git.oblomov.eu Git - git/blob - builtin/index-pack.c

   1 #include "builtin.h"
   2 #include "config.h"
   3 #include "delta.h"
   4 #include "pack.h"
   5 #include "csum-file.h"
   6 #include "blob.h"
   7 #include "commit.h"
   8 #include "tag.h"
   9 #include "tree.h"
  10 #include "progress.h"
  11 #include "fsck.h"
  12 #include "exec-cmd.h"
  13 #include "streaming.h"
  14 #include "thread-utils.h"
  15 #include "packfile.h"
  16 #include "object-store.h"
  17 #include "promisor-remote.h"
  18
  19 static const char index_pack_usage[] =
  20 "git index-pack [-v] [-o <index-file>] [--keep | --keep=<msg>] [--verify] [--strict] (<pack-file> | --stdin [--fix-thin] [<pack-file>])";
  21
  22 struct object_entry {
  23         struct pack_idx_entry idx;
  24         unsigned long size;
  25         unsigned char hdr_size;
  26         signed char type;
  27         signed char real_type;
  28 };
  29
  30 struct object_stat {
  31         unsigned delta_depth;
  32         int base_object_no;
  33 };
  34
  35 struct base_data {
  36         /* Initialized by make_base(). */
  37         struct base_data *base;
  38         struct object_entry *obj;
  39         int ref_first, ref_last;
  40         int ofs_first, ofs_last;
  41         /*
  42          * Threads should increment retain_data if they are about to call
  43          * patch_delta() using this struct's data as a base, and decrement this
  44          * when they are done. While retain_data is nonzero, this struct's data
  45          * will not be freed even if the delta base cache limit is exceeded.
  46          */
  47         int retain_data;
  48         /*
  49          * The number of direct children that have not been fully processed
  50          * (entered work_head, entered done_head, left done_head). When this
  51          * number reaches zero, this struct base_data can be freed.
  52          */
  53         int children_remaining;
  54
  55         /* Not initialized by make_base(). */
  56         struct list_head list;
  57         void *data;
  58         unsigned long size;
  59 };
  60
  61 /*
  62  * Stack of struct base_data that have unprocessed children.
  63  * threaded_second_pass() uses this as a source of work (the other being the
  64  * objects array).
  65  *
  66  * Guarded by work_mutex.
  67  */
  68 static LIST_HEAD(work_head);
  69
  70 /*
  71  * Stack of struct base_data that have children, all of whom have been
  72  * processed or are being processed, and at least one child is being processed.
  73  * These struct base_data must be kept around until the last child is
  74  * processed.
  75  *
  76  * Guarded by work_mutex.
  77  */
  78 static LIST_HEAD(done_head);
  79
  80 /*
  81  * All threads share one delta base cache.
  82  *
  83  * base_cache_used is guarded by work_mutex, and base_cache_limit is read-only
  84  * in a thread.
  85  */
  86 static size_t base_cache_used;
  87 static size_t base_cache_limit;
  88
  89 struct thread_local {
  90         pthread_t thread;
  91         int pack_fd;
  92 };
  93
  94 /* Remember to update object flag allocation in object.h */
  95 #define FLAG_LINK (1u<<20)
  96 #define FLAG_CHECKED (1u<<21)
  97
  98 struct ofs_delta_entry {
  99         off_t offset;
 100         int obj_no;
 101 };
 102
 103 struct ref_delta_entry {
 104         struct object_id oid;
 105         int obj_no;
 106 };
 107
 108 static struct object_entry *objects;
 109 static struct object_stat *obj_stat;
 110 static struct ofs_delta_entry *ofs_deltas;
 111 static struct ref_delta_entry *ref_deltas;
 112 static struct thread_local nothread_data;
 113 static int nr_objects;
 114 static int nr_ofs_deltas;
 115 static int nr_ref_deltas;
 116 static int ref_deltas_alloc;
 117 static int nr_resolved_deltas;
 118 static int nr_threads;
 119
 120 static int from_stdin;
 121 static int strict;
 122 static int do_fsck_object;
 123 static struct fsck_options fsck_options = FSCK_OPTIONS_STRICT;
 124 static int verbose;
 125 static int show_resolving_progress;
 126 static int show_stat;
 127 static int check_self_contained_and_connected;
 128
 129 static struct progress *progress;
 130
 131 /* We always read in 4kB chunks. */
 132 static unsigned char input_buffer[4096];
 133 static unsigned int input_offset, input_len;
 134 static off_t consumed_bytes;
 135 static off_t max_input_size;
 136 static unsigned deepest_delta;
 137 static git_hash_ctx input_ctx;
 138 static uint32_t input_crc32;
 139 static int input_fd, output_fd;
 140 static const char *curr_pack;
 141
 142 static struct thread_local *thread_data;
 143 static int nr_dispatched;
 144 static int threads_active;
 145
 146 static pthread_mutex_t read_mutex;
 147 #define read_lock()             lock_mutex(&read_mutex)
 148 #define read_unlock()           unlock_mutex(&read_mutex)
 149
 150 static pthread_mutex_t counter_mutex;
 151 #define counter_lock()          lock_mutex(&counter_mutex)
 152 #define counter_unlock()        unlock_mutex(&counter_mutex)
 153
 154 static pthread_mutex_t work_mutex;
 155 #define work_lock()             lock_mutex(&work_mutex)
 156 #define work_unlock()           unlock_mutex(&work_mutex)
 157
 158 static pthread_mutex_t deepest_delta_mutex;
 159 #define deepest_delta_lock()    lock_mutex(&deepest_delta_mutex)
 160 #define deepest_delta_unlock()  unlock_mutex(&deepest_delta_mutex)
 161
 162 static pthread_mutex_t type_cas_mutex;
 163 #define type_cas_lock()         lock_mutex(&type_cas_mutex)
 164 #define type_cas_unlock()       unlock_mutex(&type_cas_mutex)
 165
 166 static pthread_key_t key;
 167
 168 static inline void lock_mutex(pthread_mutex_t *mutex)
 169 {
 170         if (threads_active)
 171                 pthread_mutex_lock(mutex);
 172 }
 173
 174 static inline void unlock_mutex(pthread_mutex_t *mutex)
 175 {
 176         if (threads_active)
 177                 pthread_mutex_unlock(mutex);
 178 }
 179
 180 /*
 181  * Mutex and conditional variable can't be statically-initialized on Windows.
 182  */
 183 static void init_thread(void)
 184 {
 185         int i;
 186         init_recursive_mutex(&read_mutex);
 187         pthread_mutex_init(&counter_mutex, NULL);
 188         pthread_mutex_init(&work_mutex, NULL);
 189         pthread_mutex_init(&type_cas_mutex, NULL);
 190         if (show_stat)
 191                 pthread_mutex_init(&deepest_delta_mutex, NULL);
 192         pthread_key_create(&key, NULL);
 193         thread_data = xcalloc(nr_threads, sizeof(*thread_data));
 194         for (i = 0; i < nr_threads; i++) {
 195                 thread_data[i].pack_fd = open(curr_pack, O_RDONLY);
 196                 if (thread_data[i].pack_fd == -1)
 197                         die_errno(_("unable to open %s"), curr_pack);
 198         }
 199
 200         threads_active = 1;
 201 }
 202
 203 static void cleanup_thread(void)
 204 {
 205         int i;
 206         if (!threads_active)
 207                 return;
 208         threads_active = 0;
 209         pthread_mutex_destroy(&read_mutex);
 210         pthread_mutex_destroy(&counter_mutex);
 211         pthread_mutex_destroy(&work_mutex);
 212         pthread_mutex_destroy(&type_cas_mutex);
 213         if (show_stat)
 214                 pthread_mutex_destroy(&deepest_delta_mutex);
 215         for (i = 0; i < nr_threads; i++)
 216                 close(thread_data[i].pack_fd);
 217         pthread_key_delete(key);
 218         free(thread_data);
 219 }
 220
 221 static int mark_link(struct object *obj, int type, void *data, struct fsck_options *options)
 222 {
 223         if (!obj)
 224                 return -1;
 225
 226         if (type != OBJ_ANY && obj->type != type)
 227                 die(_("object type mismatch at %s"), oid_to_hex(&obj->oid));
 228
 229         obj->flags |= FLAG_LINK;
 230         return 0;
 231 }
 232
 233 /* The content of each linked object must have been checked
 234    or it must be already present in the object database */
 235 static unsigned check_object(struct object *obj)
 236 {
 237         if (!obj)
 238                 return 0;
 239
 240         if (!(obj->flags & FLAG_LINK))
 241                 return 0;
 242
 243         if (!(obj->flags & FLAG_CHECKED)) {
 244                 unsigned long size;
 245                 int type = oid_object_info(the_repository, &obj->oid, &size);
 246                 if (type <= 0)
 247                         die(_("did not receive expected object %s"),
 248                               oid_to_hex(&obj->oid));
 249                 if (type != obj->type)
 250                         die(_("object %s: expected type %s, found %s"),
 251                             oid_to_hex(&obj->oid),
 252                             type_name(obj->type), type_name(type));
 253                 obj->flags |= FLAG_CHECKED;
 254                 return 1;
 255         }
 256
 257         return 0;
 258 }
 259
 260 static unsigned check_objects(void)
 261 {
 262         unsigned i, max, foreign_nr = 0;
 263
 264         max = get_max_object_index();
 265
 266         if (verbose)
 267                 progress = start_delayed_progress(_("Checking objects"), max);
 268
 269         for (i = 0; i < max; i++) {
 270                 foreign_nr += check_object(get_indexed_object(i));
 271                 display_progress(progress, i + 1);
 272         }
 273
 274         stop_progress(&progress);
 275         return foreign_nr;
 276 }
 277
 278
 279 /* Discard current buffer used content. */
 280 static void flush(void)
 281 {
 282         if (input_offset) {
 283                 if (output_fd >= 0)
 284                         write_or_die(output_fd, input_buffer, input_offset);
 285                 the_hash_algo->update_fn(&input_ctx, input_buffer, input_offset);
 286                 memmove(input_buffer, input_buffer + input_offset, input_len);
 287                 input_offset = 0;
 288         }
 289 }
 290
 291 /*
 292  * Make sure at least "min" bytes are available in the buffer, and
 293  * return the pointer to the buffer.
