delta_buf;
1454
1455        /* Don't bother doing diffs between different types */
1456        if (trg_entry->type != src_entry->type)
1457                return -1;
1458
1459        /*
1460         * We do not bother to try a delta that we discarded on an
1461         * earlier try, but only when reusing delta data.  Note that
1462         * src_entry that is marked as the preferred_base should always
1463         * be considered, as even if we produce a suboptimal delta against
1464         * it, we will still save the transfer cost, as we already know
1465         * the other side has it and we won't send src_entry at all.
1466         */
1467        if (reuse_delta && trg_entry->in_pack &&
1468            trg_entry->in_pack == src_entry->in_pack &&
1469            !src_entry->preferred_base &&
1470            trg_entry->in_pack_type != OBJ_REF_DELTA &&
1471            trg_entry->in_pack_type != OBJ_OFS_DELTA)
1472                return 0;
1473
1474        /* Let's not bust the allowed depth. */
1475        if (src->depth >= max_depth)
1476                return 0;
1477
1478        /* Now some size filtering heuristics. */
1479        trg_size = trg_entry->size;
1480        if (!trg_entry->delta) {
1481                max_size = trg_size/2 - 20;
1482                ref_depth = 1;
1483        } else {
1484                max_size = trg_entry->delta_size;
1485                ref_depth = trg->depth;
1486        }
1487        max_size = (uint64_t)max_size * (max_depth - src->depth) /
1488                                                (max_depth - ref_depth + 1);
1489        if (max_size == 0)
1490                return 0;
1491        src_size = src_entry->size;
1492        sizediff = src_size < trg_size ? trg_size - src_size : 0;
1493        if (sizediff >= max_size)
1494                return 0;
1495        if (trg_size < src_size / 32)
1496                return 0;
1497
1498        /* Load data if not already done */
1499        if (!trg->data) {
1500                read_lock();
1501                trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1502                read_unlock();
1503                if (!trg->data)
1504                        die("object %s cannot be read",
1505                            sha1_to_hex(trg_entry->idx.sha1));
1506                if (sz != trg_size)
1507                        die("object %s inconsistent object length (%lu vs %lu)",
1508                            sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1509                *mem_usage += sz;
1510        }
1511        if (!src->data) {
1512                read_lock();
1513                src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1514                read_unlock();
1515                if (!src->data) {
1516                        if (src_entry->preferred_base) {
1517                                static int warned = 0;
1518                                if (!warned++)
1519                                        warning("object %s cannot be read",
1520                                                sha1_to_hex(src_entry->idx.sha1));
1521                                /*
1522                                 * Those objects are not included in the
1523                                 * resulting pack.  Be resilient and ignore
1524                                 * them if they can't be read, in case the
1525                                 * pack could be created nevertheless.
1526                                 */
1527                                return 0;
1528                        }
1529                        die("object %s cannot be read",
1530                            sha1_to_hex(src_entry->idx.sha1));
1531                }
1532                if (sz != src_size)
1533                        die("object %s inconsistent object length (%lu vs %lu)",
1534                            sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1535                *mem_usage += sz;
1536        }
1537        if (!src->index) {
1538                src->index = create_delta_index(src->data, src_size);
1539                if (!src->index) {
1540                        static int warned = 0;
1541                        if (!warned++)
1542                                warning("suboptimal pack - out of memory");
1543                        return 0;
1544                }
1545                *mem_usage += sizeof_delta_index(src->index);
1546        }
1547
1548        delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1549        if (!delta_buf)
1550                return 0;
1551
1552        if (trg_entry->delta) {
1553                /* Prefer only shallower same-sized deltas. */
1554                if (delta_size == trg_entry->delta_size &&
1555                    src->depth + 1 >= trg->depth) {
1556                        free(delta_buf);
1557                        return 0;
1558                }
1559        }
1560
1561        /*
1562         * Handle memory allocation outside of the cache
1563         * accounting lock.  Compiler will optimize the strangeness
1564         * away when NO_PTHREADS is defined.
