faf746b2a7fdc8d1bdeda79d349ae68674f3bd9b
   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
 803static int add_object_entry(const unsigned char *sha1, enum object_type type,
 804                            const char *name, int exclude)
 805{
 806        struct object_entry *entry;
 807        struct packed_git *p, *found_pack = NULL;
 808        off_t found_offset = 0;
 809        uint32_t hash = pack_name_hash(name);
 810        uint32_t index_pos;
 811
 812        entry = packlist_find(&to_pack, sha1, &index_pos);
 813        if (entry) {
 814                if (exclude) {
 815                        if (!entry->preferred_base)
 816                                nr_result--;
 817                        entry->preferred_base = 1;
 818                }
 819                return 0;
 820        }
 821
 822        if (!exclude && local && has_loose_object_nonlocal(sha1))
 823                return 0;
 824
 825        for (p = packed_git; p; p = p->next) {
 826                off_t offset = find_pack_entry_one(sha1, p);
 827                if (offset) {
 828                        if (!found_pack) {
 829                                if (!is_pack_valid(p)) {
 830                                        warning("packfile %s cannot be accessed", p->pack_name);
 831                                        continue;
 832                                }
 833                                found_offset = offset;
 834                                found_pack = p;
 835                        }
 836                        if (exclude)
 837                                break;
 838                        if (incremental)
 839                                return 0;
 840                        if (local && !p->pack_local)
 841                                return 0;
 842                        if (ignore_packed_keep && p->pack_local && p->pack_keep)
 843                                return 0;
 844                }
 845        }
 846
 847        entry = packlist_alloc(&to_pack, sha1, index_pos);
 848        entry->hash = hash;
 849        if (type)
 850                entry->type = type;
 851        if (exclude)
 852                entry->preferred_base = 1;
 853        else
 854                nr_result++;
 855        if (found_pack) {
 856                entry->in_pack = found_pack;
 857                entry->in_pack_offset = found_offset;
 858        }
 859
 860        display_progress(progress_state, to_pack.nr_objects);
 861
 862        if (name && no_try_delta(name))
 863                entry->no_try_delta = 1;
 864
 865        return 1;
 866}
 867
 868struct pbase_tree_cache {
 869        unsigned char sha1[20];
 870        int ref;
 871        int temporary;
 872        void *tree_data;
 873        unsigned long tree_size;
 874};
 875
 876static struct pbase_tree_cache *(pbase_tree_cache[256]);
 877static int pbase_tree_cache_ix(const unsigned char *sha1)
 878{
 879        return sha1[0] % ARRAY_SIZE(pbase_tree_cache);
 880}
 881static int pbase_tree_cache_ix_incr(int ix)
 882{
 883        return (ix+1) % ARRAY_SIZE(pbase_tree_cache);
 884}
 885
 886static struct pbase_tree {
 887        struct pbase_tree *next;
 888        /* This is a phony "cache" entry; we are not
 889         * going to evict it nor find it through _get()
 890         * mechanism -- this is for the toplevel node that
 891         * would almost always change with any commit.
 892         */
 893        struct pbase_tree_cache pcache;
 894} *pbase_tree;
 895
 896static struct pbase_tree_cache *pbase_tree_get(const unsigned char *sha1)
 897{
 898        struct pbase_tree_cache *ent, *nent;
 899        void *data;
 900        unsigned long size;
 901        enum object_type type;
 902        int neigh;
 903        int my_ix = pbase_tree_cache_ix(sha1);
 904        int available_ix = -1;
 905
 906        /* pbase-tree-cache acts as a limited hashtable.
 907         * your object will be found at your index or within a few
 908         * slots after that slot if it is cached.
 909         */
 910        for (neigh = 0; neigh < 8; neigh++) {
 911                ent = pbase_tree_cache[my_ix];
 912                if (ent && !hashcmp(ent->sha1, sha1)) {
 913                        ent->ref++;
 914                        return ent;
 915                }
 916                else if (((available_ix < 0) && (!ent || !ent->ref)) ||
 917                         ((0 <= available_ix) &&
 918                          (!ent && pbase_tree_cache[available_ix])))
 919                        available_ix = my_ix;
 920                if (!ent)
 921                        break;
 922                my_ix = pbase_tree_cache_ix_incr(my_ix);
 923        }
 924
 925        /* Did not find one.  Either we got a bogus request or
 926         * we need to read and perhaps cache.
 927         */
 928        data = read_sha1_file(sha1, &type, &size);
 929        if (!data)
 930                return NULL;
 931        if (type != OBJ_TREE) {
 932                free(data);
 933                return NULL;
 934        }
 935
 936        /* We need to either cache or return a throwaway copy */
 937
 938        if (available_ix < 0)
 939                ent = NULL;
 940        else {
 941                ent = pbase_tree_cache[available_ix];
 942                my_ix = available_ix;
 943        }
 944
 945        if (!ent) {
 946                nent = xmalloc(sizeof(*nent));
 947                nent->temporary = (available_ix < 0);
 948        }
 949        else {
 950                /* evict and reuse */
 951                free(ent->tree_data);
 952                nent = ent;
 953        }
 954        hashcpy(nent->sha1, sha1);
 955        nent->tree_data = data;
 956        nent->tree_size = size;
 957        nent->ref = 1;
 958        if (!nent->temporary)
 959                pbase_tree_cache[my_ix] = nent;
 960        return nent;
 961}
 962
 963static void pbase_tree_put(struct pbase_tree_cache *cache)
 964{
 965        if (!cache->temporary) {
 966                cache->ref--;
 967                return;
 968        }
 969        free(cache->tree_data);
 970        free(cache);
 971}
 972
 973static int name_cmp_len(const char *name)
 974{
 975        int i;
 976        for (i = 0; name[i] && name[i] != '\n' && name[i] != '/'; i++)
 977                ;
 978        return i;
 979}
 980
 981static void add_pbase_object(struct tree_desc *tree,
 982                             const char *name,
 983                             int cmplen,
 984                             const char *fullname)
 985{
 986        struct name_entry entry;
 987        int cmp;
 988
 989        while (tree_entry(tree,&entry)) {
 990                if (S_ISGITLINK(entry.mode))
 991                        continue;
 992                cmp = tree_entry_len(&entry) != cmplen ? 1 :
 993                      memcmp(name, entry.path, cmplen);
 994                if (cmp > 0)
 995                        continue;
 996                if (cmp < 0)
 997                        return;
 998                if (name[cmplen] != '/') {
 999                        add_object_entry(entry.sha1,
1000                                         object_type(entry.mode),
1001                                         fullname, 1);
1002                        return;
1003                }
1004                if (S_ISDIR(entry.mode)) {
1005                        struct tree_desc sub;
1006                        struct pbase_tree_cache *tree;
1007                        const char *down = name+cmplen+1;
1008                        int downlen = name_cmp_len(down);
1009
1010                        tree = pbase_tree_get(entry.sha1);
1011                        if (!tree)
1012                                return;
1013                        init_tree_desc(&sub, tree->tree_data, tree->tree_size);
1014
1015                        add_pbase_object(&sub, down, downlen, fullname);
1016                        pbase_tree_put(tree);
1017                }
1018        }
1019}
1020
1021static unsigned *done_pbase_paths;
1022static int done_pbase_paths_num;
1023static int done_pbase_paths_alloc;
1024static int done_pbase_path_pos(unsigned hash)
1025{
1026        int lo = 0;
1027        int hi = done_pbase_paths_num;
1028        while (lo < hi) {
1029                int mi = (hi + lo) / 2;
1030                if (done_pbase_paths[mi] == hash)
1031                        return mi;
1032                if (done_pbase_paths[mi] < hash)
1033                        hi = mi;
1034                else
1035                        lo = mi + 1;
1036        }
1037        return -lo-1;
1038}
1039
1040static int check_pbase_path(unsigned hash)
1041{
1042        int pos = (!done_pbase_paths) ? -1 : done_pbase_path_pos(hash);
1043        if (0 <= pos)
1044                return 1;
