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

   1 #include "cache.h"
   2 #include "list.h"
   3 #include "pack.h"
   4 #include "repository.h"
   5 #include "dir.h"
   6 #include "mergesort.h"
   7 #include "packfile.h"
   8 #include "delta.h"
   9 #include "streaming.h"
  10 #include "sha1-lookup.h"
  11 #include "commit.h"
  12 #include "object.h"
  13 #include "tag.h"
  14 #include "tree-walk.h"
  15 #include "tree.h"
  16 #include "object-store.h"
  17 #include "midx.h"
  18 #include "commit-graph.h"
  19 #include "promisor-remote.h"
  20
  21 char *odb_pack_name(struct strbuf *buf,
  22                     const unsigned char *hash,
  23                     const char *ext)
  24 {
  25         strbuf_reset(buf);
  26         strbuf_addf(buf, "%s/pack/pack-%s.%s", get_object_directory(),
  27                     hash_to_hex(hash), ext);
  28         return buf->buf;
  29 }
  30
  31 char *sha1_pack_name(const unsigned char *sha1)
  32 {
  33         static struct strbuf buf = STRBUF_INIT;
  34         return odb_pack_name(&buf, sha1, "pack");
  35 }
  36
  37 char *sha1_pack_index_name(const unsigned char *sha1)
  38 {
  39         static struct strbuf buf = STRBUF_INIT;
  40         return odb_pack_name(&buf, sha1, "idx");
  41 }
  42
  43 static unsigned int pack_used_ctr;
  44 static unsigned int pack_mmap_calls;
  45 static unsigned int peak_pack_open_windows;
  46 static unsigned int pack_open_windows;
  47 static unsigned int pack_open_fds;
  48 static unsigned int pack_max_fds;
  49 static size_t peak_pack_mapped;
  50 static size_t pack_mapped;
  51
  52 #define SZ_FMT PRIuMAX
  53 static inline uintmax_t sz_fmt(size_t s) { return s; }
  54
  55 void pack_report(void)
  56 {
  57         fprintf(stderr,
  58                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
  59                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
  60                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
  61                 sz_fmt(getpagesize()),
  62                 sz_fmt(packed_git_window_size),
  63                 sz_fmt(packed_git_limit));
  64         fprintf(stderr,
  65                 "pack_report: pack_used_ctr            = %10u\n"
  66                 "pack_report: pack_mmap_calls          = %10u\n"
  67                 "pack_report: pack_open_windows        = %10u / %10u\n"
  68                 "pack_report: pack_mapped              = "
  69                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
  70                 pack_used_ctr,
  71                 pack_mmap_calls,
  72                 pack_open_windows, peak_pack_open_windows,
  73                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
  74 }
  75
  76 /*
  77  * Open and mmap the index file at path, perform a couple of
  78  * consistency checks, then record its information to p.  Return 0 on
  79  * success.
  80  */
  81 static int check_packed_git_idx(const char *path, struct packed_git *p)
  82 {
  83         void *idx_map;
  84         size_t idx_size;
  85         int fd = git_open(path), ret;
  86         struct stat st;
  87         const unsigned int hashsz = the_hash_algo->rawsz;
  88
  89         if (fd < 0)
  90                 return -1;
  91         if (fstat(fd, &st)) {
  92                 close(fd);
  93                 return -1;
  94         }
  95         idx_size = xsize_t(st.st_size);
  96         if (idx_size < 4 * 256 + hashsz + hashsz) {
  97                 close(fd);
  98                 return error("index file %s is too small", path);
  99         }
 100         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
 101         close(fd);
 102
 103         ret = load_idx(path, hashsz, idx_map, idx_size, p);
 104
 105         if (ret)
 106                 munmap(idx_map, idx_size);
 107
 108         return ret;
 109 }
 110
 111 int load_idx(const char *path, const unsigned int hashsz, void *idx_map,
 112              size_t idx_size, struct packed_git *p)
 113 {
 114         struct pack_idx_header *hdr = idx_map;
 115         uint32_t version, nr, i, *index;
 116
 117         if (idx_size < 4 * 256 + hashsz + hashsz)
 118                 return error("index file %s is too small", path);
 119         if (idx_map == NULL)
 120                 return error("empty data");
 121
 122         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
 123                 version = ntohl(hdr->idx_version);
 124                 if (version < 2 || version > 2)
 125                         return error("index file %s is version %"PRIu32
 126                                      " and is not supported by this binary"
 127                                      " (try upgrading GIT to a newer version)",
 128                                      path, version);
 129         } else
 130                 version = 1;
 131
 132         nr = 0;
 133         index = idx_map;
 134         if (version > 1)
 135                 index += 2;  /* skip index header */
 136         for (i = 0; i < 256; i++) {
 137                 uint32_t n = ntohl(index[i]);
 138                 if (n < nr)
 139                         return error("non-monotonic index %s", path);
 140                 nr = n;
 141         }
 142
 143         if (version == 1) {
 144                 /*
 145                  * Total size:
 146                  *  - 256 index entries 4 bytes each
 147                  *  - 24-byte entries * nr (object ID + 4-byte offset)
 148                  *  - hash of the packfile
 149                  *  - file checksum
 150                  */
 151                 if (idx_size != 4 * 256 + nr * (hashsz + 4) + hashsz + hashsz)
 152                         return error("wrong index v1 file size in %s", path);
 153         } else if (version == 2) {
 154                 /*
 155                  * Minimum size:
 156                  *  - 8 bytes of header
 157                  *  - 256 index entries 4 bytes each
 158                  *  - object ID entry * nr
 159                  *  - 4-byte crc entry * nr
 160                  *  - 4-byte offset entry * nr
 161                  *  - hash of the packfile
 162                  *  - file checksum
 163                  * And after the 4-byte offset table might be a
 164                  * variable sized table containing 8-byte entries
 165                  * for offsets larger than 2^31.
 166                  */
 167                 unsigned long min_size = 8 + 4*256 + nr*(hashsz + 4 + 4) + hashsz + hashsz;
 168                 unsigned long max_size = min_size;
 169                 if (nr)
 170                         max_size += (nr - 1)*8;
 171                 if (idx_size < min_size || idx_size > max_size)
 172                         return error("wrong index v2 file size in %s", path);
 173                 if (idx_size != min_size &&
 174                     /*
 175                      * make sure we can deal with large pack offsets.
 176                      * 31-bit signed offset won't be enough, neither
 177                      * 32-bit unsigned one will be.
