git.oblomov.eu Git - git/blob - midx.c

   1 #include "cache.h"
   2 #include "config.h"
   3 #include "csum-file.h"
   4 #include "dir.h"
   5 #include "lockfile.h"
   6 #include "packfile.h"
   7 #include "object-store.h"
   8 #include "sha1-lookup.h"
   9 #include "midx.h"
  10 #include "progress.h"
  11 #include "trace2.h"
  12 #include "run-command.h"
  13
  14 #define MIDX_SIGNATURE 0x4d494458 /* "MIDX" */
  15 #define MIDX_VERSION 1
  16 #define MIDX_BYTE_FILE_VERSION 4
  17 #define MIDX_BYTE_HASH_VERSION 5
  18 #define MIDX_BYTE_NUM_CHUNKS 6
  19 #define MIDX_BYTE_NUM_PACKS 8
  20 #define MIDX_HASH_VERSION 1
  21 #define MIDX_HEADER_SIZE 12
  22 #define MIDX_HASH_LEN 20
  23 #define MIDX_MIN_SIZE (MIDX_HEADER_SIZE + MIDX_HASH_LEN)
  24
  25 #define MIDX_MAX_CHUNKS 5
  26 #define MIDX_CHUNK_ALIGNMENT 4
  27 #define MIDX_CHUNKID_PACKNAMES 0x504e414d /* "PNAM" */
  28 #define MIDX_CHUNKID_OIDFANOUT 0x4f494446 /* "OIDF" */
  29 #define MIDX_CHUNKID_OIDLOOKUP 0x4f49444c /* "OIDL" */
  30 #define MIDX_CHUNKID_OBJECTOFFSETS 0x4f4f4646 /* "OOFF" */
  31 #define MIDX_CHUNKID_LARGEOFFSETS 0x4c4f4646 /* "LOFF" */
  32 #define MIDX_CHUNKLOOKUP_WIDTH (sizeof(uint32_t) + sizeof(uint64_t))
  33 #define MIDX_CHUNK_FANOUT_SIZE (sizeof(uint32_t) * 256)
  34 #define MIDX_CHUNK_OFFSET_WIDTH (2 * sizeof(uint32_t))
  35 #define MIDX_CHUNK_LARGE_OFFSET_WIDTH (sizeof(uint64_t))
  36 #define MIDX_LARGE_OFFSET_NEEDED 0x80000000
  37
  38 #define PACK_EXPIRED UINT_MAX
  39
  40 static char *get_midx_filename(const char *object_dir)
  41 {
  42         return xstrfmt("%s/pack/multi-pack-index", object_dir);
  43 }
  44
  45 struct multi_pack_index *load_multi_pack_index(const char *object_dir, int local)
  46 {
  47         struct multi_pack_index *m = NULL;
  48         int fd;
  49         struct stat st;
  50         size_t midx_size;
  51         void *midx_map = NULL;
  52         uint32_t hash_version;
  53         char *midx_name = get_midx_filename(object_dir);
  54         uint32_t i;
  55         const char *cur_pack_name;
  56
  57         fd = git_open(midx_name);
  58
  59         if (fd < 0)
  60                 goto cleanup_fail;
  61         if (fstat(fd, &st)) {
  62                 error_errno(_("failed to read %s"), midx_name);
  63                 goto cleanup_fail;
  64         }
  65
  66         midx_size = xsize_t(st.st_size);
  67
  68         if (midx_size < MIDX_MIN_SIZE) {
  69                 error(_("multi-pack-index file %s is too small"), midx_name);
  70                 goto cleanup_fail;
  71         }
  72
  73         FREE_AND_NULL(midx_name);
  74
  75         midx_map = xmmap(NULL, midx_size, PROT_READ, MAP_PRIVATE, fd, 0);
  76
  77         FLEX_ALLOC_STR(m, object_dir, object_dir);
  78         m->fd = fd;
  79         m->data = midx_map;
  80         m->data_len = midx_size;
  81         m->local = local;
  82
  83         m->signature = get_be32(m->data);
  84         if (m->signature != MIDX_SIGNATURE)
  85                 die(_("multi-pack-index signature 0x%08x does not match signature 0x%08x"),
  86                       m->signature, MIDX_SIGNATURE);
  87
  88         m->version = m->data[MIDX_BYTE_FILE_VERSION];
  89         if (m->version != MIDX_VERSION)
  90                 die(_("multi-pack-index version %d not recognized"),
  91                       m->version);
  92
  93         hash_version = m->data[MIDX_BYTE_HASH_VERSION];
  94         if (hash_version != MIDX_HASH_VERSION)
  95                 die(_("hash version %u does not match"), hash_version);
  96         m->hash_len = MIDX_HASH_LEN;
  97
  98         m->num_chunks = m->data[MIDX_BYTE_NUM_CHUNKS];
  99
 100         m->num_packs = get_be32(m->data + MIDX_BYTE_NUM_PACKS);
 101
 102         for (i = 0; i < m->num_chunks; i++) {
 103                 uint32_t chunk_id = get_be32(m->data + MIDX_HEADER_SIZE +
 104                                              MIDX_CHUNKLOOKUP_WIDTH * i);
 105                 uint64_t chunk_offset = get_be64(m->data + MIDX_HEADER_SIZE + 4 +
 106                                                  MIDX_CHUNKLOOKUP_WIDTH * i);
 107
 108                 if (chunk_offset >= m->data_len)
 109                         die(_("invalid chunk offset (too large)"));
 110
 111                 switch (chunk_id) {
 112                         case MIDX_CHUNKID_PACKNAMES:
 113                                 m->chunk_pack_names = m->data + chunk_offset;
 114                                 break;
 115
 116                         case MIDX_CHUNKID_OIDFANOUT:
 117                                 m->chunk_oid_fanout = (uint32_t *)(m->data + chunk_offset);
 118                                 break;
 119
 120                         case MIDX_CHUNKID_OIDLOOKUP:
 121                                 m->chunk_oid_lookup = m->data + chunk_offset;
 122                                 break;
 123
 124                         case MIDX_CHUNKID_OBJECTOFFSETS:
 125                                 m->chunk_object_offsets = m->data + chunk_offset;
 126                                 break;
 127
 128                         case MIDX_CHUNKID_LARGEOFFSETS:
 129                                 m->chunk_large_offsets = m->data + chunk_offset;
 130                                 break;
 131
 132                         case 0:
 133                                 die(_("terminating multi-pack-index chunk id appears earlier than expected"));
 134                                 break;
 135
 136                         default:
 137                                 /*
 138                                  * Do nothing on unrecognized chunks, allowing future
 139                                  * extensions to add optional chunks.