1045        pos = -pos - 1;
1046        if (done_pbase_paths_alloc <= done_pbase_paths_num) {
1047                done_pbase_paths_alloc = alloc_nr(done_pbase_paths_alloc);
1048                done_pbase_paths = xrealloc(done_pbase_paths,
1049                                            done_pbase_paths_alloc *
1050                                            sizeof(unsigned));
1051        }
1052        done_pbase_paths_num++;
1053        if (pos < done_pbase_paths_num)
1054                memmove(done_pbase_paths + pos + 1,
1055                        done_pbase_paths + pos,
1056                        (done_pbase_paths_num - pos - 1) * sizeof(unsigned));
1057        done_pbase_paths[pos] = hash;
1058        return 0;
1059}
1060
1061static void add_preferred_base_object(const char *name)
1062{
1063        struct pbase_tree *it;
1064        int cmplen;
1065        unsigned hash = pack_name_hash(name);
1066
1067        if (!num_preferred_base || check_pbase_path(hash))
1068                return;
1069
1070        cmplen = name_cmp_len(name);
1071        for (it = pbase_tree; it; it = it->next) {
1072                if (cmplen == 0) {
1073                        add_object_entry(it->pcache.sha1, OBJ_TREE, NULL, 1);
1074                }
1075                else {
1076                        struct tree_desc tree;
1077                        init_tree_desc(&tree, it->pcache.tree_data, it->pcache.tree_size);
1078                        add_pbase_object(&tree, name, cmplen, name);
1079                }
1080        }
1081}
1082
1083static void add_preferred_base(unsigned char *sha1)
1084{
1085        struct pbase_tree *it;
1086        void *data;
1087        unsigned long size;
1088        unsigned char tree_sha1[20];
1089
1090        if (window <= num_preferred_base++)
1091                return;
1092
1093        data = read_object_with_reference(sha1, tree_type, &size, tree_sha1);
1094        if (!data)
1095                return;
1096
1097        for (it = pbase_tree; it; it = it->next) {
1098                if (!hashcmp(it->pcache.sha1, tree_sha1)) {
1099                        free(data);
1100                        return;
1101                }
1102        }
1103
1104        it = xcalloc(1, sizeof(*it));
1105        it->next = pbase_tree;
1106        pbase_tree = it;
1107
1108        hashcpy(it->pcache.sha1, tree_sha1);
1109        it->pcache.tree_data = data;
1110        it->pcache.tree_size = size;
1111}
1112
1113static void cleanup_preferred_base(void)
1114{
1115        struct pbase_tree *it;
1116        unsigned i;
1117
1118        it = pbase_tree;
1119        pbase_tree = NULL;
1120        while (it) {
1121                struct pbase_tree *this = it;
1122                it = this->next;
1123                free(this->pcache.tree_data);
1124                free(this);
1125        }
1126
1127        for (i = 0; i < ARRAY_SIZE(pbase_tree_cache); i++) {
1128                if (!pbase_tree_cache[i])
1129                        continue;
1130                free(pbase_tree_cache[i]->tree_data);
1131                free(pbase_tree_cache[i]);
1132                pbase_tree_cache[i] = NULL;
1133        }
1134
1135        free(done_pbase_paths);
1136        done_pbase_paths = NULL;
1137        done_pbase_paths_num = done_pbase_paths_alloc = 0;
1138}
1139
1140static void check_object(struct object_entry *entry)
1141{
1142        if (entry->in_pack) {
1143                struct packed_git *p = entry->in_pack;
1144                struct pack_window *w_curs = NULL;
1145                const unsigned char *base_ref = NULL;
1146                struct object_entry *base_entry;
1147                unsigned long used, used_0;
1148                unsigned long avail;
1149                off_t ofs;
1150                unsigned char *buf, c;
1151
1152                buf = use_pack(p, &w_curs, entry->in_pack_offset, &avail);
1153
1154                /*
1155                 * We want in_pack_type even if we do not reuse delta
1156                 * since non-delta representations could still be reused.
1157                 */
1158                used = unpack_object_header_buffer(buf, avail,
1159                                                   &entry->in_pack_type,
1160                                                   &entry->size);
1161                if (used == 0)
1162                        goto give_up;
1163
1164                /*
1165                 * Determine if this is a delta and if so whether we can
1166                 * reuse it or not.  Otherwise let's find out as cheaply as
1167                 * possible what the actual type and size for this object is.
1168                 */
1169                switch (entry->in_pack_type) {
1170                default:
1171                        /* Not a delta hence we've already got all we need. */
1172                        entry->type = entry->in_pack_type;
1173                        entry->in_pack_header_size = used;
1174                        if (entry->type < OBJ_COMMIT || entry->type > OBJ_BLOB)
1175                                goto give_up;
1176                        unuse_pack(&w_curs);
1177                        return;
1178                case OBJ_REF_DELTA:
1179                        if (reuse_delta && !entry->preferred_base)
1180                                base_ref = use_pack(p, &w_curs,
1181                                                entry->in_pack_offset + used, NULL);
1182                        entry->in_pack_header_size = used + 20;
1183                        break;
1184                case OBJ_OFS_DELTA:
1185                        buf = use_pack(p, &w_curs,
1186                                       entry->in_pack_offset + used, NULL);
1187                        used_0 = 0;
1188                        c = buf[used_0++];
1189                        ofs = c & 127;
1190                        while (c & 128) {
1191                                ofs += 1;
1192                                if (!ofs || MSB(ofs, 7)) {
1193                                        error("delta base offset overflow in pack for %s",
1194                                              sha1_to_hex(entry->idx.sha1));
1195                                        goto give_up;
1196                                }
1197                                c = buf[used_0++];
1198                                ofs = (ofs << 7) + (c & 127);
1199                        }
1200                        ofs = entry->in_pack_offset - ofs;
1201                        if (ofs <= 0 || ofs >= entry->in_pack_offset) {
1202                                error("delta base offset out of bound for %s",
1203                                      sha1_to_hex(entry->idx.sha1));
1204                                goto give_up;
1205                        }
1206                        if (reuse_delta && !entry->preferred_base) {
1207                                struct revindex_entry *revidx;
1208                                revidx = find_pack_revindex(p, ofs);
1209                                if (!revidx)
1210                                        goto give_up;
1211                                base_ref = nth_packed_object_sha1(p, revidx->nr);
1212                        }
1213                        entry->in_pack_header_size = used + used_0;
1214                        break;
1215                }
1216
1217                if (base_ref && (base_entry = packlist_find(&to_pack, base_ref, NULL))) {
1218                        /*
1219                         * If base_ref was set above that means we wish to
1220                         * reuse delta data, and we even found that base
1221                         * in the list of objects we want to pack. Goodie!
1222                         *
1223                         * Depth value does not matter - find_deltas() will
1224                         * never consider reused delta as the base object to
1225                         * deltify other objects against, in order to avoid
1226                         * circular deltas.
1227                         */
1228                        entry->type = entry->in_pack_type;
1229                        entry->delta = base_entry;
1230                        entry->delta_size = entry->size;
1231                        entry->delta_sibling = base_entry->delta_child;
1232                        base_entry->delta_child = entry;
1233                        unuse_pack(&w_curs);
1234                        return;
1235                }
1236
1237                if (entry->type) {
1238                        /*
1239                         * This must be a delta and we already know what the
1240                         * final object type is.  Let's extract the actual
1241                         * object size from the delta header.