 178                      */
 179                     (sizeof(off_t) <= 4))
 180                         return error("pack too large for current definition of off_t in %s", path);
 181         }
 182
 183         p->index_version = version;
 184         p->index_data = idx_map;
 185         p->index_size = idx_size;
 186         p->num_objects = nr;
 187         return 0;
 188 }
 189
 190 int open_pack_index(struct packed_git *p)
 191 {
 192         char *idx_name;
 193         size_t len;
 194         int ret;
 195
 196         if (p->index_data)
 197                 return 0;
 198
 199         if (!strip_suffix(p->pack_name, ".pack", &len))
 200                 BUG("pack_name does not end in .pack");
 201         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
 202         ret = check_packed_git_idx(idx_name, p);
 203         free(idx_name);
 204         return ret;
 205 }
 206
 207 uint32_t get_pack_fanout(struct packed_git *p, uint32_t value)
 208 {
 209         const uint32_t *level1_ofs = p->index_data;
 210
 211         if (!level1_ofs) {
 212                 if (open_pack_index(p))
 213                         return 0;
 214                 level1_ofs = p->index_data;
 215         }
 216
 217         if (p->index_version > 1) {
 218                 level1_ofs += 2;
 219         }
 220
 221         return ntohl(level1_ofs[value]);
 222 }
 223
 224 static struct packed_git *alloc_packed_git(int extra)
 225 {
 226         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
 227         memset(p, 0, sizeof(*p));
 228         p->pack_fd = -1;
 229         return p;
 230 }
 231
 232 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
 233 {
 234         const char *path = sha1_pack_name(sha1);
 235         size_t alloc = st_add(strlen(path), 1);
 236         struct packed_git *p = alloc_packed_git(alloc);
 237
 238         memcpy(p->pack_name, path, alloc); /* includes NUL */
 239         hashcpy(p->hash, sha1);
 240         if (check_packed_git_idx(idx_path, p)) {
 241                 free(p);
 242                 return NULL;
 243         }
 244
 245         return p;
 246 }
 247
 248 static void scan_windows(struct packed_git *p,
 249         struct packed_git **lru_p,
 250         struct pack_window **lru_w,
 251         struct pack_window **lru_l)
 252 {
 253         struct pack_window *w, *w_l;
 254
 255         for (w_l = NULL, w = p->windows; w; w = w->next) {
 256                 if (!w->inuse_cnt) {
 257                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
 258                                 *lru_p = p;
 259                                 *lru_w = w;
 260                                 *lru_l = w_l;
 261                         }
 262                 }
 263                 w_l = w;
 264         }
 265 }
 266
 267 static int unuse_one_window(struct packed_git *current)
 268 {
 269         struct packed_git *p, *lru_p = NULL;
 270         struct pack_window *lru_w = NULL, *lru_l = NULL;
 271
 272         if (current)
 273                 scan_windows(current, &lru_p, &lru_w, &lru_l);
 274         for (p = the_repository->objects->packed_git; p; p = p->next)
 275                 scan_windows(p, &lru_p, &lru_w, &lru_l);
 276         if (lru_p) {
 277                 munmap(lru_w->base, lru_w->len);
 278                 pack_mapped -= lru_w->len;
 279                 if (lru_l)
 280                         lru_l->next = lru_w->next;
 281                 else
 282                         lru_p->windows = lru_w->next;
 283                 free(lru_w);
 284                 pack_open_windows--;
 285                 return 1;
 286         }
 287         return 0;
 288 }
 289
 290 void close_pack_windows(struct packed_git *p)
 291 {
 292         while (p->windows) {
 293                 struct pack_window *w = p->windows;
 294
 295                 if (w->inuse_cnt)
 296                         die("pack '%s' still has open windows to it",
 297                             p->pack_name);
 298                 munmap(w->base, w->len);
 299                 pack_mapped -= w->len;
 300                 pack_open_windows--;
 301                 p->windows = w->next;
 302                 free(w);
 303         }
 304 }
 305
 306 int close_pack_fd(struct packed_git *p)
 307 {
 308         if (p->pack_fd < 0)
 309                 return 0;
 310
 311         close(p->pack_fd);
 312         pack_open_fds--;
 313         p->pack_fd = -1;
 314
 315         return 1;
 316 }
 317
 318 void close_pack_index(struct packed_git *p)
 319 {
 320         if (p->index_data) {
 321                 munmap((void *)p->index_data, p->index_size);
 322                 p->index_data = NULL;
 323         }
 324 }
 325
 326 void close_pack(struct packed_git *p)
 327 {
 328         close_pack_windows(p);
 329         close_pack_fd(p);
 330         close_pack_index(p);
 331 }
 332
 333 void close_object_store(struct raw_object_store *o)
 334 {
 335         struct packed_git *p;
 336
 337         for (p = o->packed_git; p; p = p->next)
 338                 if (p->do_not_close)
 339                         BUG("want to close pack marked 'do-not-close'");
 340                 else
 341                         close_pack(p);
 342
 343         if (o->multi_pack_index) {
 344                 close_midx(o->multi_pack_index);
 345                 o->multi_pack_index = NULL;
 346         }
 347
 348         close_commit_graph(o);
 349 }
 350
 351 void unlink_pack_path(const char *pack_name, int force_delete)
 352 {
 353         static const char *exts[] = {".pack", ".idx", ".keep", ".bitmap", ".promisor"};
 354         int i;
 355         struct strbuf buf = STRBUF_INIT;
 356         size_t plen;
 357
 358         strbuf_addstr(&buf, pack_name);
 359         strip_suffix_mem(buf.buf, &buf.len, ".pack");
 360         plen = buf.len;
 361
 362         if (!force_delete) {
 363                 strbuf_addstr(&buf, ".keep");
 364                 if (!access(buf.buf, F_OK)) {
 365                         strbuf_release(&buf);
 366                         return;
 367                 }
 368         }
 369
 370         for (i = 0; i < ARRAY_SIZE(exts); i++) {
 371                 strbuf_setlen(&buf, plen);
 372                 strbuf_addstr(&buf, exts[i]);
 373                 unlink(buf.buf);
 374         }
 375
 376         strbuf_release(&buf);
 377 }
 378
 379 /*
 380  * The LRU pack is the one with the oldest MRU window, preferring packs
 381  * with no used windows, or the oldest mtime if it has no windows allocated.
 382  */
 383 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
 384 {
 385         struct pack_window *w, *this_mru_w;
 386         int has_windows_inuse = 0;
 387
 388         /*
 389          * Reject this pack if it has windows and the previously selected
 390          * one does not.  If this pack does not have windows, reject
 391          * it if the pack file is newer than the previously selected one.
 392          */
 393         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
 394                 return;
 395
 396         for (w = this_mru_w = p->windows; w; w = w->next) {
 397                 /*
 398                  * Reject this pack if any of its windows are in use,
 399                  * but the previously selected pack did not have any
 400                  * inuse windows.  Otherwise, record that this pack
 401                  * has windows in use.
 402                  */
 403                 if (w->inuse_cnt) {
 404                         if (*accept_windows_inuse)
 405                                 has_windows_inuse = 1;
 406                         else
 407                                 return;
 408                 }
 409
 410                 if (w->last_used > this_mru_w->last_used)
 411                         this_mru_w = w;
 412
 413                 /*
 414                  * Reject this pack if it has windows that have been
 415                  * used more recently than the previously selected pack.
 416                  * If the previously selected pack had windows inuse and
 417                  * we have not encountered a window in this pack that is
 418                  * inuse, skip this check since we prefer a pack with no
 419                  * inuse windows to one that has inuse windows.
 420                  */
 421                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
 422                     this_mru_w->last_used > (*mru_w)->last_used)
 423                         return;
 424         }
 425
 426         /*
 427          * Select this pack.
 428          */
 429         *mru_w = this_mru_w;
 430         *lru_p = p;
 431         *accept_windows_inuse = has_windows_inuse;
 432 }
 433
 434 static int close_one_pack(void)
 435 {
 436         struct packed_git *p, *lru_p = NULL;
 437         struct pack_window *mru_w = NULL;
 438         int accept_windows_inuse = 1;
 439
 440         for (p = the_repository->objects->packed_git; p; p = p->next) {
 441                 if (p->pack_fd == -1)
 442                         continue;
 443                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
 444         }
 445
 446         if (lru_p)
 447                 return close_pack_fd(lru_p);
 448
 449         return 0;
 450 }
 451
 452 static unsigned int get_max_fd_limit(void)
 453 {
 454 #ifdef RLIMIT_NOFILE
 455         {
 456                 struct rlimit lim;
 457
 458                 if (!getrlimit(RLIMIT_NOFILE, &lim))
 459                         return lim.rlim_cur;
 460         }
 461 #endif
 462
 463 #ifdef _SC_OPEN_MAX
 464         {
 465                 long open_max = sysconf(_SC_OPEN_MAX);
 466                 if (0 < open_max)
 467                         return open_max;
 468                 /*
 469                  * Otherwise, we got -1 for one of the two
 470                  * reasons:
 471                  *
 472                  * (1) sysconf() did not understand _SC_OPEN_MAX
 473                  *     and signaled an error with -1; or
 474                  * (2) sysconf() said there is no limit.
 475                  *
 476                  * We _could_ clear errno before calling sysconf() to
 477                  * tell these two cases apart and return a huge number
 478                  * in the latter case to let the caller cap it to a
 479                  * value that is not so selfish, but letting the
 480                  * fallback OPEN_MAX codepath take care of these cases
 481                  * is a lot simpler.
 482                  */
 483         }
 484 #endif
 485
 486 #ifdef OPEN_MAX
 487         return OPEN_MAX;
 488 #else
 489         return 1; /* see the caller ;-) */
 490 #endif
 491 }
 492
 493 const char *pack_basename(struct packed_git *p)
 494 {
 495         const char *ret = strrchr(p->pack_name, '/');
 496         if (ret)
 497                 ret = ret + 1; /* skip past slash */
 498         else
 499                 ret = p->pack_name; /* we only have a base */
 500         return ret;
 501 }
 502
 503 /*
 504  * Do not call this directly as this leaks p->pack_fd on error return;
 505  * call open_packed_git() instead.