 140                                  */
 141                                 break;
 142                 }
 143         }
 144
 145         if (!m->chunk_pack_names)
 146                 die(_("multi-pack-index missing required pack-name chunk"));
 147         if (!m->chunk_oid_fanout)
 148                 die(_("multi-pack-index missing required OID fanout chunk"));
 149         if (!m->chunk_oid_lookup)
 150                 die(_("multi-pack-index missing required OID lookup chunk"));
 151         if (!m->chunk_object_offsets)
 152                 die(_("multi-pack-index missing required object offsets chunk"));
 153
 154         m->num_objects = ntohl(m->chunk_oid_fanout[255]);
 155
 156         m->pack_names = xcalloc(m->num_packs, sizeof(*m->pack_names));
 157         m->packs = xcalloc(m->num_packs, sizeof(*m->packs));
 158
 159         cur_pack_name = (const char *)m->chunk_pack_names;
 160         for (i = 0; i < m->num_packs; i++) {
 161                 m->pack_names[i] = cur_pack_name;
 162
 163                 cur_pack_name += strlen(cur_pack_name) + 1;
 164
 165                 if (i && strcmp(m->pack_names[i], m->pack_names[i - 1]) <= 0)
 166                         die(_("multi-pack-index pack names out of order: '%s' before '%s'"),
 167                               m->pack_names[i - 1],
 168                               m->pack_names[i]);
 169         }
 170
 171         trace2_data_intmax("midx", the_repository, "load/num_packs", m->num_packs);
 172         trace2_data_intmax("midx", the_repository, "load/num_objects", m->num_objects);
 173
 174         return m;
 175
 176 cleanup_fail:
 177         free(m);
 178         free(midx_name);
 179         if (midx_map)
 180                 munmap(midx_map, midx_size);
 181         if (0 <= fd)
 182                 close(fd);
 183         return NULL;
 184 }
 185
 186 void close_midx(struct multi_pack_index *m)
 187 {
 188         uint32_t i;
 189
 190         if (!m)
 191                 return;
 192
 193         munmap((unsigned char *)m->data, m->data_len);
 194         close(m->fd);
 195         m->fd = -1;
 196
 197         for (i = 0; i < m->num_packs; i++) {
 198                 if (m->packs[i])
 199                         m->packs[i]->multi_pack_index = 0;
 200         }
 201         FREE_AND_NULL(m->packs);
 202         FREE_AND_NULL(m->pack_names);
 203 }
 204
 205 int prepare_midx_pack(struct repository *r, struct multi_pack_index *m, uint32_t pack_int_id)
 206 {
 207         struct strbuf pack_name = STRBUF_INIT;
 208         struct packed_git *p;
 209
 210         if (pack_int_id >= m->num_packs)
 211                 die(_("bad pack-int-id: %u (%u total packs)"),
 212                     pack_int_id, m->num_packs);
 213
 214         if (m->packs[pack_int_id])
 215                 return 0;
 216
 217         strbuf_addf(&pack_name, "%s/pack/%s", m->object_dir,
 218                     m->pack_names[pack_int_id]);
 219
 220         p = add_packed_git(pack_name.buf, pack_name.len, m->local);
 221         strbuf_release(&pack_name);
 222
 223         if (!p)
 224                 return 1;
 225
 226         p->multi_pack_index = 1;
 227         m->packs[pack_int_id] = p;
 228         install_packed_git(r, p);
 229         list_add_tail(&p->mru, &r->objects->packed_git_mru);
 230
 231         return 0;
 232 }
 233
 234 int bsearch_midx(const struct object_id *oid, struct multi_pack_index *m, uint32_t *result)
 235 {
 236         return bsearch_hash(oid->hash, m->chunk_oid_fanout, m->chunk_oid_lookup,
 237                             MIDX_HASH_LEN, result);
 238 }
 239
 240 struct object_id *nth_midxed_object_oid(struct object_id *oid,
 241                                         struct multi_pack_index *m,
 242                                         uint32_t n)
 243 {
 244         if (n >= m->num_objects)
 245                 return NULL;
 246
 247         hashcpy(oid->hash, m->chunk_oid_lookup + m->hash_len * n);
 248         return oid;
 249 }
 250
 251 static off_t nth_midxed_offset(struct multi_pack_index *m, uint32_t pos)
 252 {
 253         const unsigned char *offset_data;
 254         uint32_t offset32;
 255
 256         offset_data = m->chunk_object_offsets + pos * MIDX_CHUNK_OFFSET_WIDTH;
 257         offset32 = get_be32(offset_data + sizeof(uint32_t));
 258
 259         if (m->chunk_large_offsets && offset32 & MIDX_LARGE_OFFSET_NEEDED) {
 260                 if (sizeof(off_t) < sizeof(uint64_t))
 261                         die(_("multi-pack-index stores a 64-bit offset, but off_t is too small"));
 262
 263                 offset32 ^= MIDX_LARGE_OFFSET_NEEDED;
 264                 return get_be64(m->chunk_large_offsets + sizeof(uint64_t) * offset32);
 265         }
 266
 267         return offset32;
 268 }
 269
 270 static uint32_t nth_midxed_pack_int_id(struct multi_pack_index *m, uint32_t pos)
 271 {
 272         return get_be32(m->chunk_object_offsets + pos * MIDX_CHUNK_OFFSET_WIDTH);
 273 }
 274
 275 static int nth_midxed_pack_entry(struct repository *r,
 276                                  struct multi_pack_index *m,
 277                                  struct pack_entry *e,
 278                                  uint32_t pos)
 279 {
 280         uint32_t pack_int_id;
 281         struct packed_git *p;
 282
 283         if (pos >= m->num_objects)
 284                 return 0;
 285
 286         pack_int_id = nth_midxed_pack_int_id(m, pos);
 287
 288         if (prepare_midx_pack(r, m, pack_int_id))
 289                 die(_("error preparing packfile from multi-pack-index"));
 290         p = m->packs[pack_int_id];
 291
 292         /*
 293         * We are about to tell the caller where they can locate the
 294         * requested object.  We better make sure the packfile is
 295         * still here and can be accessed before supplying that
 296         * answer, as it may have been deleted since the MIDX was
 297         * loaded!
 298         */
 299         if (!is_pack_valid(p))
 300                 return 0;
 301
 302         if (p->num_bad_objects) {
 303                 uint32_t i;
 304                 struct object_id oid;
 305                 nth_midxed_object_oid(&oid, m, pos);
 306                 for (i = 0; i < p->num_bad_objects; i++)
 307                         if (hasheq(oid.hash,
 308                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
 309                                 return 0;
 310         }
 311
 312         e->offset = nth_midxed_offset(m, pos);
 313         e->p = p;
 314
 315         return 1;
 316 }
 317
 318 int fill_midx_entry(struct repository * r,
 319                     const struct object_id *oid,
 320                     struct pack_entry *e,
 321                     struct multi_pack_index *m)
 322 {
 323         uint32_t pos;
 324
 325         if (!bsearch_midx(oid, m, &pos))
 326                 return 0;
 327
 328         return nth_midxed_pack_entry(r, m, e, pos);
 329 }
 330
 331 /* Match "foo.idx" against either "foo.pack" _or_ "foo.idx". */
 332 static int cmp_idx_or_pack_name(const char *idx_or_pack_name,
 333                                 const char *idx_name)
 334 {
 335         /* Skip past any initial matching prefix. */
 336         while (*idx_name && *idx_name == *idx_or_pack_name) {
 337                 idx_name++;
 338                 idx_or_pack_name++;
 339         }
 340
 341         /*
 342          * If we didn't match completely, we may have matched "pack-1234." and
 343          * be left with "idx" and "pack" respectively, which is also OK. We do
 344          * not have to check for "idx" and "idx", because that would have been
 345          * a complete match (and in that case these strcmps will be false, but
 346          * we'll correctly return 0 from the final strcmp() below.