 294  */
 295 static void *fill(int min)
 296 {
 297         if (min <= input_len)
 298                 return input_buffer + input_offset;
 299         if (min > sizeof(input_buffer))
 300                 die(Q_("cannot fill %d byte",
 301                        "cannot fill %d bytes",
 302                        min),
 303                     min);
 304         flush();
 305         do {
 306                 ssize_t ret = xread(input_fd, input_buffer + input_len,
 307                                 sizeof(input_buffer) - input_len);
 308                 if (ret <= 0) {
 309                         if (!ret)
 310                                 die(_("early EOF"));
 311                         die_errno(_("read error on input"));
 312                 }
 313                 input_len += ret;
 314                 if (from_stdin)
 315                         display_throughput(progress, consumed_bytes + input_len);
 316         } while (input_len < min);
 317         return input_buffer;
 318 }
 319
 320 static void use(int bytes)
 321 {
 322         if (bytes > input_len)
 323                 die(_("used more bytes than were available"));
 324         input_crc32 = crc32(input_crc32, input_buffer + input_offset, bytes);
 325         input_len -= bytes;
 326         input_offset += bytes;
 327
 328         /* make sure off_t is sufficiently large not to wrap */
 329         if (signed_add_overflows(consumed_bytes, bytes))
 330                 die(_("pack too large for current definition of off_t"));
 331         consumed_bytes += bytes;
 332         if (max_input_size && consumed_bytes > max_input_size)
 333                 die(_("pack exceeds maximum allowed size"));
 334 }
 335
 336 static const char *open_pack_file(const char *pack_name)
 337 {
 338         if (from_stdin) {
 339                 input_fd = 0;
 340                 if (!pack_name) {
 341                         struct strbuf tmp_file = STRBUF_INIT;
 342                         output_fd = odb_mkstemp(&tmp_file,
 343                                                 "pack/tmp_pack_XXXXXX");
 344                         pack_name = strbuf_detach(&tmp_file, NULL);
 345                 } else {
 346                         output_fd = open(pack_name, O_CREAT|O_EXCL|O_RDWR, 0600);
 347                         if (output_fd < 0)
 348                                 die_errno(_("unable to create '%s'"), pack_name);
 349                 }
 350                 nothread_data.pack_fd = output_fd;
 351         } else {
 352                 input_fd = open(pack_name, O_RDONLY);
 353                 if (input_fd < 0)
 354                         die_errno(_("cannot open packfile '%s'"), pack_name);
 355                 output_fd = -1;
 356                 nothread_data.pack_fd = input_fd;
 357         }
 358         the_hash_algo->init_fn(&input_ctx);
 359         return pack_name;
 360 }
 361
 362 static void parse_pack_header(void)
 363 {
 364         struct pack_header *hdr = fill(sizeof(struct pack_header));
 365
 366         /* Header consistency check */
 367         if (hdr->hdr_signature != htonl(PACK_SIGNATURE))
 368                 die(_("pack signature mismatch"));
 369         if (!pack_version_ok(hdr->hdr_version))
 370                 die(_("pack version %"PRIu32" unsupported"),
 371                         ntohl(hdr->hdr_version));
 372
 373         nr_objects = ntohl(hdr->hdr_entries);
 374         use(sizeof(struct pack_header));
 375 }
 376
 377 static NORETURN void bad_object(off_t offset, const char *format,
 378                        ...) __attribute__((format (printf, 2, 3)));
 379
 380 static NORETURN void bad_object(off_t offset, const char *format, ...)
 381 {
 382         va_list params;
 383         char buf[1024];
 384
 385         va_start(params, format);
 386         vsnprintf(buf, sizeof(buf), format, params);
 387         va_end(params);
 388         die(_("pack has bad object at offset %"PRIuMAX": %s"),
 389             (uintmax_t)offset, buf);
 390 }
 391
 392 static inline struct thread_local *get_thread_data(void)
 393 {
 394         if (HAVE_THREADS) {
 395                 if (threads_active)
 396                         return pthread_getspecific(key);
 397                 assert(!threads_active &&
 398                        "This should only be reached when all threads are gone");
 399         }
 400         return &nothread_data;
 401 }
 402
 403 static void set_thread_data(struct thread_local *data)
 404 {
 405         if (threads_active)
 406                 pthread_setspecific(key, data);
 407 }
 408
 409 static void free_base_data(struct base_data *c)
 410 {
 411         if (c->data) {
 412                 FREE_AND_NULL(c->data);
 413                 base_cache_used -= c->size;
 414         }
 415 }
 416
 417 static void prune_base_data(struct base_data *retain)
 418 {
 419         struct list_head *pos;
 420
 421         if (base_cache_used <= base_cache_limit)
 422                 return;
 423
 424         list_for_each_prev(pos, &done_head) {
 425                 struct base_data *b = list_entry(pos, struct base_data, list);
 426                 if (b->retain_data || b == retain)
 427                         continue;
 428                 if (b->data) {
 429                         free_base_data(b);
 430                         if (base_cache_used <= base_cache_limit)
 431                                 return;
 432                 }
 433         }
 434
 435         list_for_each_prev(pos, &work_head) {
 436                 struct base_data *b = list_entry(pos, struct base_data, list);
 437                 if (b->retain_data || b == retain)
 438                         continue;
 439                 if (b->data) {
 440                         free_base_data(b);
 441                         if (base_cache_used <= base_cache_limit)
 442                                 return;
 443                 }
 444         }
 445 }
 446
 447 static int is_delta_type(enum object_type type)
 448 {
 449         return (type == OBJ_REF_DELTA || type == OBJ_OFS_DELTA);
 450 }
 451
 452 static void *unpack_entry_data(off_t offset, unsigned long size,
 453                                enum object_type type, struct object_id *oid)
 454 {
 455         static char fixed_buf[8192];
 456         int status;
 457         git_zstream stream;
 458         void *buf;
 459         git_hash_ctx c;
 460         char hdr[32];
 461         int hdrlen;
 462
 463         if (!is_delta_type(type)) {
 464                 hdrlen = xsnprintf(hdr, sizeof(hdr), "%s %"PRIuMAX,
 465                                    type_name(type),(uintmax_t)size) + 1;
 466                 the_hash_algo->init_fn(&c);
 467                 the_hash_algo->update_fn(&c, hdr, hdrlen);
 468         } else
 469                 oid = NULL;
 470         if (type == OBJ_BLOB && size > big_file_threshold)
 471                 buf = fixed_buf;
 472         else
 473                 buf = xmallocz(size);
 474
 475         memset(&stream, 0, sizeof(stream));
 476         git_inflate_init(&stream);
 477         stream.next_out = buf;
 478         stream.avail_out = buf == fixed_buf ? sizeof(fixed_buf) : size;
 479
 480         do {
 481                 unsigned char *last_out = stream.next_out;
 482                 stream.next_in = fill(1);
 483                 stream.avail_in = input_len;
 484                 status = git_inflate(&stream, 0);
 485                 use(input_len - stream.avail_in);
 486                 if (oid)
 487                         the_hash_algo->update_fn(&c, last_out, stream.next_out - last_out);
 488                 if (buf == fixed_buf) {
 489                         stream.next_out = buf;
 490                         stream.avail_out = sizeof(fixed_buf);
 491                 }
 492         } while (status == Z_OK);
 493         if (stream.total_out != size || status != Z_STREAM_END)
 494                 bad_object(offset, _("inflate returned %d"), status);
 495         git_inflate_end(&stream);
 496         if (oid)
 497                 the_hash_algo->final_fn(oid->hash, &c);
 498         return buf == fixed_buf ? NULL : buf;
 499 }
 500
 501 static void *unpack_raw_entry(struct object_entry *obj,
 502                               off_t *ofs_offset,
 503                               struct object_id *ref_oid,
 504                               struct object_id *oid)
 505 {
 506         unsigned char *p;
 507         unsigned long size, c;
 508         off_t base_offset;
 509         unsigned shift;
 510         void *data;
 511
 512         obj->idx.offset = consumed_bytes;
 513         input_crc32 = crc32(0, NULL, 0);
 514
 515         p = fill(1);
 516         c = *p;
 517         use(1);
 518         obj->type = (c >> 4) & 7;
 519         size = (c & 15);
 520         shift = 4;
 521         while (c & 0x80) {
 522                 p = fill(1);
 523                 c = *p;
 524                 use(1);
 525                 size += (c & 0x7f) << shift;
 526                 shift += 7;
 527         }
 528         obj->size = size;
 529
 530         switch (obj->type) {
 531         case OBJ_REF_DELTA:
 532                 hashcpy(ref_oid->hash, fill(the_hash_algo->rawsz));
 533                 use(the_hash_algo->rawsz);
 534                 break;
 535         case OBJ_OFS_DELTA:
 536                 p = fill(1);
 537                 c = *p;
 538                 use(1);
 539                 base_offset = c & 127;
 540                 while (c & 128) {
 541                         base_offset += 1;
 542                         if (!base_offset || MSB(base_offset, 7))
 543                                 bad_object(obj->idx.offset, _("offset value overflow for delta base object"));
 544                         p = fill(1);
 545                         c = *p;
 546                         use(1);
 547                         base_offset = (base_offset << 7) + (c & 127);
 548                 }
 549                 *ofs_offset = obj->idx.offset - base_offset;
 550                 if (*ofs_offset <= 0 || *ofs_offset >= obj->idx.offset)