 506  */
 507 static int open_packed_git_1(struct packed_git *p)
 508 {
 509         struct stat st;
 510         struct pack_header hdr;
 511         unsigned char hash[GIT_MAX_RAWSZ];
 512         unsigned char *idx_hash;
 513         ssize_t read_result;
 514         const unsigned hashsz = the_hash_algo->rawsz;
 515
 516         if (!p->index_data) {
 517                 struct multi_pack_index *m;
 518                 const char *pack_name = pack_basename(p);
 519
 520                 for (m = the_repository->objects->multi_pack_index;
 521                      m; m = m->next) {
 522                         if (midx_contains_pack(m, pack_name))
 523                                 break;
 524                 }
 525
 526                 if (!m && open_pack_index(p))
 527                         return error("packfile %s index unavailable", p->pack_name);
 528         }
 529
 530         if (!pack_max_fds) {
 531                 unsigned int max_fds = get_max_fd_limit();
 532
 533                 /* Save 3 for stdin/stdout/stderr, 22 for work */
 534                 if (25 < max_fds)
 535                         pack_max_fds = max_fds - 25;
 536                 else
 537                         pack_max_fds = 1;
 538         }
 539
 540         while (pack_max_fds <= pack_open_fds && close_one_pack())
 541                 ; /* nothing */
 542
 543         p->pack_fd = git_open(p->pack_name);
 544         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
 545                 return -1;
 546         pack_open_fds++;
 547
 548         /* If we created the struct before we had the pack we lack size. */
 549         if (!p->pack_size) {
 550                 if (!S_ISREG(st.st_mode))
 551                         return error("packfile %s not a regular file", p->pack_name);
 552                 p->pack_size = st.st_size;
 553         } else if (p->pack_size != st.st_size)
 554                 return error("packfile %s size changed", p->pack_name);
 555
 556         /* Verify we recognize this pack file format. */
 557         read_result = read_in_full(p->pack_fd, &hdr, sizeof(hdr));
 558         if (read_result < 0)
 559                 return error_errno("error reading from %s", p->pack_name);
 560         if (read_result != sizeof(hdr))
 561                 return error("file %s is far too short to be a packfile", p->pack_name);
 562         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
 563                 return error("file %s is not a GIT packfile", p->pack_name);
 564         if (!pack_version_ok(hdr.hdr_version))
 565                 return error("packfile %s is version %"PRIu32" and not"
 566                         " supported (try upgrading GIT to a newer version)",
 567                         p->pack_name, ntohl(hdr.hdr_version));
 568
 569         /* Skip index checking if in multi-pack-index */
 570         if (!p->index_data)
 571                 return 0;
 572
 573         /* Verify the pack matches its index. */
 574         if (p->num_objects != ntohl(hdr.hdr_entries))
 575                 return error("packfile %s claims to have %"PRIu32" objects"
 576                              " while index indicates %"PRIu32" objects",
 577                              p->pack_name, ntohl(hdr.hdr_entries),
 578                              p->num_objects);
 579         read_result = pread_in_full(p->pack_fd, hash, hashsz,
 580                                         p->pack_size - hashsz);
 581         if (read_result < 0)
 582                 return error_errno("error reading from %s", p->pack_name);
 583         if (read_result != hashsz)
 584                 return error("packfile %s signature is unavailable", p->pack_name);
 585         idx_hash = ((unsigned char *)p->index_data) + p->index_size - hashsz * 2;
 586         if (!hasheq(hash, idx_hash))
 587                 return error("packfile %s does not match index", p->pack_name);
 588         return 0;
 589 }
 590
 591 static int open_packed_git(struct packed_git *p)
 592 {
 593         if (!open_packed_git_1(p))
 594                 return 0;
 595         close_pack_fd(p);
 596         return -1;
 597 }
 598
 599 static int in_window(struct pack_window *win, off_t offset)
 600 {
 601         /* We must promise at least one full hash after the
 602          * offset is available from this window, otherwise the offset
 603          * is not actually in this window and a different window (which
 604          * has that one hash excess) must be used.  This is to support
 605          * the object header and delta base parsing routines below.
 606          */
 607         off_t win_off = win->offset;
 608         return win_off <= offset
 609                 && (offset + the_hash_algo->rawsz) <= (win_off + win->len);
 610 }
 611
 612 unsigned char *use_pack(struct packed_git *p,
 613                 struct pack_window **w_cursor,
 614                 off_t offset,
 615                 unsigned long *left)
 616 {
 617         struct pack_window *win = *w_cursor;
 618
 619         /* Since packfiles end in a hash of their content and it's
 620          * pointless to ask for an offset into the middle of that
 621          * hash, and the in_window function above wouldn't match
 622          * don't allow an offset too close to the end of the file.
 623          */
 624         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
 625                 die("packfile %s cannot be accessed", p->pack_name);
 626         if (offset > (p->pack_size - the_hash_algo->rawsz))
 627                 die("offset beyond end of packfile (truncated pack?)");
 628         if (offset < 0)
 629                 die(_("offset before end of packfile (broken .idx?)"));
 630
 631         if (!win || !in_window(win, offset)) {
 632                 if (win)
 633                         win->inuse_cnt--;
 634                 for (win = p->windows; win; win = win->next) {
 635                         if (in_window(win, offset))
 636                                 break;
 637                 }
 638                 if (!win) {
 639                         size_t window_align = packed_git_window_size / 2;
 640                         off_t len;
 641
 642                         if (p->pack_fd == -1 && open_packed_git(p))
 643                                 die("packfile %s cannot be accessed", p->pack_name);
 644
 645                         win = xcalloc(1, sizeof(*win));
 646                         win->offset = (offset / window_align) * window_align;
 647                         len = p->pack_size - win->offset;
 648                         if (len > packed_git_window_size)
 649                                 len = packed_git_window_size;
 650                         win->len = (size_t)len;
 651                         pack_mapped += win->len;
 652                         while (packed_git_limit < pack_mapped
 653                                 && unuse_one_window(p))
 654                                 ; /* nothing */
 655                         win->base = xmmap_gently(NULL, win->len,
 656                                 PROT_READ, MAP_PRIVATE,
 657                                 p->pack_fd, win->offset);
 658                         if (win->base == MAP_FAILED)
 659                                 die_errno("packfile %s cannot be mapped",
 660                                           p->pack_name);
 661                         if (!win->offset && win->len == p->pack_size
 662                                 && !p->do_not_close)
 663                                 close_pack_fd(p);
 664                         pack_mmap_calls++;
 665                         pack_open_windows++;
 666                         if (pack_mapped > peak_pack_mapped)
 667                                 peak_pack_mapped = pack_mapped;
 668                         if (pack_open_windows > peak_pack_open_windows)
 669                                 peak_pack_open_windows = pack_open_windows;
 670                         win->next = p->windows;
 671                         p->windows = win;
 672                 }
 673         }
 674         if (win != *w_cursor) {
 675                 win->last_used = pack_used_ctr++;
 676                 win->inuse_cnt++;
 677                 *w_cursor = win;
 678         }
 679         offset -= win->offset;
 680         if (left)
 681                 *left = win->len - xsize_t(offset);
 682         return win->base + offset;
 683 }
 684
 685 void unuse_pack(struct pack_window **w_cursor)
 686 {
 687         struct pack_window *w = *w_cursor;
 688         if (w) {
 689                 w->inuse_cnt--;
 690                 *w_cursor = NULL;
 691         }
 692 }
 693
 694 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
 695 {
 696         struct stat st;
 697         size_t alloc;
 698         struct packed_git *p;
 699
 700         /*
 701          * Make sure a corresponding .pack file exists and that
 702          * the index looks sane.
 703          */
 704         if (!strip_suffix_mem(path, &path_len, ".idx"))
 705                 return NULL;
 706
 707         /*
 708          * ".promisor" is long enough to hold any suffix we're adding (and
 709          * the use xsnprintf double-checks that)
 710          */
 711         alloc = st_add3(path_len, strlen(".promisor"), 1);
 712         p = alloc_packed_git(alloc);
 713         memcpy(p->pack_name, path, path_len);
 714
 715         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
 716         if (!access(p->pack_name, F_OK))
 717                 p->pack_keep = 1;
 718
 719         xsnprintf(p->pack_name + path_len, alloc - path_len, ".promisor");
 720         if (!access(p->pack_name, F_OK))
 721                 p->pack_promisor = 1;
 722
 723         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
 724         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
 725                 free(p);
 726                 return NULL;
 727         }
 728
 729         /* ok, it looks sane as far as we can check without
 730          * actually mapping the pack file.
