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 "hash-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_revindex(struct packed_git *p) {
 328         if (!p->revindex_map)
 329                 return;
 330
 331         munmap((void *)p->revindex_map, p->revindex_size);
 332         p->revindex_map = NULL;
 333         p->revindex_data = NULL;
 334 }
 335
 336 void close_pack(struct packed_git *p)
 337 {
 338         close_pack_windows(p);
 339         close_pack_fd(p);
 340         close_pack_index(p);
 341         close_pack_revindex(p);
 342 }
 343
 344 void close_object_store(struct raw_object_store *o)
 345 {
 346         struct packed_git *p;
 347
 348         for (p = o->packed_git; p; p = p->next)
 349                 if (p->do_not_close)
 350                         BUG("want to close pack marked 'do-not-close'");
 351                 else
 352                         close_pack(p);
 353
 354         if (o->multi_pack_index) {
 355                 close_midx(o->multi_pack_index);
 356                 o->multi_pack_index = NULL;
 357         }
 358
 359         close_commit_graph(o);
 360 }
 361
 362 void unlink_pack_path(const char *pack_name, int force_delete)
 363 {
 364         static const char *exts[] = {".pack", ".idx", ".rev", ".keep", ".bitmap", ".promisor"};
 365         int i;
 366         struct strbuf buf = STRBUF_INIT;
 367         size_t plen;
 368
 369         strbuf_addstr(&buf, pack_name);
 370         strip_suffix_mem(buf.buf, &buf.len, ".pack");
 371         plen = buf.len;
 372
 373         if (!force_delete) {
 374                 strbuf_addstr(&buf, ".keep");
 375                 if (!access(buf.buf, F_OK)) {
 376                         strbuf_release(&buf);
 377                         return;
 378                 }
 379         }
 380
 381         for (i = 0; i < ARRAY_SIZE(exts); i++) {
 382                 strbuf_setlen(&buf, plen);
 383                 strbuf_addstr(&buf, exts[i]);
 384                 unlink(buf.buf);
 385         }
 386
 387         strbuf_release(&buf);
 388 }
 389
 390 /*
 391  * The LRU pack is the one with the oldest MRU window, preferring packs
 392  * with no used windows, or the oldest mtime if it has no windows allocated.
 393  */
 394 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
 395 {
 396         struct pack_window *w, *this_mru_w;
 397         int has_windows_inuse = 0;
 398
 399         /*
 400          * Reject this pack if it has windows and the previously selected
 401          * one does not.  If this pack does not have windows, reject
 402          * it if the pack file is newer than the previously selected one.
 403          */
 404         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
 405                 return;
 406
 407         for (w = this_mru_w = p->windows; w; w = w->next) {
 408                 /*
 409                  * Reject this pack if any of its windows are in use,
 410                  * but the previously selected pack did not have any
 411                  * inuse windows.  Otherwise, record that this pack
 412                  * has windows in use.
 413                  */
 414                 if (w->inuse_cnt) {
 415                         if (*accept_windows_inuse)
 416                                 has_windows_inuse = 1;
 417                         else
 418                                 return;
 419                 }
 420
 421                 if (w->last_used > this_mru_w->last_used)
 422                         this_mru_w = w;
 423
 424                 /*
 425                  * Reject this pack if it has windows that have been
 426                  * used more recently than the previously selected pack.
 427                  * If the previously selected pack had windows inuse and
 428                  * we have not encountered a window in this pack that is
 429                  * inuse, skip this check since we prefer a pack with no
 430                  * inuse windows to one that has inuse windows.
 431                  */
 432                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
 433                     this_mru_w->last_used > (*mru_w)->last_used)
 434                         return;
 435         }
 436
 437         /*
 438          * Select this pack.
 439          */
 440         *mru_w = this_mru_w;
 441         *lru_p = p;
 442         *accept_windows_inuse = has_windows_inuse;
 443 }
 444
 445 static int close_one_pack(void)
 446 {
 447         struct packed_git *p, *lru_p = NULL;
 448         struct pack_window *mru_w = NULL;
 449         int accept_windows_inuse = 1;
 450
 451         for (p = the_repository->objects->packed_git; p; p = p->next) {
 452                 if (p->pack_fd == -1)
 453                         continue;
 454                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
 455         }
 456
 457         if (lru_p)
 458                 return close_pack_fd(lru_p);
 459
 460         return 0;
 461 }
 462
 463 static unsigned int get_max_fd_limit(void)
 464 {
 465 #ifdef RLIMIT_NOFILE
 466         {
 467                 struct rlimit lim;
 468
 469                 if (!getrlimit(RLIMIT_NOFILE, &lim))
 470                         return lim.rlim_cur;
 471         }
 472 #endif
 473
 474 #ifdef _SC_OPEN_MAX
 475         {
 476                 long open_max = sysconf(_SC_OPEN_MAX);
 477                 if (0 < open_max)
 478                         return open_max;
 479                 /*
 480                  * Otherwise, we got -1 for one of the two
 481                  * reasons:
 482                  *
 483                  * (1) sysconf() did not understand _SC_OPEN_MAX
 484                  *     and signaled an error with -1; or
 485                  * (2) sysconf() said there is no limit.
 486                  *
 487                  * We _could_ clear errno before calling sysconf() to
 488                  * tell these two cases apart and return a huge number
 489                  * in the latter case to let the caller cap it to a
 490                  * value that is not so selfish, but letting the
 491                  * fallback OPEN_MAX codepath take care of these cases
 492                  * is a lot simpler.
 493                  */
 494         }
 495 #endif
 496
 497 #ifdef OPEN_MAX
 498         return OPEN_MAX;
 499 #else
 500         return 1; /* see the caller ;-) */
 501 #endif
 502 }
 503
 504 const char *pack_basename(struct packed_git *p)
 505 {
 506         const char *ret = strrchr(p->pack_name, '/');
 507         if (ret)
 508                 ret = ret + 1; /* skip past slash */
 509         else
 510                 ret = p->pack_name; /* we only have a base */
 511         return ret;
 512 }
 513
 514 /*
 515  * Do not call this directly as this leaks p->pack_fd on error return;
 516  * call open_packed_git() instead.
 517  */
 518 static int open_packed_git_1(struct packed_git *p)
 519 {
 520         struct stat st;
 521         struct pack_header hdr;
 522         unsigned char hash[GIT_MAX_RAWSZ];
 523         unsigned char *idx_hash;
 524         ssize_t read_result;
 525         const unsigned hashsz = the_hash_algo->rawsz;
 526
 527         if (open_pack_index(p))
 528                 return error("packfile %s index unavailable", p->pack_name);
 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         /* Verify the pack matches its index. */
 570         if (p->num_objects != ntohl(hdr.hdr_entries))
 571                 return error("packfile %s claims to have %"PRIu32" objects"
 572                              " while index indicates %"PRIu32" objects",
 573                              p->pack_name, ntohl(hdr.hdr_entries),
 574                              p->num_objects);
 575         read_result = pread_in_full(p->pack_fd, hash, hashsz,
 576                                         p->pack_size - hashsz);
 577         if (read_result < 0)
 578                 return error_errno("error reading from %s", p->pack_name);
 579         if (read_result != hashsz)
 580                 return error("packfile %s signature is unavailable", p->pack_name);
 581         idx_hash = ((unsigned char *)p->index_data) + p->index_size - hashsz * 2;
 582         if (!hasheq(hash, idx_hash))
 583                 return error("packfile %s does not match index", p->pack_name);
 584         return 0;
 585 }
 586
 587 static int open_packed_git(struct packed_git *p)
 588 {
 589         if (!open_packed_git_1(p))
 590                 return 0;
 591         close_pack_fd(p);
 592         return -1;
 593 }
 594
 595 static int in_window(struct pack_window *win, off_t offset)
 596 {
 597         /* We must promise at least one full hash after the
 598          * offset is available from this window, otherwise the offset
 599          * is not actually in this window and a different window (which
 600          * has that one hash excess) must be used.  This is to support
 601          * the object header and delta base parsing routines below.