1565         */
1566        free(trg_entry->delta_data);
1567        cache_lock();
1568        if (trg_entry->delta_data) {
1569                delta_cache_size -= trg_entry->delta_size;
1570                trg_entry->delta_data = NULL;
1571        }
1572        if (delta_cacheable(src_size, trg_size, delta_size)) {
1573                delta_cache_size += delta_size;
1574                cache_unlock();
1575                trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1576        } else {
1577                cache_unlock();
1578                free(delta_buf);
1579        }
1580
1581        trg_entry->delta = src_entry;
1582        trg_entry->delta_size = delta_size;
1583        trg->depth = src->depth + 1;
1584
1585        return 1;
1586}
1587
1588static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1589{
1590        struct object_entry *child = me->delta_child;
1591        unsigned int m = n;
1592        while (child) {
1593                unsigned int c = check_delta_limit(child, n + 1);
1594                if (m < c)
1595                        m = c;
1596                child = child->delta_sibling;
1597        }
1598        return m;
1599}
1600
1601static unsigned long free_unpacked(struct unpacked *n)
1602{
1603        unsigned long freed_mem = sizeof_delta_index(n->index);
1604        free_delta_index(n->index);
1605        n->index = NULL;
1606        if (n->data) {
1607                freed_mem += n->entry->size;
1608                free(n->data);
1609                n->data = NULL;
1610        }
1611        n->entry = NULL;
1612        n->depth = 0;
1613        return freed_mem;
1614}
1615
1616static void find_deltas(struct object_entry **list, unsigned *list_size,
1617                        int window, int depth, unsigned *processed)
1618{
1619        uint32_t i, idx = 0, count = 0;
1620        struct unpacked *array;
1621        unsigned long mem_usage = 0;
1622
1623        array = xcalloc(window, sizeof(struct unpacked));
1624
1625        for (;;) {
1626                struct object_entry *entry;
1627                struct unpacked *n = array + idx;
1628                int j, max_depth, best_base = -1;
1629
1630                progress_lock();
1631                if (!*list_size) {
1632                        progress_unlock();
1633                        break;
1634                }
1635                entry = *list++;
1636                (*list_size)--;
1637                if (!entry->preferred_base) {
1638                        (*processed)++;
1639                        display_progress(progress_state, *processed);
1640                }
1641                progress_unlock();
1642
1643                mem_usage -= free_unpacked(n);
1644                n->entry = entry;
1645
1646                while (window_memory_limit &&
1647                       mem_usage > window_memory_limit &&
1648                       count > 1) {
1649                        uint32_t tail = (idx + window - count) % window;
1650                        mem_usage -= free_unpacked(array + tail);
1651                        count--;
1652                }
1653
1654                /* We do not compute delta to *create* objects we are not
1655                 * going to pack.
1656                 */
1657                if (entry->preferred_base)
1658                        goto next;
1659
1660                /*
1661                 * If the current object is at pack edge, take the depth the
1662                 * objects that depend on the current object into account
1663                 * otherwise they would become too deep.
1664                 */
1665                max_depth = depth;
1666                if (entry->delta_child) {
1667                        max_depth -= check_delta_limit(entry, 0);
1668                        if (max_depth <= 0)
1669                                goto next;
1670                }
1671
1672                j = window;
1673                while (--j > 0) {
1674                        int ret;
1675                        uint32_t other_idx = idx + j;
1676                        struct unpacked *m;
1677                        if (other_idx >= window)
1678                                other_idx -= window;
1679                        m = array + other_idx;
1680                        if (!m->entry)
1681                                break;
1682                        ret = try_delta(n, m, max_depth, &mem_usage);
1683                        if (ret < 0)
1684                                break;
1685                        else if (ret > 0)
1686                                best_base = other_idx;
1687                }
1688
1689                /*
1690                 * If we decided to cache the delta data, then it is best
1691                 * to compress it right away.  First because we have to do
1692                 * it anyway, and doing it here while we're threaded will
1693                 * save a lot of time in the non threaded write phase,
1694                 * as well as allow for caching more deltas within
1695                 * the same cache size limit.
1696                 * ...
1697                 * But only if not writing to stdout, since in that case
1698                 * the network is most likely throttling writes anyway,
1699                 * and therefore it is best to go to the write phase ASAP
1700                 * instead, as we can afford spending more time compressing
1701                 * between writes at that moment.
1702                 */
1703                if (entry->delta_data && !pack_to_stdout) {
1704                        entry->z_delta_size = do_compress(&entry->delta_data,
1705                                                          entry->delta_size);
1706                        cache_lock();
1707                        delta_cache_size -= entry->delta_size;
1708                        delta_cache_size += entry->z_delta_size;
1709                        cache_unlock();
1710                }
1711
1712                /* if we made n a delta, and if n is already at max
1713                 * depth, leaving it in the window is pointless.  we
1714                 * should evict it first.
1715                 */
1716                if (entry->delta && max_depth <= n->depth)
1717                        continue;
1718
1719                /*
1720                 * Move the best delta base up in the window, after the
1721                 * currently deltified object, to keep it longer.  It will
1722                 * be the first base object to be attempted next.