 731          */
 732         p->pack_size = st.st_size;
 733         p->pack_local = local;
 734         p->mtime = st.st_mtime;
 735         if (path_len < the_hash_algo->hexsz ||
 736             get_sha1_hex(path + path_len - the_hash_algo->hexsz, p->hash))
 737                 hashclr(p->hash);
 738         return p;
 739 }
 740
 741 void install_packed_git(struct repository *r, struct packed_git *pack)
 742 {
 743         if (pack->pack_fd != -1)
 744                 pack_open_fds++;
 745
 746         pack->next = r->objects->packed_git;
 747         r->objects->packed_git = pack;
 748
 749         hashmap_entry_init(&pack->packmap_ent, strhash(pack->pack_name));
 750         hashmap_add(&r->objects->pack_map, &pack->packmap_ent);
 751 }
 752
 753 void (*report_garbage)(unsigned seen_bits, const char *path);
 754
 755 static void report_helper(const struct string_list *list,
 756                           int seen_bits, int first, int last)
 757 {
 758         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
 759                 return;
 760
 761         for (; first < last; first++)
 762                 report_garbage(seen_bits, list->items[first].string);
 763 }
 764
 765 static void report_pack_garbage(struct string_list *list)
 766 {
 767         int i, baselen = -1, first = 0, seen_bits = 0;
 768
 769         if (!report_garbage)
 770                 return;
 771
 772         string_list_sort(list);
 773
 774         for (i = 0; i < list->nr; i++) {
 775                 const char *path = list->items[i].string;
 776                 if (baselen != -1 &&
 777                     strncmp(path, list->items[first].string, baselen)) {
 778                         report_helper(list, seen_bits, first, i);
 779                         baselen = -1;
 780                         seen_bits = 0;
 781                 }
 782                 if (baselen == -1) {
 783                         const char *dot = strrchr(path, '.');
 784                         if (!dot) {
 785                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
 786                                 continue;
 787                         }
 788                         baselen = dot - path + 1;
 789                         first = i;
 790                 }
 791                 if (!strcmp(path + baselen, "pack"))
 792                         seen_bits |= 1;
 793                 else if (!strcmp(path + baselen, "idx"))
 794                         seen_bits |= 2;
 795         }
 796         report_helper(list, seen_bits, first, list->nr);
 797 }
 798
 799 void for_each_file_in_pack_dir(const char *objdir,
 800                                each_file_in_pack_dir_fn fn,
 801                                void *data)
 802 {
 803         struct strbuf path = STRBUF_INIT;
 804         size_t dirnamelen;
 805         DIR *dir;
 806         struct dirent *de;
 807
 808         strbuf_addstr(&path, objdir);
 809         strbuf_addstr(&path, "/pack");
 810         dir = opendir(path.buf);
 811         if (!dir) {
 812                 if (errno != ENOENT)
 813                         error_errno("unable to open object pack directory: %s",
 814                                     path.buf);
 815                 strbuf_release(&path);
 816                 return;
 817         }
 818         strbuf_addch(&path, '/');
 819         dirnamelen = path.len;
 820         while ((de = readdir(dir)) != NULL) {
 821                 if (is_dot_or_dotdot(de->d_name))
 822                         continue;
 823
 824                 strbuf_setlen(&path, dirnamelen);
 825                 strbuf_addstr(&path, de->d_name);
 826
 827                 fn(path.buf, path.len, de->d_name, data);
 828         }
 829
 830         closedir(dir);
 831         strbuf_release(&path);
 832 }
 833
 834 struct prepare_pack_data {
 835         struct repository *r;
 836         struct string_list *garbage;
 837         int local;
 838         struct multi_pack_index *m;
 839 };
 840
 841 static void prepare_pack(const char *full_name, size_t full_name_len,
 842                          const char *file_name, void *_data)
 843 {
 844         struct prepare_pack_data *data = (struct prepare_pack_data *)_data;
 845         struct packed_git *p;
 846         size_t base_len = full_name_len;
 847
 848         if (strip_suffix_mem(full_name, &base_len, ".idx") &&
 849             !(data->m && midx_contains_pack(data->m, file_name))) {
 850                 struct hashmap_entry hent;
 851                 char *pack_name = xstrfmt("%.*s.pack", (int)base_len, full_name);
 852                 unsigned int hash = strhash(pack_name);
 853                 hashmap_entry_init(&hent, hash);
 854
 855                 /* Don't reopen a pack we already have. */
 856                 if (!hashmap_get(&data->r->objects->pack_map, &hent, pack_name)) {
 857                         p = add_packed_git(full_name, full_name_len, data->local);
 858                         if (p)
 859                                 install_packed_git(data->r, p);
 860                 }
 861                 free(pack_name);
 862         }
 863
 864         if (!report_garbage)
 865                 return;
 866
 867         if (!strcmp(file_name, "multi-pack-index"))
 868                 return;
 869         if (ends_with(file_name, ".idx") ||
 870             ends_with(file_name, ".pack") ||
 871             ends_with(file_name, ".bitmap") ||
 872             ends_with(file_name, ".keep") ||
 873             ends_with(file_name, ".promisor"))
 874                 string_list_append(data->garbage, full_name);
 875         else
 876                 report_garbage(PACKDIR_FILE_GARBAGE, full_name);
 877 }
 878
 879 static void prepare_packed_git_one(struct repository *r, char *objdir, int local)
 880 {
 881         struct prepare_pack_data data;
 882         struct string_list garbage = STRING_LIST_INIT_DUP;
 883
 884         data.m = r->objects->multi_pack_index;
 885
 886         /* look for the multi-pack-index for this object directory */
 887         while (data.m && strcmp(data.m->object_dir, objdir))
 888                 data.m = data.m->next;
 889
 890         data.r = r;
 891         data.garbage = &garbage;
 892         data.local = local;
 893
 894         for_each_file_in_pack_dir(objdir, prepare_pack, &data);
 895
 896         report_pack_garbage(data.garbage);
 897         string_list_clear(data.garbage, 0);
 898 }
 899
 900 static void prepare_packed_git(struct repository *r);
 901 /*
 902  * Give a fast, rough count of the number of objects in the repository. This
 903  * ignores loose objects completely. If you have a lot of them, then either
 904  * you should repack because your performance will be awful, or they are
 905  * all unreachable objects about to be pruned, in which case they're not really
 906  * interesting as a measure of repo size in the first place.
 907  */
 908 unsigned long repo_approximate_object_count(struct repository *r)
 909 {
 910         if (!r->objects->approximate_object_count_valid) {
 911                 unsigned long count;
 912                 struct multi_pack_index *m;
 913                 struct packed_git *p;
 914
 915                 prepare_packed_git(r);
 916                 count = 0;
 917                 for (m = get_multi_pack_index(r); m; m = m->next)
 918                         count += m->num_objects;
 919                 for (p = r->objects->packed_git; p; p = p->next) {
 920                         if (open_pack_index(p))
 921                                 continue;
 922                         count += p->num_objects;
 923                 }
 924                 r->objects->approximate_object_count = count;
 925         }
 926         return r->objects->approximate_object_count;
 927 }
 928
 929 static void *get_next_packed_git(const void *p)
 930 {
 931         return ((const struct packed_git *)p)->next;
 932 }
 933
 934 static void set_next_packed_git(void *p, void *next)
 935 {
 936         ((struct packed_git *)p)->next = next;
 937 }
 938
 939 static int sort_pack(const void *a_, const void *b_)
 940 {
 941         const struct packed_git *a = a_;
 942         const struct packed_git *b = b_;
 943         int st;
 944
 945         /*
 946          * Local packs tend to contain objects specific to our
 947          * variant of the project than remote ones.  In addition,
 948          * remote ones could be on a network mounted filesystem.
 949          * Favor local ones for these reasons.
 950          */
 951         st = a->pack_local - b->pack_local;
 952         if (st)
 953                 return -st;
 954
 955         /*
 956          * Younger packs tend to contain more recent objects,
 957          * and more recent objects tend to get accessed more
 958          * often.
 959          */
 960         if (a->mtime < b->mtime)
 961                 return 1;
 962         else if (a->mtime == b->mtime)
 963                 return 0;
 964         return -1;
 965 }
 966
 967 static void rearrange_packed_git(struct repository *r)
 968 {
 969         r->objects->packed_git = llist_mergesort(
 970                 r->objects->packed_git, get_next_packed_git,
 971                 set_next_packed_git, sort_pack);
 972 }
 973
 974 static void prepare_packed_git_mru(struct repository *r)
 975 {
 976         struct packed_git *p;
 977
 978         INIT_LIST_HEAD(&r->objects->packed_git_mru);
 979
 980         for (p = r->objects->packed_git; p; p = p->next)
 981                 list_add_tail(&p->mru, &r->objects->packed_git_mru);
 982 }
 983
 984 static void prepare_packed_git(struct repository *r)
 985 {
 986         struct object_directory *odb;
 987
 988         if (r->objects->packed_git_initialized)
 989                 return;
 990
 991         prepare_alt_odb(r);
 992         for (odb = r->objects->odb; odb; odb = odb->next) {
 993                 int local = (odb == r->objects->odb);
 994                 prepare_multi_pack_index_one(r, odb->path, local);
 995                 prepare_packed_git_one(r, odb->path, local);
 996         }
 997         rearrange_packed_git(r);
 998
 999         prepare_packed_git_mru(r);
1000         r->objects->packed_git_initialized = 1;
1001 }
1002
1003 void reprepare_packed_git(struct repository *r)
1004 {
1005         struct object_directory *odb;
1006
1007         obj_read_lock();
1008         for (odb = r->objects->odb; odb; odb = odb->next)
1009                 odb_clear_loose_cache(odb);
1010
1011         r->objects->approximate_object_count_valid = 0;
1012         r->objects->packed_git_initialized = 0;
1013         prepare_packed_git(r);
1014         obj_read_unlock();
1015 }
1016
1017 struct packed_git *get_packed_git(struct repository *r)
1018 {
1019         prepare_packed_git(r);
1020         return r->objects->packed_git;
1021 }
1022
1023 struct multi_pack_index *get_multi_pack_index(struct repository *r)
1024 {
1025         prepare_packed_git(r);
1026         return r->objects->multi_pack_index;
1027 }
1028
1029 struct packed_git *get_all_packs(struct repository *r)
1030 {
1031         struct multi_pack_index *m;
1032
1033         prepare_packed_git(r);
1034         for (m = r->objects->multi_pack_index; m; m = m->next) {
1035                 uint32_t i;
1036                 for (i = 0; i < m->num_packs; i++)
1037                         prepare_midx_pack(r, m, i);
1038         }
1039
1040         return r->objects->packed_git;
1041 }
1042
1043 struct list_head *get_packed_git_mru(struct repository *r)
1044 {
1045         prepare_packed_git(r);
1046         return &r->objects->packed_git_mru;
1047 }
1048
1049 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1050                 unsigned long len, enum object_type *type, unsigned long *sizep)
1051 {
1052         unsigned shift;
1053         unsigned long size, c;
1054         unsigned long used = 0;
1055
1056         c = buf[used++];
1057         *type = (c >> 4) & 7;
1058         size = c & 15;
1059         shift = 4;
1060         while (c & 0x80) {
1061                 if (len <= used || bitsizeof(long) <= shift) {
1062                         error("bad object header");
1063                         size = used = 0;
1064                         break;
1065                 }
1066                 c = buf[used++];
1067                 size += (c & 0x7f) << shift;
1068                 shift += 7;
1069         }
1070         *sizep = size;
1071         return used;
1072 }
1073
1074 unsigned long get_size_from_delta(struct packed_git *p,
1075                                   struct pack_window **w_curs,
1076                                   off_t curpos)
1077 {
1078         const unsigned char *data;
1079         unsigned char delta_head[20], *in;
1080         git_zstream stream;
1081         int st;
1082
1083         memset(&stream, 0, sizeof(stream));
1084         stream.next_out = delta_head;
1085         stream.avail_out = sizeof(delta_head);
1086
1087         git_inflate_init(&stream);
1088         do {
1089                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1090                 stream.next_in = in;
1091                 /*
1092                  * Note: the window section returned by use_pack() must be
1093                  * available throughout git_inflate()'s unlocked execution. To
1094                  * ensure no other thread will modify the window in the
1095                  * meantime, we rely on the packed_window.inuse_cnt. This
1096                  * counter is incremented before window reading and checked
1097                  * before window disposal.