 602          */
 603         off_t win_off = win->offset;
 604         return win_off <= offset
 605                 && (offset + the_hash_algo->rawsz) <= (win_off + win->len);
 606 }
 607
 608 unsigned char *use_pack(struct packed_git *p,
 609                 struct pack_window **w_cursor,
 610                 off_t offset,
 611                 unsigned long *left)
 612 {
 613         struct pack_window *win = *w_cursor;
 614
 615         /* Since packfiles end in a hash of their content and it's
 616          * pointless to ask for an offset into the middle of that
 617          * hash, and the in_window function above wouldn't match
 618          * don't allow an offset too close to the end of the file.
 619          */
 620         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
 621                 die("packfile %s cannot be accessed", p->pack_name);
 622         if (offset > (p->pack_size - the_hash_algo->rawsz))
 623                 die("offset beyond end of packfile (truncated pack?)");
 624         if (offset < 0)
 625                 die(_("offset before end of packfile (broken .idx?)"));
 626
 627         if (!win || !in_window(win, offset)) {
 628                 if (win)
 629                         win->inuse_cnt--;
 630                 for (win = p->windows; win; win = win->next) {
 631                         if (in_window(win, offset))
 632                                 break;
 633                 }
 634                 if (!win) {
 635                         size_t window_align = packed_git_window_size / 2;
 636                         off_t len;
 637
 638                         if (p->pack_fd == -1 && open_packed_git(p))
 639                                 die("packfile %s cannot be accessed", p->pack_name);
 640
 641                         CALLOC_ARRAY(win, 1);
 642                         win->offset = (offset / window_align) * window_align;
 643                         len = p->pack_size - win->offset;
 644                         if (len > packed_git_window_size)
 645                                 len = packed_git_window_size;
 646                         win->len = (size_t)len;
 647                         pack_mapped += win->len;
 648                         while (packed_git_limit < pack_mapped
 649                                 && unuse_one_window(p))
 650                                 ; /* nothing */
 651                         win->base = xmmap_gently(NULL, win->len,
 652                                 PROT_READ, MAP_PRIVATE,
 653                                 p->pack_fd, win->offset);
 654                         if (win->base == MAP_FAILED)
 655                                 die_errno("packfile %s cannot be mapped",
 656                                           p->pack_name);
 657                         if (!win->offset && win->len == p->pack_size
 658                                 && !p->do_not_close)
 659                                 close_pack_fd(p);
 660                         pack_mmap_calls++;
 661                         pack_open_windows++;
 662                         if (pack_mapped > peak_pack_mapped)
 663                                 peak_pack_mapped = pack_mapped;
 664                         if (pack_open_windows > peak_pack_open_windows)
 665                                 peak_pack_open_windows = pack_open_windows;
 666                         win->next = p->windows;
 667                         p->windows = win;
 668                 }
 669         }
 670         if (win != *w_cursor) {
 671                 win->last_used = pack_used_ctr++;
 672                 win->inuse_cnt++;
 673                 *w_cursor = win;
 674         }
 675         offset -= win->offset;
 676         if (left)
 677                 *left = win->len - xsize_t(offset);
 678         return win->base + offset;
 679 }
 680
 681 void unuse_pack(struct pack_window **w_cursor)
 682 {
 683         struct pack_window *w = *w_cursor;
 684         if (w) {
 685                 w->inuse_cnt--;
 686                 *w_cursor = NULL;
 687         }
 688 }
 689
 690 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
 691 {
 692         struct stat st;
 693         size_t alloc;
 694         struct packed_git *p;
 695
 696         /*
 697          * Make sure a corresponding .pack file exists and that
 698          * the index looks sane.
 699          */
 700         if (!strip_suffix_mem(path, &path_len, ".idx"))
 701                 return NULL;
 702
 703         /*
 704          * ".promisor" is long enough to hold any suffix we're adding (and
 705          * the use xsnprintf double-checks that)
 706          */
 707         alloc = st_add3(path_len, strlen(".promisor"), 1);
 708         p = alloc_packed_git(alloc);
 709         memcpy(p->pack_name, path, path_len);
 710
 711         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
 712         if (!access(p->pack_name, F_OK))
 713                 p->pack_keep = 1;
 714
 715         xsnprintf(p->pack_name + path_len, alloc - path_len, ".promisor");
 716         if (!access(p->pack_name, F_OK))
 717                 p->pack_promisor = 1;
 718
 719         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
 720         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
 721                 free(p);
 722                 return NULL;
 723         }
 724
 725         /* ok, it looks sane as far as we can check without
 726          * actually mapping the pack file.
 727          */
 728         p->pack_size = st.st_size;
 729         p->pack_local = local;
 730         p->mtime = st.st_mtime;
 731         if (path_len < the_hash_algo->hexsz ||
 732             get_sha1_hex(path + path_len - the_hash_algo->hexsz, p->hash))
 733                 hashclr(p->hash);
 734         return p;
 735 }
 736
 737 void install_packed_git(struct repository *r, struct packed_git *pack)
 738 {
 739         if (pack->pack_fd != -1)
 740                 pack_open_fds++;
 741
 742         pack->next = r->objects->packed_git;
 743         r->objects->packed_git = pack;
 744
 745         hashmap_entry_init(&pack->packmap_ent, strhash(pack->pack_name));
 746         hashmap_add(&r->objects->pack_map, &pack->packmap_ent);
 747 }
 748
 749 void (*report_garbage)(unsigned seen_bits, const char *path);
 750
 751 static void report_helper(const struct string_list *list,
 752                           int seen_bits, int first, int last)
 753 {
 754         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
 755                 return;
 756
 757         for (; first < last; first++)
 758                 report_garbage(seen_bits, list->items[first].string);
 759 }
 760
 761 static void report_pack_garbage(struct string_list *list)
 762 {
 763         int i, baselen = -1, first = 0, seen_bits = 0;
 764
 765         if (!report_garbage)
 766                 return;
 767
 768         string_list_sort(list);
 769
 770         for (i = 0; i < list->nr; i++) {
 771                 const char *path = list->items[i].string;
 772                 if (baselen != -1 &&
 773                     strncmp(path, list->items[first].string, baselen)) {
 774                         report_helper(list, seen_bits, first, i);
 775                         baselen = -1;
 776                         seen_bits = 0;
 777                 }
 778                 if (baselen == -1) {
 779                         const char *dot = strrchr(path, '.');
 780                         if (!dot) {
 781                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
 782                                 continue;
 783                         }
 784                         baselen = dot - path + 1;
 785                         first = i;
 786                 }
 787                 if (!strcmp(path + baselen, "pack"))
 788                         seen_bits |= 1;
 789                 else if (!strcmp(path + baselen, "idx"))
 790                         seen_bits |= 2;
 791         }
 792         report_helper(list, seen_bits, first, list->nr);
 793 }
 794
 795 void for_each_file_in_pack_dir(const char *objdir,
 796                                each_file_in_pack_dir_fn fn,
 797                                void *data)
 798 {
 799         struct strbuf path = STRBUF_INIT;
 800         size_t dirnamelen;
 801         DIR *dir;
 802         struct dirent *de;
 803
 804         strbuf_addstr(&path, objdir);
 805         strbuf_addstr(&path, "/pack");
 806         dir = opendir(path.buf);
 807         if (!dir) {
 808                 if (errno != ENOENT)
 809                         error_errno("unable to open object pack directory: %s",
 810                                     path.buf);
 811                 strbuf_release(&path);
 812                 return;
 813         }
 814         strbuf_addch(&path, '/');
 815         dirnamelen = path.len;
 816         while ((de = readdir_skip_dot_and_dotdot(dir)) != NULL) {
 817                 strbuf_setlen(&path, dirnamelen);
 818                 strbuf_addstr(&path, de->d_name);
 819
 820                 fn(path.buf, path.len, de->d_name, data);
 821         }
 822
 823         closedir(dir);
 824         strbuf_release(&path);
 825 }
 826
 827 struct prepare_pack_data {
 828         struct repository *r;
 829         struct string_list *garbage;
 830         int local;
 831         struct multi_pack_index *m;
 832 };
 833
 834 static void prepare_pack(const char *full_name, size_t full_name_len,
 835                          const char *file_name, void *_data)
 836 {
 837         struct prepare_pack_data *data = (struct prepare_pack_data *)_data;
 838         struct packed_git *p;
 839         size_t base_len = full_name_len;
 840
 841         if (strip_suffix_mem(full_name, &base_len, ".idx") &&