1723                 */
1724                if (entry->delta) {
1725                        struct unpacked swap = array[best_base];
1726                        int dist = (window + idx - best_base) % window;
1727                        int dst = best_base;
1728                        while (dist--) {
1729                                int src = (dst + 1) % window;
1730                                array[dst] = array[src];
1731                                dst = src;
1732                        }
1733                        array[dst] = swap;
1734                }
1735
1736                next:
1737                idx++;
1738                if (count + 1 < window)
1739                        count++;
1740                if (idx >= window)
1741                        idx = 0;
1742        }
1743
1744        for (i = 0; i < window; ++i) {
1745                free_delta_index(array[i].index);
1746                free(array[i].data);
1747        }
1748        free(array);
1749}
1750
1751#ifndef NO_PTHREADS
1752
1753static void try_to_free_from_threads(size_t size)
1754{
1755        read_lock();
1756        release_pack_memory(size);
1757        read_unlock();
1758}
1759
1760static try_to_free_t old_try_to_free_routine;
1761
1762/*
1763 * The main thread waits on the condition that (at least) one of the workers
1764 * has stopped working (which is indicated in the .working member of
1765 * struct thread_params).
1766 * When a work thread has completed its work, it sets .working to 0 and
1767 * signals the main thread and waits on the condition that .data_ready
1768 * becomes 1.
1769 */
1770
1771struct thread_params {
1772        pthread_t thread;
1773        struct object_entry **list;
1774        unsigned list_size;
1775        unsigned remaining;
1776        int window;
1777        int depth;
1778        int working;
1779        int data_ready;
1780        pthread_mutex_t mutex;
1781        pthread_cond_t cond;
1782        unsigned *processed;
1783};
1784
1785static pthread_cond_t progress_cond;
1786
1787/*
1788 * Mutex and conditional variable can't be statically-initialized on Windows.
1789 */
1790static void init_threaded_search(void)
1791{
1792        init_recursive_mutex(&read_mutex);
1793        pthread_mutex_init(&cache_mutex, NULL);
1794        pthread_mutex_init(&progress_mutex, NULL);
1795        pthread_cond_init(&progress_cond, NULL);
1796        old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1797}
1798
1799static void cleanup_threaded_search(void)
1800{
1801        set_try_to_free_routine(old_try_to_free_routine);
1802        pthread_cond_destroy(&progress_cond);
1803        pthread_mutex_destroy(&read_mutex);
1804        pthread_mutex_destroy(&cache_mutex);
1805        pthread_mutex_destroy(&progress_mutex);
1806}
1807
1808static void *threaded_find_deltas(void *arg)
1809{
1810        struct thread_params *me = arg;
1811
1812        while (me->remaining) {
1813                find_deltas(me->list, &me->remaining,
1814                            me->window, me->depth, me->processed);
1815
1816                progress_lock();
1817                me->working = 0;
1818                pthread_cond_signal(&progress_cond);
1819                progress_unlock();
1820
1821                /*
1822                 * We must not set ->data_ready before we wait on the
1823                 * condition because the main thread may have set it to 1
1824                 * before we get here. In order to be sure that new
1825                 * work is available if we see 1 in ->data_ready, it
1826                 * was initialized to 0 before this thread was spawned
1827                 * and we reset it to 0 right away.
1828                 */
1829                pthread_mutex_lock(&me->mutex);
1830                while (!me->data_ready)
1831                        pthread_cond_wait(&me->cond, &me->mutex);
1832                me->data_ready = 0;
1833                pthread_mutex_unlock(&me->mutex);
1834        }
1835        /* leave ->working 1 so that this doesn't get more work assigned */
1836        return NULL;
1837}
1838
1839static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1840                           int window, int depth, unsigned *processed)
1841{
1842        struct thread_params *p;
1843        int i, ret, active_threads = 0;
1844
1845        init_threaded_search();
1846
1847        if (!delta_search_threads)      /* --threads=0 means autodetect */
1848                delta_search_threads = online_cpus();
1849        if (delta_search_threads <= 1) {
1850                find_deltas(list, &list_size, window, depth, processed);
1851                cleanup_threaded_search();
1852                return;
1853        }
1854        if (progress > pack_to_stdout)
1855                fprintf(stderr, "Delta compression using up to %d threads.\n",
1856                                delta_search_threads);
1857        p = xcalloc(delta_search_threads, sizeof(*p));
1858
1859        /* Partition the work amongst work threads. */
1860        for (i = 0; i < delta_search_threads; i++) {
1861                unsigned sub_size = list_size / (delta_search_threads - i);
1862
1863                /* don't use too small segments or no deltas will be found */
1864                if (sub_size < 2*window && i+1 < delta_search_threads)
1865                        sub_size = 0;
1866
1867                p[i].window = window;
1868                p[i].depth = depth;
1869                p[i].processed = processed;
1870                p[i].working = 1;
1871                p[i].data_ready = 0;
1872
1873                /* try to split chunks on "path" boundaries */
1874                while (sub_size && sub_size < list_size &&
1875                       list[sub_size]->hash &&
1876                       list[sub_size]->hash == list[sub_size-1]->hash)
1877                        sub_size++;
1878
1879                p[i].list = list;
1880                p[i].list_size = sub_size;
1881                p[i].remaining = sub_size;
1882
1883                list += sub_size;
1884                list_size -= sub_size;
1885        }
1886
1887        /* Start work threads. */
1888        for (i = 0; i < delta_search_threads; i++) {
1889                if (!p[i].list_size)
1890                        continue;
1891                pthread_mutex_init(&p[i].mutex, NULL);
1892                pthread_cond_init(&p[i].cond, NULL);
1893                ret = pthread_create(&p[i].thread, NULL,
1894                                     threaded_find_deltas, &p[i]);
1895                if (ret)
1896                        die("unable to create thread: %s", strerror(ret));
1897                active_threads++;
1898        }
1899
1900        /*
1901         * Now let's wait for work completion.  Each time a thread is done
1902         * with its work, we steal half of the remaining work from the
1903         * thread with the largest number of unprocessed objects and give
1904         * it to that newly idle thread.  This ensure good load balancing
1905         * until the remaining object list segments are simply too short
1906         * to be worth splitting anymore.
