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 "list.h"
  10 #include "streaming.h"
  11 #include "sha1-lookup.h"
  12 #include "commit.h"
  13 #include "object.h"
  14 #include "tag.h"
  15 #include "tree-walk.h"
  16 #include "tree.h"
  17 #include "object-store.h"
  18 #include "midx.h"
  19
  20 char *odb_pack_name(struct strbuf *buf,
  21                     const unsigned char *sha1,
  22                     const char *ext)
  23 {
  24         strbuf_reset(buf);
  25         strbuf_addf(buf, "%s/pack/pack-%s.%s", get_object_directory(),
  26                     sha1_to_hex(sha1), ext);
  27         return buf->buf;
  28 }
  29
  30 char *sha1_pack_name(const unsigned char *sha1)
  31 {
  32         static struct strbuf buf = STRBUF_INIT;
  33         return odb_pack_name(&buf, sha1, "pack");
  34 }
  35
  36 char *sha1_pack_index_name(const unsigned char *sha1)
  37 {
  38         static struct strbuf buf = STRBUF_INIT;
  39         return odb_pack_name(&buf, sha1, "idx");
  40 }
  41
  42 static unsigned int pack_used_ctr;
  43 static unsigned int pack_mmap_calls;
  44 static unsigned int peak_pack_open_windows;
  45 static unsigned int pack_open_windows;
  46 static unsigned int pack_open_fds;
  47 static unsigned int pack_max_fds;
  48 static size_t peak_pack_mapped;
  49 static size_t pack_mapped;
  50
  51 #define SZ_FMT PRIuMAX
  52 static inline uintmax_t sz_fmt(size_t s) { return s; }
  53
  54 void pack_report(void)
  55 {
  56         fprintf(stderr,
  57                 "pack_report: getpagesize()            = %10" SZ_FMT "\n"
  58                 "pack_report: core.packedGitWindowSize = %10" SZ_FMT "\n"
  59                 "pack_report: core.packedGitLimit      = %10" SZ_FMT "\n",
  60                 sz_fmt(getpagesize()),
  61                 sz_fmt(packed_git_window_size),
  62                 sz_fmt(packed_git_limit));
  63         fprintf(stderr,
  64                 "pack_report: pack_used_ctr            = %10u\n"
  65                 "pack_report: pack_mmap_calls          = %10u\n"
  66                 "pack_report: pack_open_windows        = %10u / %10u\n"
  67                 "pack_report: pack_mapped              = "
  68                         "%10" SZ_FMT " / %10" SZ_FMT "\n",
  69                 pack_used_ctr,
  70                 pack_mmap_calls,
  71                 pack_open_windows, peak_pack_open_windows,
  72                 sz_fmt(pack_mapped), sz_fmt(peak_pack_mapped));
  73 }
  74
  75 /*
  76  * Open and mmap the index file at path, perform a couple of
  77  * consistency checks, then record its information to p.  Return 0 on
  78  * success.
  79  */
  80 static int check_packed_git_idx(const char *path, struct packed_git *p)
  81 {
  82         void *idx_map;
  83         struct pack_idx_header *hdr;
  84         size_t idx_size;
  85         uint32_t version, nr, i, *index;
  86         int fd = git_open(path);
  87         struct stat st;
  88         const unsigned int hashsz = the_hash_algo->rawsz;
  89
  90         if (fd < 0)
  91                 return -1;
  92         if (fstat(fd, &st)) {
  93                 close(fd);
  94                 return -1;
  95         }
  96         idx_size = xsize_t(st.st_size);
  97         if (idx_size < 4 * 256 + hashsz + hashsz) {
  98                 close(fd);
  99                 return error("index file %s is too small", path);
 100         }
 101         idx_map = xmmap(NULL, idx_size, PROT_READ, MAP_PRIVATE, fd, 0);
 102         close(fd);
 103
 104         hdr = idx_map;
 105         if (hdr->idx_signature == htonl(PACK_IDX_SIGNATURE)) {
 106                 version = ntohl(hdr->idx_version);
 107                 if (version < 2 || version > 2) {
 108                         munmap(idx_map, idx_size);
 109                         return error("index file %s is version %"PRIu32
 110                                      " and is not supported by this binary"
 111                                      " (try upgrading GIT to a newer version)",
 112                                      path, version);
 113                 }
 114         } else
 115                 version = 1;
 116
 117         nr = 0;
 118         index = idx_map;
 119         if (version > 1)
 120                 index += 2;  /* skip index header */
 121         for (i = 0; i < 256; i++) {
 122                 uint32_t n = ntohl(index[i]);
 123                 if (n < nr) {
 124                         munmap(idx_map, idx_size);
 125                         return error("non-monotonic index %s", path);
 126                 }
 127                 nr = n;
 128         }
 129
 130         if (version == 1) {
 131                 /*
 132                  * Total size:
 133                  *  - 256 index entries 4 bytes each
 134                  *  - 24-byte entries * nr (object ID + 4-byte offset)
 135                  *  - hash of the packfile
 136                  *  - file checksum
 137                  */
 138                 if (idx_size != 4*256 + nr * (hashsz + 4) + hashsz + hashsz) {
 139                         munmap(idx_map, idx_size);
 140                         return error("wrong index v1 file size in %s", path);
 141                 }
 142         } else if (version == 2) {
 143                 /*
 144                  * Minimum size:
 145                  *  - 8 bytes of header
 146                  *  - 256 index entries 4 bytes each
 147                  *  - object ID entry * nr
 148                  *  - 4-byte crc entry * nr
 149                  *  - 4-byte offset entry * nr
 150                  *  - hash of the packfile
 151                  *  - file checksum
 152                  * And after the 4-byte offset table might be a
 153                  * variable sized table containing 8-byte entries
 154                  * for offsets larger than 2^31.