1098                  *
1099                  * Other worrying sections could be the call to close_pack_fd(),
1100                  * which can close packs even with in-use windows, and to
1101                  * reprepare_packed_git(). Regarding the former, mmap doc says:
1102                  * "closing the file descriptor does not unmap the region". And
1103                  * for the latter, it won't re-open already available packs.
1104                  */
1105                 obj_read_unlock();
1106                 st = git_inflate(&stream, Z_FINISH);
1107                 obj_read_lock();
1108                 curpos += stream.next_in - in;
1109         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1110                  stream.total_out < sizeof(delta_head));
1111         git_inflate_end(&stream);
1112         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1113                 error("delta data unpack-initial failed");
1114                 return 0;
1115         }
1116
1117         /* Examine the initial part of the delta to figure out
1118          * the result size.
1119          */
1120         data = delta_head;
1121
1122         /* ignore base size */
1123         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1124
1125         /* Read the result size */
1126         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1127 }
1128
1129 int unpack_object_header(struct packed_git *p,
1130                          struct pack_window **w_curs,
1131                          off_t *curpos,
1132                          unsigned long *sizep)
1133 {
1134         unsigned char *base;
1135         unsigned long left;
1136         unsigned long used;
1137         enum object_type type;
1138
1139         /* use_pack() assures us we have [base, base + 20) available
1140          * as a range that we can look at.  (Its actually the hash
1141          * size that is assured.)  With our object header encoding
1142          * the maximum deflated object size is 2^137, which is just
1143          * insane, so we know won't exceed what we have been given.
1144          */
1145         base = use_pack(p, w_curs, *curpos, &left);
1146         used = unpack_object_header_buffer(base, left, &type, sizep);
1147         if (!used) {
1148                 type = OBJ_BAD;
1149         } else
1150                 *curpos += used;
1151
1152         return type;
1153 }
1154
1155 void mark_bad_packed_object(struct packed_git *p, const unsigned char *sha1)
1156 {
1157         unsigned i;
1158         const unsigned hashsz = the_hash_algo->rawsz;
1159         for (i = 0; i < p->num_bad_objects; i++)
1160                 if (hasheq(sha1, p->bad_object_sha1 + hashsz * i))
1161                         return;
1162         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1163                                       st_mult(GIT_MAX_RAWSZ,
1164                                               st_add(p->num_bad_objects, 1)));
1165         hashcpy(p->bad_object_sha1 + hashsz * p->num_bad_objects, sha1);
1166         p->num_bad_objects++;
1167 }
1168
1169 const struct packed_git *has_packed_and_bad(struct repository *r,
1170                                             const unsigned char *sha1)
1171 {
1172         struct packed_git *p;
1173         unsigned i;
1174
1175         for (p = r->objects->packed_git; p; p = p->next)
1176                 for (i = 0; i < p->num_bad_objects; i++)
1177                         if (hasheq(sha1,
1178                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
1179                                 return p;
1180         return NULL;
1181 }
1182
1183 off_t get_delta_base(struct packed_git *p,
1184                      struct pack_window **w_curs,
1185                      off_t *curpos,
1186                      enum object_type type,
1187                      off_t delta_obj_offset)
1188 {
1189         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1190         off_t base_offset;
1191
1192         /* use_pack() assured us we have [base_info, base_info + 20)
1193          * as a range that we can look at without walking off the
1194          * end of the mapped window.  Its actually the hash size
1195          * that is assured.  An OFS_DELTA longer than the hash size
1196          * is stupid, as then a REF_DELTA would be smaller to store.
1197          */
1198         if (type == OBJ_OFS_DELTA) {
1199                 unsigned used = 0;
1200                 unsigned char c = base_info[used++];
1201                 base_offset = c & 127;
1202                 while (c & 128) {
1203                         base_offset += 1;
1204                         if (!base_offset || MSB(base_offset, 7))
1205                                 return 0;  /* overflow */
1206                         c = base_info[used++];
1207                         base_offset = (base_offset << 7) + (c & 127);
1208                 }
1209                 base_offset = delta_obj_offset - base_offset;
1210                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1211                         return 0;  /* out of bound */
1212                 *curpos += used;
1213         } else if (type == OBJ_REF_DELTA) {
1214                 /* The base entry _must_ be in the same pack */
1215                 base_offset = find_pack_entry_one(base_info, p);
1216                 *curpos += the_hash_algo->rawsz;
1217         } else
1218                 die("I am totally screwed");
1219         return base_offset;
1220 }
1221
1222 /*
1223  * Like get_delta_base above, but we return the sha1 instead of the pack
1224  * offset. This means it is cheaper for REF deltas (we do not have to do
1225  * the final object lookup), but more expensive for OFS deltas (we
1226  * have to load the revidx to convert the offset back into a sha1).
1227  */
1228 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1229                                                 struct pack_window **w_curs,
1230                                                 off_t curpos,
1231                                                 enum object_type type,
1232                                                 off_t delta_obj_offset)
1233 {
1234         if (type == OBJ_REF_DELTA) {
1235                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1236                 return base;
1237         } else if (type == OBJ_OFS_DELTA) {
1238                 struct revindex_entry *revidx;
1239                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1240                                                    type, delta_obj_offset);
1241
1242                 if (!base_offset)
1243                         return NULL;
1244
1245                 revidx = find_pack_revindex(p, base_offset);
1246                 if (!revidx)
1247                         return NULL;
1248
1249                 return nth_packed_object_sha1(p, revidx->nr);
1250         } else
1251                 return NULL;
1252 }
1253
1254 static int retry_bad_packed_offset(struct repository *r,
1255                                    struct packed_git *p,
1256                                    off_t obj_offset)
1257 {
1258         int type;
1259         struct revindex_entry *revidx;
1260         struct object_id oid;
1261         revidx = find_pack_revindex(p, obj_offset);
1262         if (!revidx)
1263                 return OBJ_BAD;
1264         nth_packed_object_id(&oid, p, revidx->nr);
1265         mark_bad_packed_object(p, oid.hash);
1266         type = oid_object_info(r, &oid, NULL);
1267         if (type <= OBJ_NONE)
1268                 return OBJ_BAD;
1269         return type;
1270 }
1271
1272 #define POI_STACK_PREALLOC 64
1273
1274 static enum object_type packed_to_object_type(struct repository *r,
1275                                               struct packed_git *p,
1276                                               off_t obj_offset,
1277                                               enum object_type type,
1278                                               struct pack_window **w_curs,
1279                                               off_t curpos)
1280 {
1281         off_t small_poi_stack[POI_STACK_PREALLOC];
1282         off_t *poi_stack = small_poi_stack;
1283         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1284
1285         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1286                 off_t base_offset;
1287                 unsigned long size;
1288                 /* Push the object we're going to leave behind */
1289                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1290                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1291                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
1292                         COPY_ARRAY(poi_stack, small_poi_stack, poi_stack_nr);
1293                 } else {
1294                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1295                 }
1296                 poi_stack[poi_stack_nr++] = obj_offset;
1297                 /* If parsing the base offset fails, just unwind */
1298                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1299                 if (!base_offset)
1300                         goto unwind;
1301                 curpos = obj_offset = base_offset;
1302                 type = unpack_object_header(p, w_curs, &curpos, &size);
1303                 if (type <= OBJ_NONE) {
1304                         /* If getting the base itself fails, we first
1305                          * retry the base, otherwise unwind */
1306                         type = retry_bad_packed_offset(r, p, base_offset);
1307                         if (type > OBJ_NONE)
1308                                 goto out;
1309                         goto unwind;
1310                 }
1311         }
1312
1313         switch (type) {
1314         case OBJ_BAD:
1315         case OBJ_COMMIT:
1316         case OBJ_TREE:
1317         case OBJ_BLOB:
1318         case OBJ_TAG:
1319                 break;
1320         default:
1321                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1322                       type, (uintmax_t)obj_offset, p->pack_name);
1323                 type = OBJ_BAD;
1324         }
1325
1326 out:
1327         if (poi_stack != small_poi_stack)
1328                 free(poi_stack);
1329         return type;
1330
1331 unwind:
1332         while (poi_stack_nr) {
1333                 obj_offset = poi_stack[--poi_stack_nr];
1334                 type = retry_bad_packed_offset(r, p, obj_offset);