1907         */
1908        while (active_threads) {
1909                struct thread_params *target = NULL;
1910                struct thread_params *victim = NULL;
1911                unsigned sub_size = 0;
1912
1913                progress_lock();
1914                for (;;) {
1915                        for (i = 0; !target && i < delta_search_threads; i++)
1916                                if (!p[i].working)
1917                                        target = &p[i];
1918                        if (target)
1919                                break;
1920                        pthread_cond_wait(&progress_cond, &progress_mutex);
1921                }
1922
1923                for (i = 0; i < delta_search_threads; i++)
1924                        if (p[i].remaining > 2*window &&
1925                            (!victim || victim->remaining < p[i].remaining))
1926                                victim = &p[i];
1927                if (victim) {
1928                        sub_size = victim->remaining / 2;
1929                        list = victim->list + victim->list_size - sub_size;
1930                        while (sub_size && list[0]->hash &&
1931                               list[0]->hash == list[-1]->hash) {
1932                                list++;
1933                                sub_size--;
1934                        }
1935                        if (!sub_size) {
1936                                /*
1937                                 * It is possible for some "paths" to have
1938                                 * so many objects that no hash boundary
1939                                 * might be found.  Let's just steal the
1940                                 * exact half in that case.
1941                                 */
1942                                sub_size = victim->remaining / 2;
1943                                list -= sub_size;
1944                        }
1945                        target->list = list;
1946                        victim->list_size -= sub_size;
1947                        victim->remaining -= sub_size;
1948                }
1949                target->list_size = sub_size;
1950                target->remaining = sub_size;
1951                target->working = 1;
1952                progress_unlock();
1953
1954                pthread_mutex_lock(&target->mutex);
1955                target->data_ready = 1;
1956                pthread_cond_signal(&target->cond);
1957                pthread_mutex_unlock(&target->mutex);
1958
1959                if (!sub_size) {
1960                        pthread_join(target->thread, NULL);
1961                        pthread_cond_destroy(&target->cond);
1962                        pthread_mutex_destroy(&target->mutex);
1963                        active_threads--;
1964                }
1965        }
1966        cleanup_threaded_search();
1967        free(p);
1968}
1969
1970#else
1971#define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
1972#endif
1973
1974static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1975{
1976        unsigned char peeled[20];
1977
1978        if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1979            !peel_ref(path, peeled)        && /* peelable? */
1980            packlist_find(&to_pack, peeled, NULL))      /* object packed? */
1981                add_object_entry(sha1, OBJ_TAG, NULL, 0);
1982        return 0;
1983}
1984
1985static void prepare_pack(int window, int depth)
1986{
1987        struct object_entry **delta_list;
1988        uint32_t i, nr_deltas;
1989        unsigned n;
1990
1991        get_object_details();
1992
1993        /*
1994         * If we're locally repacking then we need to be doubly careful
1995         * from now on in order to make sure no stealth corruption gets
1996         * propagated to the new pack.  Clients receiving streamed packs
1997         * should validate everything they get anyway so no need to incur
1998         * the additional cost here in that case.
1999         */
2000        if (!pack_to_stdout)
2001                do_check_packed_object_crc = 1;
2002
2003        if (!to_pack.nr_objects || !window || !depth)
2004                return;
2005
2006        delta_list = xmalloc(to_pack.nr_objects * sizeof(*delta_list));
2007        nr_deltas = n = 0;
2008
2009        for (i = 0; i < to_pack.nr_objects; i++) {
2010                struct object_entry *entry = to_pack.objects + i;
2011
2012                if (entry->delta)
2013                        /* This happens if we decided to reuse existing
2014                         * delta from a pack.  "reuse_delta &&" is implied.