 842             !(data->m && midx_contains_pack(data->m, file_name))) {
 843                 struct hashmap_entry hent;
 844                 char *pack_name = xstrfmt("%.*s.pack", (int)base_len, full_name);
 845                 unsigned int hash = strhash(pack_name);
 846                 hashmap_entry_init(&hent, hash);
 847
 848                 /* Don't reopen a pack we already have. */
 849                 if (!hashmap_get(&data->r->objects->pack_map, &hent, pack_name)) {
 850                         p = add_packed_git(full_name, full_name_len, data->local);
 851                         if (p)
 852                                 install_packed_git(data->r, p);
 853                 }
 854                 free(pack_name);
 855         }
 856
 857         if (!report_garbage)
 858                 return;
 859
 860         if (!strcmp(file_name, "multi-pack-index"))
 861                 return;
 862         if (ends_with(file_name, ".idx") ||
 863             ends_with(file_name, ".rev") ||
 864             ends_with(file_name, ".pack") ||
 865             ends_with(file_name, ".bitmap") ||
 866             ends_with(file_name, ".keep") ||
 867             ends_with(file_name, ".promisor"))
 868                 string_list_append(data->garbage, full_name);
 869         else
 870                 report_garbage(PACKDIR_FILE_GARBAGE, full_name);
 871 }
 872
 873 static void prepare_packed_git_one(struct repository *r, char *objdir, int local)
 874 {
 875         struct prepare_pack_data data;
 876         struct string_list garbage = STRING_LIST_INIT_DUP;
 877
 878         data.m = r->objects->multi_pack_index;
 879
 880         /* look for the multi-pack-index for this object directory */
 881         while (data.m && strcmp(data.m->object_dir, objdir))
 882                 data.m = data.m->next;
 883
 884         data.r = r;
 885         data.garbage = &garbage;
 886         data.local = local;
 887
 888         for_each_file_in_pack_dir(objdir, prepare_pack, &data);
 889
 890         report_pack_garbage(data.garbage);
 891         string_list_clear(data.garbage, 0);
 892 }
 893
 894 static void prepare_packed_git(struct repository *r);
 895 /*
 896  * Give a fast, rough count of the number of objects in the repository. This
 897  * ignores loose objects completely. If you have a lot of them, then either
 898  * you should repack because your performance will be awful, or they are
 899  * all unreachable objects about to be pruned, in which case they're not really
 900  * interesting as a measure of repo size in the first place.
 901  */
 902 unsigned long repo_approximate_object_count(struct repository *r)
 903 {
 904         if (!r->objects->approximate_object_count_valid) {
 905                 unsigned long count;
 906                 struct multi_pack_index *m;
 907                 struct packed_git *p;
 908
 909                 prepare_packed_git(r);
 910                 count = 0;
 911                 for (m = get_multi_pack_index(r); m; m = m->next)
 912                         count += m->num_objects;
 913                 for (p = r->objects->packed_git; p; p = p->next) {
 914                         if (open_pack_index(p))
 915                                 continue;
 916                         count += p->num_objects;
 917                 }
 918                 r->objects->approximate_object_count = count;
 919                 r->objects->approximate_object_count_valid = 1;
 920         }
 921         return r->objects->approximate_object_count;
 922 }
 923
 924 static void *get_next_packed_git(const void *p)
 925 {
 926         return ((const struct packed_git *)p)->next;
 927 }
 928
 929 static void set_next_packed_git(void *p, void *next)
 930 {
 931         ((struct packed_git *)p)->next = next;
 932 }
 933
 934 static int sort_pack(const void *a_, const void *b_)
 935 {
 936         const struct packed_git *a = a_;
 937         const struct packed_git *b = b_;
 938         int st;
 939
 940         /*
 941          * Local packs tend to contain objects specific to our
 942          * variant of the project than remote ones.  In addition,
 943          * remote ones could be on a network mounted filesystem.
 944          * Favor local ones for these reasons.
 945          */
 946         st = a->pack_local - b->pack_local;
 947         if (st)
 948                 return -st;
 949
 950         /*
 951          * Younger packs tend to contain more recent objects,
 952          * and more recent objects tend to get accessed more
 953          * often.
 954          */
 955         if (a->mtime < b->mtime)
 956                 return 1;
 957         else if (a->mtime == b->mtime)
 958                 return 0;
 959         return -1;
 960 }
 961
 962 static void rearrange_packed_git(struct repository *r)
 963 {
 964         r->objects->packed_git = llist_mergesort(
 965                 r->objects->packed_git, get_next_packed_git,
 966                 set_next_packed_git, sort_pack);
 967 }
 968
 969 static void prepare_packed_git_mru(struct repository *r)
 970 {
 971         struct packed_git *p;
 972
 973         INIT_LIST_HEAD(&r->objects->packed_git_mru);
 974
 975         for (p = r->objects->packed_git; p; p = p->next)
 976                 list_add_tail(&p->mru, &r->objects->packed_git_mru);
 977 }
 978
 979 static void prepare_packed_git(struct repository *r)
 980 {
 981         struct object_directory *odb;
 982
 983         if (r->objects->packed_git_initialized)
 984                 return;
 985
 986         prepare_alt_odb(r);
 987         for (odb = r->objects->odb; odb; odb = odb->next) {
 988                 int local = (odb == r->objects->odb);
 989                 prepare_multi_pack_index_one(r, odb->path, local);
 990                 prepare_packed_git_one(r, odb->path, local);
 991         }
 992         rearrange_packed_git(r);
 993
 994         prepare_packed_git_mru(r);
 995         r->objects->packed_git_initialized = 1;
 996 }
 997
 998 void reprepare_packed_git(struct repository *r)
 999 {
1000         struct object_directory *odb;
1001
1002         obj_read_lock();
1003         for (odb = r->objects->odb; odb; odb = odb->next)
1004                 odb_clear_loose_cache(odb);
1005
1006         r->objects->approximate_object_count_valid = 0;
1007         r->objects->packed_git_initialized = 0;
1008         prepare_packed_git(r);
1009         obj_read_unlock();
1010 }
1011
1012 struct packed_git *get_packed_git(struct repository *r)
1013 {
1014         prepare_packed_git(r);
1015         return r->objects->packed_git;
1016 }
1017
1018 struct multi_pack_index *get_multi_pack_index(struct repository *r)
1019 {
1020         prepare_packed_git(r);
1021         return r->objects->multi_pack_index;
1022 }
1023
1024 struct multi_pack_index *get_local_multi_pack_index(struct repository *r)
1025 {
1026         struct multi_pack_index *m = get_multi_pack_index(r);
1027
1028         /* no need to iterate; we always put the local one first (if any) */
1029         if (m && m->local)
1030                 return m;
1031
1032         return NULL;
1033 }
1034
1035 struct packed_git *get_all_packs(struct repository *r)
1036 {
1037         struct multi_pack_index *m;
1038
1039         prepare_packed_git(r);
1040         for (m = r->objects->multi_pack_index; m; m = m->next) {
1041                 uint32_t i;
1042                 for (i = 0; i < m->num_packs; i++)
1043                         prepare_midx_pack(r, m, i);
1044         }
1045
1046         return r->objects->packed_git;
1047 }
1048
1049 struct list_head *get_packed_git_mru(struct repository *r)
1050 {
1051         prepare_packed_git(r);
1052         return &r->objects->packed_git_mru;
1053 }
1054
1055 unsigned long unpack_object_header_buffer(const unsigned char *buf,
1056                 unsigned long len, enum object_type *type, unsigned long *sizep)
1057 {
1058         unsigned shift;
1059         unsigned long size, c;
1060         unsigned long used = 0;
1061
1062         c = buf[used++];
1063         *type = (c >> 4) & 7;
1064         size = c & 15;
1065         shift = 4;
1066         while (c & 0x80) {
1067                 if (len <= used || bitsizeof(long) <= shift) {
1068                         error("bad object header");
1069                         size = used = 0;
1070                         break;
1071                 }
1072                 c = buf[used++];
1073                 size += (c & 0x7f) << shift;
1074                 shift += 7;
1075         }
1076         *sizep = size;
1077         return used;
1078 }
1079
1080 unsigned long get_size_from_delta(struct packed_git *p,
1081                                   struct pack_window **w_curs,
1082                                   off_t curpos)
1083 {
1084         const unsigned char *data;
1085         unsigned char delta_head[20], *in;
1086         git_zstream stream;
1087         int st;
1088
1089         memset(&stream, 0, sizeof(stream));
1090         stream.next_out = delta_head;
1091         stream.avail_out = sizeof(delta_head);
1092
1093         git_inflate_init(&stream);
1094         do {
1095                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1096                 stream.next_in = in;
1097                 /*
1098                  * Note: the window section returned by use_pack() must be
1099                  * available throughout git_inflate()'s unlocked execution. To
1100                  * ensure no other thread will modify the window in the
1101                  * meantime, we rely on the packed_window.inuse_cnt. This
1102                  * counter is incremented before window reading and checked
1103                  * before window disposal.