 347          *
 348          * Technically this matches "fooidx" and "foopack", but we'd never have
 349          * such names in the first place.
 350          */
 351         if (!strcmp(idx_name, "idx") && !strcmp(idx_or_pack_name, "pack"))
 352                 return 0;
 353
 354         /*
 355          * This not only checks for a complete match, but also orders based on
 356          * the first non-identical character, which means our ordering will
 357          * match a raw strcmp(). That makes it OK to use this to binary search
 358          * a naively-sorted list.
 359          */
 360         return strcmp(idx_or_pack_name, idx_name);
 361 }
 362
 363 int midx_contains_pack(struct multi_pack_index *m, const char *idx_or_pack_name)
 364 {
 365         uint32_t first = 0, last = m->num_packs;
 366
 367         while (first < last) {
 368                 uint32_t mid = first + (last - first) / 2;
 369                 const char *current;
 370                 int cmp;
 371
 372                 current = m->pack_names[mid];
 373                 cmp = cmp_idx_or_pack_name(idx_or_pack_name, current);
 374                 if (!cmp)
 375                         return 1;
 376                 if (cmp > 0) {
 377                         first = mid + 1;
 378                         continue;
 379                 }
 380                 last = mid;
 381         }
 382
 383         return 0;
 384 }
 385
 386 int prepare_multi_pack_index_one(struct repository *r, const char *object_dir, int local)
 387 {
 388         struct multi_pack_index *m;
 389         struct multi_pack_index *m_search;
 390         int config_value;
 391         static int env_value = -1;
 392
 393         if (env_value < 0)
 394                 env_value = git_env_bool(GIT_TEST_MULTI_PACK_INDEX, 0);
 395
 396         if (!env_value &&
 397             (repo_config_get_bool(r, "core.multipackindex", &config_value) ||
 398             !config_value))
 399                 return 0;
 400
 401         for (m_search = r->objects->multi_pack_index; m_search; m_search = m_search->next)
 402                 if (!strcmp(object_dir, m_search->object_dir))
 403                         return 1;
 404
 405         m = load_multi_pack_index(object_dir, local);
 406
 407         if (m) {
 408                 m->next = r->objects->multi_pack_index;
 409                 r->objects->multi_pack_index = m;
 410                 return 1;
 411         }
 412
 413         return 0;
 414 }
 415
 416 static size_t write_midx_header(struct hashfile *f,
 417                                 unsigned char num_chunks,
 418                                 uint32_t num_packs)
 419 {
 420         unsigned char byte_values[4];
 421
 422         hashwrite_be32(f, MIDX_SIGNATURE);
 423         byte_values[0] = MIDX_VERSION;
 424         byte_values[1] = MIDX_HASH_VERSION;
 425         byte_values[2] = num_chunks;
 426         byte_values[3] = 0; /* unused */
 427         hashwrite(f, byte_values, sizeof(byte_values));
 428         hashwrite_be32(f, num_packs);
 429
 430         return MIDX_HEADER_SIZE;
 431 }
 432
 433 struct pack_info {
 434         uint32_t orig_pack_int_id;
 435         char *pack_name;
 436         struct packed_git *p;
 437         unsigned expired : 1;
 438 };
 439
 440 static int pack_info_compare(const void *_a, const void *_b)
 441 {
 442         struct pack_info *a = (struct pack_info *)_a;
 443         struct pack_info *b = (struct pack_info *)_b;
 444         return strcmp(a->pack_name, b->pack_name);
 445 }
 446
 447 struct pack_list {
 448         struct pack_info *info;
 449         uint32_t nr;
 450         uint32_t alloc;
 451         struct multi_pack_index *m;
 452 };
 453
 454 static void add_pack_to_midx(const char *full_path, size_t full_path_len,
 455                              const char *file_name, void *data)
 456 {
 457         struct pack_list *packs = (struct pack_list *)data;
 458
 459         if (ends_with(file_name, ".idx")) {
 460                 if (packs->m && midx_contains_pack(packs->m, file_name))
 461                         return;
 462
 463                 ALLOC_GROW(packs->info, packs->nr + 1, packs->alloc);
 464
 465                 packs->info[packs->nr].p = add_packed_git(full_path,
 466                                                           full_path_len,
 467                                                           0);
 468
 469                 if (!packs->info[packs->nr].p) {
 470                         warning(_("failed to add packfile '%s'"),
 471                                 full_path);
 472                         return;
 473                 }
 474
 475                 if (open_pack_index(packs->info[packs->nr].p)) {
 476                         warning(_("failed to open pack-index '%s'"),
 477                                 full_path);
 478                         close_pack(packs->info[packs->nr].p);
 479                         FREE_AND_NULL(packs->info[packs->nr].p);
 480                         return;
 481                 }
 482
 483                 packs->info[packs->nr].pack_name = xstrdup(file_name);
 484                 packs->info[packs->nr].orig_pack_int_id = packs->nr;
 485                 packs->info[packs->nr].expired = 0;
 486                 packs->nr++;
 487         }
 488 }
 489
 490 struct pack_midx_entry {
 491         struct object_id oid;
 492         uint32_t pack_int_id;
 493         time_t pack_mtime;
 494         uint64_t offset;
 495 };
 496
 497 static int midx_oid_compare(const void *_a, const void *_b)
 498 {
 499         const struct pack_midx_entry *a = (const struct pack_midx_entry *)_a;
 500         const struct pack_midx_entry *b = (const struct pack_midx_entry *)_b;
 501         int cmp = oidcmp(&a->oid, &b->oid);
 502
 503         if (cmp)
 504                 return cmp;
 505
 506         if (a->pack_mtime > b->pack_mtime)
 507                 return -1;
 508         else if (a->pack_mtime < b->pack_mtime)
 509                 return 1;
 510
 511         return a->pack_int_id - b->pack_int_id;
 512 }
 513
 514 static int nth_midxed_pack_midx_entry(struct multi_pack_index *m,
 515                                       struct pack_midx_entry *e,
 516                                       uint32_t pos)
 517 {
 518         if (pos >= m->num_objects)
 519                 return 1;
 520
 521         nth_midxed_object_oid(&e->oid, m, pos);
 522         e->pack_int_id = nth_midxed_pack_int_id(m, pos);
 523         e->offset = nth_midxed_offset(m, pos);
 524
 525         /* consider objects in midx to be from "old" packs */
 526         e->pack_mtime = 0;
 527         return 0;
 528 }
 529
 530 static void fill_pack_entry(uint32_t pack_int_id,
 531                             struct packed_git *p,
 532                             uint32_t cur_object,
 533                             struct pack_midx_entry *entry)
 534 {
 535         if (!nth_packed_object_oid(&entry->oid, p, cur_object))
 536                 die(_("failed to locate object %d in packfile"), cur_object);
 537
 538         entry->pack_int_id = pack_int_id;
 539         entry->pack_mtime = p->mtime;
 540
 541         entry->offset = nth_packed_object_offset(p, cur_object);
 542 }
 543
 544 /*
 545  * It is possible to artificially get into a state where there are many
 546  * duplicate copies of objects. That can create high memory pressure if
 547  * we are to create a list of all objects before de-duplication. To reduce
 548  * this memory pressure without a significant performance drop, automatically
 549  * group objects by the first byte of their object id. Use the IDX fanout
 550  * tables to group the data, copy to a local array, then sort.