 551                         bad_object(obj->idx.offset, _("delta base offset is out of bound"));
 552                 break;
 553         case OBJ_COMMIT:
 554         case OBJ_TREE:
 555         case OBJ_BLOB:
 556         case OBJ_TAG:
 557                 break;
 558         default:
 559                 bad_object(obj->idx.offset, _("unknown object type %d"), obj->type);
 560         }
 561         obj->hdr_size = consumed_bytes - obj->idx.offset;
 562
 563         data = unpack_entry_data(obj->idx.offset, obj->size, obj->type, oid);
 564         obj->idx.crc32 = input_crc32;
 565         return data;
 566 }
 567
 568 static void *unpack_data(struct object_entry *obj,
 569                          int (*consume)(const unsigned char *, unsigned long, void *),
 570                          void *cb_data)
 571 {
 572         off_t from = obj[0].idx.offset + obj[0].hdr_size;
 573         off_t len = obj[1].idx.offset - from;
 574         unsigned char *data, *inbuf;
 575         git_zstream stream;
 576         int status;
 577
 578         data = xmallocz(consume ? 64*1024 : obj->size);
 579         inbuf = xmalloc((len < 64*1024) ? (int)len : 64*1024);
 580
 581         memset(&stream, 0, sizeof(stream));
 582         git_inflate_init(&stream);
 583         stream.next_out = data;
 584         stream.avail_out = consume ? 64*1024 : obj->size;
 585
 586         do {
 587                 ssize_t n = (len < 64*1024) ? (ssize_t)len : 64*1024;
 588                 n = xpread(get_thread_data()->pack_fd, inbuf, n, from);
 589                 if (n < 0)
 590                         die_errno(_("cannot pread pack file"));
 591                 if (!n)
 592                         die(Q_("premature end of pack file, %"PRIuMAX" byte missing",
 593                                "premature end of pack file, %"PRIuMAX" bytes missing",
 594                                (unsigned int)len),
 595                             (uintmax_t)len);
 596                 from += n;
 597                 len -= n;
 598                 stream.next_in = inbuf;
 599                 stream.avail_in = n;
 600                 if (!consume)
 601                         status = git_inflate(&stream, 0);
 602                 else {
 603                         do {
 604                                 status = git_inflate(&stream, 0);
 605                                 if (consume(data, stream.next_out - data, cb_data)) {
 606                                         free(inbuf);
 607                                         free(data);
 608                                         return NULL;
 609                                 }
 610                                 stream.next_out = data;
 611                                 stream.avail_out = 64*1024;
 612                         } while (status == Z_OK && stream.avail_in);
 613                 }
 614         } while (len && status == Z_OK && !stream.avail_in);
 615
 616         /* This has been inflated OK when first encountered, so... */
 617         if (status != Z_STREAM_END || stream.total_out != obj->size)
 618                 die(_("serious inflate inconsistency"));
 619
 620         git_inflate_end(&stream);
 621         free(inbuf);
 622         if (consume) {
 623                 FREE_AND_NULL(data);
 624         }
 625         return data;
 626 }
 627
 628 static void *get_data_from_pack(struct object_entry *obj)
 629 {
 630         return unpack_data(obj, NULL, NULL);
 631 }
 632
 633 static int compare_ofs_delta_bases(off_t offset1, off_t offset2,
 634                                    enum object_type type1,
 635                                    enum object_type type2)
 636 {
 637         int cmp = type1 - type2;
 638         if (cmp)
 639                 return cmp;
 640         return offset1 < offset2 ? -1 :
 641                offset1 > offset2 ?  1 :
 642                0;
 643 }
 644
 645 static int find_ofs_delta(const off_t offset)
 646 {
 647         int first = 0, last = nr_ofs_deltas;
 648
 649         while (first < last) {
 650                 int next = first + (last - first) / 2;
 651                 struct ofs_delta_entry *delta = &ofs_deltas[next];
 652                 int cmp;
 653
 654                 cmp = compare_ofs_delta_bases(offset, delta->offset,
 655                                               OBJ_OFS_DELTA,
 656                                               objects[delta->obj_no].type);
 657                 if (!cmp)
 658                         return next;
 659                 if (cmp < 0) {
 660                         last = next;
 661                         continue;
 662                 }
 663                 first = next+1;
 664         }
 665         return -first-1;
 666 }
 667
 668 static void find_ofs_delta_children(off_t offset,
 669                                     int *first_index, int *last_index)
 670 {
 671         int first = find_ofs_delta(offset);
 672         int last = first;
 673         int end = nr_ofs_deltas - 1;
 674
 675         if (first < 0) {
 676                 *first_index = 0;
 677                 *last_index = -1;
 678                 return;
 679         }
 680         while (first > 0 && ofs_deltas[first - 1].offset == offset)
 681                 --first;
 682         while (last < end && ofs_deltas[last + 1].offset == offset)
 683                 ++last;
 684         *first_index = first;
 685         *last_index = last;
 686 }
 687
 688 static int compare_ref_delta_bases(const struct object_id *oid1,
 689                                    const struct object_id *oid2,
 690                                    enum object_type type1,
 691                                    enum object_type type2)
 692 {
 693         int cmp = type1 - type2;
 694         if (cmp)
 695                 return cmp;
 696         return oidcmp(oid1, oid2);
 697 }
 698
 699 static int find_ref_delta(const struct object_id *oid)
 700 {
 701         int first = 0, last = nr_ref_deltas;
 702
 703         while (first < last) {
 704                 int next = first + (last - first) / 2;
 705                 struct ref_delta_entry *delta = &ref_deltas[next];
 706                 int cmp;
 707
 708                 cmp = compare_ref_delta_bases(oid, &delta->oid,
 709                                               OBJ_REF_DELTA,
 710                                               objects[delta->obj_no].type);
 711                 if (!cmp)
 712                         return next;
 713                 if (cmp < 0) {
 714                         last = next;
 715                         continue;
 716                 }
 717                 first = next+1;
 718         }
 719         return -first-1;
 720 }
 721
 722 static void find_ref_delta_children(const struct object_id *oid,
 723                                     int *first_index, int *last_index)
 724 {
 725         int first = find_ref_delta(oid);
 726         int last = first;
 727         int end = nr_ref_deltas - 1;
 728
 729         if (first < 0) {
 730                 *first_index = 0;
 731                 *last_index = -1;
 732                 return;
 733         }
 734         while (first > 0 && oideq(&ref_deltas[first - 1].oid, oid))
 735                 --first;
 736         while (last < end && oideq(&ref_deltas[last + 1].oid, oid))
 737                 ++last;
 738         *first_index = first;
 739         *last_index = last;
 740 }
 741
 742 struct compare_data {
 743         struct object_entry *entry;
 744         struct git_istream *st;
 745         unsigned char *buf;
 746         unsigned long buf_size;
 747 };
 748
 749 static int compare_objects(const unsigned char *buf, unsigned long size,
 750                            void *cb_data)
 751 {
 752         struct compare_data *data = cb_data;
 753
 754         if (data->buf_size < size) {
 755                 free(data->buf);
 756                 data->buf = xmalloc(size);
 757                 data->buf_size = size;
 758         }
 759
 760         while (size) {
 761                 ssize_t len = read_istream(data->st, data->buf, size);
 762                 if (len == 0)
 763                         die(_("SHA1 COLLISION FOUND WITH %s !"),
 764                             oid_to_hex(&data->entry->idx.oid));
 765                 if (len < 0)
 766                         die(_("unable to read %s"),
 767                             oid_to_hex(&data->entry->idx.oid));
 768                 if (memcmp(buf, data->buf, len))
 769                         die(_("SHA1 COLLISION FOUND WITH %s !"),
 770                             oid_to_hex(&data->entry->idx.oid));
 771                 size -= len;
 772                 buf += len;
 773         }
 774         return 0;
 775 }
 776
 777 static int check_collison(struct object_entry *entry)
 778 {
 779         struct compare_data data;
 780         enum object_type type;
 781         unsigned long size;
 782
 783         if (entry->size <= big_file_threshold || entry->type != OBJ_BLOB)
 784                 return -1;
 785
 786         memset(&data, 0, sizeof(data));
 787         data.entry = entry;
 788         data.st = open_istream(the_repository, &entry->idx.oid, &type, &size,
 789                                NULL);
 790         if (!data.st)
 791                 return -1;
 792         if (size != entry->size || type != entry->type)
 793                 die(_("SHA1 COLLISION FOUND WITH %s !"),
 794                     oid_to_hex(&entry->idx.oid));
 795         unpack_data(entry, compare_objects, &data);
 796         close_istream(data.st);
 797         free(data.buf);
 798         return 0;
 799 }
 800
 801 static void sha1_object(const void *data, struct object_entry *obj_entry,
 802                         unsigned long size, enum object_type type,
 803                         const struct object_id *oid)
 804 {
 805         void *new_data = NULL;
 806         int collision_test_needed = 0;
 807
 808         assert(data || obj_entry);
 809
 810         if (startup_info->have_repository) {
 811                 read_lock();
 812                 collision_test_needed =
 813                         has_object_file_with_flags(oid, OBJECT_INFO_QUICK);