2015                         */
2016                        continue;
2017
2018                if (entry->size < 50)
2019                        continue;
2020
2021                if (entry->no_try_delta)
2022                        continue;
2023
2024                if (!entry->preferred_base) {
2025                        nr_deltas++;
2026                        if (entry->type < 0)
2027                                die("unable to get type of object %s",
2028                                    sha1_to_hex(entry->idx.sha1));
2029                } else {
2030                        if (entry->type < 0) {
2031                                /*
2032                                 * This object is not found, but we
2033                                 * don't have to include it anyway.
2034                                 */
2035                                continue;
2036                        }
2037                }
2038
2039                delta_list[n++] = entry;
2040        }
2041
2042        if (nr_deltas && n > 1) {
2043                unsigned nr_done = 0;
2044                if (progress)
2045                        progress_state = start_progress("Compressing objects",
2046                                                        nr_deltas);
2047                qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
2048                ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
2049                stop_progress(&progress_state);
2050                if (nr_done != nr_deltas)
2051                        die("inconsistency with delta count");
2052        }
2053        free(delta_list);
2054}
2055
2056static int git_pack_config(const char *k, const char *v, void *cb)
2057{
2058        if (!strcmp(k, "pack.window")) {
2059                window = git_config_int(k, v);
2060                return 0;
2061        }
2062        if (!strcmp(k, "pack.windowmemory")) {
2063                window_memory_limit = git_config_ulong(k, v);
2064                return 0;
2065        }
2066        if (!strcmp(k, "pack.depth")) {
2067                depth = git_config_int(k, v);
2068                return 0;
2069        }
2070        if (!strcmp(k, "pack.compression")) {
2071                int level = git_config_int(k, v);
2072                if (level == -1)
2073                        level = Z_DEFAULT_COMPRESSION;
2074                else if (level < 0 || level > Z_BEST_COMPRESSION)
2075                        die("bad pack compression level %d", level);
2076                pack_compression_level = level;
2077                pack_compression_seen = 1;
2078                return 0;
2079        }
2080        if (!strcmp(k, "pack.deltacachesize")) {
2081                max_delta_cache_size = git_config_int(k, v);
2082                return 0;
2083        }
2084        if (!strcmp(k, "pack.deltacachelimit")) {
2085                cache_max_small_delta_size = git_config_int(k, v);
2086                return 0;
2087        }
2088        if (!strcmp(k, "pack.threads")) {
2089                delta_search_threads = git_config_int(k, v);
2090                if (delta_search_threads < 0)
2091                        die("invalid number of threads specified (%d)",
2092                            delta_search_threads);
2093#ifdef NO_PTHREADS
2094                if (delta_search_threads != 1)
2095                        warning("no threads support, ignoring %s", k);
2096#endif
2097                return 0;
2098        }
2099        if (!strcmp(k, "pack.indexversion")) {
2100                pack_idx_opts.version = git_config_int(k, v);
2101                if (pack_idx_opts.version > 2)
2102                        die("bad pack.indexversion=%"PRIu32,
2103                            pack_idx_opts.version);
2104                return 0;
2105        }
2106        return git_default_config(k, v, cb);
2107}
2108
2109static void read_object_list_from_stdin(void)
2110{
2111        char line[40 + 1 + PATH_MAX + 2];
2112        unsigned char sha1[20];
2113
2114        for (;;) {
2115                if (!fgets(line, sizeof(line), stdin)) {
2116                        if (feof(stdin))
2117                                break;
2118                        if (!ferror(stdin))
2119                                die("fgets returned NULL, not EOF, not error!");
2120                        if (errno != EINTR)
2121                                die_errno("fgets");
2122                        clearerr(stdin);
2123                        continue;
2124                }
2125                if (line[0] == '-') {
2126                        if (get_sha1_hex(line+1, sha1))
2127                                die("expected edge sha1, got garbage:\n %s",
2128                                    line);
2129                        add_preferred_base(sha1);
2130                        continue;
2131                }
2132                if (get_sha1_hex(line, sha1))
2133                        die("expected sha1, got garbage:\n %s", line);
2134
2135                add_preferred_base_object(line+41);
2136                add_object_entry(sha1, 0, line+41, 0);
2137        }
2138}
2139
2140#define OBJECT_ADDED (1u<<20)
2141
2142static void show_commit(struct commit *commit, void *data)
2143{
2144        add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
2145        commit->object.flags |= OBJECT_ADDED;
2146}
2147
2148static void show_object(struct object *obj,
2149                        const struct name_path *path, const char *last,
2150                        void *data)
2151{
2152        char *name = path_name(path, last);
2153
2154        add_preferred_base_object(name);
2155        add_object_entry(obj->sha1, obj->type, name, 0);
2156        obj->flags |= OBJECT_ADDED;
2157
2158        /*
2159         * We will have generated the hash from the name,
2160         * but not saved a pointer to it - we can free it
2161         */
2162        free((char *)name);
2163}
2164
2165static void show_edge(struct commit *commit)
2166{
2167        add_preferred_base(commit->object.sha1);
2168}
2169
2170struct in_pack_object {
2171        off_t offset;
2172        struct object *object;
2173};
2174
2175struct in_pack {
2176        int alloc;
2177        int nr;
2178        struct in_pack_object *array;
2179};
2180
2181static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2182{
2183        in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
2184        in_pack->array[in_pack->nr].object = object;
2185        in_pack->nr++;
2186}
2187
2188/*
2189 * Compare the objects in the offset order, in order to emulate the
2190 * "git rev-list --objects" output that produced the pack originally.