 551  *
 552  * Copy only the de-duplicated entries (selected by most-recent modified time
 553  * of a packfile containing the object).
 554  */
 555 static struct pack_midx_entry *get_sorted_entries(struct multi_pack_index *m,
 556                                                   struct pack_info *info,
 557                                                   uint32_t nr_packs,
 558                                                   uint32_t *nr_objects)
 559 {
 560         uint32_t cur_fanout, cur_pack, cur_object;
 561         uint32_t alloc_fanout, alloc_objects, total_objects = 0;
 562         struct pack_midx_entry *entries_by_fanout = NULL;
 563         struct pack_midx_entry *deduplicated_entries = NULL;
 564         uint32_t start_pack = m ? m->num_packs : 0;
 565
 566         for (cur_pack = start_pack; cur_pack < nr_packs; cur_pack++)
 567                 total_objects += info[cur_pack].p->num_objects;
 568
 569         /*
 570          * As we de-duplicate by fanout value, we expect the fanout
 571          * slices to be evenly distributed, with some noise. Hence,
 572          * allocate slightly more than one 256th.
 573          */
 574         alloc_objects = alloc_fanout = total_objects > 3200 ? total_objects / 200 : 16;
 575
 576         ALLOC_ARRAY(entries_by_fanout, alloc_fanout);
 577         ALLOC_ARRAY(deduplicated_entries, alloc_objects);
 578         *nr_objects = 0;
 579
 580         for (cur_fanout = 0; cur_fanout < 256; cur_fanout++) {
 581                 uint32_t nr_fanout = 0;
 582
 583                 if (m) {
 584                         uint32_t start = 0, end;
 585
 586                         if (cur_fanout)
 587                                 start = ntohl(m->chunk_oid_fanout[cur_fanout - 1]);
 588                         end = ntohl(m->chunk_oid_fanout[cur_fanout]);
 589
 590                         for (cur_object = start; cur_object < end; cur_object++) {
 591                                 ALLOC_GROW(entries_by_fanout, nr_fanout + 1, alloc_fanout);
 592                                 nth_midxed_pack_midx_entry(m,
 593                                                            &entries_by_fanout[nr_fanout],
 594                                                            cur_object);
 595                                 nr_fanout++;
 596                         }
 597                 }
 598
 599                 for (cur_pack = start_pack; cur_pack < nr_packs; cur_pack++) {
 600                         uint32_t start = 0, end;
 601
 602                         if (cur_fanout)
 603                                 start = get_pack_fanout(info[cur_pack].p, cur_fanout - 1);
 604                         end = get_pack_fanout(info[cur_pack].p, cur_fanout);
 605
 606                         for (cur_object = start; cur_object < end; cur_object++) {
 607                                 ALLOC_GROW(entries_by_fanout, nr_fanout + 1, alloc_fanout);
 608                                 fill_pack_entry(cur_pack, info[cur_pack].p, cur_object, &entries_by_fanout[nr_fanout]);
 609                                 nr_fanout++;
 610                         }
 611                 }
 612
 613                 QSORT(entries_by_fanout, nr_fanout, midx_oid_compare);
 614
 615                 /*
 616                  * The batch is now sorted by OID and then mtime (descending).
 617                  * Take only the first duplicate.
 618                  */
 619                 for (cur_object = 0; cur_object < nr_fanout; cur_object++) {
 620                         if (cur_object && oideq(&entries_by_fanout[cur_object - 1].oid,
 621                                                 &entries_by_fanout[cur_object].oid))
 622                                 continue;
 623
 624                         ALLOC_GROW(deduplicated_entries, *nr_objects + 1, alloc_objects);
 625                         memcpy(&deduplicated_entries[*nr_objects],
 626                                &entries_by_fanout[cur_object],
 627                                sizeof(struct pack_midx_entry));
 628                         (*nr_objects)++;
 629                 }
 630         }
 631
 632         free(entries_by_fanout);
 633         return deduplicated_entries;
 634 }
 635
 636 static size_t write_midx_pack_names(struct hashfile *f,
 637                                     struct pack_info *info,
 638                                     uint32_t num_packs)
 639 {
 640         uint32_t i;
 641         unsigned char padding[MIDX_CHUNK_ALIGNMENT];
 642         size_t written = 0;
 643
 644         for (i = 0; i < num_packs; i++) {
 645                 size_t writelen;
 646
 647                 if (info[i].expired)
 648                         continue;
 649
 650                 if (i && strcmp(info[i].pack_name, info[i - 1].pack_name) <= 0)
 651                         BUG("incorrect pack-file order: %s before %s",
 652                             info[i - 1].pack_name,
 653                             info[i].pack_name);
 654
 655                 writelen = strlen(info[i].pack_name) + 1;
 656                 hashwrite(f, info[i].pack_name, writelen);
 657                 written += writelen;
 658         }
 659
 660         /* add padding to be aligned */
 661         i = MIDX_CHUNK_ALIGNMENT - (written % MIDX_CHUNK_ALIGNMENT);
 662         if (i < MIDX_CHUNK_ALIGNMENT) {
 663                 memset(padding, 0, sizeof(padding));
 664                 hashwrite(f, padding, i);
 665                 written += i;
 666         }
 667
 668         return written;
 669 }
 670
 671 static size_t write_midx_oid_fanout(struct hashfile *f,
 672                                     struct pack_midx_entry *objects,
 673                                     uint32_t nr_objects)
 674 {
 675         struct pack_midx_entry *list = objects;
 676         struct pack_midx_entry *last = objects + nr_objects;
 677         uint32_t count = 0;
 678         uint32_t i;
 679
 680         /*
 681         * Write the first-level table (the list is sorted,
 682         * but we use a 256-entry lookup to be able to avoid
 683         * having to do eight extra binary search iterations).