1335                 if (type > OBJ_NONE)
1336                         goto out;
1337         }
1338         type = OBJ_BAD;
1339         goto out;
1340 }
1341
1342 static struct hashmap delta_base_cache;
1343 static size_t delta_base_cached;
1344
1345 static LIST_HEAD(delta_base_cache_lru);
1346
1347 struct delta_base_cache_key {
1348         struct packed_git *p;
1349         off_t base_offset;
1350 };
1351
1352 struct delta_base_cache_entry {
1353         struct hashmap_entry ent;
1354         struct delta_base_cache_key key;
1355         struct list_head lru;
1356         void *data;
1357         unsigned long size;
1358         enum object_type type;
1359 };
1360
1361 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
1362 {
1363         unsigned int hash;
1364
1365         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
1366         hash += (hash >> 8) + (hash >> 16);
1367         return hash;
1368 }
1369
1370 static struct delta_base_cache_entry *
1371 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
1372 {
1373         struct hashmap_entry entry, *e;
1374         struct delta_base_cache_key key;
1375
1376         if (!delta_base_cache.cmpfn)
1377                 return NULL;
1378
1379         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
1380         key.p = p;
1381         key.base_offset = base_offset;
1382         e = hashmap_get(&delta_base_cache, &entry, &key);
1383         return e ? container_of(e, struct delta_base_cache_entry, ent) : NULL;
1384 }
1385
1386 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
1387                                    const struct delta_base_cache_key *b)
1388 {
1389         return a->p == b->p && a->base_offset == b->base_offset;
1390 }
1391
1392 static int delta_base_cache_hash_cmp(const void *unused_cmp_data,
1393                                      const struct hashmap_entry *va,
1394                                      const struct hashmap_entry *vb,
1395                                      const void *vkey)
1396 {
1397         const struct delta_base_cache_entry *a, *b;
1398         const struct delta_base_cache_key *key = vkey;
1399
1400         a = container_of(va, const struct delta_base_cache_entry, ent);
1401         b = container_of(vb, const struct delta_base_cache_entry, ent);
1402
1403         if (key)
1404                 return !delta_base_cache_key_eq(&a->key, key);
1405         else
1406                 return !delta_base_cache_key_eq(&a->key, &b->key);
1407 }
1408
1409 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
1410 {
1411         return !!get_delta_base_cache_entry(p, base_offset);
1412 }
1413
1414 /*
1415  * Remove the entry from the cache, but do _not_ free the associated
1416  * entry data. The caller takes ownership of the "data" buffer, and
1417  * should copy out any fields it wants before detaching.
1418  */
1419 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
1420 {
1421         hashmap_remove(&delta_base_cache, &ent->ent, &ent->key);
1422         list_del(&ent->lru);
1423         delta_base_cached -= ent->size;
1424         free(ent);
1425 }
1426
1427 static void *cache_or_unpack_entry(struct repository *r, struct packed_git *p,
1428                                    off_t base_offset, unsigned long *base_size,
1429                                    enum object_type *type)
1430 {
1431         struct delta_base_cache_entry *ent;
1432
1433         ent = get_delta_base_cache_entry(p, base_offset);
1434         if (!ent)
1435                 return unpack_entry(r, p, base_offset, type, base_size);
1436
1437         if (type)
1438                 *type = ent->type;
1439         if (base_size)
1440                 *base_size = ent->size;
1441         return xmemdupz(ent->data, ent->size);
1442 }
1443
1444 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1445 {
1446         free(ent->data);
1447         detach_delta_base_cache_entry(ent);
1448 }
1449
1450 void clear_delta_base_cache(void)
1451 {
1452         struct list_head *lru, *tmp;
1453         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1454                 struct delta_base_cache_entry *entry =
1455                         list_entry(lru, struct delta_base_cache_entry, lru);
1456                 release_delta_base_cache(entry);
1457         }
1458 }
1459
1460 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1461         void *base, unsigned long base_size, enum object_type type)
1462 {
1463         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
1464         struct list_head *lru, *tmp;
1465
1466         /*
1467          * Check required to avoid redundant entries when more than one thread
1468          * is unpacking the same object, in unpack_entry() (since its phases I
1469          * and III might run concurrently across multiple threads).
1470          */
1471         if (in_delta_base_cache(p, base_offset))
1472                 return;
1473
1474         delta_base_cached += base_size;
1475
1476         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1477                 struct delta_base_cache_entry *f =
1478                         list_entry(lru, struct delta_base_cache_entry, lru);
1479                 if (delta_base_cached <= delta_base_cache_limit)
1480                         break;
1481                 release_delta_base_cache(f);
1482         }
1483
1484         ent->key.p = p;
1485         ent->key.base_offset = base_offset;
1486         ent->type = type;
1487         ent->data = base;
1488         ent->size = base_size;
1489         list_add_tail(&ent->lru, &delta_base_cache_lru);
1490
1491         if (!delta_base_cache.cmpfn)
1492                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, NULL, 0);
1493         hashmap_entry_init(&ent->ent, pack_entry_hash(p, base_offset));
1494         hashmap_add(&delta_base_cache, &ent->ent);
1495 }
1496
1497 int packed_object_info(struct repository *r, struct packed_git *p,
1498                        off_t obj_offset, struct object_info *oi)
1499 {
1500         struct pack_window *w_curs = NULL;
1501         unsigned long size;
1502         off_t curpos = obj_offset;
1503         enum object_type type;
1504
1505         /*
1506          * We always get the representation type, but only convert it to
1507          * a "real" type later if the caller is interested.
1508          */
1509         if (oi->contentp) {
1510                 *oi->contentp = cache_or_unpack_entry(r, p, obj_offset, oi->sizep,
1511                                                       &type);
1512                 if (!*oi->contentp)
1513                         type = OBJ_BAD;
1514         } else {
1515                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1516         }
1517
1518         if (!oi->contentp && oi->sizep) {
1519                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1520                         off_t tmp_pos = curpos;
1521                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1522                                                            type, obj_offset);
1523                         if (!base_offset) {
1524                                 type = OBJ_BAD;
1525                                 goto out;
1526                         }
1527                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1528                         if (*oi->sizep == 0) {
1529                                 type = OBJ_BAD;
1530                                 goto out;
1531                         }
1532                 } else {
1533                         *oi->sizep = size;
1534                 }
1535         }
1536
1537         if (oi->disk_sizep) {
1538                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1539                 *oi->disk_sizep = revidx[1].offset - obj_offset;
1540         }
1541
1542         if (oi->typep || oi->type_name) {
1543                 enum object_type ptot;
1544                 ptot = packed_to_object_type(r, p, obj_offset,
1545                                              type, &w_curs, curpos);
1546                 if (oi->typep)
1547                         *oi->typep = ptot;
1548                 if (oi->type_name) {
1549                         const char *tn = type_name(ptot);
1550                         if (tn)
1551                                 strbuf_addstr(oi->type_name, tn);
1552                 }
1553                 if (ptot < 0) {
1554                         type = OBJ_BAD;
1555                         goto out;
1556                 }
1557         }
1558
1559         if (oi->delta_base_oid) {
1560                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1561                         const unsigned char *base;
1562
1563                         base = get_delta_base_sha1(p, &w_curs, curpos,
1564                                                    type, obj_offset);
1565                         if (!base) {
1566                                 type = OBJ_BAD;
1567                                 goto out;
1568                         }
1569
1570                         hashcpy(oi->delta_base_oid->hash, base);
1571                 } else
1572                         oidclr(oi->delta_base_oid);
1573         }
1574
1575         oi->whence = in_delta_base_cache(p, obj_offset) ? OI_DBCACHED :
1576                                                           OI_PACKED;
1577
1578 out:
1579         unuse_pack(&w_curs);
1580         return type;
1581 }
1582
1583 static void *unpack_compressed_entry(struct packed_git *p,
1584                                     struct pack_window **w_curs,
1585                                     off_t curpos,
1586                                     unsigned long size)
1587 {
1588         int st;
1589         git_zstream stream;
1590         unsigned char *buffer, *in;
1591
1592         buffer = xmallocz_gently(size);
1593         if (!buffer)
1594                 return NULL;
1595         memset(&stream, 0, sizeof(stream));
1596         stream.next_out = buffer;
1597         stream.avail_out = size + 1;
1598
1599         git_inflate_init(&stream);
1600         do {
1601                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1602                 stream.next_in = in;
1603                 /*
1604                  * Note: we must ensure the window section returned by
1605                  * use_pack() will be available throughout git_inflate()'s
1606                  * unlocked execution. Please refer to the comment at
1607                  * get_size_from_delta() to see how this is done.