1104                  *
1105                  * Other worrying sections could be the call to close_pack_fd(),
1106                  * which can close packs even with in-use windows, and to
1107                  * reprepare_packed_git(). Regarding the former, mmap doc says:
1108                  * "closing the file descriptor does not unmap the region". And
1109                  * for the latter, it won't re-open already available packs.
1110                  */
1111                 obj_read_unlock();
1112                 st = git_inflate(&stream, Z_FINISH);
1113                 obj_read_lock();
1114                 curpos += stream.next_in - in;
1115         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1116                  stream.total_out < sizeof(delta_head));
1117         git_inflate_end(&stream);
1118         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1119                 error("delta data unpack-initial failed");
1120                 return 0;
1121         }
1122
1123         /* Examine the initial part of the delta to figure out
1124          * the result size.
1125          */
1126         data = delta_head;
1127
1128         /* ignore base size */
1129         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1130
1131         /* Read the result size */
1132         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1133 }
1134
1135 int unpack_object_header(struct packed_git *p,
1136                          struct pack_window **w_curs,
1137                          off_t *curpos,
1138                          unsigned long *sizep)
1139 {
1140         unsigned char *base;
1141         unsigned long left;
1142         unsigned long used;
1143         enum object_type type;
1144
1145         /* use_pack() assures us we have [base, base + 20) available
1146          * as a range that we can look at.  (Its actually the hash
1147          * size that is assured.)  With our object header encoding
1148          * the maximum deflated object size is 2^137, which is just
1149          * insane, so we know won't exceed what we have been given.
1150          */
1151         base = use_pack(p, w_curs, *curpos, &left);
1152         used = unpack_object_header_buffer(base, left, &type, sizep);
1153         if (!used) {
1154                 type = OBJ_BAD;
1155         } else
1156                 *curpos += used;
1157
1158         return type;
1159 }
1160
1161 void mark_bad_packed_object(struct packed_git *p, const unsigned char *sha1)
1162 {
1163         unsigned i;
1164         const unsigned hashsz = the_hash_algo->rawsz;
1165         for (i = 0; i < p->num_bad_objects; i++)
1166                 if (hasheq(sha1, p->bad_object_sha1 + hashsz * i))
1167                         return;
1168         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1169                                       st_mult(GIT_MAX_RAWSZ,
1170                                               st_add(p->num_bad_objects, 1)));
1171         hashcpy(p->bad_object_sha1 + hashsz * p->num_bad_objects, sha1);
1172         p->num_bad_objects++;
1173 }
1174
1175 const struct packed_git *has_packed_and_bad(struct repository *r,
1176                                             const unsigned char *sha1)
1177 {
1178         struct packed_git *p;
1179         unsigned i;
1180
1181         for (p = r->objects->packed_git; p; p = p->next)
1182                 for (i = 0; i < p->num_bad_objects; i++)
1183                         if (hasheq(sha1,
1184                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
1185                                 return p;
1186         return NULL;
1187 }
1188
1189 off_t get_delta_base(struct packed_git *p,
1190                      struct pack_window **w_curs,
1191                      off_t *curpos,
1192                      enum object_type type,
1193                      off_t delta_obj_offset)
1194 {
1195         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1196         off_t base_offset;
1197
1198         /* use_pack() assured us we have [base_info, base_info + 20)
1199          * as a range that we can look at without walking off the
1200          * end of the mapped window.  Its actually the hash size
1201          * that is assured.  An OFS_DELTA longer than the hash size
1202          * is stupid, as then a REF_DELTA would be smaller to store.
1203          */
1204         if (type == OBJ_OFS_DELTA) {
1205                 unsigned used = 0;
1206                 unsigned char c = base_info[used++];
1207                 base_offset = c & 127;
1208                 while (c & 128) {
1209                         base_offset += 1;
1210                         if (!base_offset || MSB(base_offset, 7))
1211                                 return 0;  /* overflow */
1212                         c = base_info[used++];
1213                         base_offset = (base_offset << 7) + (c & 127);
1214                 }
1215                 base_offset = delta_obj_offset - base_offset;
1216                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1217                         return 0;  /* out of bound */
1218                 *curpos += used;
1219         } else if (type == OBJ_REF_DELTA) {
1220                 /* The base entry _must_ be in the same pack */
1221                 base_offset = find_pack_entry_one(base_info, p);
1222                 *curpos += the_hash_algo->rawsz;
1223         } else
1224                 die("I am totally screwed");
1225         return base_offset;
1226 }
1227
1228 /*
1229  * Like get_delta_base above, but we return the sha1 instead of the pack
1230  * offset. This means it is cheaper for REF deltas (we do not have to do
1231  * the final object lookup), but more expensive for OFS deltas (we
1232  * have to load the revidx to convert the offset back into a sha1).