 155                  */
 156                 unsigned long min_size = 8 + 4*256 + nr*(hashsz + 4 + 4) + hashsz + hashsz;
 157                 unsigned long max_size = min_size;
 158                 if (nr)
 159                         max_size += (nr - 1)*8;
 160                 if (idx_size < min_size || idx_size > max_size) {
 161                         munmap(idx_map, idx_size);
 162                         return error("wrong index v2 file size in %s", path);
 163                 }
 164                 if (idx_size != min_size &&
 165                     /*
 166                      * make sure we can deal with large pack offsets.
 167                      * 31-bit signed offset won't be enough, neither
 168                      * 32-bit unsigned one will be.
 169                      */
 170                     (sizeof(off_t) <= 4)) {
 171                         munmap(idx_map, idx_size);
 172                         return error("pack too large for current definition of off_t in %s", path);
 173                 }
 174         }
 175
 176         p->index_version = version;
 177         p->index_data = idx_map;
 178         p->index_size = idx_size;
 179         p->num_objects = nr;
 180         return 0;
 181 }
 182
 183 int open_pack_index(struct packed_git *p)
 184 {
 185         char *idx_name;
 186         size_t len;
 187         int ret;
 188
 189         if (p->index_data)
 190                 return 0;
 191
 192         if (!strip_suffix(p->pack_name, ".pack", &len))
 193                 BUG("pack_name does not end in .pack");
 194         idx_name = xstrfmt("%.*s.idx", (int)len, p->pack_name);
 195         ret = check_packed_git_idx(idx_name, p);
 196         free(idx_name);
 197         return ret;
 198 }
 199
 200 uint32_t get_pack_fanout(struct packed_git *p, uint32_t value)
 201 {
 202         const uint32_t *level1_ofs = p->index_data;
 203
 204         if (!level1_ofs) {
 205                 if (open_pack_index(p))
 206                         return 0;
 207                 level1_ofs = p->index_data;
 208         }
 209
 210         if (p->index_version > 1) {
 211                 level1_ofs += 2;
 212         }
 213
 214         return ntohl(level1_ofs[value]);
 215 }
 216
 217 static struct packed_git *alloc_packed_git(int extra)
 218 {
 219         struct packed_git *p = xmalloc(st_add(sizeof(*p), extra));
 220         memset(p, 0, sizeof(*p));
 221         p->pack_fd = -1;
 222         return p;
 223 }
 224
 225 struct packed_git *parse_pack_index(unsigned char *sha1, const char *idx_path)
 226 {
 227         const char *path = sha1_pack_name(sha1);
 228         size_t alloc = st_add(strlen(path), 1);
 229         struct packed_git *p = alloc_packed_git(alloc);
 230
 231         memcpy(p->pack_name, path, alloc); /* includes NUL */
 232         hashcpy(p->sha1, sha1);
 233         if (check_packed_git_idx(idx_path, p)) {
 234                 free(p);
 235                 return NULL;
 236         }
 237
 238         return p;
 239 }
 240
 241 static void scan_windows(struct packed_git *p,
 242         struct packed_git **lru_p,
 243         struct pack_window **lru_w,
 244         struct pack_window **lru_l)
 245 {
 246         struct pack_window *w, *w_l;
 247
 248         for (w_l = NULL, w = p->windows; w; w = w->next) {
 249                 if (!w->inuse_cnt) {
 250                         if (!*lru_w || w->last_used < (*lru_w)->last_used) {
 251                                 *lru_p = p;
 252                                 *lru_w = w;
 253                                 *lru_l = w_l;
 254                         }
 255                 }
 256                 w_l = w;
 257         }
 258 }
 259
 260 static int unuse_one_window(struct packed_git *current)
 261 {
 262         struct packed_git *p, *lru_p = NULL;
 263         struct pack_window *lru_w = NULL, *lru_l = NULL;
 264
 265         if (current)
 266                 scan_windows(current, &lru_p, &lru_w, &lru_l);
 267         for (p = the_repository->objects->packed_git; p; p = p->next)
 268                 scan_windows(p, &lru_p, &lru_w, &lru_l);
 269         if (lru_p) {
 270                 munmap(lru_w->base, lru_w->len);
 271                 pack_mapped -= lru_w->len;
 272                 if (lru_l)
 273                         lru_l->next = lru_w->next;
 274                 else
 275                         lru_p->windows = lru_w->next;
 276                 free(lru_w);
 277                 pack_open_windows--;
 278                 return 1;
 279         }
 280         return 0;
 281 }
 282
 283 void release_pack_memory(size_t need)
 284 {
 285         size_t cur = pack_mapped;
 286         while (need >= (cur - pack_mapped) && unuse_one_window(NULL))
 287                 ; /* nothing */
 288 }
 289
 290 void close_pack_windows(struct packed_git *p)
 291 {
 292         while (p->windows) {
 293                 struct pack_window *w = p->windows;
 294
 295                 if (w->inuse_cnt)
 296                         die("pack '%s' still has open windows to it",
 297                             p->pack_name);
 298                 munmap(w->base, w->len);
 299                 pack_mapped -= w->len;
 300                 pack_open_windows--;
 301                 p->windows = w->next;
 302                 free(w);
 303         }
 304 }
 305
 306 static int close_pack_fd(struct packed_git *p)
 307 {
 308         if (p->pack_fd < 0)
 309                 return 0;
 310
 311         close(p->pack_fd);
 312         pack_open_fds--;
 313         p->pack_fd = -1;
 314
 315         return 1;
 316 }
 317
 318 void close_pack_index(struct packed_git *p)
 319 {
 320         if (p->index_data) {
 321                 munmap((void *)p->index_data, p->index_size);
 322                 p->index_data = NULL;
 323         }
 324 }
 325
 326 void close_pack(struct packed_git *p)
 327 {
 328         close_pack_windows(p);
 329         close_pack_fd(p);
 330         close_pack_index(p);
 331 }
 332
 333 void close_all_packs(struct raw_object_store *o)
 334 {
 335         struct packed_git *p;
 336
 337         for (p = o->packed_git; p; p = p->next)
 338                 if (p->do_not_close)
 339                         BUG("want to close pack marked 'do-not-close'");
 340                 else
 341                         close_pack(p);
 342 }
 343
 344 /*
 345  * The LRU pack is the one with the oldest MRU window, preferring packs
 346  * with no used windows, or the oldest mtime if it has no windows allocated.