1242                         */
1243                        entry->size = get_size_from_delta(p, &w_curs,
1244                                        entry->in_pack_offset + entry->in_pack_header_size);
1245                        if (entry->size == 0)
1246                                goto give_up;
1247                        unuse_pack(&w_curs);
1248                        return;
1249                }
1250
1251                /*
1252                 * No choice but to fall back to the recursive delta walk
1253                 * with sha1_object_info() to find about the object type
1254                 * at this point...
1255                 */
1256                give_up:
1257                unuse_pack(&w_curs);
1258        }
1259
1260        entry->type = sha1_object_info(entry->idx.sha1, &entry->size);
1261        /*
1262         * The error condition is checked in prepare_pack().  This is
1263         * to permit a missing preferred base object to be ignored
1264         * as a preferred base.  Doing so can result in a larger
1265         * pack file, but the transfer will still take place.
1266         */
1267}
1268
1269static int pack_offset_sort(const void *_a, const void *_b)
1270{
1271        const struct object_entry *a = *(struct object_entry **)_a;
1272        const struct object_entry *b = *(struct object_entry **)_b;
1273
1274        /* avoid filesystem trashing with loose objects */
1275        if (!a->in_pack && !b->in_pack)
1276                return hashcmp(a->idx.sha1, b->idx.sha1);
1277
1278        if (a->in_pack < b->in_pack)
1279                return -1;
1280        if (a->in_pack > b->in_pack)
1281                return 1;
1282        return a->in_pack_offset < b->in_pack_offset ? -1 :
1283                        (a->in_pack_offset > b->in_pack_offset);
1284}
1285
1286static void get_object_details(void)
1287{
1288        uint32_t i;
1289        struct object_entry **sorted_by_offset;
1290
1291        sorted_by_offset = xcalloc(to_pack.nr_objects, sizeof(struct object_entry *));
1292        for (i = 0; i < to_pack.nr_objects; i++)
1293                sorted_by_offset[i] = to_pack.objects + i;
1294        qsort(sorted_by_offset, to_pack.nr_objects, sizeof(*sorted_by_offset), pack_offset_sort);
1295
1296        for (i = 0; i < to_pack.nr_objects; i++) {
1297                struct object_entry *entry = sorted_by_offset[i];
1298                check_object(entry);
1299                if (big_file_threshold < entry->size)
1300                        entry->no_try_delta = 1;
1301        }
1302
1303        free(sorted_by_offset);
1304}
1305
1306/*
1307 * We search for deltas in a list sorted by type, by filename hash, and then
1308 * by size, so that we see progressively smaller and smaller files.
1309 * That's because we prefer deltas to be from the bigger file
1310 * to the smaller -- deletes are potentially cheaper, but perhaps
1311 * more importantly, the bigger file is likely the more recent
1312 * one.  The deepest deltas are therefore the oldest objects which are
1313 * less susceptible to be accessed often.
1314 */
1315static int type_size_sort(const void *_a, const void *_b)
1316{
1317        const struct object_entry *a = *(struct object_entry **)_a;
1318        const struct object_entry *b = *(struct object_entry **)_b;
1319
1320        if (a->type > b->type)
1321                return -1;
1322        if (a->type < b->type)
1323                return 1;
1324        if (a->hash > b->hash)
1325                return -1;
1326        if (a->hash < b->hash)
1327                return 1;
1328        if (a->preferred_base > b->preferred_base)
1329                return -1;
1330        if (a->preferred_base < b->preferred_base)
1331                return 1;
1332        if (a->size > b->size)
1333                return -1;
1334        if (a->size < b->size)
1335                return 1;
1336        return a < b ? -1 : (a > b);  /* newest first */
1337}
1338
1339struct unpacked {
1340        struct object_entry *entry;
1341        void *data;
1342        struct delta_index *index;
1343        unsigned depth;
1344};
1345
1346static int delta_cacheable(unsigned long src_size, unsigned long trg_size,
1347                           unsigned long delta_size)
1348{
1349        if (max_delta_cache_size && delta_cache_size + delta_size > max_delta_cache_size)
1350                return 0;
1351
1352        if (delta_size < cache_max_small_delta_size)
1353                return 1;
1354
1355        /* cache delta, if objects are large enough compared to delta size */
1356        if ((src_size >> 20) + (trg_size >> 21) > (delta_size >> 10))
1357                return 1;
1358
1359        return 0;
1360}
1361
1362#ifndef NO_PTHREADS
1363
1364static pthread_mutex_t read_mutex;
1365#define read_lock()             pthread_mutex_lock(&read_mutex)
1366#define read_unlock()           pthread_mutex_unlock(&read_mutex)
1367
1368static pthread_mutex_t cache_mutex;
1369#define cache_lock()            pthread_mutex_lock(&cache_mutex)
1370#define cache_unlock()          pthread_mutex_unlock(&cache_mutex)
1371
1372static pthread_mutex_t progress_mutex;
1373#define progress_lock()         pthread_mutex_lock(&progress_mutex)
1374#define progress_unlock()       pthread_mutex_unlock(&progress_mutex)
1375
1376#else
1377
1378#define read_lock()             (void)0
1379#define read_unlock()           (void)0
1380#define cache_lock()            (void)0
1381#define cache_unlock()          (void)0
1382#define progress_lock()         (void)0
1383#define progress_unlock()       (void)0
1384
1385#endif
1386
1387static int try_delta(struct unpacked *trg, struct unpacked *src,
1388                     unsigned max_depth, unsigned long *mem_usage)
1389{
1390        struct object_entry *trg_entry = trg->entry;
1391        struct object_entry *src_entry = src->entry;
1392        unsigned long trg_size, src_size, delta_size, sizediff, max_size, sz;
1393        unsigned ref_depth;
1394        enum object_type type;
1395        void *delta_buf;
1396
1397        /* Don't bother doing diffs between different types */
1398        if (trg_entry->type != src_entry->type)
1399                return -1;
1400
1401        /*
1402         * We do not bother to try a delta that we discarded on an
1403         * earlier try, but only when reusing delta data.  Note that
1404         * src_entry that is marked as the preferred_base should always
1405         * be considered, as even if we produce a suboptimal delta against
1406         * it, we will still save the transfer cost, as we already know
1407         * the other side has it and we won't send src_entry at all.
1408         */
1409        if (reuse_delta && trg_entry->in_pack &&
1410            trg_entry->in_pack == src_entry->in_pack &&
1411            !src_entry->preferred_base &&
1412            trg_entry->in_pack_type != OBJ_REF_DELTA &&
1413            trg_entry->in_pack_type != OBJ_OFS_DELTA)
1414                return 0;
1415
1416        /* Let's not bust the allowed depth. */
1417        if (src->depth >= max_depth)
1418                return 0;
1419
1420        /* Now some size filtering heuristics. */
1421        trg_size = trg_entry->size;
1422        if (!trg_entry->delta) {
1423                max_size = trg_size/2 - 20;
1424                ref_depth = 1;
1425        } else {
1426                max_size = trg_entry->delta_size;
1427                ref_depth = trg->depth;
1428        }
1429        max_size = (uint64_t)max_size * (max_depth - src->depth) /
1430                                                (max_depth - ref_depth + 1);
1431        if (max_size == 0)
1432                return 0;
1433        src_size = src_entry->size;
1434        sizediff = src_size < trg_size ? trg_size - src_size : 0;
1435        if (sizediff >= max_size)
1436                return 0;
1437        if (trg_size < src_size / 32)
1438                return 0;
1439
1440        /* Load data if not already done */
1441        if (!trg->data) {
1442                read_lock();
1443                trg->data = read_sha1_file(trg_entry->idx.sha1, &type, &sz);
1444                read_unlock();
1445                if (!trg->data)
1446                        die("object %s cannot be read",
1447                            sha1_to_hex(trg_entry->idx.sha1));
1448                if (sz != trg_size)
1449                        die("object %s inconsistent object length (%lu vs %lu)",
1450                            sha1_to_hex(trg_entry->idx.sha1), sz, trg_size);
1451                *mem_usage += sz;
1452        }
1453        if (!src->data) {
1454                read_lock();
1455                src->data = read_sha1_file(src_entry->idx.sha1, &type, &sz);
1456                read_unlock();
1457                if (!src->data) {
1458                        if (src_entry->preferred_base) {
1459                                static int warned = 0;
1460                                if (!warned++)
1461                                        warning("object %s cannot be read",
1462                                                sha1_to_hex(src_entry->idx.sha1));
1463                                /*
1464                                 * Those objects are not included in the
1465                                 * resulting pack.  Be resilient and ignore
1466                                 * them if they can't be read, in case the
1467                                 * pack could be created nevertheless.