1233  */
1234 static int get_delta_base_oid(struct packed_git *p,
1235                               struct pack_window **w_curs,
1236                               off_t curpos,
1237                               struct object_id *oid,
1238                               enum object_type type,
1239                               off_t delta_obj_offset)
1240 {
1241         if (type == OBJ_REF_DELTA) {
1242                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1243                 oidread(oid, base);
1244                 return 0;
1245         } else if (type == OBJ_OFS_DELTA) {
1246                 uint32_t base_pos;
1247                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1248                                                    type, delta_obj_offset);
1249
1250                 if (!base_offset)
1251                         return -1;
1252
1253                 if (offset_to_pack_pos(p, base_offset, &base_pos) < 0)
1254                         return -1;
1255
1256                 return nth_packed_object_id(oid, p,
1257                                             pack_pos_to_index(p, base_pos));
1258         } else
1259                 return -1;
1260 }
1261
1262 static int retry_bad_packed_offset(struct repository *r,
1263                                    struct packed_git *p,
1264                                    off_t obj_offset)
1265 {
1266         int type;
1267         uint32_t pos;
1268         struct object_id oid;
1269         if (offset_to_pack_pos(p, obj_offset, &pos) < 0)
1270                 return OBJ_BAD;
1271         nth_packed_object_id(&oid, p, pack_pos_to_index(p, pos));
1272         mark_bad_packed_object(p, oid.hash);
1273         type = oid_object_info(r, &oid, NULL);
1274         if (type <= OBJ_NONE)
1275                 return OBJ_BAD;
1276         return type;
1277 }
1278
1279 #define POI_STACK_PREALLOC 64
1280
1281 static enum object_type packed_to_object_type(struct repository *r,
1282                                               struct packed_git *p,
1283                                               off_t obj_offset,
1284                                               enum object_type type,
1285                                               struct pack_window **w_curs,
1286                                               off_t curpos)
1287 {
1288         off_t small_poi_stack[POI_STACK_PREALLOC];
1289         off_t *poi_stack = small_poi_stack;
1290         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1291
1292         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1293                 off_t base_offset;
1294                 unsigned long size;
1295                 /* Push the object we're going to leave behind */
1296                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1297                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1298                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
1299                         COPY_ARRAY(poi_stack, small_poi_stack, poi_stack_nr);
1300                 } else {
1301                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1302                 }
1303                 poi_stack[poi_stack_nr++] = obj_offset;
1304                 /* If parsing the base offset fails, just unwind */
1305                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1306                 if (!base_offset)
1307                         goto unwind;
1308                 curpos = obj_offset = base_offset;
1309                 type = unpack_object_header(p, w_curs, &curpos, &size);
1310                 if (type <= OBJ_NONE) {
1311                         /* If getting the base itself fails, we first
1312                          * retry the base, otherwise unwind */
1313                         type = retry_bad_packed_offset(r, p, base_offset);
1314                         if (type > OBJ_NONE)
1315                                 goto out;
1316                         goto unwind;
1317                 }
1318         }
1319
1320         switch (type) {
1321         case OBJ_BAD:
1322         case OBJ_COMMIT:
1323         case OBJ_TREE:
1324         case OBJ_BLOB:
1325         case OBJ_TAG:
1326                 break;
1327         default:
1328                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1329                       type, (uintmax_t)obj_offset, p->pack_name);
1330                 type = OBJ_BAD;
1331         }
1332
1333 out:
1334         if (poi_stack != small_poi_stack)
1335                 free(poi_stack);
1336         return type;
1337
1338 unwind:
1339         while (poi_stack_nr) {
1340                 obj_offset = poi_stack[--poi_stack_nr];
1341                 type = retry_bad_packed_offset(r, p, obj_offset);
1342                 if (type > OBJ_NONE)
1343                         goto out;
1344         }
1345         type = OBJ_BAD;
1346         goto out;
1347 }
1348
1349 static struct hashmap delta_base_cache;
1350 static size_t delta_base_cached;
1351
1352 static LIST_HEAD(delta_base_cache_lru);
1353
1354 struct delta_base_cache_key {
1355         struct packed_git *p;
1356         off_t base_offset;
1357 };
1358
1359 struct delta_base_cache_entry {
1360         struct hashmap_entry ent;
1361         struct delta_base_cache_key key;
1362         struct list_head lru;
1363         void *data;
1364         unsigned long size;
1365         enum object_type type;
1366 };
1367
1368 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
1369 {
1370         unsigned int hash;
1371
1372         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
1373         hash += (hash >> 8) + (hash >> 16);
1374         return hash;
1375 }
1376
1377 static struct delta_base_cache_entry *
1378 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
1379 {
1380         struct hashmap_entry entry, *e;
1381         struct delta_base_cache_key key;
1382
1383         if (!delta_base_cache.cmpfn)
1384                 return NULL;
1385
1386         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
1387         key.p = p;
1388         key.base_offset = base_offset;
1389         e = hashmap_get(&delta_base_cache, &entry, &key);
1390         return e ? container_of(e, struct delta_base_cache_entry, ent) : NULL;
1391 }
1392
1393 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
1394                                    const struct delta_base_cache_key *b)
1395 {
1396         return a->p == b->p && a->base_offset == b->base_offset;
1397 }
1398
1399 static int delta_base_cache_hash_cmp(const void *unused_cmp_data,
1400                                      const struct hashmap_entry *va,
1401                                      const struct hashmap_entry *vb,
1402                                      const void *vkey)
1403 {
1404         const struct delta_base_cache_entry *a, *b;
1405         const struct delta_base_cache_key *key = vkey;
1406
1407         a = container_of(va, const struct delta_base_cache_entry, ent);
1408         b = container_of(vb, const struct delta_base_cache_entry, ent);
1409
1410         if (key)
1411                 return !delta_base_cache_key_eq(&a->key, key);
1412         else
1413                 return !delta_base_cache_key_eq(&a->key, &b->key);
1414 }
1415
1416 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
1417 {
1418         return !!get_delta_base_cache_entry(p, base_offset);
1419 }
1420
1421 /*
1422  * Remove the entry from the cache, but do _not_ free the associated
1423  * entry data. The caller takes ownership of the "data" buffer, and
1424  * should copy out any fields it wants before detaching.
1425  */
1426 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
1427 {
1428         hashmap_remove(&delta_base_cache, &ent->ent, &ent->key);
1429         list_del(&ent->lru);
1430         delta_base_cached -= ent->size;
1431         free(ent);
1432 }
1433
1434 static void *cache_or_unpack_entry(struct repository *r, struct packed_git *p,
1435                                    off_t base_offset, unsigned long *base_size,
1436                                    enum object_type *type)
1437 {
1438         struct delta_base_cache_entry *ent;
1439
1440         ent = get_delta_base_cache_entry(p, base_offset);
1441         if (!ent)
1442                 return unpack_entry(r, p, base_offset, type, base_size);
1443
1444         if (type)
1445                 *type = ent->type;
1446         if (base_size)
1447                 *base_size = ent->size;
1448         return xmemdupz(ent->data, ent->size);
1449 }
1450
1451 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1452 {
1453         free(ent->data);
1454         detach_delta_base_cache_entry(ent);
1455 }
1456
1457 void clear_delta_base_cache(void)
1458 {
1459         struct list_head *lru, *tmp;
1460         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1461                 struct delta_base_cache_entry *entry =
1462                         list_entry(lru, struct delta_base_cache_entry, lru);
1463                 release_delta_base_cache(entry);
1464         }
1465 }
1466
1467 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1468         void *base, unsigned long base_size, enum object_type type)
1469 {
1470         struct delta_base_cache_entry *ent;
1471         struct list_head *lru, *tmp;
1472
1473         /*
1474          * Check required to avoid redundant entries when more than one thread
1475          * is unpacking the same object, in unpack_entry() (since its phases I
1476          * and III might run concurrently across multiple threads).
1477          */
1478         if (in_delta_base_cache(p, base_offset)) {
1479                 free(base);
1480                 return;
1481         }
1482
1483         delta_base_cached += base_size;
1484
1485         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1486                 struct delta_base_cache_entry *f =
1487                         list_entry(lru, struct delta_base_cache_entry, lru);
1488                 if (delta_base_cached <= delta_base_cache_limit)
1489                         break;
1490                 release_delta_base_cache(f);
1491         }
1492
1493         ent = xmalloc(sizeof(*ent));
1494         ent->key.p = p;
1495         ent->key.base_offset = base_offset;
1496         ent->type = type;
1497         ent->data = base;
1498         ent->size = base_size;
1499         list_add_tail(&ent->lru, &delta_base_cache_lru);
1500
1501         if (!delta_base_cache.cmpfn)
1502                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, NULL, 0);
1503         hashmap_entry_init(&ent->ent, pack_entry_hash(p, base_offset));
1504         hashmap_add(&delta_base_cache, &ent->ent);
1505 }
1506
1507 int packed_object_info(struct repository *r, struct packed_git *p,
1508                        off_t obj_offset, struct object_info *oi)
1509 {
1510         struct pack_window *w_curs = NULL;
1511         unsigned long size;
1512         off_t curpos = obj_offset;
1513         enum object_type type;
1514
1515         /*
1516          * We always get the representation type, but only convert it to
1517          * a "real" type later if the caller is interested.