 814                 read_unlock();
 815         }
 816
 817         if (collision_test_needed && !data) {
 818                 read_lock();
 819                 if (!check_collison(obj_entry))
 820                         collision_test_needed = 0;
 821                 read_unlock();
 822         }
 823         if (collision_test_needed) {
 824                 void *has_data;
 825                 enum object_type has_type;
 826                 unsigned long has_size;
 827                 read_lock();
 828                 has_type = oid_object_info(the_repository, oid, &has_size);
 829                 if (has_type < 0)
 830                         die(_("cannot read existing object info %s"), oid_to_hex(oid));
 831                 if (has_type != type || has_size != size)
 832                         die(_("SHA1 COLLISION FOUND WITH %s !"), oid_to_hex(oid));
 833                 has_data = read_object_file(oid, &has_type, &has_size);
 834                 read_unlock();
 835                 if (!data)
 836                         data = new_data = get_data_from_pack(obj_entry);
 837                 if (!has_data)
 838                         die(_("cannot read existing object %s"), oid_to_hex(oid));
 839                 if (size != has_size || type != has_type ||
 840                     memcmp(data, has_data, size) != 0)
 841                         die(_("SHA1 COLLISION FOUND WITH %s !"), oid_to_hex(oid));
 842                 free(has_data);
 843         }
 844
 845         if (strict || do_fsck_object) {
 846                 read_lock();
 847                 if (type == OBJ_BLOB) {
 848                         struct blob *blob = lookup_blob(the_repository, oid);
 849                         if (blob)
 850                                 blob->object.flags |= FLAG_CHECKED;
 851                         else
 852                                 die(_("invalid blob object %s"), oid_to_hex(oid));
 853                         if (do_fsck_object &&
 854                             fsck_object(&blob->object, (void *)data, size, &fsck_options))
 855                                 die(_("fsck error in packed object"));
 856                 } else {
 857                         struct object *obj;
 858                         int eaten;
 859                         void *buf = (void *) data;
 860
 861                         assert(data && "data can only be NULL for large _blobs_");
 862
 863                         /*
 864                          * we do not need to free the memory here, as the
 865                          * buf is deleted by the caller.
 866                          */
 867                         obj = parse_object_buffer(the_repository, oid, type,
 868                                                   size, buf,
 869                                                   &eaten);
 870                         if (!obj)
 871                                 die(_("invalid %s"), type_name(type));
 872                         if (do_fsck_object &&
 873                             fsck_object(obj, buf, size, &fsck_options))
 874                                 die(_("fsck error in packed object"));
 875                         if (strict && fsck_walk(obj, NULL, &fsck_options))
 876                                 die(_("Not all child objects of %s are reachable"), oid_to_hex(&obj->oid));
 877
 878                         if (obj->type == OBJ_TREE) {
 879                                 struct tree *item = (struct tree *) obj;
 880                                 item->buffer = NULL;
 881                                 obj->parsed = 0;
 882                         }
 883                         if (obj->type == OBJ_COMMIT) {
 884                                 struct commit *commit = (struct commit *) obj;
 885                                 if (detach_commit_buffer(commit, NULL) != data)
 886                                         BUG("parse_object_buffer transmogrified our buffer");
 887                         }
 888                         obj->flags |= FLAG_CHECKED;
 889                 }
 890                 read_unlock();
 891         }
 892
 893         free(new_data);
 894 }
 895
 896 /*
 897  * Walk from current node up
 898  * to top parent if necessary to deflate the node. In normal
 899  * situation, its parent node would be already deflated, so it just
 900  * needs to apply delta.
 901  *
 902  * In the worst case scenario, parent node is no longer deflated because
 903  * we're running out of delta_base_cache_limit; we need to re-deflate
 904  * parents, possibly up to the top base.
 905  *
 906  * All deflated objects here are subject to be freed if we exceed
 907  * delta_base_cache_limit, just like in find_unresolved_deltas(), we
 908  * just need to make sure the last node is not freed.
 909  */
 910 static void *get_base_data(struct base_data *c)
 911 {
 912         if (!c->data) {
 913                 struct object_entry *obj = c->obj;
 914                 struct base_data **delta = NULL;
 915                 int delta_nr = 0, delta_alloc = 0;
 916
 917                 while (is_delta_type(c->obj->type) && !c->data) {
 918                         ALLOC_GROW(delta, delta_nr + 1, delta_alloc);
 919                         delta[delta_nr++] = c;
 920                         c = c->base;
 921                 }
 922                 if (!delta_nr) {
 923                         c->data = get_data_from_pack(obj);
 924                         c->size = obj->size;
 925                         base_cache_used += c->size;
 926                         prune_base_data(c);
 927                 }
 928                 for (; delta_nr > 0; delta_nr--) {
 929                         void *base, *raw;
 930                         c = delta[delta_nr - 1];
 931                         obj = c->obj;
 932                         base = get_base_data(c->base);
 933                         raw = get_data_from_pack(obj);
 934                         c->data = patch_delta(
 935                                 base, c->base->size,
 936                                 raw, obj->size,
 937                                 &c->size);
 938                         free(raw);
 939                         if (!c->data)
 940                                 bad_object(obj->idx.offset, _("failed to apply delta"));
 941                         base_cache_used += c->size;
 942                         prune_base_data(c);
 943                 }
 944                 free(delta);
 945         }
 946         return c->data;
 947 }
 948
 949 static struct base_data *make_base(struct object_entry *obj,
 950                                    struct base_data *parent)
 951 {
 952         struct base_data *base = xcalloc(1, sizeof(struct base_data));
 953         base->base = parent;
 954         base->obj = obj;
 955         find_ref_delta_children(&obj->idx.oid,
 956                                 &base->ref_first, &base->ref_last);
 957         find_ofs_delta_children(obj->idx.offset,
 958                                 &base->ofs_first, &base->ofs_last);
 959         base->children_remaining = base->ref_last - base->ref_first +
 960                 base->ofs_last - base->ofs_first + 2;
 961         return base;
 962 }
 963
 964 static struct base_data *resolve_delta(struct object_entry *delta_obj,
 965                                        struct base_data *base)
 966 {
 967         void *delta_data, *result_data;
 968         struct base_data *result;
 969         unsigned long result_size;
 970
 971         if (show_stat) {
 972                 int i = delta_obj - objects;
 973                 int j = base->obj - objects;
 974                 obj_stat[i].delta_depth = obj_stat[j].delta_depth + 1;
 975                 deepest_delta_lock();
 976                 if (deepest_delta < obj_stat[i].delta_depth)
 977                         deepest_delta = obj_stat[i].delta_depth;
 978                 deepest_delta_unlock();
 979                 obj_stat[i].base_object_no = j;
 980         }
 981         delta_data = get_data_from_pack(delta_obj);
 982         assert(base->data);
 983         result_data = patch_delta(base->data, base->size,
 984                                   delta_data, delta_obj->size, &result_size);
 985         free(delta_data);
 986         if (!result_data)
 987                 bad_object(delta_obj->idx.offset, _("failed to apply delta"));
 988         hash_object_file(the_hash_algo, result_data, result_size,
 989                          type_name(delta_obj->real_type), &delta_obj->idx.oid);
 990         sha1_object(result_data, NULL, result_size, delta_obj->real_type,
 991                     &delta_obj->idx.oid);
 992
 993         result = make_base(delta_obj, base);
 994         result->data = result_data;
 995         result->size = result_size;
 996
 997         counter_lock();
 998         nr_resolved_deltas++;
 999         counter_unlock();
1000
1001         return result;
1002 }
1003
1004 static int compare_ofs_delta_entry(const void *a, const void *b)
1005 {
1006         const struct ofs_delta_entry *delta_a = a;
1007         const struct ofs_delta_entry *delta_b = b;
1008
1009         return delta_a->offset < delta_b->offset ? -1 :
1010                delta_a->offset > delta_b->offset ?  1 :
1011                0;
1012 }
1013
1014 static int compare_ref_delta_entry(const void *a, const void *b)
1015 {
1016         const struct ref_delta_entry *delta_a = a;
1017         const struct ref_delta_entry *delta_b = b;
1018
1019         return oidcmp(&delta_a->oid, &delta_b->oid);
1020 }
1021
1022 static void *threaded_second_pass(void *data)
1023 {
1024         if (data)
1025                 set_thread_data(data);
1026         for (;;) {
1027                 struct base_data *parent = NULL;
1028                 struct object_entry *child_obj;
1029                 struct base_data *child;
1030
1031                 work_lock();
1032                 if (list_empty(&work_head)) {
1033                         /*
1034                          * Take an object from the object array.