1608                  */
1609                 obj_read_unlock();
1610                 st = git_inflate(&stream, Z_FINISH);
1611                 obj_read_lock();
1612                 if (!stream.avail_out)
1613                         break; /* the payload is larger than it should be */
1614                 curpos += stream.next_in - in;
1615         } while (st == Z_OK || st == Z_BUF_ERROR);
1616         git_inflate_end(&stream);
1617         if ((st != Z_STREAM_END) || stream.total_out != size) {
1618                 free(buffer);
1619                 return NULL;
1620         }
1621
1622         /* versions of zlib can clobber unconsumed portion of outbuf */
1623         buffer[size] = '\0';
1624
1625         return buffer;
1626 }
1627
1628 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
1629 {
1630         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
1631         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
1632                          p->pack_name, (uintmax_t)obj_offset);
1633 }
1634
1635 int do_check_packed_object_crc;
1636
1637 #define UNPACK_ENTRY_STACK_PREALLOC 64
1638 struct unpack_entry_stack_ent {
1639         off_t obj_offset;
1640         off_t curpos;
1641         unsigned long size;
1642 };
1643
1644 static void *read_object(struct repository *r,
1645                          const struct object_id *oid,
1646                          enum object_type *type,
1647                          unsigned long *size)
1648 {
1649         struct object_info oi = OBJECT_INFO_INIT;
1650         void *content;
1651         oi.typep = type;
1652         oi.sizep = size;
1653         oi.contentp = &content;
1654
1655         if (oid_object_info_extended(r, oid, &oi, 0) < 0)
1656                 return NULL;
1657         return content;
1658 }
1659
1660 void *unpack_entry(struct repository *r, struct packed_git *p, off_t obj_offset,
1661                    enum object_type *final_type, unsigned long *final_size)
1662 {
1663         struct pack_window *w_curs = NULL;
1664         off_t curpos = obj_offset;
1665         void *data = NULL;
1666         unsigned long size;
1667         enum object_type type;
1668         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
1669         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
1670         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
1671         int base_from_cache = 0;
1672
1673         write_pack_access_log(p, obj_offset);
1674
1675         /* PHASE 1: drill down to the innermost base object */
1676         for (;;) {
1677                 off_t base_offset;
1678                 int i;
1679                 struct delta_base_cache_entry *ent;
1680
1681                 ent = get_delta_base_cache_entry(p, curpos);
1682                 if (ent) {
1683                         type = ent->type;
1684                         data = ent->data;
1685                         size = ent->size;
1686                         detach_delta_base_cache_entry(ent);
1687                         base_from_cache = 1;
1688                         break;
1689                 }
1690
1691                 if (do_check_packed_object_crc && p->index_version > 1) {
1692                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1693                         off_t len = revidx[1].offset - obj_offset;
1694                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1695                                 struct object_id oid;
1696                                 nth_packed_object_id(&oid, p, revidx->nr);
1697                                 error("bad packed object CRC for %s",
1698                                       oid_to_hex(&oid));
1699                                 mark_bad_packed_object(p, oid.hash);
1700                                 data = NULL;
1701                                 goto out;
1702                         }
1703                 }
1704
1705                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1706                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
1707                         break;
1708
1709                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1710                 if (!base_offset) {
1711                         error("failed to validate delta base reference "
1712                               "at offset %"PRIuMAX" from %s",
1713                               (uintmax_t)curpos, p->pack_name);
1714                         /* bail to phase 2, in hopes of recovery */
1715                         data = NULL;
1716                         break;
1717                 }
1718
1719                 /* push object, proceed to base */
1720                 if (delta_stack_nr >= delta_stack_alloc
1721                     && delta_stack == small_delta_stack) {
1722                         delta_stack_alloc = alloc_nr(delta_stack_nr);
1723                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
1724                         COPY_ARRAY(delta_stack, small_delta_stack,
1725                                    delta_stack_nr);
1726                 } else {
1727                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
1728                 }
1729                 i = delta_stack_nr++;
1730                 delta_stack[i].obj_offset = obj_offset;
1731                 delta_stack[i].curpos = curpos;
1732                 delta_stack[i].size = size;
1733
1734                 curpos = obj_offset = base_offset;
1735         }
1736
1737         /* PHASE 2: handle the base */
1738         switch (type) {
1739         case OBJ_OFS_DELTA:
1740         case OBJ_REF_DELTA:
1741                 if (data)
1742                         BUG("unpack_entry: left loop at a valid delta");
1743                 break;
1744         case OBJ_COMMIT:
1745         case OBJ_TREE:
1746         case OBJ_BLOB:
1747         case OBJ_TAG:
1748                 if (!base_from_cache)
1749                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
1750                 break;
1751         default:
1752                 data = NULL;
1753                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1754                       type, (uintmax_t)obj_offset, p->pack_name);
1755         }
1756
1757         /* PHASE 3: apply deltas in order */
1758
1759         /* invariants:
1760          *   'data' holds the base data, or NULL if there was corruption
1761          */
1762         while (delta_stack_nr) {
1763                 void *delta_data;
1764                 void *base = data;
1765                 void *external_base = NULL;
1766                 unsigned long delta_size, base_size = size;
1767                 int i;
1768
1769                 data = NULL;
1770
1771                 if (base)
1772                         add_delta_base_cache(p, obj_offset, base, base_size, type);
1773
1774                 if (!base) {
1775                         /*
1776                          * We're probably in deep shit, but let's try to fetch
1777                          * the required base anyway from another pack or loose.
1778                          * This is costly but should happen only in the presence
1779                          * of a corrupted pack, and is better than failing outright.
1780                          */
1781                         struct revindex_entry *revidx;
1782                         struct object_id base_oid;
1783                         revidx = find_pack_revindex(p, obj_offset);
1784                         if (revidx) {
1785                                 nth_packed_object_id(&base_oid, p, revidx->nr);
1786                                 error("failed to read delta base object %s"
1787                                       " at offset %"PRIuMAX" from %s",
1788                                       oid_to_hex(&base_oid), (uintmax_t)obj_offset,
1789                                       p->pack_name);
1790                                 mark_bad_packed_object(p, base_oid.hash);
1791                                 base = read_object(r, &base_oid, &type, &base_size);
1792                                 external_base = base;
1793                         }
1794                 }
1795
1796                 i = --delta_stack_nr;
1797                 obj_offset = delta_stack[i].obj_offset;
1798                 curpos = delta_stack[i].curpos;
1799                 delta_size = delta_stack[i].size;
1800
1801                 if (!base)
1802                         continue;
1803
1804                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
1805
1806                 if (!delta_data) {
1807                         error("failed to unpack compressed delta "
1808                               "at offset %"PRIuMAX" from %s",
1809                               (uintmax_t)curpos, p->pack_name);
1810                         data = NULL;
1811                         free(external_base);
1812                         continue;
1813                 }
1814
1815                 data = patch_delta(base, base_size,
1816                                    delta_data, delta_size,
1817                                    &size);
1818
1819                 /*
1820                  * We could not apply the delta; warn the user, but keep going.
1821                  * Our failure will be noticed either in the next iteration of
1822                  * the loop, or if this is the final delta, in the caller when
1823                  * we return NULL. Those code paths will take care of making
1824                  * a more explicit warning and retrying with another copy of
1825                  * the object.