 684         */
 685         for (i = 0; i < 256; i++) {
 686                 struct pack_midx_entry *next = list;
 687
 688                 while (next < last && next->oid.hash[0] == i) {
 689                         count++;
 690                         next++;
 691                 }
 692
 693                 hashwrite_be32(f, count);
 694                 list = next;
 695         }
 696
 697         return MIDX_CHUNK_FANOUT_SIZE;
 698 }
 699
 700 static size_t write_midx_oid_lookup(struct hashfile *f, unsigned char hash_len,
 701                                     struct pack_midx_entry *objects,
 702                                     uint32_t nr_objects)
 703 {
 704         struct pack_midx_entry *list = objects;
 705         uint32_t i;
 706         size_t written = 0;
 707
 708         for (i = 0; i < nr_objects; i++) {
 709                 struct pack_midx_entry *obj = list++;
 710
 711                 if (i < nr_objects - 1) {
 712                         struct pack_midx_entry *next = list;
 713                         if (oidcmp(&obj->oid, &next->oid) >= 0)
 714                                 BUG("OIDs not in order: %s >= %s",
 715                                     oid_to_hex(&obj->oid),
 716                                     oid_to_hex(&next->oid));
 717                 }
 718
 719                 hashwrite(f, obj->oid.hash, (int)hash_len);
 720                 written += hash_len;
 721         }
 722
 723         return written;
 724 }
 725
 726 static size_t write_midx_object_offsets(struct hashfile *f, int large_offset_needed,
 727                                         uint32_t *perm,
 728                                         struct pack_midx_entry *objects, uint32_t nr_objects)
 729 {
 730         struct pack_midx_entry *list = objects;
 731         uint32_t i, nr_large_offset = 0;
 732         size_t written = 0;
 733
 734         for (i = 0; i < nr_objects; i++) {
 735                 struct pack_midx_entry *obj = list++;
 736
 737                 if (perm[obj->pack_int_id] == PACK_EXPIRED)
 738                         BUG("object %s is in an expired pack with int-id %d",
 739                             oid_to_hex(&obj->oid),
 740                             obj->pack_int_id);
 741
 742                 hashwrite_be32(f, perm[obj->pack_int_id]);
 743
 744                 if (large_offset_needed && obj->offset >> 31)
 745                         hashwrite_be32(f, MIDX_LARGE_OFFSET_NEEDED | nr_large_offset++);
 746                 else if (!large_offset_needed && obj->offset >> 32)
 747                         BUG("object %s requires a large offset (%"PRIx64") but the MIDX is not writing large offsets!",
 748                             oid_to_hex(&obj->oid),
 749                             obj->offset);
 750                 else
 751                         hashwrite_be32(f, (uint32_t)obj->offset);
 752
 753                 written += MIDX_CHUNK_OFFSET_WIDTH;
 754         }
 755
 756         return written;
 757 }
 758
 759 static size_t write_midx_large_offsets(struct hashfile *f, uint32_t nr_large_offset,
 760                                        struct pack_midx_entry *objects, uint32_t nr_objects)
 761 {
 762         struct pack_midx_entry *list = objects, *end = objects + nr_objects;
 763         size_t written = 0;
 764
 765         while (nr_large_offset) {
 766                 struct pack_midx_entry *obj;
 767                 uint64_t offset;
 768
 769                 if (list >= end)
 770                         BUG("too many large-offset objects");
 771
 772                 obj = list++;
 773                 offset = obj->offset;
 774
 775                 if (!(offset >> 31))
 776                         continue;
 777
 778                 hashwrite_be32(f, offset >> 32);
 779                 hashwrite_be32(f, offset & 0xffffffffUL);
 780                 written += 2 * sizeof(uint32_t);
 781
 782                 nr_large_offset--;
 783         }
 784
 785         return written;
 786 }
 787
 788 static int write_midx_internal(const char *object_dir, struct multi_pack_index *m,
 789                                struct string_list *packs_to_drop)
 790 {
 791         unsigned char cur_chunk, num_chunks = 0;
 792         char *midx_name;
 793         uint32_t i;
 794         struct hashfile *f = NULL;
 795         struct lock_file lk;
 796         struct pack_list packs;
 797         uint32_t *pack_perm = NULL;
 798         uint64_t written = 0;
 799         uint32_t chunk_ids[MIDX_MAX_CHUNKS + 1];
 800         uint64_t chunk_offsets[MIDX_MAX_CHUNKS + 1];
 801         uint32_t nr_entries, num_large_offsets = 0;
 802         struct pack_midx_entry *entries = NULL;
 803         int large_offsets_needed = 0;
 804         int pack_name_concat_len = 0;
 805         int dropped_packs = 0;
 806         int result = 0;
 807
 808         midx_name = get_midx_filename(object_dir);
 809         if (safe_create_leading_directories(midx_name)) {
 810                 UNLEAK(midx_name);
 811                 die_errno(_("unable to create leading directories of %s"),
 812                           midx_name);
 813         }
 814
 815         if (m)
 816                 packs.m = m;
 817         else
 818                 packs.m = load_multi_pack_index(object_dir, 1);
 819
 820         packs.nr = 0;
 821         packs.alloc = packs.m ? packs.m->num_packs : 16;
 822         packs.info = NULL;
 823         ALLOC_ARRAY(packs.info, packs.alloc);
 824
 825         if (packs.m) {
 826                 for (i = 0; i < packs.m->num_packs; i++) {
 827                         ALLOC_GROW(packs.info, packs.nr + 1, packs.alloc);
 828
 829                         packs.info[packs.nr].orig_pack_int_id = i;
 830                         packs.info[packs.nr].pack_name = xstrdup(packs.m->pack_names[i]);
 831                         packs.info[packs.nr].p = NULL;
 832                         packs.info[packs.nr].expired = 0;
 833                         packs.nr++;
 834                 }
 835         }
 836
 837         for_each_file_in_pack_dir(object_dir, add_pack_to_midx, &packs);
 838
 839         if (packs.m && packs.nr == packs.m->num_packs && !packs_to_drop)
 840                 goto cleanup;
 841
 842         entries = get_sorted_entries(packs.m, packs.info, packs.nr, &nr_entries);
 843
 844         for (i = 0; i < nr_entries; i++) {
 845                 if (entries[i].offset > 0x7fffffff)
 846                         num_large_offsets++;
 847                 if (entries[i].offset > 0xffffffff)
 848                         large_offsets_needed = 1;
 849         }
 850
 851         QSORT(packs.info, packs.nr, pack_info_compare);
 852
 853         if (packs_to_drop && packs_to_drop->nr) {
 854                 int drop_index = 0;
 855                 int missing_drops = 0;
 856
 857                 for (i = 0; i < packs.nr && drop_index < packs_to_drop->nr; i++) {
 858                         int cmp = strcmp(packs.info[i].pack_name,
 859                                          packs_to_drop->items[drop_index].string);
 860
 861                         if (!cmp) {
 862                                 drop_index++;
 863                                 packs.info[i].expired = 1;
 864                         } else if (cmp > 0) {
 865                                 error(_("did not see pack-file %s to drop"),