1035                          */
1036                         while (nr_dispatched < nr_objects &&
1037                                is_delta_type(objects[nr_dispatched].type))
1038                                 nr_dispatched++;
1039                         if (nr_dispatched >= nr_objects) {
1040                                 work_unlock();
1041                                 break;
1042                         }
1043                         child_obj = &objects[nr_dispatched++];
1044                 } else {
1045                         /*
1046                          * Peek at the top of the stack, and take a child from
1047                          * it.
1048                          */
1049                         parent = list_first_entry(&work_head, struct base_data,
1050                                                   list);
1051
1052                         if (parent->ref_first <= parent->ref_last) {
1053                                 int offset = ref_deltas[parent->ref_first++].obj_no;
1054                                 child_obj = objects + offset;
1055                                 if (child_obj->real_type != OBJ_REF_DELTA)
1056                                         die("REF_DELTA at offset %"PRIuMAX" already resolved (duplicate base %s?)",
1057                                             (uintmax_t) child_obj->idx.offset,
1058                                             oid_to_hex(&parent->obj->idx.oid));
1059                                 child_obj->real_type = parent->obj->real_type;
1060                         } else {
1061                                 child_obj = objects +
1062                                         ofs_deltas[parent->ofs_first++].obj_no;
1063                                 assert(child_obj->real_type == OBJ_OFS_DELTA);
1064                                 child_obj->real_type = parent->obj->real_type;
1065                         }
1066
1067                         if (parent->ref_first > parent->ref_last &&
1068                             parent->ofs_first > parent->ofs_last) {
1069                                 /*
1070                                  * This parent has run out of children, so move
1071                                  * it to done_head.
1072                                  */
1073                                 list_del(&parent->list);
1074                                 list_add(&parent->list, &done_head);
1075                         }
1076
1077                         /*
1078                          * Ensure that the parent has data, since we will need
1079                          * it later.
1080                          *
1081                          * NEEDSWORK: If parent data needs to be reloaded, this
1082                          * prolongs the time that the current thread spends in
1083                          * the mutex. A mitigating factor is that parent data
1084                          * needs to be reloaded only if the delta base cache
1085                          * limit is exceeded, so in the typical case, this does
1086                          * not happen.
1087                          */
1088                         get_base_data(parent);
1089                         parent->retain_data++;
1090                 }
1091                 work_unlock();
1092
1093                 if (parent) {
1094                         child = resolve_delta(child_obj, parent);
1095                         if (!child->children_remaining)
1096                                 FREE_AND_NULL(child->data);
1097                 } else {
1098                         child = make_base(child_obj, NULL);
1099                         if (child->children_remaining) {
1100                                 /*
1101                                  * Since this child has its own delta children,
1102                                  * we will need this data in the future.
1103                                  * Inflate now so that future iterations will
1104                                  * have access to this object's data while
1105                                  * outside the work mutex.
1106                                  */
1107                                 child->data = get_data_from_pack(child_obj);
1108                                 child->size = child_obj->size;
1109                         }
1110                 }
1111
1112                 work_lock();
1113                 if (parent)
1114                         parent->retain_data--;
1115                 if (child->data) {
1116                         /*
1117                          * This child has its own children, so add it to
1118                          * work_head.
1119                          */
1120                         list_add(&child->list, &work_head);
1121                         base_cache_used += child->size;
1122                         prune_base_data(NULL);
1123                 } else {
1124                         /*
1125                          * This child does not have its own children. It may be
1126                          * the last descendant of its ancestors; free those
1127                          * that we can.
1128                          */
1129                         struct base_data *p = parent;
1130
1131                         while (p) {
1132                                 struct base_data *next_p;
1133
1134                                 p->children_remaining--;
1135                                 if (p->children_remaining)
1136                                         break;
1137
1138                                 next_p = p->base;
1139                                 free_base_data(p);
1140                                 list_del(&p->list);
1141                                 free(p);
1142
1143                                 p = next_p;
1144                         }
1145                 }
1146                 work_unlock();
1147         }
1148         return NULL;
1149 }
1150
1151 /*
1152  * First pass:
1153  * - find locations of all objects;
1154  * - calculate SHA1 of all non-delta objects;
1155  * - remember base (SHA1 or offset) for all deltas.
1156  */
1157 static void parse_pack_objects(unsigned char *hash)
1158 {
1159         int i, nr_delays = 0;
1160         struct ofs_delta_entry *ofs_delta = ofs_deltas;
1161         struct object_id ref_delta_oid;
1162         struct stat st;
1163
1164         if (verbose)
1165                 progress = start_progress(
1166                                 from_stdin ? _("Receiving objects") : _("Indexing objects"),
1167                                 nr_objects);
1168         for (i = 0; i < nr_objects; i++) {
1169                 struct object_entry *obj = &objects[i];
1170                 void *data = unpack_raw_entry(obj, &ofs_delta->offset,
1171                                               &ref_delta_oid,
1172                                               &obj->idx.oid);
1173                 obj->real_type = obj->type;
1174                 if (obj->type == OBJ_OFS_DELTA) {
1175                         nr_ofs_deltas++;
1176                         ofs_delta->obj_no = i;
1177                         ofs_delta++;
1178                 } else if (obj->type == OBJ_REF_DELTA) {
1179                         ALLOC_GROW(ref_deltas, nr_ref_deltas + 1, ref_deltas_alloc);
1180                         oidcpy(&ref_deltas[nr_ref_deltas].oid, &ref_delta_oid);
1181                         ref_deltas[nr_ref_deltas].obj_no = i;
1182                         nr_ref_deltas++;
1183                 } else if (!data) {
1184                         /* large blobs, check later */
1185                         obj->real_type = OBJ_BAD;
1186                         nr_delays++;
1187                 } else
1188                         sha1_object(data, NULL, obj->size, obj->type,
1189                                     &obj->idx.oid);
1190                 free(data);
1191                 display_progress(progress, i+1);
1192         }
1193         objects[i].idx.offset = consumed_bytes;
1194         stop_progress(&progress);
1195
1196         /* Check pack integrity */
1197         flush();
1198         the_hash_algo->final_fn(hash, &input_ctx);
1199         if (!hasheq(fill(the_hash_algo->rawsz), hash))
1200                 die(_("pack is corrupted (SHA1 mismatch)"));
1201         use(the_hash_algo->rawsz);
1202
1203         /* If input_fd is a file, we should have reached its end now. */
1204         if (fstat(input_fd, &st))
1205                 die_errno(_("cannot fstat packfile"));
1206         if (S_ISREG(st.st_mode) &&
1207                         lseek(input_fd, 0, SEEK_CUR) - input_len != st.st_size)
1208                 die(_("pack has junk at the end"));
1209
1210         for (i = 0; i < nr_objects; i++) {
1211                 struct object_entry *obj = &objects[i];
1212                 if (obj->real_type != OBJ_BAD)
1213                         continue;
1214                 obj->real_type = obj->type;
1215                 sha1_object(NULL, obj, obj->size, obj->type,
1216                             &obj->idx.oid);
1217                 nr_delays--;
1218         }
1219         if (nr_delays)
1220                 die(_("confusion beyond insanity in parse_pack_objects()"));
1221 }
1222
1223 /*
1224  * Second pass:
1225  * - for all non-delta objects, look if it is used as a base for
1226  *   deltas;
1227  * - if used as a base, uncompress the object and apply all deltas,
1228  *   recursively checking if the resulting object is used as a base
1229  *   for some more deltas.