1468                                 */
1469                                return 0;
1470                        }
1471                        die("object %s cannot be read",
1472                            sha1_to_hex(src_entry->idx.sha1));
1473                }
1474                if (sz != src_size)
1475                        die("object %s inconsistent object length (%lu vs %lu)",
1476                            sha1_to_hex(src_entry->idx.sha1), sz, src_size);
1477                *mem_usage += sz;
1478        }
1479        if (!src->index) {
1480                src->index = create_delta_index(src->data, src_size);
1481                if (!src->index) {
1482                        static int warned = 0;
1483                        if (!warned++)
1484                                warning("suboptimal pack - out of memory");
1485                        return 0;
1486                }
1487                *mem_usage += sizeof_delta_index(src->index);
1488        }
1489
1490        delta_buf = create_delta(src->index, trg->data, trg_size, &delta_size, max_size);
1491        if (!delta_buf)
1492                return 0;
1493
1494        if (trg_entry->delta) {
1495                /* Prefer only shallower same-sized deltas. */
1496                if (delta_size == trg_entry->delta_size &&
1497                    src->depth + 1 >= trg->depth) {
1498                        free(delta_buf);
1499                        return 0;
1500                }
1501        }
1502
1503        /*
1504         * Handle memory allocation outside of the cache
1505         * accounting lock.  Compiler will optimize the strangeness
1506         * away when NO_PTHREADS is defined.
1507         */
1508        free(trg_entry->delta_data);
1509        cache_lock();
1510        if (trg_entry->delta_data) {
1511                delta_cache_size -= trg_entry->delta_size;
1512                trg_entry->delta_data = NULL;
1513        }
1514        if (delta_cacheable(src_size, trg_size, delta_size)) {
1515                delta_cache_size += delta_size;
1516                cache_unlock();
1517                trg_entry->delta_data = xrealloc(delta_buf, delta_size);
1518        } else {
1519                cache_unlock();
1520                free(delta_buf);
1521        }
1522
1523        trg_entry->delta = src_entry;
1524        trg_entry->delta_size = delta_size;
1525        trg->depth = src->depth + 1;
1526
1527        return 1;
1528}
1529
1530static unsigned int check_delta_limit(struct object_entry *me, unsigned int n)
1531{
1532        struct object_entry *child = me->delta_child;
1533        unsigned int m = n;
1534        while (child) {
1535                unsigned int c = check_delta_limit(child, n + 1);
1536                if (m < c)
1537                        m = c;
1538                child = child->delta_sibling;
1539        }
1540        return m;
1541}
1542
1543static unsigned long free_unpacked(struct unpacked *n)
1544{
1545        unsigned long freed_mem = sizeof_delta_index(n->index);
1546        free_delta_index(n->index);
1547        n->index = NULL;
1548        if (n->data) {
1549                freed_mem += n->entry->size;
1550                free(n->data);
1551                n->data = NULL;
1552        }
1553        n->entry = NULL;
1554        n->depth = 0;
1555        return freed_mem;
1556}
1557
1558static void find_deltas(struct object_entry **list, unsigned *list_size,
1559                        int window, int depth, unsigned *processed)
1560{
1561        uint32_t i, idx = 0, count = 0;
1562        struct unpacked *array;
1563        unsigned long mem_usage = 0;
1564
1565        array = xcalloc(window, sizeof(struct unpacked));
1566
1567        for (;;) {
1568                struct object_entry *entry;
1569                struct unpacked *n = array + idx;
1570                int j, max_depth, best_base = -1;
1571
1572                progress_lock();
1573                if (!*list_size) {
1574                        progress_unlock();
1575                        break;
1576                }
1577                entry = *list++;
1578                (*list_size)--;
1579                if (!entry->preferred_base) {
1580                        (*processed)++;
1581                        display_progress(progress_state, *processed);
1582                }
1583                progress_unlock();
1584
1585                mem_usage -= free_unpacked(n);
1586                n->entry = entry;
1587
1588                while (window_memory_limit &&
1589                       mem_usage > window_memory_limit &&
1590                       count > 1) {
1591                        uint32_t tail = (idx + window - count) % window;
1592                        mem_usage -= free_unpacked(array + tail);
1593                        count--;
1594                }
1595
1596                /* We do not compute delta to *create* objects we are not
1597                 * going to pack.
1598                 */
1599                if (entry->preferred_base)
1600                        goto next;
1601
1602                /*
1603                 * If the current object is at pack edge, take the depth the
1604                 * objects that depend on the current object into account
1605                 * otherwise they would become too deep.
1606                 */
1607                max_depth = depth;
1608                if (entry->delta_child) {
1609                        max_depth -= check_delta_limit(entry, 0);
1610                        if (max_depth <= 0)
1611                                goto next;
1612                }
1613
1614                j = window;
1615                while (--j > 0) {
1616                        int ret;
1617                        uint32_t other_idx = idx + j;
1618                        struct unpacked *m;
1619                        if (other_idx >= window)
1620                                other_idx -= window;
1621                        m = array + other_idx;
1622                        if (!m->entry)
1623                                break;
1624                        ret = try_delta(n, m, max_depth, &mem_usage);
1625                        if (ret < 0)
1626                                break;
1627                        else if (ret > 0)
1628                                best_base = other_idx;
1629                }
1630
1631                /*
1632                 * If we decided to cache the delta data, then it is best
1633                 * to compress it right away.  First because we have to do
1634                 * it anyway, and doing it here while we're threaded will
1635                 * save a lot of time in the non threaded write phase,
1636                 * as well as allow for caching more deltas within
1637                 * the same cache size limit.
1638                 * ...
1639                 * But only if not writing to stdout, since in that case
1640                 * the network is most likely throttling writes anyway,
1641                 * and therefore it is best to go to the write phase ASAP
1642                 * instead, as we can afford spending more time compressing
1643                 * between writes at that moment.
1644                 */
1645                if (entry->delta_data && !pack_to_stdout) {
1646                        entry->z_delta_size = do_compress(&entry->delta_data,
1647                                                          entry->delta_size);
1648                        cache_lock();
1649                        delta_cache_size -= entry->delta_size;
1650                        delta_cache_size += entry->z_delta_size;
1651                        cache_unlock();
1652                }
1653
1654                /* if we made n a delta, and if n is already at max
1655                 * depth, leaving it in the window is pointless.  we
1656                 * should evict it first.
1657                 */
1658                if (entry->delta && max_depth <= n->depth)
1659                        continue;
1660
1661                /*
1662                 * Move the best delta base up in the window, after the
1663                 * currently deltified object, to keep it longer.  It will
1664                 * be the first base object to be attempted next.