2191 */
2192static int ofscmp(const void *a_, const void *b_)
2193{
2194        struct in_pack_object *a = (struct in_pack_object *)a_;
2195        struct in_pack_object *b = (struct in_pack_object *)b_;
2196
2197        if (a->offset < b->offset)
2198                return -1;
2199        else if (a->offset > b->offset)
2200                return 1;
2201        else
2202                return hashcmp(a->object->sha1, b->object->sha1);
2203}
2204
2205static void add_objects_in_unpacked_packs(struct rev_info *revs)
2206{
2207        struct packed_git *p;
2208        struct in_pack in_pack;
2209        uint32_t i;
2210
2211        memset(&in_pack, 0, sizeof(in_pack));
2212
2213        for (p = packed_git; p; p = p->next) {
2214                const unsigned char *sha1;
2215                struct object *o;
2216
2217                if (!p->pack_local || p->pack_keep)
2218                        continue;
2219                if (open_pack_index(p))
2220                        die("cannot open pack index");
2221
2222                ALLOC_GROW(in_pack.array,
2223                           in_pack.nr + p->num_objects,
2224                           in_pack.alloc);
2225
2226                for (i = 0; i < p->num_objects; i++) {
2227                        sha1 = nth_packed_object_sha1(p, i);
2228                        o = lookup_unknown_object(sha1);
2229                        if (!(o->flags & OBJECT_ADDED))
2230                                mark_in_pack_object(o, p, &in_pack);
2231                        o->flags |= OBJECT_ADDED;
2232                }
2233        }
2234
2235        if (in_pack.nr) {
2236                qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2237                      ofscmp);
2238                for (i = 0; i < in_pack.nr; i++) {
2239                        struct object *o = in_pack.array[i].object;
2240                        add_object_entry(o->sha1, o->type, "", 0);
2241                }
2242        }
2243        free(in_pack.array);
2244}
2245
2246static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2247{
2248        static struct packed_git *last_found = (void *)1;
2249        struct packed_git *p;
2250
2251        p = (last_found != (void *)1) ? last_found : packed_git;
2252
2253        while (p) {
2254                if ((!p->pack_local || p->pack_keep) &&
2255                        find_pack_entry_one(sha1, p)) {
2256                        last_found = p;
2257                        return 1;
2258                }
2259                if (p == last_found)
2260                        p = packed_git;
2261                else
2262                        p = p->next;
2263                if (p == last_found)
2264                        p = p->next;
2265        }
2266        return 0;
2267}
2268
2269static void loosen_unused_packed_objects(struct rev_info *revs)
2270{
2271        struct packed_git *p;
2272        uint32_t i;
2273        const unsigned char *sha1;
2274
2275        for (p = packed_git; p; p = p->next) {
2276                if (!p->pack_local || p->pack_keep)
2277                        continue;
2278
2279                if (unpack_unreachable_expiration &&
2280                    p->mtime < unpack_unreachable_expiration)
2281                        continue;
2282
2283                if (open_pack_index(p))
2284                        die("cannot open pack index");
2285
2286                for (i = 0; i < p->num_objects; i++) {
2287                        sha1 = nth_packed_object_sha1(p, i);
2288                        if (!packlist_find(&to_pack, sha1, NULL) &&
2289                                !has_sha1_pack_kept_or_nonlocal(sha1))
2290                                if (force_object_loose(sha1, p->mtime))
2291                                        die("unable to force loose object");
2292                }
2293        }
2294}
2295
2296static void get_object_list(int ac, const char **av)
2297{
2298        struct rev_info revs;
2299        char line[1000];
2300        int flags = 0;
2301
2302        init_revisions(&revs, NULL);
2303        save_commit_buffer = 0;
2304        setup_revisions(ac, av, &revs, NULL);
2305
2306        while (fgets(line, sizeof(line), stdin) != NULL) {
2307                int len = strlen(line);
2308                if (len && line[len - 1] == '\n')
2309                        line[--len] = 0;
2310                if (!len)
2311                        break;
2312                if (*line == '-') {
2313                        if (!strcmp(line, "--not")) {
2314                                flags ^= UNINTERESTING;
2315                                continue;
2316                        }
2317                        die("not a rev '%s'", line);
2318                }
2319                if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2320                        die("bad revision '%s'", line);
2321        }
2322
2323        if (prepare_revision_walk(&revs))
2324                die("revision walk setup failed");
2325        mark_edges_uninteresting(&revs, show_edge);
2326        traverse_commit_list(&revs, show_commit, show_object, NULL);
2327
2328        if (keep_unreachable)