 347  */
 348 static void find_lru_pack(struct packed_git *p, struct packed_git **lru_p, struct pack_window **mru_w, int *accept_windows_inuse)
 349 {
 350         struct pack_window *w, *this_mru_w;
 351         int has_windows_inuse = 0;
 352
 353         /*
 354          * Reject this pack if it has windows and the previously selected
 355          * one does not.  If this pack does not have windows, reject
 356          * it if the pack file is newer than the previously selected one.
 357          */
 358         if (*lru_p && !*mru_w && (p->windows || p->mtime > (*lru_p)->mtime))
 359                 return;
 360
 361         for (w = this_mru_w = p->windows; w; w = w->next) {
 362                 /*
 363                  * Reject this pack if any of its windows are in use,
 364                  * but the previously selected pack did not have any
 365                  * inuse windows.  Otherwise, record that this pack
 366                  * has windows in use.
 367                  */
 368                 if (w->inuse_cnt) {
 369                         if (*accept_windows_inuse)
 370                                 has_windows_inuse = 1;
 371                         else
 372                                 return;
 373                 }
 374
 375                 if (w->last_used > this_mru_w->last_used)
 376                         this_mru_w = w;
 377
 378                 /*
 379                  * Reject this pack if it has windows that have been
 380                  * used more recently than the previously selected pack.
 381                  * If the previously selected pack had windows inuse and
 382                  * we have not encountered a window in this pack that is
 383                  * inuse, skip this check since we prefer a pack with no
 384                  * inuse windows to one that has inuse windows.
 385                  */
 386                 if (*mru_w && *accept_windows_inuse == has_windows_inuse &&
 387                     this_mru_w->last_used > (*mru_w)->last_used)
 388                         return;
 389         }
 390
 391         /*
 392          * Select this pack.
 393          */
 394         *mru_w = this_mru_w;
 395         *lru_p = p;
 396         *accept_windows_inuse = has_windows_inuse;
 397 }
 398
 399 static int close_one_pack(void)
 400 {
 401         struct packed_git *p, *lru_p = NULL;
 402         struct pack_window *mru_w = NULL;
 403         int accept_windows_inuse = 1;
 404
 405         for (p = the_repository->objects->packed_git; p; p = p->next) {
 406                 if (p->pack_fd == -1)
 407                         continue;
 408                 find_lru_pack(p, &lru_p, &mru_w, &accept_windows_inuse);
 409         }
 410
 411         if (lru_p)
 412                 return close_pack_fd(lru_p);
 413
 414         return 0;
 415 }
 416
 417 static unsigned int get_max_fd_limit(void)
 418 {
 419 #ifdef RLIMIT_NOFILE
 420         {
 421                 struct rlimit lim;
 422
 423                 if (!getrlimit(RLIMIT_NOFILE, &lim))
 424                         return lim.rlim_cur;
 425         }
 426 #endif
 427
 428 #ifdef _SC_OPEN_MAX
 429         {
 430                 long open_max = sysconf(_SC_OPEN_MAX);
 431                 if (0 < open_max)
 432                         return open_max;
 433                 /*
 434                  * Otherwise, we got -1 for one of the two
 435                  * reasons:
 436                  *
 437                  * (1) sysconf() did not understand _SC_OPEN_MAX
 438                  *     and signaled an error with -1; or
 439                  * (2) sysconf() said there is no limit.
 440                  *
 441                  * We _could_ clear errno before calling sysconf() to
 442                  * tell these two cases apart and return a huge number
 443                  * in the latter case to let the caller cap it to a
 444                  * value that is not so selfish, but letting the
 445                  * fallback OPEN_MAX codepath take care of these cases
 446                  * is a lot simpler.
 447                  */
 448         }
 449 #endif
 450
 451 #ifdef OPEN_MAX
 452         return OPEN_MAX;
 453 #else
 454         return 1; /* see the caller ;-) */
 455 #endif
 456 }
 457
 458 /*
 459  * Do not call this directly as this leaks p->pack_fd on error return;
 460  * call open_packed_git() instead.