1518          */
1519         if (oi->contentp) {
1520                 *oi->contentp = cache_or_unpack_entry(r, p, obj_offset, oi->sizep,
1521                                                       &type);
1522                 if (!*oi->contentp)
1523                         type = OBJ_BAD;
1524         } else {
1525                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1526         }
1527
1528         if (!oi->contentp && oi->sizep) {
1529                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1530                         off_t tmp_pos = curpos;
1531                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1532                                                            type, obj_offset);
1533                         if (!base_offset) {
1534                                 type = OBJ_BAD;
1535                                 goto out;
1536                         }
1537                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1538                         if (*oi->sizep == 0) {
1539                                 type = OBJ_BAD;
1540                                 goto out;
1541                         }
1542                 } else {
1543                         *oi->sizep = size;
1544                 }
1545         }
1546
1547         if (oi->disk_sizep) {
1548                 uint32_t pos;
1549                 if (offset_to_pack_pos(p, obj_offset, &pos) < 0) {
1550                         error("could not find object at offset %"PRIuMAX" "
1551                               "in pack %s", (uintmax_t)obj_offset, p->pack_name);
1552                         type = OBJ_BAD;
1553                         goto out;
1554                 }
1555
1556                 *oi->disk_sizep = pack_pos_to_offset(p, pos + 1) - obj_offset;
1557         }
1558
1559         if (oi->typep || oi->type_name) {
1560                 enum object_type ptot;
1561                 ptot = packed_to_object_type(r, p, obj_offset,
1562                                              type, &w_curs, curpos);
1563                 if (oi->typep)
1564                         *oi->typep = ptot;
1565                 if (oi->type_name) {
1566                         const char *tn = type_name(ptot);
1567                         if (tn)
1568                                 strbuf_addstr(oi->type_name, tn);
1569                 }
1570                 if (ptot < 0) {
1571                         type = OBJ_BAD;
1572                         goto out;
1573                 }
1574         }
1575
1576         if (oi->delta_base_oid) {
1577                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1578                         if (get_delta_base_oid(p, &w_curs, curpos,
1579                                                oi->delta_base_oid,
1580                                                type, obj_offset) < 0) {
1581                                 type = OBJ_BAD;
1582                                 goto out;
1583                         }
1584                 } else
1585                         oidclr(oi->delta_base_oid);
1586         }
1587
1588         oi->whence = in_delta_base_cache(p, obj_offset) ? OI_DBCACHED :
1589                                                           OI_PACKED;
1590
1591 out:
1592         unuse_pack(&w_curs);
1593         return type;
1594 }
1595
1596 static void *unpack_compressed_entry(struct packed_git *p,
1597                                     struct pack_window **w_curs,
1598                                     off_t curpos,
1599                                     unsigned long size)
1600 {
1601         int st;
1602         git_zstream stream;
1603         unsigned char *buffer, *in;
1604
1605         buffer = xmallocz_gently(size);
1606         if (!buffer)
1607                 return NULL;
1608         memset(&stream, 0, sizeof(stream));
1609         stream.next_out = buffer;
1610         stream.avail_out = size + 1;
1611
1612         git_inflate_init(&stream);
1613         do {
1614                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1615                 stream.next_in = in;
1616                 /*
1617                  * Note: we must ensure the window section returned by
1618                  * use_pack() will be available throughout git_inflate()'s
1619                  * unlocked execution. Please refer to the comment at
1620                  * get_size_from_delta() to see how this is done.
1621                  */
1622                 obj_read_unlock();
1623                 st = git_inflate(&stream, Z_FINISH);
1624                 obj_read_lock();
1625                 if (!stream.avail_out)
1626                         break; /* the payload is larger than it should be */
1627                 curpos += stream.next_in - in;
1628         } while (st == Z_OK || st == Z_BUF_ERROR);
1629         git_inflate_end(&stream);
1630         if ((st != Z_STREAM_END) || stream.total_out != size) {
1631                 free(buffer);
1632                 return NULL;
1633         }
1634
1635         /* versions of zlib can clobber unconsumed portion of outbuf */
1636         buffer[size] = '\0';
1637
1638         return buffer;
1639 }
1640
1641 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
1642 {
1643         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
1644         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
1645                          p->pack_name, (uintmax_t)obj_offset);
1646 }
1647
1648 int do_check_packed_object_crc;
1649
1650 #define UNPACK_ENTRY_STACK_PREALLOC 64
1651 struct unpack_entry_stack_ent {
1652         off_t obj_offset;
1653         off_t curpos;
1654         unsigned long size;
1655 };
1656
1657 static void *read_object(struct repository *r,
1658                          const struct object_id *oid,
1659                          enum object_type *type,
1660                          unsigned long *size)
1661 {
1662         struct object_info oi = OBJECT_INFO_INIT;
1663         void *content;
1664         oi.typep = type;
1665         oi.sizep = size;
1666         oi.contentp = &content;
1667
1668         if (oid_object_info_extended(r, oid, &oi, 0) < 0)
1669                 return NULL;
1670         return content;
1671 }
1672
1673 void *unpack_entry(struct repository *r, struct packed_git *p, off_t obj_offset,
1674                    enum object_type *final_type, unsigned long *final_size)
1675 {
1676         struct pack_window *w_curs = NULL;
1677         off_t curpos = obj_offset;
1678         void *data = NULL;
1679         unsigned long size;
1680         enum object_type type;
1681         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
1682         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
1683         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
1684         int base_from_cache = 0;
1685
1686         write_pack_access_log(p, obj_offset);
1687
1688         /* PHASE 1: drill down to the innermost base object */
1689         for (;;) {
1690                 off_t base_offset;
1691                 int i;
1692                 struct delta_base_cache_entry *ent;
1693
1694                 ent = get_delta_base_cache_entry(p, curpos);
1695                 if (ent) {
1696                         type = ent->type;
1697                         data = ent->data;
1698                         size = ent->size;
1699                         detach_delta_base_cache_entry(ent);
1700                         base_from_cache = 1;
1701                         break;
1702                 }
1703
1704                 if (do_check_packed_object_crc && p->index_version > 1) {
1705                         uint32_t pack_pos, index_pos;
1706                         off_t len;
1707
1708                         if (offset_to_pack_pos(p, obj_offset, &pack_pos) < 0) {
1709                                 error("could not find object at offset %"PRIuMAX" in pack %s",
1710                                       (uintmax_t)obj_offset, p->pack_name);
1711                                 data = NULL;
1712                                 goto out;
1713                         }
1714
1715                         len = pack_pos_to_offset(p, pack_pos + 1) - obj_offset;
1716                         index_pos = pack_pos_to_index(p, pack_pos);
1717                         if (check_pack_crc(p, &w_curs, obj_offset, len, index_pos)) {
1718                                 struct object_id oid;
1719                                 nth_packed_object_id(&oid, p, index_pos);
1720                                 error("bad packed object CRC for %s",
1721                                       oid_to_hex(&oid));
1722                                 mark_bad_packed_object(p, oid.hash);
1723                                 data = NULL;
1724                                 goto out;
1725                         }
1726                 }
1727
1728                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1729                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
1730                         break;
1731
1732                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1733                 if (!base_offset) {
1734                         error("failed to validate delta base reference "
1735                               "at offset %"PRIuMAX" from %s",
1736                               (uintmax_t)curpos, p->pack_name);
1737                         /* bail to phase 2, in hopes of recovery */
1738                         data = NULL;
1739                         break;
1740                 }
1741
1742                 /* push object, proceed to base */
1743                 if (delta_stack_nr >= delta_stack_alloc
1744                     && delta_stack == small_delta_stack) {
1745                         delta_stack_alloc = alloc_nr(delta_stack_nr);
1746                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
1747                         COPY_ARRAY(delta_stack, small_delta_stack,
1748                                    delta_stack_nr);
1749                 } else {
1750                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
1751                 }
1752                 i = delta_stack_nr++;
1753                 delta_stack[i].obj_offset = obj_offset;
1754                 delta_stack[i].curpos = curpos;
1755                 delta_stack[i].size = size;
1756
1757                 curpos = obj_offset = base_offset;
1758         }
1759
1760         /* PHASE 2: handle the base */
1761         switch (type) {
1762         case OBJ_OFS_DELTA:
1763         case OBJ_REF_DELTA:
1764                 if (data)
1765                         BUG("unpack_entry: left loop at a valid delta");
1766                 break;
1767         case OBJ_COMMIT:
1768         case OBJ_TREE:
1769         case OBJ_BLOB:
1770         case OBJ_TAG:
1771                 if (!base_from_cache)
1772                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
1773                 break;
1774         default:
1775                 data = NULL;
1776                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1777                       type, (uintmax_t)obj_offset, p->pack_name);
1778         }
1779
1780         /* PHASE 3: apply deltas in order */
1781
1782         /* invariants:
1783          *   'data' holds the base data, or NULL if there was corruption
1784          */
1785         while (delta_stack_nr) {
1786                 void *delta_data;
1787                 void *base = data;
1788                 void *external_base = NULL;
1789                 unsigned long delta_size, base_size = size;
1790                 int i;
1791                 off_t base_obj_offset = obj_offset;
1792
1793                 data = NULL;
1794
1795                 if (!base) {
1796                         /*
1797                          * We're probably in deep shit, but let's try to fetch
1798                          * the required base anyway from another pack or loose.