 866                                       packs_to_drop->items[drop_index].string);
 867                                 drop_index++;
 868                                 missing_drops++;
 869                                 i--;
 870                         } else {
 871                                 packs.info[i].expired = 0;
 872                         }
 873                 }
 874
 875                 if (missing_drops) {
 876                         result = 1;
 877                         goto cleanup;
 878                 }
 879         }
 880
 881         /*
 882          * pack_perm stores a permutation between pack-int-ids from the
 883          * previous multi-pack-index to the new one we are writing:
 884          *
 885          * pack_perm[old_id] = new_id
 886          */
 887         ALLOC_ARRAY(pack_perm, packs.nr);
 888         for (i = 0; i < packs.nr; i++) {
 889                 if (packs.info[i].expired) {
 890                         dropped_packs++;
 891                         pack_perm[packs.info[i].orig_pack_int_id] = PACK_EXPIRED;
 892                 } else {
 893                         pack_perm[packs.info[i].orig_pack_int_id] = i - dropped_packs;
 894                 }
 895         }
 896
 897         for (i = 0; i < packs.nr; i++) {
 898                 if (!packs.info[i].expired)
 899                         pack_name_concat_len += strlen(packs.info[i].pack_name) + 1;
 900         }
 901
 902         if (pack_name_concat_len % MIDX_CHUNK_ALIGNMENT)
 903                 pack_name_concat_len += MIDX_CHUNK_ALIGNMENT -
 904                                         (pack_name_concat_len % MIDX_CHUNK_ALIGNMENT);
 905
 906         hold_lock_file_for_update(&lk, midx_name, LOCK_DIE_ON_ERROR);
 907         f = hashfd(lk.tempfile->fd, lk.tempfile->filename.buf);
 908         FREE_AND_NULL(midx_name);
 909
 910         if (packs.m)
 911                 close_midx(packs.m);
 912
 913         cur_chunk = 0;
 914         num_chunks = large_offsets_needed ? 5 : 4;
 915
 916         written = write_midx_header(f, num_chunks, packs.nr - dropped_packs);
 917
 918         chunk_ids[cur_chunk] = MIDX_CHUNKID_PACKNAMES;
 919         chunk_offsets[cur_chunk] = written + (num_chunks + 1) * MIDX_CHUNKLOOKUP_WIDTH;
 920
 921         cur_chunk++;
 922         chunk_ids[cur_chunk] = MIDX_CHUNKID_OIDFANOUT;
 923         chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] + pack_name_concat_len;
 924
 925         cur_chunk++;
 926         chunk_ids[cur_chunk] = MIDX_CHUNKID_OIDLOOKUP;
 927         chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] + MIDX_CHUNK_FANOUT_SIZE;
 928
 929         cur_chunk++;
 930         chunk_ids[cur_chunk] = MIDX_CHUNKID_OBJECTOFFSETS;
 931         chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] + nr_entries * MIDX_HASH_LEN;
 932
 933         cur_chunk++;
 934         chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] + nr_entries * MIDX_CHUNK_OFFSET_WIDTH;
 935         if (large_offsets_needed) {
 936                 chunk_ids[cur_chunk] = MIDX_CHUNKID_LARGEOFFSETS;
 937
 938                 cur_chunk++;
 939                 chunk_offsets[cur_chunk] = chunk_offsets[cur_chunk - 1] +
 940                                            num_large_offsets * MIDX_CHUNK_LARGE_OFFSET_WIDTH;
 941         }
 942
 943         chunk_ids[cur_chunk] = 0;
 944
 945         for (i = 0; i <= num_chunks; i++) {
 946                 if (i && chunk_offsets[i] < chunk_offsets[i - 1])
 947                         BUG("incorrect chunk offsets: %"PRIu64" before %"PRIu64,
 948                             chunk_offsets[i - 1],
 949                             chunk_offsets[i]);
 950
 951                 if (chunk_offsets[i] % MIDX_CHUNK_ALIGNMENT)
 952                         BUG("chunk offset %"PRIu64" is not properly aligned",
 953                             chunk_offsets[i]);
 954
 955                 hashwrite_be32(f, chunk_ids[i]);
 956                 hashwrite_be32(f, chunk_offsets[i] >> 32);
 957                 hashwrite_be32(f, chunk_offsets[i]);
 958
 959                 written += MIDX_CHUNKLOOKUP_WIDTH;
 960         }
 961
 962         for (i = 0; i < num_chunks; i++) {
 963                 if (written != chunk_offsets[i])
 964                         BUG("incorrect chunk offset (%"PRIu64" != %"PRIu64") for chunk id %"PRIx32,
 965                             chunk_offsets[i],
 966                             written,
 967                             chunk_ids[i]);
 968
 969                 switch (chunk_ids[i]) {
 970                         case MIDX_CHUNKID_PACKNAMES:
 971                                 written += write_midx_pack_names(f, packs.info, packs.nr);
 972                                 break;
 973
 974                         case MIDX_CHUNKID_OIDFANOUT:
 975                                 written += write_midx_oid_fanout(f, entries, nr_entries);
 976                                 break;
 977
 978                         case MIDX_CHUNKID_OIDLOOKUP:
 979                                 written += write_midx_oid_lookup(f, MIDX_HASH_LEN, entries, nr_entries);
 980                                 break;
 981
 982                         case MIDX_CHUNKID_OBJECTOFFSETS:
 983                                 written += write_midx_object_offsets(f, large_offsets_needed, pack_perm, entries, nr_entries);
 984                                 break;
 985
 986                         case MIDX_CHUNKID_LARGEOFFSETS:
 987                                 written += write_midx_large_offsets(f, num_large_offsets, entries, nr_entries);
 988                                 break;
 989
 990                         default:
 991                                 BUG("trying to write unknown chunk id %"PRIx32,
 992                                     chunk_ids[i]);
 993                 }
 994         }
 995
 996         if (written != chunk_offsets[num_chunks])
 997                 BUG("incorrect final offset %"PRIu64" != %"PRIu64,
 998                     written,
 999                     chunk_offsets[num_chunks]);
1000
1001         finalize_hashfile(f, NULL, CSUM_FSYNC | CSUM_HASH_IN_STREAM);
1002         commit_lock_file(&lk);
1003
1004 cleanup:
1005         for (i = 0; i < packs.nr; i++) {
1006                 if (packs.info[i].p) {
1007                         close_pack(packs.info[i].p);
1008                         free(packs.info[i].p);
1009                 }
1010                 free(packs.info[i].pack_name);
1011         }
1012
1013         free(packs.info);
1014         free(entries);
1015         free(pack_perm);
1016         free(midx_name);
1017         return result;
1018 }
1019
1020 int write_midx_file(const char *object_dir)
1021 {
1022         return write_midx_internal(object_dir, NULL, NULL);
1023 }
1024
1025 void clear_midx_file(struct repository *r)
1026 {
1027         char *midx = get_midx_filename(r->objects->odb->path);
1028
1029         if (r->objects && r->objects->multi_pack_index) {
1030                 close_midx(r->objects->multi_pack_index);
1031                 r->objects->multi_pack_index = NULL;
1032         }
1033
1034         if (remove_path(midx)) {
1035                 UNLEAK(midx);
1036                 die(_("failed to clear multi-pack-index at %s"), midx);
1037         }
1038
1039         free(midx);
1040 }
1041
1042 static int verify_midx_error;
1043
1044 static void midx_report(const char *fmt, ...)