 461  */
 462 static int open_packed_git_1(struct packed_git *p)
 463 {
 464         struct stat st;
 465         struct pack_header hdr;
 466         unsigned char hash[GIT_MAX_RAWSZ];
 467         unsigned char *idx_hash;
 468         long fd_flag;
 469         ssize_t read_result;
 470         const unsigned hashsz = the_hash_algo->rawsz;
 471
 472         if (!p->index_data && open_pack_index(p))
 473                 return error("packfile %s index unavailable", p->pack_name);
 474
 475         if (!pack_max_fds) {
 476                 unsigned int max_fds = get_max_fd_limit();
 477
 478                 /* Save 3 for stdin/stdout/stderr, 22 for work */
 479                 if (25 < max_fds)
 480                         pack_max_fds = max_fds - 25;
 481                 else
 482                         pack_max_fds = 1;
 483         }
 484
 485         while (pack_max_fds <= pack_open_fds && close_one_pack())
 486                 ; /* nothing */
 487
 488         p->pack_fd = git_open(p->pack_name);
 489         if (p->pack_fd < 0 || fstat(p->pack_fd, &st))
 490                 return -1;
 491         pack_open_fds++;
 492
 493         /* If we created the struct before we had the pack we lack size. */
 494         if (!p->pack_size) {
 495                 if (!S_ISREG(st.st_mode))
 496                         return error("packfile %s not a regular file", p->pack_name);
 497                 p->pack_size = st.st_size;
 498         } else if (p->pack_size != st.st_size)
 499                 return error("packfile %s size changed", p->pack_name);
 500
 501         /* We leave these file descriptors open with sliding mmap;
 502          * there is no point keeping them open across exec(), though.
 503          */
 504         fd_flag = fcntl(p->pack_fd, F_GETFD, 0);
 505         if (fd_flag < 0)
 506                 return error("cannot determine file descriptor flags");
 507         fd_flag |= FD_CLOEXEC;
 508         if (fcntl(p->pack_fd, F_SETFD, fd_flag) == -1)
 509                 return error("cannot set FD_CLOEXEC");
 510
 511         /* Verify we recognize this pack file format. */
 512         read_result = read_in_full(p->pack_fd, &hdr, sizeof(hdr));
 513         if (read_result < 0)
 514                 return error_errno("error reading from %s", p->pack_name);
 515         if (read_result != sizeof(hdr))
 516                 return error("file %s is far too short to be a packfile", p->pack_name);
 517         if (hdr.hdr_signature != htonl(PACK_SIGNATURE))
 518                 return error("file %s is not a GIT packfile", p->pack_name);
 519         if (!pack_version_ok(hdr.hdr_version))
 520                 return error("packfile %s is version %"PRIu32" and not"
 521                         " supported (try upgrading GIT to a newer version)",
 522                         p->pack_name, ntohl(hdr.hdr_version));
 523
 524         /* Verify the pack matches its index. */
 525         if (p->num_objects != ntohl(hdr.hdr_entries))
 526                 return error("packfile %s claims to have %"PRIu32" objects"
 527                              " while index indicates %"PRIu32" objects",
 528                              p->pack_name, ntohl(hdr.hdr_entries),
 529                              p->num_objects);
 530         if (lseek(p->pack_fd, p->pack_size - hashsz, SEEK_SET) == -1)
 531                 return error("end of packfile %s is unavailable", p->pack_name);
 532         read_result = read_in_full(p->pack_fd, hash, hashsz);
 533         if (read_result < 0)
 534                 return error_errno("error reading from %s", p->pack_name);
 535         if (read_result != hashsz)
 536                 return error("packfile %s signature is unavailable", p->pack_name);
 537         idx_hash = ((unsigned char *)p->index_data) + p->index_size - hashsz * 2;
 538         if (hashcmp(hash, idx_hash))
 539                 return error("packfile %s does not match index", p->pack_name);
 540         return 0;
 541 }
 542
 543 static int open_packed_git(struct packed_git *p)
 544 {
 545         if (!open_packed_git_1(p))
 546                 return 0;
 547         close_pack_fd(p);
 548         return -1;
 549 }
 550
 551 static int in_window(struct pack_window *win, off_t offset)
 552 {
 553         /* We must promise at least one full hash after the
 554          * offset is available from this window, otherwise the offset
 555          * is not actually in this window and a different window (which
 556          * has that one hash excess) must be used.  This is to support
 557          * the object header and delta base parsing routines below.
 558          */
 559         off_t win_off = win->offset;
 560         return win_off <= offset
 561                 && (offset + the_hash_algo->rawsz) <= (win_off + win->len);
 562 }
 563
 564 unsigned char *use_pack(struct packed_git *p,
 565                 struct pack_window **w_cursor,
 566                 off_t offset,
 567                 unsigned long *left)
 568 {
 569         struct pack_window *win = *w_cursor;
 570
 571         /* Since packfiles end in a hash of their content and it's
 572          * pointless to ask for an offset into the middle of that
 573          * hash, and the in_window function above wouldn't match
 574          * don't allow an offset too close to the end of the file.