1665                 */
1666                if (entry->delta) {
1667                        struct unpacked swap = array[best_base];
1668                        int dist = (window + idx - best_base) % window;
1669                        int dst = best_base;
1670                        while (dist--) {
1671                                int src = (dst + 1) % window;
1672                                array[dst] = array[src];
1673                                dst = src;
1674                        }
1675                        array[dst] = swap;
1676                }
1677
1678                next:
1679                idx++;
1680                if (count + 1 < window)
1681                        count++;
1682                if (idx >= window)
1683                        idx = 0;
1684        }
1685
1686        for (i = 0; i < window; ++i) {
1687                free_delta_index(array[i].index);
1688                free(array[i].data);
1689        }
1690        free(array);
1691}
1692
1693#ifndef NO_PTHREADS
1694
1695static void try_to_free_from_threads(size_t size)
1696{
1697        read_lock();
1698        release_pack_memory(size);
1699        read_unlock();
1700}
1701
1702static try_to_free_t old_try_to_free_routine;
1703
1704/*
1705 * The main thread waits on the condition that (at least) one of the workers
1706 * has stopped working (which is indicated in the .working member of
1707 * struct thread_params).
1708 * When a work thread has completed its work, it sets .working to 0 and
1709 * signals the main thread and waits on the condition that .data_ready
1710 * becomes 1.
1711 */
1712
1713struct thread_params {
1714        pthread_t thread;
1715        struct object_entry **list;
1716        unsigned list_size;
1717        unsigned remaining;
1718        int window;
1719        int depth;
1720        int working;
1721        int data_ready;
1722        pthread_mutex_t mutex;
1723        pthread_cond_t cond;
1724        unsigned *processed;
1725};
1726
1727static pthread_cond_t progress_cond;
1728
1729/*
1730 * Mutex and conditional variable can't be statically-initialized on Windows.
1731 */
1732static void init_threaded_search(void)
1733{
1734        init_recursive_mutex(&read_mutex);
1735        pthread_mutex_init(&cache_mutex, NULL);
1736        pthread_mutex_init(&progress_mutex, NULL);
1737        pthread_cond_init(&progress_cond, NULL);
1738        old_try_to_free_routine = set_try_to_free_routine(try_to_free_from_threads);
1739}
1740
1741static void cleanup_threaded_search(void)
1742{
1743        set_try_to_free_routine(old_try_to_free_routine);
1744        pthread_cond_destroy(&progress_cond);
1745        pthread_mutex_destroy(&read_mutex);
1746        pthread_mutex_destroy(&cache_mutex);
1747        pthread_mutex_destroy(&progress_mutex);
1748}
1749
1750static void *threaded_find_deltas(void *arg)
1751{
1752        struct thread_params *me = arg;
1753
1754        while (me->remaining) {
1755                find_deltas(me->list, &me->remaining,
1756                            me->window, me->depth, me->processed);
1757
1758                progress_lock();
1759                me->working = 0;
1760                pthread_cond_signal(&progress_cond);
1761                progress_unlock();
1762
1763                /*
1764                 * We must not set ->data_ready before we wait on the
1765                 * condition because the main thread may have set it to 1
1766                 * before we get here. In order to be sure that new
1767                 * work is available if we see 1 in ->data_ready, it
1768                 * was initialized to 0 before this thread was spawned
1769                 * and we reset it to 0 right away.
1770                 */
1771                pthread_mutex_lock(&me->mutex);
1772                while (!me->data_ready)
1773                        pthread_cond_wait(&me->cond, &me->mutex);
1774                me->data_ready = 0;
1775                pthread_mutex_unlock(&me->mutex);
1776        }
1777        /* leave ->working 1 so that this doesn't get more work assigned */
1778        return NULL;
1779}
1780
1781static void ll_find_deltas(struct object_entry **list, unsigned list_size,
1782                           int window, int depth, unsigned *processed)
1783{
1784        struct thread_params *p;
1785        int i, ret, active_threads = 0;
1786
1787        init_threaded_search();
1788
1789        if (!delta_search_threads)      /* --threads=0 means autodetect */
1790                delta_search_threads = online_cpus();
1791        if (delta_search_threads <= 1) {
1792                find_deltas(list, &list_size, window, depth, processed);
1793                cleanup_threaded_search();
1794                return;
1795        }
1796        if (progress > pack_to_stdout)
1797                fprintf(stderr, "Delta compression using up to %d threads.\n",
1798                                delta_search_threads);
1799        p = xcalloc(delta_search_threads, sizeof(*p));
1800
1801        /* Partition the work amongst work threads. */
1802        for (i = 0; i < delta_search_threads; i++) {
1803                unsigned sub_size = list_size / (delta_search_threads - i);
1804
1805                /* don't use too small segments or no deltas will be found */
1806                if (sub_size < 2*window && i+1 < delta_search_threads)
1807                        sub_size = 0;
1808
1809                p[i].window = window;
1810                p[i].depth = depth;
1811                p[i].processed = processed;
1812                p[i].working = 1;
1813                p[i].data_ready = 0;
1814
1815                /* try to split chunks on "path" boundaries */
1816                while (sub_size && sub_size < list_size &&
1817                       list[sub_size]->hash &&
1818                       list[sub_size]->hash == list[sub_size-1]->hash)
1819                        sub_size++;
1820
1821                p[i].list = list;
1822                p[i].list_size = sub_size;
1823                p[i].remaining = sub_size;
1824
1825                list += sub_size;
1826                list_size -= sub_size;
1827        }
1828
1829        /* Start work threads. */
1830        for (i = 0; i < delta_search_threads; i++) {
1831                if (!p[i].list_size)
1832                        continue;
1833                pthread_mutex_init(&p[i].mutex, NULL);
1834                pthread_cond_init(&p[i].cond, NULL);
1835                ret = pthread_create(&p[i].thread, NULL,
1836                                     threaded_find_deltas, &p[i]);
1837                if (ret)
1838                        die("unable to create thread: %s", strerror(ret));
1839                active_threads++;
1840        }
1841
1842        /*
1843         * Now let's wait for work completion.  Each time a thread is done
1844         * with its work, we steal half of the remaining work from the
1845         * thread with the largest number of unprocessed objects and give
1846         * it to that newly idle thread.  This ensure good load balancing
1847         * until the remaining object list segments are simply too short
1848         * to be worth splitting anymore.
1849         */
1850        while (active_threads) {
1851                struct thread_params *target = NULL;
1852                struct thread_params *victim = NULL;
1853                unsigned sub_size = 0;
1854
1855                progress_lock();
1856                for (;;) {
1857                        for (i = 0; !target && i < delta_search_threads; i++)
1858                                if (!p[i].working)
1859                                        target = &p[i];
1860                        if (target)
1861                                break;
1862                        pthread_cond_wait(&progress_cond, &progress_mutex);
1863                }
1864
1865                for (i = 0; i < delta_search_threads; i++)
1866                        if (p[i].remaining > 2*window &&
1867                            (!victim || victim->remaining < p[i].remaining))
1868                                victim = &p[i];
1869                if (victim) {
1870                        sub_size = victim->remaining / 2;
1871                        list = victim->list + victim->list_size - sub_size;
1872                        while (sub_size && list[0]->hash &&
1873                               list[0]->hash == list[-1]->hash) {
1874                                list++;
1875                                sub_size--;
1876                        }
1877                        if (!sub_size) {
1878                                /*
1879                                 * It is possible for some "paths" to have
1880                                 * so many objects that no hash boundary
1881                                 * might be found.  Let's just steal the
1882                                 * exact half in that case.