1230  */
1231 static void resolve_deltas(void)
1232 {
1233         int i;
1234
1235         if (!nr_ofs_deltas && !nr_ref_deltas)
1236                 return;
1237
1238         /* Sort deltas by base SHA1/offset for fast searching */
1239         QSORT(ofs_deltas, nr_ofs_deltas, compare_ofs_delta_entry);
1240         QSORT(ref_deltas, nr_ref_deltas, compare_ref_delta_entry);
1241
1242         if (verbose || show_resolving_progress)
1243                 progress = start_progress(_("Resolving deltas"),
1244                                           nr_ref_deltas + nr_ofs_deltas);
1245
1246         nr_dispatched = 0;
1247         base_cache_limit = delta_base_cache_limit * nr_threads;
1248         if (nr_threads > 1 || getenv("GIT_FORCE_THREADS")) {
1249                 init_thread();
1250                 for (i = 0; i < nr_threads; i++) {
1251                         int ret = pthread_create(&thread_data[i].thread, NULL,
1252                                                  threaded_second_pass, thread_data + i);
1253                         if (ret)
1254                                 die(_("unable to create thread: %s"),
1255                                     strerror(ret));
1256                 }
1257                 for (i = 0; i < nr_threads; i++)
1258                         pthread_join(thread_data[i].thread, NULL);
1259                 cleanup_thread();
1260                 return;
1261         }
1262         threaded_second_pass(&nothread_data);
1263 }
1264
1265 /*
1266  * Third pass:
1267  * - append objects to convert thin pack to full pack if required
1268  * - write the final pack hash
1269  */
1270 static void fix_unresolved_deltas(struct hashfile *f);
1271 static void conclude_pack(int fix_thin_pack, const char *curr_pack, unsigned char *pack_hash)
1272 {
1273         if (nr_ref_deltas + nr_ofs_deltas == nr_resolved_deltas) {
1274                 stop_progress(&progress);
1275                 /* Flush remaining pack final hash. */
1276                 flush();
1277                 return;
1278         }
1279
1280         if (fix_thin_pack) {
1281                 struct hashfile *f;
1282                 unsigned char read_hash[GIT_MAX_RAWSZ], tail_hash[GIT_MAX_RAWSZ];
1283                 struct strbuf msg = STRBUF_INIT;
1284                 int nr_unresolved = nr_ofs_deltas + nr_ref_deltas - nr_resolved_deltas;
1285                 int nr_objects_initial = nr_objects;
1286                 if (nr_unresolved <= 0)
1287                         die(_("confusion beyond insanity"));
1288                 REALLOC_ARRAY(objects, nr_objects + nr_unresolved + 1);
1289                 memset(objects + nr_objects + 1, 0,
1290                        nr_unresolved * sizeof(*objects));
1291                 f = hashfd(output_fd, curr_pack);
1292                 fix_unresolved_deltas(f);
1293                 strbuf_addf(&msg, Q_("completed with %d local object",
1294                                      "completed with %d local objects",
1295                                      nr_objects - nr_objects_initial),
1296                             nr_objects - nr_objects_initial);
1297                 stop_progress_msg(&progress, msg.buf);
1298                 strbuf_release(&msg);
1299                 finalize_hashfile(f, tail_hash, 0);
1300                 hashcpy(read_hash, pack_hash);
1301                 fixup_pack_header_footer(output_fd, pack_hash,
1302                                          curr_pack, nr_objects,
1303                                          read_hash, consumed_bytes-the_hash_algo->rawsz);
1304                 if (!hasheq(read_hash, tail_hash))
1305                         die(_("Unexpected tail checksum for %s "
1306                               "(disk corruption?)"), curr_pack);
1307         }
1308         if (nr_ofs_deltas + nr_ref_deltas != nr_resolved_deltas)
1309                 die(Q_("pack has %d unresolved delta",
1310                        "pack has %d unresolved deltas",
1311                        nr_ofs_deltas + nr_ref_deltas - nr_resolved_deltas),
1312                     nr_ofs_deltas + nr_ref_deltas - nr_resolved_deltas);
1313 }
1314
1315 static int write_compressed(struct hashfile *f, void *in, unsigned int size)
1316 {
1317         git_zstream stream;
1318         int status;
1319         unsigned char outbuf[4096];
1320
1321         git_deflate_init(&stream, zlib_compression_level);
1322         stream.next_in = in;
1323         stream.avail_in = size;
1324
1325         do {
1326                 stream.next_out = outbuf;
1327                 stream.avail_out = sizeof(outbuf);
1328                 status = git_deflate(&stream, Z_FINISH);
1329                 hashwrite(f, outbuf, sizeof(outbuf) - stream.avail_out);
1330         } while (status == Z_OK);
1331
1332         if (status != Z_STREAM_END)
1333                 die(_("unable to deflate appended object (%d)"), status);
1334         size = stream.total_out;
1335         git_deflate_end(&stream);
1336         return size;
1337 }
1338
1339 static struct object_entry *append_obj_to_pack(struct hashfile *f,
1340                                const unsigned char *sha1, void *buf,
1341                                unsigned long size, enum object_type type)
1342 {
1343         struct object_entry *obj = &objects[nr_objects++];
1344         unsigned char header[10];
1345         unsigned long s = size;
1346         int n = 0;
1347         unsigned char c = (type << 4) | (s & 15);
1348         s >>= 4;
1349         while (s) {
1350                 header[n++] = c | 0x80;
1351                 c = s & 0x7f;
1352                 s >>= 7;
1353         }
1354         header[n++] = c;
1355         crc32_begin(f);
1356         hashwrite(f, header, n);
1357         obj[0].size = size;
1358         obj[0].hdr_size = n;
1359         obj[0].type = type;
1360         obj[0].real_type = type;
1361         obj[1].idx.offset = obj[0].idx.offset + n;
1362         obj[1].idx.offset += write_compressed(f, buf, size);
1363         obj[0].idx.crc32 = crc32_end(f);
1364         hashflush(f);
1365         hashcpy(obj->idx.oid.hash, sha1);
1366         return obj;
1367 }
1368
1369 static int delta_pos_compare(const void *_a, const void *_b)
1370 {
1371         struct ref_delta_entry *a = *(struct ref_delta_entry **)_a;
1372         struct ref_delta_entry *b = *(struct ref_delta_entry **)_b;
1373         return a->obj_no - b->obj_no;
1374 }
1375
1376 static void fix_unresolved_deltas(struct hashfile *f)
1377 {
1378         struct ref_delta_entry **sorted_by_pos;
1379         int i;
1380
1381         /*
1382          * Since many unresolved deltas may well be themselves base objects
1383          * for more unresolved deltas, we really want to include the
1384          * smallest number of base objects that would cover as much delta
1385          * as possible by picking the
1386          * trunc deltas first, allowing for other deltas to resolve without
1387          * additional base objects.  Since most base objects are to be found
1388          * before deltas depending on them, a good heuristic is to start
1389          * resolving deltas in the same order as their position in the pack.
1390          */
1391         ALLOC_ARRAY(sorted_by_pos, nr_ref_deltas);
1392         for (i = 0; i < nr_ref_deltas; i++)
1393                 sorted_by_pos[i] = &ref_deltas[i];
1394         QSORT(sorted_by_pos, nr_ref_deltas, delta_pos_compare);
1395
1396         if (has_promisor_remote()) {
1397                 /*
1398                  * Prefetch the delta bases.
1399                  */
1400                 struct oid_array to_fetch = OID_ARRAY_INIT;
1401                 for (i = 0; i < nr_ref_deltas; i++) {
1402                         struct ref_delta_entry *d = sorted_by_pos[i];
1403                         if (!oid_object_info_extended(the_repository, &d->oid,
1404                                                       NULL,
1405                                                       OBJECT_INFO_FOR_PREFETCH))
1406                                 continue;
1407                         oid_array_append(&to_fetch, &d->oid);
1408                 }
1409                 promisor_remote_get_direct(the_repository,
1410                                            to_fetch.oid, to_fetch.nr);
1411                 oid_array_clear(&to_fetch);
1412         }
1413
1414         for (i = 0; i < nr_ref_deltas; i++) {
1415                 struct ref_delta_entry *d = sorted_by_pos[i];
1416                 enum object_type type;
1417                 void *data;
1418                 unsigned long size;
1419
1420                 if (objects[d->obj_no].real_type != OBJ_REF_DELTA)
1421                         continue;
1422                 data = read_object_file(&d->oid, &type, &size);
1423                 if (!data)
1424                         continue;
1425
1426                 if (check_object_signature(the_repository, &d->oid,
1427                                            data, size,
1428                                            type_name(type)))
1429                         die(_("local object %s is corrupt"), oid_to_hex(&d->oid));
1430
1431                 /*
1432                  * Add this as an object to the objects array and call
1433                  * threaded_second_pass() (which will pick up the added
1434                  * object).