 575          */
 576         if (!p->pack_size && p->pack_fd == -1 && open_packed_git(p))
 577                 die("packfile %s cannot be accessed", p->pack_name);
 578         if (offset > (p->pack_size - the_hash_algo->rawsz))
 579                 die("offset beyond end of packfile (truncated pack?)");
 580         if (offset < 0)
 581                 die(_("offset before end of packfile (broken .idx?)"));
 582
 583         if (!win || !in_window(win, offset)) {
 584                 if (win)
 585                         win->inuse_cnt--;
 586                 for (win = p->windows; win; win = win->next) {
 587                         if (in_window(win, offset))
 588                                 break;
 589                 }
 590                 if (!win) {
 591                         size_t window_align = packed_git_window_size / 2;
 592                         off_t len;
 593
 594                         if (p->pack_fd == -1 && open_packed_git(p))
 595                                 die("packfile %s cannot be accessed", p->pack_name);
 596
 597                         win = xcalloc(1, sizeof(*win));
 598                         win->offset = (offset / window_align) * window_align;
 599                         len = p->pack_size - win->offset;
 600                         if (len > packed_git_window_size)
 601                                 len = packed_git_window_size;
 602                         win->len = (size_t)len;
 603                         pack_mapped += win->len;
 604                         while (packed_git_limit < pack_mapped
 605                                 && unuse_one_window(p))
 606                                 ; /* nothing */
 607                         win->base = xmmap(NULL, win->len,
 608                                 PROT_READ, MAP_PRIVATE,
 609                                 p->pack_fd, win->offset);
 610                         if (win->base == MAP_FAILED)
 611                                 die_errno("packfile %s cannot be mapped",
 612                                           p->pack_name);
 613                         if (!win->offset && win->len == p->pack_size
 614                                 && !p->do_not_close)
 615                                 close_pack_fd(p);
 616                         pack_mmap_calls++;
 617                         pack_open_windows++;
 618                         if (pack_mapped > peak_pack_mapped)
 619                                 peak_pack_mapped = pack_mapped;
 620                         if (pack_open_windows > peak_pack_open_windows)
 621                                 peak_pack_open_windows = pack_open_windows;
 622                         win->next = p->windows;
 623                         p->windows = win;
 624                 }
 625         }
 626         if (win != *w_cursor) {
 627                 win->last_used = pack_used_ctr++;
 628                 win->inuse_cnt++;
 629                 *w_cursor = win;
 630         }
 631         offset -= win->offset;
 632         if (left)
 633                 *left = win->len - xsize_t(offset);
 634         return win->base + offset;
 635 }
 636
 637 void unuse_pack(struct pack_window **w_cursor)
 638 {
 639         struct pack_window *w = *w_cursor;
 640         if (w) {
 641                 w->inuse_cnt--;
 642                 *w_cursor = NULL;
 643         }
 644 }
 645
 646 static void try_to_free_pack_memory(size_t size)
 647 {
 648         release_pack_memory(size);
 649 }
 650
 651 struct packed_git *add_packed_git(const char *path, size_t path_len, int local)
 652 {
 653         static int have_set_try_to_free_routine;
 654         struct stat st;
 655         size_t alloc;
 656         struct packed_git *p;
 657
 658         if (!have_set_try_to_free_routine) {
 659                 have_set_try_to_free_routine = 1;
 660                 set_try_to_free_routine(try_to_free_pack_memory);
 661         }
 662
 663         /*
 664          * Make sure a corresponding .pack file exists and that
 665          * the index looks sane.
 666          */
 667         if (!strip_suffix_mem(path, &path_len, ".idx"))
 668                 return NULL;
 669
 670         /*
 671          * ".promisor" is long enough to hold any suffix we're adding (and
 672          * the use xsnprintf double-checks that)
 673          */
 674         alloc = st_add3(path_len, strlen(".promisor"), 1);
 675         p = alloc_packed_git(alloc);
 676         memcpy(p->pack_name, path, path_len);
 677
 678         xsnprintf(p->pack_name + path_len, alloc - path_len, ".keep");
 679         if (!access(p->pack_name, F_OK))
 680                 p->pack_keep = 1;
 681
 682         xsnprintf(p->pack_name + path_len, alloc - path_len, ".promisor");
 683         if (!access(p->pack_name, F_OK))
 684                 p->pack_promisor = 1;
 685
 686         xsnprintf(p->pack_name + path_len, alloc - path_len, ".pack");
 687         if (stat(p->pack_name, &st) || !S_ISREG(st.st_mode)) {
 688                 free(p);
 689                 return NULL;
 690         }
 691
 692         /* ok, it looks sane as far as we can check without
 693          * actually mapping the pack file.