1883                                 */
1884                                sub_size = victim->remaining / 2;
1885                                list -= sub_size;
1886                        }
1887                        target->list = list;
1888                        victim->list_size -= sub_size;
1889                        victim->remaining -= sub_size;
1890                }
1891                target->list_size = sub_size;
1892                target->remaining = sub_size;
1893                target->working = 1;
1894                progress_unlock();
1895
1896                pthread_mutex_lock(&target->mutex);
1897                target->data_ready = 1;
1898                pthread_cond_signal(&target->cond);
1899                pthread_mutex_unlock(&target->mutex);
1900
1901                if (!sub_size) {
1902                        pthread_join(target->thread, NULL);
1903                        pthread_cond_destroy(&target->cond);
1904                        pthread_mutex_destroy(&target->mutex);
1905                        active_threads--;
1906                }
1907        }
1908        cleanup_threaded_search();
1909        free(p);
1910}
1911
1912#else
1913#define ll_find_deltas(l, s, w, d, p)   find_deltas(l, &s, w, d, p)
1914#endif
1915
1916static int add_ref_tag(const char *path, const unsigned char *sha1, int flag, void *cb_data)
1917{
1918        unsigned char peeled[20];
1919
1920        if (!prefixcmp(path, "refs/tags/") && /* is a tag? */
1921            !peel_ref(path, peeled)        && /* peelable? */
1922            packlist_find(&to_pack, peeled, NULL))      /* object packed? */
1923                add_object_entry(sha1, OBJ_TAG, NULL, 0);
1924        return 0;
1925}
1926
1927static void prepare_pack(int window, int depth)
1928{
1929        struct object_entry **delta_list;
1930        uint32_t i, nr_deltas;
1931        unsigned n;
1932
1933        get_object_details();
1934
1935        /*
1936         * If we're locally repacking then we need to be doubly careful
1937         * from now on in order to make sure no stealth corruption gets
1938         * propagated to the new pack.  Clients receiving streamed packs
1939         * should validate everything they get anyway so no need to incur
1940         * the additional cost here in that case.
1941         */
1942        if (!pack_to_stdout)
1943                do_check_packed_object_crc = 1;
1944
1945        if (!to_pack.nr_objects || !window || !depth)
1946                return;
1947
1948        delta_list = xmalloc(to_pack.nr_objects * sizeof(*delta_list));
1949        nr_deltas = n = 0;
1950
1951        for (i = 0; i < to_pack.nr_objects; i++) {
1952                struct object_entry *entry = to_pack.objects + i;
1953
1954                if (entry->delta)
1955                        /* This happens if we decided to reuse existing
1956                         * delta from a pack.  "reuse_delta &&" is implied.
1957                         */
1958                        continue;
1959
1960                if (entry->size < 50)
1961                        continue;
1962
1963                if (entry->no_try_delta)
1964                        continue;
1965
1966                if (!entry->preferred_base) {
1967                        nr_deltas++;
1968                        if (entry->type < 0)
1969                                die("unable to get type of object %s",
1970                                    sha1_to_hex(entry->idx.sha1));
1971                } else {
1972                        if (entry->type < 0) {
1973                                /*
1974                                 * This object is not found, but we
1975                                 * don't have to include it anyway.
1976                                 */
1977                                continue;
1978                        }
1979                }
1980
1981                delta_list[n++] = entry;
1982        }
1983
1984        if (nr_deltas && n > 1) {
1985                unsigned nr_done = 0;
1986                if (progress)
1987                        progress_state = start_progress("Compressing objects",
1988                                                        nr_deltas);
1989                qsort(delta_list, n, sizeof(*delta_list), type_size_sort);
1990                ll_find_deltas(delta_list, n, window+1, depth, &nr_done);
1991                stop_progress(&progress_state);
1992                if (nr_done != nr_deltas)
1993                        die("inconsistency with delta count");
1994        }
1995        free(delta_list);
1996}
1997
1998static int git_pack_config(const char *k, const char *v, void *cb)
1999{
2000        if (!strcmp(k, "pack.window")) {
2001                window = git_config_int(k, v);
2002                return 0;
2003        }
2004        if (!strcmp(k, "pack.windowmemory")) {
2005                window_memory_limit = git_config_ulong(k, v);
2006                return 0;
2007        }
2008        if (!strcmp(k, "pack.depth")) {
2009                depth = git_config_int(k, v);
2010                return 0;
2011        }
2012        if (!strcmp(k, "pack.compression")) {
2013                int level = git_config_int(k, v);
2014                if (level == -1)
2015                        level = Z_DEFAULT_COMPRESSION;
2016                else if (level < 0 || level > Z_BEST_COMPRESSION)
2017                        die("bad pack compression level %d", level);
2018                pack_compression_level = level;
2019                pack_compression_seen = 1;
2020                return 0;
2021        }
2022        if (!strcmp(k, "pack.deltacachesize")) {
2023                max_delta_cache_size = git_config_int(k, v);
2024                return 0;
2025        }
2026        if (!strcmp(k, "pack.deltacachelimit")) {
2027                cache_max_small_delta_size = git_config_int(k, v);
2028                return 0;
2029        }
2030        if (!strcmp(k, "pack.threads")) {
2031                delta_search_threads = git_config_int(k, v);
2032                if (delta_search_threads < 0)
2033                        die("invalid number of threads specified (%d)",
2034                            delta_search_threads);
2035#ifdef NO_PTHREADS
2036                if (delta_search_threads != 1)
2037                        warning("no threads support, ignoring %s", k);
2038#endif
2039                return 0;
2040        }
2041        if (!strcmp(k, "pack.indexversion")) {
2042                pack_idx_opts.version = git_config_int(k, v);
2043                if (pack_idx_opts.version > 2)
2044                        die("bad pack.indexversion=%"PRIu32,
2045                            pack_idx_opts.version);
2046                return 0;
2047        }
2048        return git_default_config(k, v, cb);
2049}
2050
2051static void read_object_list_from_stdin(void)
2052{
2053        char line[40 + 1 + PATH_MAX + 2];
2054        unsigned char sha1[20];
2055
2056        for (;;) {
2057                if (!fgets(line, sizeof(line), stdin)) {
2058                        if (feof(stdin))
2059                                break;
2060                        if (!ferror(stdin))
2061                                die("fgets returned NULL, not EOF, not error!");
2062                        if (errno != EINTR)
2063                                die_errno("fgets");
2064                        clearerr(stdin);
2065                        continue;
2066                }
2067                if (line[0] == '-') {
2068                        if (get_sha1_hex(line+1, sha1))
2069                                die("expected edge sha1, got garbage:\n %s",
2070                                    line);
2071                        add_preferred_base(sha1);
2072                        continue;
2073                }
2074                if (get_sha1_hex(line, sha1))
2075                        die("expected sha1, got garbage:\n %s", line);
2076
2077                add_preferred_base_object(line+41);
2078                add_object_entry(sha1, 0, line+41, 0);
2079        }
2080}
2081
2082#define OBJECT_ADDED (1u<<20)
2083
2084static void show_commit(struct commit *commit, void *data)
2085{
2086        add_object_entry(commit->object.sha1, OBJ_COMMIT, NULL, 0);
2087        commit->object.flags |= OBJECT_ADDED;
2088}
2089
2090static void show_object(struct object *obj,
2091                        const struct name_path *path, const char *last,
2092                        void *data)
2093{
2094        char *name = path_name(path, last);
2095
2096        add_preferred_base_object(name);
2097        add_object_entry(obj->sha1, obj->type, name, 0);
2098        obj->flags |= OBJECT_ADDED;
2099
2100        /*
2101         * We will have generated the hash from the name,
2102         * but not saved a pointer to it - we can free it
2103         */
2104        free((char *)name);
2105}
2106
2107static void show_edge(struct commit *commit)
2108{
2109        add_preferred_base(commit->object.sha1);
2110}
2111
2112struct in_pack_object {
2113        off_t offset;
2114        struct object *object;
2115};
2116
2117struct in_pack {
2118        int alloc;
2119        int nr;
2120        struct in_pack_object *array;
2121};
2122
2123static void mark_in_pack_object(struct object *object, struct packed_git *p, struct in_pack *in_pack)
2124{
2125        in_pack->array[in_pack->nr].offset = find_pack_entry_one(object->sha1, p);
2126        in_pack->array[in_pack->nr].object = object;
2127        in_pack->nr++;
2128}
2129
2130/*
2131 * Compare the objects in the offset order, in order to emulate the
2132 * "git rev-list --objects" output that produced the pack originally.