1045 {
1046         va_list ap;
1047         verify_midx_error = 1;
1048         va_start(ap, fmt);
1049         vfprintf(stderr, fmt, ap);
1050         fprintf(stderr, "\n");
1051         va_end(ap);
1052 }
1053
1054 struct pair_pos_vs_id
1055 {
1056         uint32_t pos;
1057         uint32_t pack_int_id;
1058 };
1059
1060 static int compare_pair_pos_vs_id(const void *_a, const void *_b)
1061 {
1062         struct pair_pos_vs_id *a = (struct pair_pos_vs_id *)_a;
1063         struct pair_pos_vs_id *b = (struct pair_pos_vs_id *)_b;
1064
1065         return b->pack_int_id - a->pack_int_id;
1066 }
1067
1068 /*
1069  * Limit calls to display_progress() for performance reasons.
1070  * The interval here was arbitrarily chosen.
1071  */
1072 #define SPARSE_PROGRESS_INTERVAL (1 << 12)
1073 #define midx_display_sparse_progress(progress, n) \
1074         do { \
1075                 uint64_t _n = (n); \
1076                 if ((_n & (SPARSE_PROGRESS_INTERVAL - 1)) == 0) \
1077                         display_progress(progress, _n); \
1078         } while (0)
1079
1080 int verify_midx_file(struct repository *r, const char *object_dir)
1081 {
1082         struct pair_pos_vs_id *pairs = NULL;
1083         uint32_t i;
1084         struct progress *progress;
1085         struct multi_pack_index *m = load_multi_pack_index(object_dir, 1);
1086         verify_midx_error = 0;
1087
1088         if (!m)
1089                 return 0;
1090
1091         progress = start_progress(_("Looking for referenced packfiles"),
1092                                   m->num_packs);
1093         for (i = 0; i < m->num_packs; i++) {
1094                 if (prepare_midx_pack(r, m, i))
1095                         midx_report("failed to load pack in position %d", i);
1096
1097                 display_progress(progress, i + 1);
1098         }
1099         stop_progress(&progress);
1100
1101         for (i = 0; i < 255; i++) {
1102                 uint32_t oid_fanout1 = ntohl(m->chunk_oid_fanout[i]);
1103                 uint32_t oid_fanout2 = ntohl(m->chunk_oid_fanout[i + 1]);
1104
1105                 if (oid_fanout1 > oid_fanout2)
1106                         midx_report(_("oid fanout out of order: fanout[%d] = %"PRIx32" > %"PRIx32" = fanout[%d]"),
1107                                     i, oid_fanout1, oid_fanout2, i + 1);
1108         }
1109
1110         progress = start_sparse_progress(_("Verifying OID order in MIDX"),
1111                                          m->num_objects - 1);
1112         for (i = 0; i < m->num_objects - 1; i++) {
1113                 struct object_id oid1, oid2;
1114
1115                 nth_midxed_object_oid(&oid1, m, i);
1116                 nth_midxed_object_oid(&oid2, m, i + 1);
1117
1118                 if (oidcmp(&oid1, &oid2) >= 0)
1119                         midx_report(_("oid lookup out of order: oid[%d] = %s >= %s = oid[%d]"),
1120                                     i, oid_to_hex(&oid1), oid_to_hex(&oid2), i + 1);
1121
1122                 midx_display_sparse_progress(progress, i + 1);
1123         }
1124         stop_progress(&progress);
1125
1126         /*
1127          * Create an array mapping each object to its packfile id.  Sort it
1128          * to group the objects by packfile.  Use this permutation to visit
1129          * each of the objects and only require 1 packfile to be open at a
1130          * time.
1131          */
1132         ALLOC_ARRAY(pairs, m->num_objects);
1133         for (i = 0; i < m->num_objects; i++) {
1134                 pairs[i].pos = i;
1135                 pairs[i].pack_int_id = nth_midxed_pack_int_id(m, i);
1136         }
1137
1138         progress = start_sparse_progress(_("Sorting objects by packfile"),
1139                                          m->num_objects);
1140         display_progress(progress, 0); /* TODO: Measure QSORT() progress */
1141         QSORT(pairs, m->num_objects, compare_pair_pos_vs_id);
1142         stop_progress(&progress);
1143
1144         progress = start_sparse_progress(_("Verifying object offsets"), m->num_objects);
1145         for (i = 0; i < m->num_objects; i++) {
1146                 struct object_id oid;
1147                 struct pack_entry e;
1148                 off_t m_offset, p_offset;
1149
1150                 if (i > 0 && pairs[i-1].pack_int_id != pairs[i].pack_int_id &&
1151                     m->packs[pairs[i-1].pack_int_id])
1152                 {
1153                         close_pack_fd(m->packs[pairs[i-1].pack_int_id]);
1154                         close_pack_index(m->packs[pairs[i-1].pack_int_id]);
1155                 }
1156
1157                 nth_midxed_object_oid(&oid, m, pairs[i].pos);
1158
1159                 if (!fill_midx_entry(r, &oid, &e, m)) {
1160                         midx_report(_("failed to load pack entry for oid[%d] = %s"),
1161                                     pairs[i].pos, oid_to_hex(&oid));
1162                         continue;
1163                 }
1164
1165                 if (open_pack_index(e.p)) {
1166                         midx_report(_("failed to load pack-index for packfile %s"),
1167                                     e.p->pack_name);
1168                         break;
1169                 }
1170
1171                 m_offset = e.offset;
1172                 p_offset = find_pack_entry_one(oid.hash, e.p);
1173
1174                 if (m_offset != p_offset)
1175                         midx_report(_("incorrect object offset for oid[%d] = %s: %"PRIx64" != %"PRIx64),
1176                                     pairs[i].pos, oid_to_hex(&oid), m_offset, p_offset);
1177
1178                 midx_display_sparse_progress(progress, i + 1);
1179         }
1180         stop_progress(&progress);
1181
1182         free(pairs);
1183
1184         return verify_midx_error;
1185 }
1186
1187 int expire_midx_packs(struct repository *r, const char *object_dir)
1188 {
1189         uint32_t i, *count, result = 0;
1190         struct string_list packs_to_drop = STRING_LIST_INIT_DUP;
1191         struct multi_pack_index *m = load_multi_pack_index(object_dir, 1);
1192
1193         if (!m)
1194                 return 0;
1195
1196         count = xcalloc(m->num_packs, sizeof(uint32_t));
1197         for (i = 0; i < m->num_objects; i++) {