 694          */
 695         p->pack_size = st.st_size;
 696         p->pack_local = local;
 697         p->mtime = st.st_mtime;
 698         if (path_len < the_hash_algo->hexsz ||
 699             get_sha1_hex(path + path_len - the_hash_algo->hexsz, p->sha1))
 700                 hashclr(p->sha1);
 701         return p;
 702 }
 703
 704 void install_packed_git(struct repository *r, struct packed_git *pack)
 705 {
 706         if (pack->pack_fd != -1)
 707                 pack_open_fds++;
 708
 709         pack->next = r->objects->packed_git;
 710         r->objects->packed_git = pack;
 711 }
 712
 713 void (*report_garbage)(unsigned seen_bits, const char *path);
 714
 715 static void report_helper(const struct string_list *list,
 716                           int seen_bits, int first, int last)
 717 {
 718         if (seen_bits == (PACKDIR_FILE_PACK|PACKDIR_FILE_IDX))
 719                 return;
 720
 721         for (; first < last; first++)
 722                 report_garbage(seen_bits, list->items[first].string);
 723 }
 724
 725 static void report_pack_garbage(struct string_list *list)
 726 {
 727         int i, baselen = -1, first = 0, seen_bits = 0;
 728
 729         if (!report_garbage)
 730                 return;
 731
 732         string_list_sort(list);
 733
 734         for (i = 0; i < list->nr; i++) {
 735                 const char *path = list->items[i].string;
 736                 if (baselen != -1 &&
 737                     strncmp(path, list->items[first].string, baselen)) {
 738                         report_helper(list, seen_bits, first, i);
 739                         baselen = -1;
 740                         seen_bits = 0;
 741                 }
 742                 if (baselen == -1) {
 743                         const char *dot = strrchr(path, '.');
 744                         if (!dot) {
 745                                 report_garbage(PACKDIR_FILE_GARBAGE, path);
 746                                 continue;
 747                         }
 748                         baselen = dot - path + 1;
 749                         first = i;
 750                 }
 751                 if (!strcmp(path + baselen, "pack"))
 752                         seen_bits |= 1;
 753                 else if (!strcmp(path + baselen, "idx"))
 754                         seen_bits |= 2;
 755         }
 756         report_helper(list, seen_bits, first, list->nr);
 757 }
 758
 759 void for_each_file_in_pack_dir(const char *objdir,
 760                                each_file_in_pack_dir_fn fn,
 761                                void *data)
 762 {
 763         struct strbuf path = STRBUF_INIT;
 764         size_t dirnamelen;
 765         DIR *dir;
 766         struct dirent *de;
 767
 768         strbuf_addstr(&path, objdir);
 769         strbuf_addstr(&path, "/pack");
 770         dir = opendir(path.buf);
 771         if (!dir) {
 772                 if (errno != ENOENT)
 773                         error_errno("unable to open object pack directory: %s",
 774                                     path.buf);
 775                 strbuf_release(&path);
 776                 return;
 777         }
 778         strbuf_addch(&path, '/');
 779         dirnamelen = path.len;
 780         while ((de = readdir(dir)) != NULL) {
 781                 if (is_dot_or_dotdot(de->d_name))
 782                         continue;
 783
 784                 strbuf_setlen(&path, dirnamelen);
 785                 strbuf_addstr(&path, de->d_name);
 786
 787                 fn(path.buf, path.len, de->d_name, data);
 788         }
 789
 790         closedir(dir);
 791         strbuf_release(&path);
 792 }
 793
 794 struct prepare_pack_data {
 795         struct repository *r;
 796         struct string_list *garbage;
 797         int local;
 798 };
 799
 800 static void prepare_pack(const char *full_name, size_t full_name_len,
 801                          const char *file_name, void *_data)
 802 {
 803         struct prepare_pack_data *data = (struct prepare_pack_data *)_data;
 804         struct packed_git *p;
 805         size_t base_len = full_name_len;
 806
 807         if (strip_suffix_mem(full_name, &base_len, ".idx")) {
 808                 /* Don't reopen a pack we already have. */
 809                 for (p = data->r->objects->packed_git; p; p = p->next) {
 810                         size_t len;
 811                         if (strip_suffix(p->pack_name, ".pack", &len) &&
 812                             len == base_len &&
 813                             !memcmp(p->pack_name, full_name, len))
 814                                 break;
 815                 }
 816
 817                 if (!p) {
 818                         p = add_packed_git(full_name, full_name_len, data->local);
 819                         if (p)
 820                                 install_packed_git(data->r, p);
 821                 }
 822         }
 823
 824         if (!report_garbage)
 825                 return;
 826
 827         if (ends_with(file_name, ".idx") ||
 828             ends_with(file_name, ".pack") ||
 829             ends_with(file_name, ".bitmap") ||
 830             ends_with(file_name, ".keep") ||
 831             ends_with(file_name, ".promisor"))
 832                 string_list_append(data->garbage, full_name);
 833         else
 834                 report_garbage(PACKDIR_FILE_GARBAGE, full_name);
 835 }
 836
 837 static void prepare_packed_git_one(struct repository *r, char *objdir, int local)
 838 {
 839         struct prepare_pack_data data;
 840         struct string_list garbage = STRING_LIST_INIT_DUP;
 841
 842         data.r = r;
 843         data.garbage = &garbage;
 844         data.local = local;
 845
 846         for_each_file_in_pack_dir(objdir, prepare_pack, &data);
 847
 848         report_pack_garbage(data.garbage);
 849         string_list_clear(data.garbage, 0);
 850 }
 851
 852 static void prepare_packed_git(struct repository *r);
 853 /*
 854  * Give a fast, rough count of the number of objects in the repository. This
 855  * ignores loose objects completely. If you have a lot of them, then either
 856  * you should repack because your performance will be awful, or they are
 857  * all unreachable objects about to be pruned, in which case they're not really
 858  * interesting as a measure of repo size in the first place.