1435                  */
1436                 append_obj_to_pack(f, d->oid.hash, data, size, type);
1437                 threaded_second_pass(NULL);
1438
1439                 display_progress(progress, nr_resolved_deltas);
1440         }
1441         free(sorted_by_pos);
1442 }
1443
1444 static const char *derive_filename(const char *pack_name, const char *suffix,
1445                                    struct strbuf *buf)
1446 {
1447         size_t len;
1448         if (!strip_suffix(pack_name, ".pack", &len))
1449                 die(_("packfile name '%s' does not end with '.pack'"),
1450                     pack_name);
1451         strbuf_add(buf, pack_name, len);
1452         strbuf_addch(buf, '.');
1453         strbuf_addstr(buf, suffix);
1454         return buf->buf;
1455 }
1456
1457 static void write_special_file(const char *suffix, const char *msg,
1458                                const char *pack_name, const unsigned char *hash,
1459                                const char **report)
1460 {
1461         struct strbuf name_buf = STRBUF_INIT;
1462         const char *filename;
1463         int fd;
1464         int msg_len = strlen(msg);
1465
1466         if (pack_name)
1467                 filename = derive_filename(pack_name, suffix, &name_buf);
1468         else
1469                 filename = odb_pack_name(&name_buf, hash, suffix);
1470
1471         fd = odb_pack_keep(filename);
1472         if (fd < 0) {
1473                 if (errno != EEXIST)
1474                         die_errno(_("cannot write %s file '%s'"),
1475                                   suffix, filename);
1476         } else {
1477                 if (msg_len > 0) {
1478                         write_or_die(fd, msg, msg_len);
1479                         write_or_die(fd, "\n", 1);
1480                 }
1481                 if (close(fd) != 0)
1482                         die_errno(_("cannot close written %s file '%s'"),
1483                                   suffix, filename);
1484                 if (report)
1485                         *report = suffix;
1486         }
1487         strbuf_release(&name_buf);
1488 }
1489
1490 static void final(const char *final_pack_name, const char *curr_pack_name,
1491                   const char *final_index_name, const char *curr_index_name,
1492                   const char *keep_msg, const char *promisor_msg,
1493                   unsigned char *hash)
1494 {
1495         const char *report = "pack";
1496         struct strbuf pack_name = STRBUF_INIT;
1497         struct strbuf index_name = STRBUF_INIT;
1498         int err;
1499
1500         if (!from_stdin) {
1501                 close(input_fd);
1502         } else {
1503                 fsync_or_die(output_fd, curr_pack_name);
1504                 err = close(output_fd);
1505                 if (err)
1506                         die_errno(_("error while closing pack file"));
1507         }
1508
1509         if (keep_msg)
1510                 write_special_file("keep", keep_msg, final_pack_name, hash,
1511                                    &report);
1512         if (promisor_msg)
1513                 write_special_file("promisor", promisor_msg, final_pack_name,
1514                                    hash, NULL);
1515
1516         if (final_pack_name != curr_pack_name) {
1517                 if (!final_pack_name)
1518                         final_pack_name = odb_pack_name(&pack_name, hash, "pack");
1519                 if (finalize_object_file(curr_pack_name, final_pack_name))
1520                         die(_("cannot store pack file"));
1521         } else if (from_stdin)
1522                 chmod(final_pack_name, 0444);
1523
1524         if (final_index_name != curr_index_name) {
1525                 if (!final_index_name)
1526                         final_index_name = odb_pack_name(&index_name, hash, "idx");
1527                 if (finalize_object_file(curr_index_name, final_index_name))
1528                         die(_("cannot store index file"));
1529         } else
1530                 chmod(final_index_name, 0444);
1531
1532         if (do_fsck_object) {
1533                 struct packed_git *p;
1534                 p = add_packed_git(final_index_name, strlen(final_index_name), 0);
1535                 if (p)
1536                         install_packed_git(the_repository, p);
1537         }
1538
1539         if (!from_stdin) {
1540                 printf("%s\n", hash_to_hex(hash));
1541         } else {
1542                 struct strbuf buf = STRBUF_INIT;
1543
1544                 strbuf_addf(&buf, "%s\t%s\n", report, hash_to_hex(hash));
1545                 write_or_die(1, buf.buf, buf.len);
1546                 strbuf_release(&buf);
1547
1548                 /*
1549                  * Let's just mimic git-unpack-objects here and write
1550                  * the last part of the input buffer to stdout.
1551                  */
1552                 while (input_len) {
1553                         err = xwrite(1, input_buffer + input_offset, input_len);
1554                         if (err <= 0)
1555                                 break;
1556                         input_len -= err;
1557                         input_offset += err;
1558                 }
1559         }
1560
1561         strbuf_release(&index_name);
1562         strbuf_release(&pack_name);
1563 }
1564
1565 static int git_index_pack_config(const char *k, const char *v, void *cb)
1566 {
1567         struct pack_idx_option *opts = cb;
1568
1569         if (!strcmp(k, "pack.indexversion")) {
1570                 opts->version = git_config_int(k, v);
1571                 if (opts->version > 2)
1572                         die(_("bad pack.indexversion=%"PRIu32), opts->version);
1573                 return 0;
1574         }
1575         if (!strcmp(k, "pack.threads")) {
1576                 nr_threads = git_config_int(k, v);
1577                 if (nr_threads < 0)
1578                         die(_("invalid number of threads specified (%d)"),
1579                             nr_threads);
1580                 if (!HAVE_THREADS && nr_threads != 1) {
1581                         warning(_("no threads support, ignoring %s"), k);
1582                         nr_threads = 1;
1583                 }
1584                 return 0;
1585         }
1586         return git_default_config(k, v, cb);
1587 }
1588
1589 static int cmp_uint32(const void *a_, const void *b_)
1590 {
1591         uint32_t a = *((uint32_t *)a_);
1592         uint32_t b = *((uint32_t *)b_);
1593
1594         return (a < b) ? -1 : (a != b);
1595 }
1596
1597 static void read_v2_anomalous_offsets(struct packed_git *p,
1598                                       struct pack_idx_option *opts)
1599 {
1600         const uint32_t *idx1, *idx2;
1601         uint32_t i;
1602
1603         /* The address of the 4-byte offset table */
1604         idx1 = (((const uint32_t *)((const uint8_t *)p->index_data + p->crc_offset))
1605                 + p->num_objects /* CRC32 table */
1606                 );
1607
1608         /* The address of the 8-byte offset table */
1609         idx2 = idx1 + p->num_objects;
1610
1611         for (i = 0; i < p->num_objects; i++) {
1612                 uint32_t off = ntohl(idx1[i]);
1613                 if (!(off & 0x80000000))
1614                         continue;
1615                 off = off & 0x7fffffff;
1616                 check_pack_index_ptr(p, &idx2[off * 2]);
1617                 if (idx2[off * 2])
1618                         continue;
1619                 /*
1620                  * The real offset is ntohl(idx2[off * 2]) in high 4
1621                  * octets, and ntohl(idx2[off * 2 + 1]) in low 4
1622                  * octets.  But idx2[off * 2] is Zero!!!
1623                  */
1624                 ALLOC_GROW(opts->anomaly, opts->anomaly_nr + 1, opts->anomaly_alloc);
1625                 opts->anomaly[opts->anomaly_nr++] = ntohl(idx2[off * 2 + 1]);
1626         }
1627
1628         QSORT(opts->anomaly, opts->anomaly_nr, cmp_uint32);
1629 }
1630
1631 static void read_idx_option(struct pack_idx_option *opts, const char *pack_name)
1632 {
1633         struct packed_git *p = add_packed_git(pack_name, strlen(pack_name), 1);
1634
1635         if (!p)
1636                 die(_("Cannot open existing pack file '%s'"), pack_name);
1637         if (open_pack_index(p))
1638                 die(_("Cannot open existing pack idx file for '%s'"), pack_name);
1639
1640         /* Read the attributes from the existing idx file */
1641         opts->version = p->index_version;
1642
1643         if (opts->version == 2)
1644                 read_v2_anomalous_offsets(p, opts);
1645
1646         /*
1647          * Get rid of the idx file as we do not need it anymore.
1648          * NEEDSWORK: extract this bit from free_pack_by_name() in
1649          * sha1-file.c, perhaps?  It shouldn't matter very much as we
1650          * know we haven't installed this pack (hence we never have
1651          * read anything from it).
1652          */
1653         close_pack_index(p);
1654         free(p);
1655 }
1656
1657 static void show_pack_info(int stat_only)
1658 {
1659         int i, baseobjects = nr_objects - nr_ref_deltas - nr_ofs_deltas;
1660         unsigned long *chain_histogram = NULL;
1661
1662         if (deepest_delta)
1663                 chain_histogram = xcalloc(deepest_delta, sizeof(unsigned long));
1664
1665         for (i = 0; i < nr_objects; i++) {
1666                 struct object_entry *obj = &objects[i];
1667
1668                 if (is_delta_type(obj->type))
1669                         chain_histogram[obj_stat[i].delta_depth - 1]++;
1670                 if (stat_only)
1671                         continue;
1672                 printf("%s %-6s %"PRIuMAX" %"PRIuMAX" %"PRIuMAX,
1673                        oid_to_hex(&obj->idx.oid),
1674                        type_name(obj->real_type), (uintmax_t)obj->size,
1675                        (uintmax_t)(obj[1].idx.offset - obj->idx.offset),
1676                        (uintmax_t)obj->idx.offset);
1677                 if (is_delta_type(obj->type)) {
1678                         struct object_entry *bobj = &objects[obj_stat[i].base_object_no];
1679                         printf(" %u %s", obj_stat[i].delta_depth,
1680                                oid_to_hex(&bobj->idx.oid));
1681                 }
1682                 putchar('\n');
1683         }
1684
1685         if (baseobjects)
1686                 printf_ln(Q_("non delta: %d object",
1687                              "non delta: %d objects",
1688                              baseobjects),
1689                           baseobjects);
1690         for (i = 0; i < deepest_delta; i++) {
1691                 if (!chain_histogram[i])
1692                         continue;
1693                 printf_ln(Q_("chain length = %d: %lu object",
1694                              "chain length = %d: %lu objects",
1695                              chain_histogram[i]),
1696                           i + 1,
1697                           chain_histogram[i]);
1698         }
1699 }
1700
1701 int cmd_index_pack(int argc, const char **argv, const char *prefix)
1702 {
1703         int i, fix_thin_pack = 0, verify = 0, stat_only = 0;
1704         const char *curr_index;
1705         const char *index_name = NULL, *pack_name = NULL;
1706         const char *keep_msg = NULL;
1707         const char *promisor_msg = NULL;
1708         struct strbuf index_name_buf = STRBUF_INIT;
1709         struct pack_idx_entry **idx_objects;
1710         struct pack_idx_option opts;
1711         unsigned char pack_hash[GIT_MAX_RAWSZ];
1712         unsigned foreign_nr = 1;        /* zero is a "good" value, assume bad */
1713         int report_end_of_input = 0;
1714         int hash_algo = 0;
1715
1716         /*
1717          * index-pack never needs to fetch missing objects except when
1718          * REF_DELTA bases are missing (which are explicitly handled). It only
1719          * accesses the repo to do hash collision checks and to check which
1720          * REF_DELTA bases need to be fetched.