1799                          * This is costly but should happen only in the presence
1800                          * of a corrupted pack, and is better than failing outright.
1801                          */
1802                         uint32_t pos;
1803                         struct object_id base_oid;
1804                         if (!(offset_to_pack_pos(p, obj_offset, &pos))) {
1805                                 nth_packed_object_id(&base_oid, p,
1806                                                      pack_pos_to_index(p, pos));
1807                                 error("failed to read delta base object %s"
1808                                       " at offset %"PRIuMAX" from %s",
1809                                       oid_to_hex(&base_oid), (uintmax_t)obj_offset,
1810                                       p->pack_name);
1811                                 mark_bad_packed_object(p, base_oid.hash);
1812                                 base = read_object(r, &base_oid, &type, &base_size);
1813                                 external_base = base;
1814                         }
1815                 }
1816
1817                 i = --delta_stack_nr;
1818                 obj_offset = delta_stack[i].obj_offset;
1819                 curpos = delta_stack[i].curpos;
1820                 delta_size = delta_stack[i].size;
1821
1822                 if (!base)
1823                         continue;
1824
1825                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
1826
1827                 if (!delta_data) {
1828                         error("failed to unpack compressed delta "
1829                               "at offset %"PRIuMAX" from %s",
1830                               (uintmax_t)curpos, p->pack_name);
1831                         data = NULL;
1832                 } else {
1833                         data = patch_delta(base, base_size, delta_data,
1834                                            delta_size, &size);
1835
1836                         /*
1837                          * We could not apply the delta; warn the user, but
1838                          * keep going. Our failure will be noticed either in
1839                          * the next iteration of the loop, or if this is the
1840                          * final delta, in the caller when we return NULL.
1841                          * Those code paths will take care of making a more
1842                          * explicit warning and retrying with another copy of
1843                          * the object.
1844                          */
1845                         if (!data)
1846                                 error("failed to apply delta");
1847                 }
1848
1849                 /*
1850                  * We delay adding `base` to the cache until the end of the loop
1851                  * because unpack_compressed_entry() momentarily releases the
1852                  * obj_read_mutex, giving another thread the chance to access
1853                  * the cache. Therefore, if `base` was already there, this other
1854                  * thread could free() it (e.g. to make space for another entry)
1855                  * before we are done using it.
1856                  */
1857                 if (!external_base)
1858                         add_delta_base_cache(p, base_obj_offset, base, base_size, type);
1859
1860                 free(delta_data);
1861                 free(external_base);
1862         }
1863
1864         if (final_type)
1865                 *final_type = type;
1866         if (final_size)
1867                 *final_size = size;
1868
1869 out:
1870         unuse_pack(&w_curs);
1871
1872         if (delta_stack != small_delta_stack)
1873                 free(delta_stack);
1874
1875         return data;
1876 }
1877
1878 int bsearch_pack(const struct object_id *oid, const struct packed_git *p, uint32_t *result)
1879 {
1880         const unsigned char *index_fanout = p->index_data;
1881         const unsigned char *index_lookup;
1882         const unsigned int hashsz = the_hash_algo->rawsz;
1883         int index_lookup_width;
1884
1885         if (!index_fanout)
1886                 BUG("bsearch_pack called without a valid pack-index");
1887
1888         index_lookup = index_fanout + 4 * 256;
1889         if (p->index_version == 1) {
1890                 index_lookup_width = hashsz + 4;
1891                 index_lookup += 4;
1892         } else {
1893                 index_lookup_width = hashsz;
1894                 index_fanout += 8;
1895                 index_lookup += 8;
1896         }
1897
1898         return bsearch_hash(oid->hash, (const uint32_t*)index_fanout,
1899                             index_lookup, index_lookup_width, result);
1900 }
1901
1902 int nth_packed_object_id(struct object_id *oid,
1903                          struct packed_git *p,
1904                          uint32_t n)
1905 {
1906         const unsigned char *index = p->index_data;
1907         const unsigned int hashsz = the_hash_algo->rawsz;
1908         if (!index) {
1909                 if (open_pack_index(p))
1910                         return -1;
1911                 index = p->index_data;
1912         }
1913         if (n >= p->num_objects)
1914                 return -1;
1915         index += 4 * 256;
1916         if (p->index_version == 1) {
1917                 oidread(oid, index + (hashsz + 4) * n + 4);
1918         } else {
1919                 index += 8;
1920                 oidread(oid, index + hashsz * n);
1921         }
1922         return 0;
1923 }
1924
1925 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
1926 {
1927         const unsigned char *ptr = vptr;
1928         const unsigned char *start = p->index_data;
1929         const unsigned char *end = start + p->index_size;
1930         if (ptr < start)
1931                 die(_("offset before start of pack index for %s (corrupt index?)"),
1932                     p->pack_name);
1933         /* No need to check for underflow; .idx files must be at least 8 bytes */
1934         if (ptr >= end - 8)
1935                 die(_("offset beyond end of pack index for %s (truncated index?)"),
1936                     p->pack_name);
1937 }
1938
1939 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1940 {
1941         const unsigned char *index = p->index_data;
1942         const unsigned int hashsz = the_hash_algo->rawsz;
1943         index += 4 * 256;
1944         if (p->index_version == 1) {
1945                 return ntohl(*((uint32_t *)(index + (hashsz + 4) * (size_t)n)));
1946         } else {
1947                 uint32_t off;
1948                 index += 8 + (size_t)p->num_objects * (hashsz + 4);
1949                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1950                 if (!(off & 0x80000000))
1951                         return off;
1952                 index += (size_t)p->num_objects * 4 + (off & 0x7fffffff) * 8;
1953                 check_pack_index_ptr(p, index);
1954                 return get_be64(index);
1955         }
1956 }
1957
1958 off_t find_pack_entry_one(const unsigned char *sha1,
1959                                   struct packed_git *p)
1960 {
1961         const unsigned char *index = p->index_data;
1962         struct object_id oid;
1963         uint32_t result;
1964
1965         if (!index) {
1966                 if (open_pack_index(p))
1967                         return 0;
1968         }
1969
1970         hashcpy(oid.hash, sha1);
1971         if (bsearch_pack(&oid, p, &result))
1972                 return nth_packed_object_offset(p, result);
1973         return 0;
1974 }
1975
1976 int is_pack_valid(struct packed_git *p)
1977 {
1978         /* An already open pack is known to be valid. */
1979         if (p->pack_fd != -1)
1980                 return 1;
1981
1982         /* If the pack has one window completely covering the
1983          * file size, the pack is known to be valid even if
1984          * the descriptor is not currently open.