 859  */
 860 unsigned long approximate_object_count(void)
 861 {
 862         if (!the_repository->objects->approximate_object_count_valid) {
 863                 unsigned long count;
 864                 struct packed_git *p;
 865
 866                 prepare_packed_git(the_repository);
 867                 count = 0;
 868                 for (p = the_repository->objects->packed_git; p; p = p->next) {
 869                         if (open_pack_index(p))
 870                                 continue;
 871                         count += p->num_objects;
 872                 }
 873                 the_repository->objects->approximate_object_count = count;
 874         }
 875         return the_repository->objects->approximate_object_count;
 876 }
 877
 878 static void *get_next_packed_git(const void *p)
 879 {
 880         return ((const struct packed_git *)p)->next;
 881 }
 882
 883 static void set_next_packed_git(void *p, void *next)
 884 {
 885         ((struct packed_git *)p)->next = next;
 886 }
 887
 888 static int sort_pack(const void *a_, const void *b_)
 889 {
 890         const struct packed_git *a = a_;
 891         const struct packed_git *b = b_;
 892         int st;
 893
 894         /*
 895          * Local packs tend to contain objects specific to our
 896          * variant of the project than remote ones.  In addition,
 897          * remote ones could be on a network mounted filesystem.
 898          * Favor local ones for these reasons.
 899          */
 900         st = a->pack_local - b->pack_local;
 901         if (st)
 902                 return -st;
 903
 904         /*
 905          * Younger packs tend to contain more recent objects,
 906          * and more recent objects tend to get accessed more
 907          * often.
 908          */
 909         if (a->mtime < b->mtime)
 910                 return 1;
 911         else if (a->mtime == b->mtime)
 912                 return 0;
 913         return -1;
 914 }
 915
 916 static void rearrange_packed_git(struct repository *r)
 917 {
 918         r->objects->packed_git = llist_mergesort(
 919                 r->objects->packed_git, get_next_packed_git,
 920                 set_next_packed_git, sort_pack);
 921 }
 922
 923 static void prepare_packed_git_mru(struct repository *r)
 924 {
 925         struct packed_git *p;
 926
 927         INIT_LIST_HEAD(&r->objects->packed_git_mru);
 928
 929         for (p = r->objects->packed_git; p; p = p->next)
 930                 list_add_tail(&p->mru, &r->objects->packed_git_mru);
 931 }
 932
 933 static void prepare_packed_git(struct repository *r)
 934 {
 935         struct alternate_object_database *alt;
 936
 937         if (r->objects->packed_git_initialized)
 938                 return;
 939         prepare_multi_pack_index_one(r, r->objects->objectdir);
 940         prepare_packed_git_one(r, r->objects->objectdir, 1);
 941         prepare_alt_odb(r);
 942         for (alt = r->objects->alt_odb_list; alt; alt = alt->next) {
 943                 prepare_multi_pack_index_one(r, alt->path);
 944                 prepare_packed_git_one(r, alt->path, 0);
 945         }
 946         rearrange_packed_git(r);
 947         prepare_packed_git_mru(r);
 948         r->objects->packed_git_initialized = 1;
 949 }
 950
 951 void reprepare_packed_git(struct repository *r)
 952 {
 953         r->objects->approximate_object_count_valid = 0;
 954         r->objects->packed_git_initialized = 0;
 955         prepare_packed_git(r);
 956 }
 957
 958 struct packed_git *get_packed_git(struct repository *r)
 959 {
 960         prepare_packed_git(r);
 961         return r->objects->packed_git;
 962 }
 963
 964 struct list_head *get_packed_git_mru(struct repository *r)
 965 {
 966         prepare_packed_git(r);
 967         return &r->objects->packed_git_mru;
 968 }
 969
 970 unsigned long unpack_object_header_buffer(const unsigned char *buf,
 971                 unsigned long len, enum object_type *type, unsigned long *sizep)
 972 {
 973         unsigned shift;
 974         unsigned long size, c;
 975         unsigned long used = 0;
 976
 977         c = buf[used++];
 978         *type = (c >> 4) & 7;
 979         size = c & 15;
 980         shift = 4;
 981         while (c & 0x80) {
 982                 if (len <= used || bitsizeof(long) <= shift) {
 983                         error("bad object header");
 984                         size = used = 0;
 985                         break;
 986                 }
 987                 c = buf[used++];
 988                 size += (c & 0x7f) << shift;
 989                 shift += 7;
 990         }
 991         *sizep = size;
 992         return used;
 993 }
 994
 995 unsigned long get_size_from_delta(struct packed_git *p,
 996                                   struct pack_window **w_curs,
 997                                   off_t curpos)
 998 {
 999         const unsigned char *data;
1000         unsigned char delta_head[20], *in;
1001         git_zstream stream;
1002         int st;
1003
1004         memset(&stream, 0, sizeof(stream));
1005         stream.next_out = delta_head;
1006         stream.avail_out = sizeof(delta_head);
1007
1008         git_inflate_init(&stream);
1009         do {
1010                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1011                 stream.next_in = in;
1012                 st = git_inflate(&stream, Z_FINISH);
1013                 curpos += stream.next_in - in;
1014         } while ((st == Z_OK || st == Z_BUF_ERROR) &&
1015                  stream.total_out < sizeof(delta_head));
1016         git_inflate_end(&stream);
1017         if ((st != Z_STREAM_END) && stream.total_out != sizeof(delta_head)) {
1018                 error("delta data unpack-initial failed");
1019                 return 0;
1020         }
1021
1022         /* Examine the initial part of the delta to figure out
1023          * the result size.