1826                  */
1827                 if (!data)
1828                         error("failed to apply delta");
1829
1830                 free(delta_data);
1831                 free(external_base);
1832         }
1833
1834         if (final_type)
1835                 *final_type = type;
1836         if (final_size)
1837                 *final_size = size;
1838
1839 out:
1840         unuse_pack(&w_curs);
1841
1842         if (delta_stack != small_delta_stack)
1843                 free(delta_stack);
1844
1845         return data;
1846 }
1847
1848 int bsearch_pack(const struct object_id *oid, const struct packed_git *p, uint32_t *result)
1849 {
1850         const unsigned char *index_fanout = p->index_data;
1851         const unsigned char *index_lookup;
1852         const unsigned int hashsz = the_hash_algo->rawsz;
1853         int index_lookup_width;
1854
1855         if (!index_fanout)
1856                 BUG("bsearch_pack called without a valid pack-index");
1857
1858         index_lookup = index_fanout + 4 * 256;
1859         if (p->index_version == 1) {
1860                 index_lookup_width = hashsz + 4;
1861                 index_lookup += 4;
1862         } else {
1863                 index_lookup_width = hashsz;
1864                 index_fanout += 8;
1865                 index_lookup += 8;
1866         }
1867
1868         return bsearch_hash(oid->hash, (const uint32_t*)index_fanout,
1869                             index_lookup, index_lookup_width, result);
1870 }
1871
1872 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1873                                             uint32_t n)
1874 {
1875         const unsigned char *index = p->index_data;
1876         const unsigned int hashsz = the_hash_algo->rawsz;
1877         if (!index) {
1878                 if (open_pack_index(p))
1879                         return NULL;
1880                 index = p->index_data;
1881         }
1882         if (n >= p->num_objects)
1883                 return NULL;
1884         index += 4 * 256;
1885         if (p->index_version == 1) {
1886                 return index + (hashsz + 4) * n + 4;
1887         } else {
1888                 index += 8;
1889                 return index + hashsz * n;
1890         }
1891 }
1892
1893 int nth_packed_object_id(struct object_id *oid,
1894                          struct packed_git *p,
1895                          uint32_t n)
1896 {
1897         const unsigned char *hash = nth_packed_object_sha1(p, n);
1898         if (!hash)
1899                 return -1;
1900         hashcpy(oid->hash, hash);
1901         return 0;
1902 }
1903
1904 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
1905 {
1906         const unsigned char *ptr = vptr;
1907         const unsigned char *start = p->index_data;
1908         const unsigned char *end = start + p->index_size;
1909         if (ptr < start)
1910                 die(_("offset before start of pack index for %s (corrupt index?)"),
1911                     p->pack_name);
1912         /* No need to check for underflow; .idx files must be at least 8 bytes */
1913         if (ptr >= end - 8)
1914                 die(_("offset beyond end of pack index for %s (truncated index?)"),
1915                     p->pack_name);
1916 }
1917
1918 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1919 {
1920         const unsigned char *index = p->index_data;
1921         const unsigned int hashsz = the_hash_algo->rawsz;
1922         index += 4 * 256;
1923         if (p->index_version == 1) {
1924                 return ntohl(*((uint32_t *)(index + (hashsz + 4) * n)));
1925         } else {
1926                 uint32_t off;
1927                 index += 8 + p->num_objects * (hashsz + 4);
1928                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1929                 if (!(off & 0x80000000))
1930                         return off;
1931                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1932                 check_pack_index_ptr(p, index);
1933                 return get_be64(index);
1934         }
1935 }
1936
1937 off_t find_pack_entry_one(const unsigned char *sha1,
1938                                   struct packed_git *p)
1939 {
1940         const unsigned char *index = p->index_data;
1941         struct object_id oid;
1942         uint32_t result;
1943
1944         if (!index) {
1945                 if (open_pack_index(p))
1946                         return 0;
1947         }
1948
1949         hashcpy(oid.hash, sha1);
1950         if (bsearch_pack(&oid, p, &result))
1951                 return nth_packed_object_offset(p, result);
1952         return 0;
1953 }
1954
1955 int is_pack_valid(struct packed_git *p)
1956 {
1957         /* An already open pack is known to be valid. */
1958         if (p->pack_fd != -1)
1959                 return 1;
1960
1961         /* If the pack has one window completely covering the
1962          * file size, the pack is known to be valid even if
1963          * the descriptor is not currently open.
1964          */
1965         if (p->windows) {
1966                 struct pack_window *w = p->windows;
1967
1968                 if (!w->offset && w->len == p->pack_size)
1969                         return 1;
1970         }
1971
1972         /* Force the pack to open to prove its valid. */
1973         return !open_packed_git(p);
1974 }
1975
1976 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1977                                   struct packed_git *packs)
1978 {
1979         struct packed_git *p;
1980
1981         for (p = packs; p; p = p->next) {
1982                 if (find_pack_entry_one(sha1, p))
1983                         return p;
1984         }
1985         return NULL;
1986
1987 }
1988
1989 static int fill_pack_entry(const struct object_id *oid,
1990                            struct pack_entry *e,
1991                            struct packed_git *p)
1992 {
1993         off_t offset;
1994
1995         if (p->num_bad_objects) {
1996                 unsigned i;
1997                 for (i = 0; i < p->num_bad_objects; i++)
1998                         if (hasheq(oid->hash,
1999                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
2000                                 return 0;
2001         }
2002
2003         offset = find_pack_entry_one(oid->hash, p);
2004         if (!offset)
2005                 return 0;
2006
2007         /*
2008          * We are about to tell the caller where they can locate the
2009          * requested object.  We better make sure the packfile is
2010          * still here and can be accessed before supplying that
2011          * answer, as it may have been deleted since the index was
2012          * loaded!
2013          */
2014         if (!is_pack_valid(p))
2015                 return 0;
2016         e->offset = offset;
2017         e->p = p;
2018         return 1;
2019 }
2020
2021 int find_pack_entry(struct repository *r, const struct object_id *oid, struct pack_entry *e)
2022 {
2023         struct list_head *pos;
2024         struct multi_pack_index *m;
2025
2026         prepare_packed_git(r);
2027         if (!r->objects->packed_git && !r->objects->multi_pack_index)
2028                 return 0;
2029
2030         for (m = r->objects->multi_pack_index; m; m = m->next) {
2031                 if (fill_midx_entry(r, oid, e, m))
2032                         return 1;
2033         }
2034
2035         list_for_each(pos, &r->objects->packed_git_mru) {
2036                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
2037                 if (!p->multi_pack_index && fill_pack_entry(oid, e, p)) {
2038                         list_move(&p->mru, &r->objects->packed_git_mru);
2039                         return 1;
2040                 }
2041         }
2042         return 0;
2043 }
2044
2045 int has_object_pack(const struct object_id *oid)
2046 {
2047         struct pack_entry e;
2048         return find_pack_entry(the_repository, oid, &e);
2049 }
2050
2051 int has_pack_index(const unsigned char *sha1)
2052 {
2053         struct stat st;
2054         if (stat(sha1_pack_index_name(sha1), &st))
2055                 return 0;
2056         return 1;
2057 }
2058
2059 int for_each_object_in_pack(struct packed_git *p,
2060                             each_packed_object_fn cb, void *data,
2061                             enum for_each_object_flags flags)
2062 {
2063         uint32_t i;
2064         int r = 0;
2065
2066         if (flags & FOR_EACH_OBJECT_PACK_ORDER) {
2067                 if (load_pack_revindex(p))
2068                         return -1;
2069         }
2070
2071         for (i = 0; i < p->num_objects; i++) {
2072                 uint32_t pos;
2073                 struct object_id oid;
2074
2075                 if (flags & FOR_EACH_OBJECT_PACK_ORDER)
2076                         pos = p->revindex[i].nr;
2077                 else
2078                         pos = i;
2079
2080                 if (nth_packed_object_id(&oid, p, pos) < 0)
2081                         return error("unable to get sha1 of object %u in %s",
2082                                      pos, p->pack_name);
2083
2084                 r = cb(&oid, p, pos, data);
2085                 if (r)
2086                         break;
2087         }
2088         return r;
2089 }
2090
2091 int for_each_packed_object(each_packed_object_fn cb, void *data,
2092                            enum for_each_object_flags flags)
2093 {
2094         struct packed_git *p;
2095         int r = 0;
2096         int pack_errors = 0;
2097
2098         prepare_packed_git(the_repository);
2099         for (p = get_all_packs(the_repository); p; p = p->next) {
2100                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
2101                         continue;
2102                 if ((flags & FOR_EACH_OBJECT_PROMISOR_ONLY) &&
2103                     !p->pack_promisor)
2104                         continue;
2105                 if (open_pack_index(p)) {
2106                         pack_errors = 1;
2107                         continue;
2108                 }
2109                 r = for_each_object_in_pack(p, cb, data, flags);
2110                 if (r)
2111                         break;
2112         }
2113         return r ? r : pack_errors;
2114 }
2115
2116 static int add_promisor_object(const struct object_id *oid,
2117                                struct packed_git *pack,
2118                                uint32_t pos,
2119                                void *set_)
2120 {
2121         struct oidset *set = set_;
2122         struct object *obj = parse_object(the_repository, oid);
2123         if (!obj)
2124                 return 1;
2125
2126         oidset_insert(set, oid);
2127
2128         /*
2129          * If this is a tree, commit, or tag, the objects it refers
2130          * to are also promisor objects. (Blobs refer to no objects->)
2131          */
2132         if (obj->type == OBJ_TREE) {
2133                 struct tree *tree = (struct tree *)obj;
2134                 struct tree_desc desc;
2135                 struct name_entry entry;
2136                 if (init_tree_desc_gently(&desc, tree->buffer, tree->size))
2137                         /*
2138                          * Error messages are given when packs are
2139                          * verified, so do not print any here.
2140                          */
2141                         return 0;
2142                 while (tree_entry_gently(&desc, &entry))
2143                         oidset_insert(set, &entry.oid);
2144         } else if (obj->type == OBJ_COMMIT) {
2145                 struct commit *commit = (struct commit *) obj;
2146                 struct commit_list *parents = commit->parents;
2147
2148                 oidset_insert(set, get_commit_tree_oid(commit));
2149                 for (; parents; parents = parents->next)
2150                         oidset_insert(set, &parents->item->object.oid);
2151         } else if (obj->type == OBJ_TAG) {
2152                 struct tag *tag = (struct tag *) obj;
2153                 oidset_insert(set, get_tagged_oid(tag));
2154         }
2155         return 0;
2156 }
2157
2158 int is_promisor_object(const struct object_id *oid)
2159 {
2160         static struct oidset promisor_objects;
2161         static int promisor_objects_prepared;
2162
2163         if (!promisor_objects_prepared) {
2164                 if (has_promisor_remote()) {
2165                         for_each_packed_object(add_promisor_object,
2166                                                &promisor_objects,
2167                                                FOR_EACH_OBJECT_PROMISOR_ONLY);
2168                 }
2169                 promisor_objects_prepared = 1;
2170         }
2171         return oidset_contains(&promisor_objects, oid);
2172 }