2133 */
2134static int ofscmp(const void *a_, const void *b_)
2135{
2136        struct in_pack_object *a = (struct in_pack_object *)a_;
2137        struct in_pack_object *b = (struct in_pack_object *)b_;
2138
2139        if (a->offset < b->offset)
2140                return -1;
2141        else if (a->offset > b->offset)
2142                return 1;
2143        else
2144                return hashcmp(a->object->sha1, b->object->sha1);
2145}
2146
2147static void add_objects_in_unpacked_packs(struct rev_info *revs)
2148{
2149        struct packed_git *p;
2150        struct in_pack in_pack;
2151        uint32_t i;
2152
2153        memset(&in_pack, 0, sizeof(in_pack));
2154
2155        for (p = packed_git; p; p = p->next) {
2156                const unsigned char *sha1;
2157                struct object *o;
2158
2159                if (!p->pack_local || p->pack_keep)
2160                        continue;
2161                if (open_pack_index(p))
2162                        die("cannot open pack index");
2163
2164                ALLOC_GROW(in_pack.array,
2165                           in_pack.nr + p->num_objects,
2166                           in_pack.alloc);
2167
2168                for (i = 0; i < p->num_objects; i++) {
2169                        sha1 = nth_packed_object_sha1(p, i);
2170                        o = lookup_unknown_object(sha1);
2171                        if (!(o->flags & OBJECT_ADDED))
2172                                mark_in_pack_object(o, p, &in_pack);
2173                        o->flags |= OBJECT_ADDED;
2174                }
2175        }
2176
2177        if (in_pack.nr) {
2178                qsort(in_pack.array, in_pack.nr, sizeof(in_pack.array[0]),
2179                      ofscmp);
2180                for (i = 0; i < in_pack.nr; i++) {
2181                        struct object *o = in_pack.array[i].object;
2182                        add_object_entry(o->sha1, o->type, "", 0);
2183                }
2184        }
2185        free(in_pack.array);
2186}
2187
2188static int has_sha1_pack_kept_or_nonlocal(const unsigned char *sha1)
2189{
2190        static struct packed_git *last_found = (void *)1;
2191        struct packed_git *p;
2192
2193        p = (last_found != (void *)1) ? last_found : packed_git;
2194
2195        while (p) {
2196                if ((!p->pack_local || p->pack_keep) &&
2197                        find_pack_entry_one(sha1, p)) {
2198                        last_found = p;
2199                        return 1;
2200                }
2201                if (p == last_found)
2202                        p = packed_git;
2203                else
2204                        p = p->next;
2205                if (p == last_found)
2206                        p = p->next;
2207        }
2208        return 0;
2209}
2210
2211static void loosen_unused_packed_objects(struct rev_info *revs)
2212{
2213        struct packed_git *p;
2214        uint32_t i;
2215        const unsigned char *sha1;
2216
2217        for (p = packed_git; p; p = p->next) {
2218                if (!p->pack_local || p->pack_keep)
2219                        continue;
2220
2221                if (unpack_unreachable_expiration &&
2222                    p->mtime < unpack_unreachable_expiration)
2223                        continue;
2224
2225                if (open_pack_index(p))
2226                        die("cannot open pack index");
2227
2228                for (i = 0; i < p->num_objects; i++) {
2229                        sha1 = nth_packed_object_sha1(p, i);
2230                        if (!packlist_find(&to_pack, sha1, NULL) &&
2231                                !has_sha1_pack_kept_or_nonlocal(sha1))
2232                                if (force_object_loose(sha1, p->mtime))
2233                                        die("unable to force loose object");
2234                }
2235        }
2236}
2237
2238static void get_object_list(int ac, const char **av)
2239{
2240        struct rev_info revs;
2241        char line[1000];
2242        int flags = 0;
2243
2244        init_revisions(&revs, NULL);
2245        save_commit_buffer = 0;
2246        setup_revisions(ac, av, &revs, NULL);
2247
2248        while (fgets(line, sizeof(line), stdin) != NULL) {
2249                int len = strlen(line);
2250                if (len && line[len - 1] == '\n')
2251                        line[--len] = 0;
2252                if (!len)
2253                        break;
2254                if (*line == '-') {
2255                        if (!strcmp(line, "--not")) {
2256                                flags ^= UNINTERESTING;
2257                                continue;
2258                        }
2259                        die("not a rev '%s'", line);
2260                }
2261                if (handle_revision_arg(line, &revs, flags, REVARG_CANNOT_BE_FILENAME))
2262                        die("bad revision '%s'", line);
2263        }
2264
2265        if (prepare_revision_walk(&revs))
2266                die("revision walk setup failed");
2267        mark_edges_uninteresting(&revs, show_edge);
2268        traverse_commit_list(&revs, show_commit, show_object, NULL);
2269
2270        if (keep_unreachable)
2271                add_objects_in_unpacked_packs(&revs);
2272        if (unpack_unreachable)
2273                loosen_unused_packed_objects(&revs);
2274}
2275
2276static int option_parse_index_version(const struct option *opt,
2277                                      const char *arg, int unset)
2278{
2279        char *c;
2280        const char *val = arg;
2281        pack_idx_opts.version = strtoul(val, &c, 10);
2282        if (pack_idx_opts.version > 2)
2283                die(_("unsupported index version %s"), val);
2284        if (*c == ',' && c[1])
2285                pack_idx_opts.off32_limit = strtoul(c+1, &c, 0);
2286        if (*c || pack_idx_opts.off32_limit & 0x80000000)
2287                die(_("bad index version '%s'"), val);
2288        return 0;
2289}
2290
2291static int option_parse_unpack_unreachable(const struct option *opt,
2292                                           const char *arg, int unset)
2293{
2294        if (unset) {
2295                unpack_unreachable = 0;
2296                unpack_unreachable_expiration = 0;
2297        }
2298        else {
2299                unpack_unreachable = 1;
2300                if (arg)
2301                        unpack_unreachable_expiration = approxidate(arg);
2302        }
2303        return 0;
2304}
2305
2306static int option_parse_ulong(const struct option *opt,
2307                              const char *arg, int unset)
2308{
2309        if (unset)
2310                die(_("option %s does not accept negative form"),
2311                    opt->long_name);
2312
2313        if (!git_parse_ulong(arg, opt->value))
2314                die(_("unable to parse value '%s' for option %s"),
2315                    arg, opt->long_name);
2316        return 0;
2317}
2318
2319#define OPT_ULONG(s, l, v, h) \
2320        { OPTION_CALLBACK, (s), (l), (v), "n", (h),     \
2321          PARSE_OPT_NONEG, option_parse_ulong }
2322
2323int cmd_pack_objects(int argc, const char **argv, const char *prefix)
2324{
2325        int use_internal_rev_list = 0;
2326        int thin = 0;
2327        int all_progress_implied = 0;
2328        const char *rp_av[6];
2329        int rp_ac = 0;
2330        int rev_list_unpacked = 0, rev_list_all = 0, rev_list_reflog = 0;
2331        struct option pack_objects_options[] = {
2332                OPT_SET_INT('q', "quiet", &progress,
2333                            N_("do not show progress meter"), 0),
2334                OPT_SET_INT(0, "progress", &progress,
2335                            N_("show progress meter"), 1),
2336                OPT_SET_INT(0, "all-progress", &progress,