1198                 int pack_int_id = nth_midxed_pack_int_id(m, i);
1199                 count[pack_int_id]++;
1200         }
1201
1202         for (i = 0; i < m->num_packs; i++) {
1203                 char *pack_name;
1204
1205                 if (count[i])
1206                         continue;
1207
1208                 if (prepare_midx_pack(r, m, i))
1209                         continue;
1210
1211                 if (m->packs[i]->pack_keep)
1212                         continue;
1213
1214                 pack_name = xstrdup(m->packs[i]->pack_name);
1215                 close_pack(m->packs[i]);
1216
1217                 string_list_insert(&packs_to_drop, m->pack_names[i]);
1218                 unlink_pack_path(pack_name, 0);
1219                 free(pack_name);
1220         }
1221
1222         free(count);
1223
1224         if (packs_to_drop.nr)
1225                 result = write_midx_internal(object_dir, m, &packs_to_drop);
1226
1227         string_list_clear(&packs_to_drop, 0);
1228         return result;
1229 }
1230
1231 struct repack_info {
1232         timestamp_t mtime;
1233         uint32_t referenced_objects;
1234         uint32_t pack_int_id;
1235 };
1236
1237 static int compare_by_mtime(const void *a_, const void *b_)
1238 {
1239         const struct repack_info *a, *b;
1240
1241         a = (const struct repack_info *)a_;
1242         b = (const struct repack_info *)b_;
1243
1244         if (a->mtime < b->mtime)
1245                 return -1;
1246         if (a->mtime > b->mtime)
1247                 return 1;
1248         return 0;
1249 }
1250
1251 static int fill_included_packs_all(struct multi_pack_index *m,
1252                                    unsigned char *include_pack)
1253 {
1254         uint32_t i;
1255
1256         for (i = 0; i < m->num_packs; i++)
1257                 include_pack[i] = 1;
1258
1259         return m->num_packs < 2;
1260 }
1261
1262 static int fill_included_packs_batch(struct repository *r,
1263                                      struct multi_pack_index *m,
1264                                      unsigned char *include_pack,
1265                                      size_t batch_size)
1266 {
1267         uint32_t i, packs_to_repack;
1268         size_t total_size;
1269         struct repack_info *pack_info = xcalloc(m->num_packs, sizeof(struct repack_info));
1270
1271         for (i = 0; i < m->num_packs; i++) {
1272                 pack_info[i].pack_int_id = i;
1273
1274                 if (prepare_midx_pack(r, m, i))
1275                         continue;
1276
1277                 pack_info[i].mtime = m->packs[i]->mtime;
1278         }
1279
1280         for (i = 0; batch_size && i < m->num_objects; i++) {
1281                 uint32_t pack_int_id = nth_midxed_pack_int_id(m, i);
1282                 pack_info[pack_int_id].referenced_objects++;
1283         }
1284
1285         QSORT(pack_info, m->num_packs, compare_by_mtime);
1286
1287         total_size = 0;
1288         packs_to_repack = 0;
1289         for (i = 0; total_size < batch_size && i < m->num_packs; i++) {
1290                 int pack_int_id = pack_info[i].pack_int_id;
1291                 struct packed_git *p = m->packs[pack_int_id];
1292                 size_t expected_size;
1293
1294                 if (!p)
1295                         continue;
1296                 if (open_pack_index(p) || !p->num_objects)
1297                         continue;
1298
1299                 expected_size = (size_t)(p->pack_size
1300                                          * pack_info[i].referenced_objects);
1301                 expected_size /= p->num_objects;
1302
1303                 if (expected_size >= batch_size)
1304                         continue;
1305
1306                 packs_to_repack++;
1307                 total_size += expected_size;
1308                 include_pack[pack_int_id] = 1;
1309         }
1310
1311         free(pack_info);
1312
1313         if (total_size < batch_size || packs_to_repack < 2)
1314                 return 1;
1315
1316         return 0;
1317 }
1318
1319 int midx_repack(struct repository *r, const char *object_dir, size_t batch_size)
1320 {
1321         int result = 0;
1322         uint32_t i;
1323         unsigned char *include_pack;
1324         struct child_process cmd = CHILD_PROCESS_INIT;
1325         struct strbuf base_name = STRBUF_INIT;
1326         struct multi_pack_index *m = load_multi_pack_index(object_dir, 1);
1327
1328         if (!m)
1329                 return 0;
1330
1331         include_pack = xcalloc(m->num_packs, sizeof(unsigned char));
1332
1333         if (batch_size) {
1334                 if (fill_included_packs_batch(r, m, include_pack, batch_size))
1335                         goto cleanup;
1336         } else if (fill_included_packs_all(m, include_pack))
1337                 goto cleanup;
1338
1339         argv_array_push(&cmd.args, "pack-objects");
1340
1341         strbuf_addstr(&base_name, object_dir);
1342         strbuf_addstr(&base_name, "/pack/pack");
1343         argv_array_push(&cmd.args, base_name.buf);
1344         strbuf_release(&base_name);
1345
1346         cmd.git_cmd = 1;
1347         cmd.in = cmd.out = -1;
1348
1349         if (start_command(&cmd)) {
1350                 error(_("could not start pack-objects"));
1351                 result = 1;
1352                 goto cleanup;
1353         }
1354
1355         for (i = 0; i < m->num_objects; i++) {
1356                 struct object_id oid;
1357                 uint32_t pack_int_id = nth_midxed_pack_int_id(m, i);
1358
1359                 if (!include_pack[pack_int_id])
1360                         continue;
1361
1362                 nth_midxed_object_oid(&oid, m, i);
1363                 xwrite(cmd.in, oid_to_hex(&oid), the_hash_algo->hexsz);
1364                 xwrite(cmd.in, "\n", 1);
1365         }
1366         close(cmd.in);
1367
1368         if (finish_command(&cmd)) {
1369                 error(_("could not finish pack-objects"));
1370                 result = 1;
1371                 goto cleanup;
1372         }
1373
1374         result = write_midx_internal(object_dir, m, NULL);
1375         m = NULL;
1376
1377 cleanup:
1378         if (m)
1379                 close_midx(m);
1380         free(include_pack);
1381         return result;
1382 }