2329                add_objects_in_unpacked_packs(&revs);
2330        if (unpack_unreachable)
2331                loosen_unused_packed_objects(&revs);
2332}
2333
2334static int option_parse_index_version(const struct option *opt,
2335                                      const char *arg, int unset)
2336{
2337        char *c;
2338        const char *val = arg;
2339        pack_idx_opts.version = strtoul(val, &c, 10);
2340        if (pack_idx_opts.version > 2)
2341                die(_("unsupported index version %s"), val);
2342        if (*c == ',' && c[1])
2343                pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2344        if (*c || pack_idx_opts.off32_limit & 0x80000000)
2345                die(_("bad index version '%s'"), val);
2346        return 0;
2347}
2348
2349static int option_parse_unpack_unreachable(const struct option *opt,
2350                                           const char *arg, int unset)
2351{
2352        if (unset) {
2353                unpack_unreachable = 0;
2354                unpack_unreachable_expiration = 0;
2355        }
2356        else {
2357                unpack_unreachable = 1;
2358                if (arg)
2359                        unpack_unreachable_expiration = approxidate(arg);
2360        }
2361        return 0;
2362}
2363
2364static int option_parse_ulong(const struct option *opt,
2365                              const char *arg, int unset)
2366{
2367        if (unset)
2368                die(_("option %s does not accept negative form"),
2369                    opt->long_name);
2370
2371        if (!git_parse_ulong(arg, opt->value))
2372                die(_("unable to parse value '%s' for option %s"),
2373                    arg, opt->long_name);
2374        return 0;
2375}
2376
2377#define OPT_ULONG(s, l, v, h) \
2378        { OPTION_CALLBACK, (s), (l), (v), "n", (h),     \
2379          PARSE_OPT_NONEG, option_parse_ulong }
2380
2381int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2382{
2383        int use_internal_rev_list = 0;
2384        int thin = 0;
2385        int all_progress_implied = 0;
2386        const char *rp_av[6];
2387        int rp_ac = 0;
2388        int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2389        struct option pack_objects_options[] = {
2390                OPT_SET_INT('q', "quiet", &progress,
2391                            N_("do not show progress meter"), 0),
2392                OPT_SET_INT(0, "progress", &progress,
2393                            N_("show progress meter"), 1),
2394                OPT_SET_INT(0, "all-progress", &progress,
2395                            N_("show progress meter during object writing phase"), 2),
2396                OPT_BOOL(0, "all-progress-implied",
2397                         &all_progress_implied,
2398                         N_("similar to --all-progress when progress meter is shown")),
2399                { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2400                  N_("write the pack index file in the specified idx format version"),
2401                  0, option_parse_index_version },
2402                OPT_ULONG(0, "max-pack-size", &pack_size_limit,
2403                          N_("maximum size of each output pack file")),
2404                OPT_BOOL(0, "local", &local,
2405                         N_("ignore borrowed objects from alternate object store")),
2406                OPT_BOOL(0, "incremental", &incremental,
2407                         N_("ignore packed objects")),
2408                OPT_INTEGER(0, "window", &window,
2409                            N_("limit pack window by objects")),
2410                OPT_ULONG(0, "window-memory", &window_memory_limit,
2411                          N_("limit pack window by memory in addition to object limit")),
2412                OPT_INTEGER(0, "depth", &depth,
2413                            N_("maximum length of delta chain allowed in the resulting pack")),
2414                OPT_BOOL(0, "reuse-delta", &reuse_delta,
2415                         N_("reuse existing deltas")),
2416                OPT_BOOL(0, "reuse-object", &reuse_object,
2417                         N_("reuse existing objects")),
2418                OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2419                         N_("use OFS_DELTA objects")),
2420                OPT_INTEGER(0, "threads", &delta_search_threads,
2421                            N_("use threads when searching for best delta matches")),
2422                OPT_BOOL(0, "non-empty", &non_empty,
2423                         N_("do not create an empty pack output")),
2424                OPT_BOOL(0, "revs", &use_internal_rev_list,
2425                         N_("read revision arguments from standard input")),
2426                { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2427                  N_("limit the objects to those that are not yet packed"),
2428                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2429                { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2430                  N_("include objects reachable from any reference"),
