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