2337                            N_("show progress meter during object writing phase"), 2),
2338                OPT_BOOL(0, "all-progress-implied",
2339                         &all_progress_implied,
2340                         N_("similar to --all-progress when progress meter is shown")),
2341                { OPTION_CALLBACK, 0, "index-version", NULL, N_("version[,offset]"),
2342                  N_("write the pack index file in the specified idx format version"),
2343                  0, option_parse_index_version },
2344                OPT_ULONG(0, "max-pack-size", &pack_size_limit,
2345                          N_("maximum size of each output pack file")),
2346                OPT_BOOL(0, "local", &local,
2347                         N_("ignore borrowed objects from alternate object store")),
2348                OPT_BOOL(0, "incremental", &incremental,
2349                         N_("ignore packed objects")),
2350                OPT_INTEGER(0, "window", &window,
2351                            N_("limit pack window by objects")),
2352                OPT_ULONG(0, "window-memory", &window_memory_limit,
2353                          N_("limit pack window by memory in addition to object limit")),
2354                OPT_INTEGER(0, "depth", &depth,
2355                            N_("maximum length of delta chain allowed in the resulting pack")),
2356                OPT_BOOL(0, "reuse-delta", &reuse_delta,
2357                         N_("reuse existing deltas")),
2358                OPT_BOOL(0, "reuse-object", &reuse_object,
2359                         N_("reuse existing objects")),
2360                OPT_BOOL(0, "delta-base-offset", &allow_ofs_delta,
2361                         N_("use OFS_DELTA objects")),
2362                OPT_INTEGER(0, "threads", &delta_search_threads,
2363                            N_("use threads when searching for best delta matches")),
2364                OPT_BOOL(0, "non-empty", &non_empty,
2365                         N_("do not create an empty pack output")),
2366                OPT_BOOL(0, "revs", &use_internal_rev_list,
2367                         N_("read revision arguments from standard input")),
2368                { OPTION_SET_INT, 0, "unpacked", &rev_list_unpacked, NULL,
2369                  N_("limit the objects to those that are not yet packed"),
2370                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2371                { OPTION_SET_INT, 0, "all", &rev_list_all, NULL,
2372                  N_("include objects reachable from any reference"),
2373                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2374                { OPTION_SET_INT, 0, "reflog", &rev_list_reflog, NULL,
2375                  N_("include objects referred by reflog entries"),
2376                  PARSE_OPT_NOARG | PARSE_OPT_NONEG, NULL, 1 },
2377                OPT_BOOL(0, "stdout", &pack_to_stdout,
2378                         N_("output pack to stdout")),
2379                OPT_BOOL(0, "include-tag", &include_tag,
2380                         N_("include tag objects that refer to objects to be packed")),
2381                OPT_BOOL(0, "keep-unreachable", &keep_unreachable,
2382                         N_("keep unreachable objects")),
2383                { OPTION_CALLBACK, 0, "unpack-unreachable", NULL, N_("time"),
2384                  N_("unpack unreachable objects newer than <time>"),
2385                  PARSE_OPT_OPTARG, option_parse_unpack_unreachable },
2386                OPT_BOOL(0, "thin", &thin,
2387                         N_("create thin packs")),
2388                OPT_BOOL(0, "honor-pack-keep", &ignore_packed_keep,
2389                         N_("ignore packs that have companion .keep file")),
2390                OPT_INTEGER(0, "compression", &pack_compression_level,
2391                            N_("pack compression level")),
2392                OPT_SET_INT(0, "keep-true-parents", &grafts_replace_parents,
2393                            N_("do not hide commits by grafts"), 0),
2394                OPT_END(),
2395        };
2396
2397        read_replace_refs = 0;
2398
2399        reset_pack_idx_option(&pack_idx_opts);
2400        git_config(git_pack_config, NULL);
2401        if (!pack_compression_seen && core_compression_seen)
2402                pack_compression_level = core_compression_level;
2403
2404        progress = isatty(2);
2405        argc = parse_options(argc, argv, prefix, pack_objects_options,
2406                             pack_usage, 0);
2407
2408        if (argc) {
2409                base_name = argv[0];
2410                argc--;
2411        }
2412        if (pack_to_stdout != !base_name || argc)
2413                usage_with_options(pack_usage, pack_objects_options);
2414
2415        rp_av[rp_ac++] = "pack-objects";
2416        if (thin) {
2417                use_internal_rev_list = 1;
2418                rp_av[rp_ac++] = "--objects-edge";
2419        } else
2420                rp_av[rp_ac++] = "--objects";
2421
2422        if (rev_list_all) {
2423                use_internal_rev_list = 1;
2424                rp_av[rp_ac++] = "--all";
2425        }
2426        if (rev_list_reflog) {
2427                use_internal_rev_list = 1;
2428                rp_av[rp_ac++] = "--reflog";
2429        }
2430        if (rev_list_unpacked) {
2431                use_internal_rev_list = 1;
2432                rp_av[rp_ac++] = "--unpacked";
2433        }
2434
2435        if (!reuse_object)
2436                reuse_delta = 0;
2437        if (pack_compression_level == -1)
2438                pack_compression_level = Z_DEFAULT_COMPRESSION;
2439        else if (pack_compression_level < 0 || pack_compression_level > Z_BEST_COMPRESSION)
2440                die("bad pack compression level %d", pack_compression_level);
2441#ifdef NO_PTHREADS
2442        if (delta_search_threads != 1)
2443                warning("no threads support, ignoring --threads");
2444#endif
2445        if (!pack_to_stdout && !pack_size_limit)
2446                pack_size_limit = pack_size_limit_cfg;
2447        if (pack_to_stdout && pack_size_limit)
2448                die("--max-pack-size cannot be used to build a pack for transfer.");
2449        if (pack_size_limit && pack_size_limit < 1024*1024) {
2450                warning("minimum pack size limit is 1 MiB");
2451                pack_size_limit = 1024*1024;
2452        }
2453
2454        if (!pack_to_stdout && thin)
2455                die("--thin cannot be used to build an indexable pack.");
2456
2457        if (keep_unreachable && unpack_unreachable)
2458                die("--keep-unreachable and --unpack-unreachable are incompatible.");
2459
2460        if (progress && all_progress_implied)
2461                progress = 2;
2462
2463        prepare_packed_git();
2464
2465        if (progress)
2466                progress_state = start_progress("Counting objects", 0);
2467        if (!use_internal_rev_list)
2468                read_object_list_from_stdin();
2469        else {
2470                rp_av[rp_ac] = NULL;
2471                get_object_list(rp_ac, rp_av);
2472        }
2473        cleanup_preferred_base();
2474        if (include_tag && nr_result)
2475                for_each_ref(add_ref_tag, NULL);
2476        stop_progress(&progress_state);
2477
2478        if (non_empty && !nr_result)
2479                return 0;
2480        if (nr_result)
2481                prepare_pack(window, depth);
2482        write_pack_file();
2483        if (progress)
2484                fprintf(stderr, "Total %"PRIu32" (delta %"PRIu32"),"
2485                        " reused %"PRIu32" (delta %"PRIu32")\n",
2486                        written, written_delta, reused, reused_delta);
2487        return 0;
2488}