1721          */
1722         fetch_if_missing = 0;
1723
1724         if (argc == 2 && !strcmp(argv[1], "-h"))
1725                 usage(index_pack_usage);
1726
1727         read_replace_refs = 0;
1728         fsck_options.walk = mark_link;
1729
1730         reset_pack_idx_option(&opts);
1731         git_config(git_index_pack_config, &opts);
1732         if (prefix && chdir(prefix))
1733                 die(_("Cannot come back to cwd"));
1734
1735         for (i = 1; i < argc; i++) {
1736                 const char *arg = argv[i];
1737
1738                 if (*arg == '-') {
1739                         if (!strcmp(arg, "--stdin")) {
1740                                 from_stdin = 1;
1741                         } else if (!strcmp(arg, "--fix-thin")) {
1742                                 fix_thin_pack = 1;
1743                         } else if (skip_to_optional_arg(arg, "--strict", &arg)) {
1744                                 strict = 1;
1745                                 do_fsck_object = 1;
1746                                 fsck_set_msg_types(&fsck_options, arg);
1747                         } else if (!strcmp(arg, "--check-self-contained-and-connected")) {
1748                                 strict = 1;
1749                                 check_self_contained_and_connected = 1;
1750                         } else if (!strcmp(arg, "--fsck-objects")) {
1751                                 do_fsck_object = 1;
1752                         } else if (!strcmp(arg, "--verify")) {
1753                                 verify = 1;
1754                         } else if (!strcmp(arg, "--verify-stat")) {
1755                                 verify = 1;
1756                                 show_stat = 1;
1757                         } else if (!strcmp(arg, "--verify-stat-only")) {
1758                                 verify = 1;
1759                                 show_stat = 1;
1760                                 stat_only = 1;
1761                         } else if (skip_to_optional_arg(arg, "--keep", &keep_msg)) {
1762                                 ; /* nothing to do */
1763                         } else if (skip_to_optional_arg(arg, "--promisor", &promisor_msg)) {
1764                                 ; /* already parsed */
1765                         } else if (starts_with(arg, "--threads=")) {
1766                                 char *end;
1767                                 nr_threads = strtoul(arg+10, &end, 0);
1768                                 if (!arg[10] || *end || nr_threads < 0)
1769                                         usage(index_pack_usage);
1770                                 if (!HAVE_THREADS && nr_threads != 1) {
1771                                         warning(_("no threads support, ignoring %s"), arg);
1772                                         nr_threads = 1;
1773                                 }
1774                         } else if (starts_with(arg, "--pack_header=")) {
1775                                 struct pack_header *hdr;
1776                                 char *c;
1777
1778                                 hdr = (struct pack_header *)input_buffer;
1779                                 hdr->hdr_signature = htonl(PACK_SIGNATURE);
1780                                 hdr->hdr_version = htonl(strtoul(arg + 14, &c, 10));
1781                                 if (*c != ',')
1782                                         die(_("bad %s"), arg);
1783                                 hdr->hdr_entries = htonl(strtoul(c + 1, &c, 10));
1784                                 if (*c)
1785                                         die(_("bad %s"), arg);
1786                                 input_len = sizeof(*hdr);
1787                         } else if (!strcmp(arg, "-v")) {
1788                                 verbose = 1;
1789                         } else if (!strcmp(arg, "--show-resolving-progress")) {
1790                                 show_resolving_progress = 1;
1791                         } else if (!strcmp(arg, "--report-end-of-input")) {
1792                                 report_end_of_input = 1;
1793                         } else if (!strcmp(arg, "-o")) {
1794                                 if (index_name || (i+1) >= argc)
1795                                         usage(index_pack_usage);
1796                                 index_name = argv[++i];
1797                         } else if (starts_with(arg, "--index-version=")) {
1798                                 char *c;
1799                                 opts.version = strtoul(arg + 16, &c, 10);
1800                                 if (opts.version > 2)
1801                                         die(_("bad %s"), arg);
1802                                 if (*c == ',')
1803                                         opts.off32_limit = strtoul(c+1, &c, 0);
1804                                 if (*c || opts.off32_limit & 0x80000000)
1805                                         die(_("bad %s"), arg);
1806                         } else if (skip_prefix(arg, "--max-input-size=", &arg)) {
1807                                 max_input_size = strtoumax(arg, NULL, 10);
1808                         } else if (skip_prefix(arg, "--object-format=", &arg)) {
1809                                 hash_algo = hash_algo_by_name(arg);
1810                                 if (hash_algo == GIT_HASH_UNKNOWN)
1811                                         die(_("unknown hash algorithm '%s'"), arg);
1812                                 repo_set_hash_algo(the_repository, hash_algo);
1813                         } else
1814                                 usage(index_pack_usage);
1815                         continue;
1816                 }
1817
1818                 if (pack_name)
1819                         usage(index_pack_usage);
1820                 pack_name = arg;
1821         }
1822
1823         if (!pack_name && !from_stdin)
1824                 usage(index_pack_usage);
1825         if (fix_thin_pack && !from_stdin)
1826                 die(_("--fix-thin cannot be used without --stdin"));
1827         if (from_stdin && !startup_info->have_repository)
1828                 die(_("--stdin requires a git repository"));
1829         if (from_stdin && hash_algo)
1830                 die(_("--object-format cannot be used with --stdin"));
1831         if (!index_name && pack_name)
1832                 index_name = derive_filename(pack_name, "idx", &index_name_buf);
1833
1834         if (verify) {
1835                 if (!index_name)
1836                         die(_("--verify with no packfile name given"));
1837                 read_idx_option(&opts, index_name);
1838                 opts.flags |= WRITE_IDX_VERIFY | WRITE_IDX_STRICT;
1839         }
1840         if (strict)
1841                 opts.flags |= WRITE_IDX_STRICT;
1842
1843         if (HAVE_THREADS && !nr_threads) {
1844                 nr_threads = online_cpus();
1845                 /*
1846                  * Experiments show that going above 20 threads doesn't help,
1847                  * no matter how many cores you have. Below that, we tend to
1848                  * max at half the number of online_cpus(), presumably because
1849                  * half of those are hyperthreads rather than full cores. We'll
1850                  * never reduce the level below "3", though, to match a
1851                  * historical value that nobody complained about.
1852                  */
1853                 if (nr_threads < 4)
1854                         ; /* too few cores to consider capping */
1855                 else if (nr_threads < 6)
1856                         nr_threads = 3; /* historic cap */
1857                 else if (nr_threads < 40)
1858                         nr_threads /= 2;
1859                 else
1860                         nr_threads = 20; /* hard cap */
1861         }
1862
1863         curr_pack = open_pack_file(pack_name);
1864         parse_pack_header();
1865         objects = xcalloc(st_add(nr_objects, 1), sizeof(struct object_entry));
1866         if (show_stat)
1867                 obj_stat = xcalloc(st_add(nr_objects, 1), sizeof(struct object_stat));
1868         ofs_deltas = xcalloc(nr_objects, sizeof(struct ofs_delta_entry));
1869         parse_pack_objects(pack_hash);
1870         if (report_end_of_input)
1871                 write_in_full(2, "\0", 1);
1872         resolve_deltas();
1873         conclude_pack(fix_thin_pack, curr_pack, pack_hash);
1874         free(ofs_deltas);
1875         free(ref_deltas);
1876         if (strict)
1877                 foreign_nr = check_objects();
1878
1879         if (show_stat)
1880                 show_pack_info(stat_only);
1881
1882         ALLOC_ARRAY(idx_objects, nr_objects);
1883         for (i = 0; i < nr_objects; i++)
1884                 idx_objects[i] = &objects[i].idx;
1885         curr_index = write_idx_file(index_name, idx_objects, nr_objects, &opts, pack_hash);
1886         free(idx_objects);
1887
1888         if (!verify)
1889                 final(pack_name, curr_pack,
1890                       index_name, curr_index,
1891                       keep_msg, promisor_msg,
1892                       pack_hash);
1893         else
1894                 close(input_fd);
1895
1896         if (do_fsck_object && fsck_finish(&fsck_options))
1897                 die(_("fsck error in pack objects"));
1898
1899         free(objects);
1900         strbuf_release(&index_name_buf);
1901         if (pack_name == NULL)
1902                 free((void *) curr_pack);
1903         if (index_name == NULL)
1904                 free((void *) curr_index);
1905
1906         /*
1907          * Let the caller know this pack is not self contained
1908          */
1909         if (check_self_contained_and_connected && foreign_nr)
1910                 return 1;
1911
1912         return 0;
1913 }