1985          */
1986         if (p->windows) {
1987                 struct pack_window *w = p->windows;
1988
1989                 if (!w->offset && w->len == p->pack_size)
1990                         return 1;
1991         }
1992
1993         /* Force the pack to open to prove its valid. */
1994         return !open_packed_git(p);
1995 }
1996
1997 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1998                                   struct packed_git *packs)
1999 {
2000         struct packed_git *p;
2001
2002         for (p = packs; p; p = p->next) {
2003                 if (find_pack_entry_one(sha1, p))
2004                         return p;
2005         }
2006         return NULL;
2007
2008 }
2009
2010 static int fill_pack_entry(const struct object_id *oid,
2011                            struct pack_entry *e,
2012                            struct packed_git *p)
2013 {
2014         off_t offset;
2015
2016         if (p->num_bad_objects) {
2017                 unsigned i;
2018                 for (i = 0; i < p->num_bad_objects; i++)
2019                         if (hasheq(oid->hash,
2020                                    p->bad_object_sha1 + the_hash_algo->rawsz * i))
2021                                 return 0;
2022         }
2023
2024         offset = find_pack_entry_one(oid->hash, p);
2025         if (!offset)
2026                 return 0;
2027
2028         /*
2029          * We are about to tell the caller where they can locate the
2030          * requested object.  We better make sure the packfile is
2031          * still here and can be accessed before supplying that
2032          * answer, as it may have been deleted since the index was
2033          * loaded!
2034          */
2035         if (!is_pack_valid(p))
2036                 return 0;
2037         e->offset = offset;
2038         e->p = p;
2039         return 1;
2040 }
2041
2042 int find_pack_entry(struct repository *r, const struct object_id *oid, struct pack_entry *e)
2043 {
2044         struct list_head *pos;
2045         struct multi_pack_index *m;
2046
2047         prepare_packed_git(r);
2048         if (!r->objects->packed_git && !r->objects->multi_pack_index)
2049                 return 0;
2050
2051         for (m = r->objects->multi_pack_index; m; m = m->next) {
2052                 if (fill_midx_entry(r, oid, e, m))
2053                         return 1;
2054         }
2055
2056         list_for_each(pos, &r->objects->packed_git_mru) {
2057                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
2058                 if (!p->multi_pack_index && fill_pack_entry(oid, e, p)) {
2059                         list_move(&p->mru, &r->objects->packed_git_mru);
2060                         return 1;
2061                 }
2062         }
2063         return 0;
2064 }
2065
2066 int has_object_pack(const struct object_id *oid)
2067 {
2068         struct pack_entry e;
2069         return find_pack_entry(the_repository, oid, &e);
2070 }
2071
2072 int has_pack_index(const unsigned char *sha1)
2073 {
2074         struct stat st;
2075         if (stat(sha1_pack_index_name(sha1), &st))
2076                 return 0;
2077         return 1;
2078 }
2079
2080 int for_each_object_in_pack(struct packed_git *p,
2081                             each_packed_object_fn cb, void *data,
2082                             enum for_each_object_flags flags)
2083 {
2084         uint32_t i;
2085         int r = 0;
2086
2087         if (flags & FOR_EACH_OBJECT_PACK_ORDER) {
2088                 if (load_pack_revindex(p))
2089                         return -1;
2090         }
2091
2092         for (i = 0; i < p->num_objects; i++) {
2093                 uint32_t index_pos;
2094                 struct object_id oid;
2095
2096                 /*
2097                  * We are iterating "i" from 0 up to num_objects, but its
2098                  * meaning may be different, depending on the requested output
2099                  * order:
2100                  *
2101                  *   - in object-name order, it is the same as the index order
2102                  *     used by nth_packed_object_id(), so we can pass it
2103                  *     directly
2104                  *
2105                  *   - in pack-order, it is pack position, which we must
2106                  *     convert to an index position in order to get the oid.
2107                  */
2108                 if (flags & FOR_EACH_OBJECT_PACK_ORDER)
2109                         index_pos = pack_pos_to_index(p, i);
2110                 else
2111                         index_pos = i;
2112
2113                 if (nth_packed_object_id(&oid, p, index_pos) < 0)
2114                         return error("unable to get sha1 of object %u in %s",
2115                                      index_pos, p->pack_name);
2116
2117                 r = cb(&oid, p, index_pos, data);
2118                 if (r)
2119                         break;
2120         }
2121         return r;
2122 }
2123
2124 int for_each_packed_object(each_packed_object_fn cb, void *data,
2125                            enum for_each_object_flags flags)
2126 {
2127         struct packed_git *p;
2128         int r = 0;
2129         int pack_errors = 0;
2130
2131         prepare_packed_git(the_repository);
2132         for (p = get_all_packs(the_repository); p; p = p->next) {
2133                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
2134                         continue;
2135                 if ((flags & FOR_EACH_OBJECT_PROMISOR_ONLY) &&
2136                     !p->pack_promisor)
2137                         continue;
2138                 if (open_pack_index(p)) {
2139                         pack_errors = 1;
2140                         continue;
2141                 }
2142                 r = for_each_object_in_pack(p, cb, data, flags);
2143                 if (r)
2144                         break;
2145         }
2146         return r ? r : pack_errors;
2147 }
2148
2149 static int add_promisor_object(const struct object_id *oid,
2150                                struct packed_git *pack,
2151                                uint32_t pos,
2152                                void *set_)
2153 {
2154         struct oidset *set = set_;
2155         struct object *obj = parse_object(the_repository, oid);
2156         if (!obj)
2157                 return 1;
2158
2159         oidset_insert(set, oid);
2160
2161         /*
2162          * If this is a tree, commit, or tag, the objects it refers
2163          * to are also promisor objects. (Blobs refer to no objects->)
2164          */
2165         if (obj->type == OBJ_TREE) {
2166                 struct tree *tree = (struct tree *)obj;
2167                 struct tree_desc desc;
2168                 struct name_entry entry;
2169                 if (init_tree_desc_gently(&desc, tree->buffer, tree->size))
2170                         /*
2171                          * Error messages are given when packs are
2172                          * verified, so do not print any here.
2173                          */
2174                         return 0;
2175                 while (tree_entry_gently(&desc, &entry))
2176                         oidset_insert(set, &entry.oid);
2177         } else if (obj->type == OBJ_COMMIT) {
2178                 struct commit *commit = (struct commit *) obj;
2179                 struct commit_list *parents = commit->parents;
2180
2181                 oidset_insert(set, get_commit_tree_oid(commit));
2182                 for (; parents; parents = parents->next)
2183                         oidset_insert(set, &parents->item->object.oid);
2184         } else if (obj->type == OBJ_TAG) {
2185                 struct tag *tag = (struct tag *) obj;
2186                 oidset_insert(set, get_tagged_oid(tag));
2187         }
2188         return 0;
2189 }
2190
2191 int is_promisor_object(const struct object_id *oid)
2192 {
2193         static struct oidset promisor_objects;
2194         static int promisor_objects_prepared;
2195
2196         if (!promisor_objects_prepared) {
2197                 if (has_promisor_remote()) {
2198                         for_each_packed_object(add_promisor_object,
2199                                                &promisor_objects,
2200                                                FOR_EACH_OBJECT_PROMISOR_ONLY);
2201                 }
2202                 promisor_objects_prepared = 1;
2203         }
2204         return oidset_contains(&promisor_objects, oid);
2205 }