2431                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2432                { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2433                  N_("include objects referred by reflog entries"),
2434                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2435                OPT_BOOL(0, "stdout", &pack_to_stdout,
2436                         N_("output pack to stdout")),
2437                OPT_BOOL(0, "include-tag", &include_tag,
2438                         N_("include tag objects that refer to objects to be packed")),
2439                OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2440                         N_("keep unreachable objects")),
2441                { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2442                  N_("unpack unreachable objects newer than <time>"),
2443                  PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2444                OPT_BOOL(0, "thin", &thin,
2445                         N_("create thin packs")),
2446                OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2447                         N_("ignore packs that have companion .keep file")),
2448                OPT_INTEGER(0, "compression", &pack_compression_level,
2449                            N_("pack compression level")),
2450                OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2451                            N_("do not hide commits by grafts"), 0),
2452                OPT_END(),
2453        };
2454
2455        read_replace_refs = 0;
2456
2457        reset_pack_idx_option(&pack_idx_opts);
2458        git_config(git_pack_config, NULL);
2459        if (!pack_compression_seen && core_compression_seen)
2460                pack_compression_level = core_compression_level;
2461
2462        progress = isatty(2);
2463        argc = parse_options(argc, argv, prefix, pack_objects_options,
2464                             pack_usage, 0);
2465
2466        if (argc) {
2467                base_name = argv[0];
2468                argc--;
2469        }
2470        if (pack_to_stdout != !base_name || argc)
2471                usage_with_options(pack_usage, pack_objects_options);
2472
2473        rp_av[rp_ac++] = "pack-objects";
2474        if (thin) {
2475                use_internal_rev_list = 1;
2476                rp_av[rp_ac++] = "--objects-edge";
2477        } else
2478                rp_av[rp_ac++] = "--objects";
2479
2480        if (rev_list_all) {
2481                use_internal_rev_list = 1;
2482                rp_av[rp_ac++] = "--all";
2483        }
2484        if (rev_list_reflog) {
2485                use_internal_rev_list = 1;
2486                rp_av[rp_ac++] = "--reflog";
2487        }
2488        if (rev_list_unpacked) {
2489                use_internal_rev_list = 1;
2490                rp_av[rp_ac++] = "--unpacked";
2491        }
2492
2493        if (!reuse_object)
2494                reuse_delta = 0;
2495        if (pack_compression_level == -1)
2496                pack_compression_level = Z_DEFAULT_COMPRESSION;
2497        else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2498                die("bad pack compression level %d", pack_compression_level);
2499#ifdef NO_PTHREADS
2500        if (delta_search_threads != 1)
2501                warning("no threads support, ignoring --threads");
2502#endif
2503        if (!pack_to_stdout && !pack_size_limit)
2504                pack_size_limit = pack_size_limit_cfg;
2505        if (pack_to_stdout && pack_size_limit)
2506                die("--max-pack-size cannot be used to build a pack for transfer.");
2507        if (pack_size_limit && pack_size_limit < 1024*1024) {
2508                warning("minimum pack size limit is 1 MiB");
2509                pack_size_limit = 1024*1024;
2510        }
2511
2512        if (!pack_to_stdout && thin)
2513                die("--thin cannot be used to build an indexable pack.");
2514
2515        if (keep_unreachable && unpack_unreachable)
2516                die("--keep-unreachable and --unpack-unreachable are incompatible.");
2517
2518        if (progress && all_progress_implied)
2519                progress = 2;
2520
2521        prepare_packed_git();
2522
2523        if (progress)
2524                progress_state = start_progress("Counting objects", 0);
2525        if (!use_internal_rev_list)
2526                read_object_list_from_stdin();
2527        else {
2528                rp_av[rp_ac] = NULL;
2529                get_object_list(rp_ac, rp_av);
2530        }
2531        cleanup_preferred_base();
2532        if (include_tag && nr_result)
2533                for_each_ref(add_ref_tag, NULL);
2534        stop_progress(&progress_state);
2535
2536        if (non_empty && !nr_result)
2537                return 0;
2538        if (nr_result)
2539                prepare_pack(window, depth);
2540        write_pack_file();
2541        if (progress)
2542                fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2543                        " reused %"PRIu32" (delta %"PRIu32")\n",
2544                        written, written_delta, reused, reused_delta);
2545        return 0;
2546}