1024          */
1025         data = delta_head;
1026
1027         /* ignore base size */
1028         get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1029
1030         /* Read the result size */
1031         return get_delta_hdr_size(&data, delta_head+sizeof(delta_head));
1032 }
1033
1034 int unpack_object_header(struct packed_git *p,
1035                          struct pack_window **w_curs,
1036                          off_t *curpos,
1037                          unsigned long *sizep)
1038 {
1039         unsigned char *base;
1040         unsigned long left;
1041         unsigned long used;
1042         enum object_type type;
1043
1044         /* use_pack() assures us we have [base, base + 20) available
1045          * as a range that we can look at.  (Its actually the hash
1046          * size that is assured.)  With our object header encoding
1047          * the maximum deflated object size is 2^137, which is just
1048          * insane, so we know won't exceed what we have been given.
1049          */
1050         base = use_pack(p, w_curs, *curpos, &left);
1051         used = unpack_object_header_buffer(base, left, &type, sizep);
1052         if (!used) {
1053                 type = OBJ_BAD;
1054         } else
1055                 *curpos += used;
1056
1057         return type;
1058 }
1059
1060 void mark_bad_packed_object(struct packed_git *p, const unsigned char *sha1)
1061 {
1062         unsigned i;
1063         for (i = 0; i < p->num_bad_objects; i++)
1064                 if (!hashcmp(sha1, p->bad_object_sha1 + GIT_SHA1_RAWSZ * i))
1065                         return;
1066         p->bad_object_sha1 = xrealloc(p->bad_object_sha1,
1067                                       st_mult(GIT_MAX_RAWSZ,
1068                                               st_add(p->num_bad_objects, 1)));
1069         hashcpy(p->bad_object_sha1 + GIT_SHA1_RAWSZ * p->num_bad_objects, sha1);
1070         p->num_bad_objects++;
1071 }
1072
1073 const struct packed_git *has_packed_and_bad(const unsigned char *sha1)
1074 {
1075         struct packed_git *p;
1076         unsigned i;
1077
1078         for (p = the_repository->objects->packed_git; p; p = p->next)
1079                 for (i = 0; i < p->num_bad_objects; i++)
1080                         if (!hashcmp(sha1,
1081                                      p->bad_object_sha1 + the_hash_algo->rawsz * i))
1082                                 return p;
1083         return NULL;
1084 }
1085
1086 static off_t get_delta_base(struct packed_git *p,
1087                                     struct pack_window **w_curs,
1088                                     off_t *curpos,
1089                                     enum object_type type,
1090                                     off_t delta_obj_offset)
1091 {
1092         unsigned char *base_info = use_pack(p, w_curs, *curpos, NULL);
1093         off_t base_offset;
1094
1095         /* use_pack() assured us we have [base_info, base_info + 20)
1096          * as a range that we can look at without walking off the
1097          * end of the mapped window.  Its actually the hash size
1098          * that is assured.  An OFS_DELTA longer than the hash size
1099          * is stupid, as then a REF_DELTA would be smaller to store.
1100          */
1101         if (type == OBJ_OFS_DELTA) {
1102                 unsigned used = 0;
1103                 unsigned char c = base_info[used++];
1104                 base_offset = c & 127;
1105                 while (c & 128) {
1106                         base_offset += 1;
1107                         if (!base_offset || MSB(base_offset, 7))
1108                                 return 0;  /* overflow */
1109                         c = base_info[used++];
1110                         base_offset = (base_offset << 7) + (c & 127);
1111                 }
1112                 base_offset = delta_obj_offset - base_offset;
1113                 if (base_offset <= 0 || base_offset >= delta_obj_offset)
1114                         return 0;  /* out of bound */
1115                 *curpos += used;
1116         } else if (type == OBJ_REF_DELTA) {
1117                 /* The base entry _must_ be in the same pack */
1118                 base_offset = find_pack_entry_one(base_info, p);
1119                 *curpos += the_hash_algo->rawsz;
1120         } else
1121                 die("I am totally screwed");
1122         return base_offset;
1123 }
1124
1125 /*
1126  * Like get_delta_base above, but we return the sha1 instead of the pack
1127  * offset. This means it is cheaper for REF deltas (we do not have to do
1128  * the final object lookup), but more expensive for OFS deltas (we
1129  * have to load the revidx to convert the offset back into a sha1).
1130  */
1131 static const unsigned char *get_delta_base_sha1(struct packed_git *p,
1132                                                 struct pack_window **w_curs,
1133                                                 off_t curpos,
1134                                                 enum object_type type,
1135                                                 off_t delta_obj_offset)
1136 {
1137         if (type == OBJ_REF_DELTA) {
1138                 unsigned char *base = use_pack(p, w_curs, curpos, NULL);
1139                 return base;
1140         } else if (type == OBJ_OFS_DELTA) {
1141                 struct revindex_entry *revidx;
1142                 off_t base_offset = get_delta_base(p, w_curs, &curpos,
1143                                                    type, delta_obj_offset);
1144
1145                 if (!base_offset)
1146                         return NULL;
1147
1148                 revidx = find_pack_revindex(p, base_offset);
1149                 if (!revidx)
1150                         return NULL;
1151
1152                 return nth_packed_object_sha1(p, revidx->nr);
1153         } else
1154                 return NULL;
1155 }
1156
1157 static int retry_bad_packed_offset(struct repository *r,
1158                                    struct packed_git *p,
1159                                    off_t obj_offset)
1160 {
1161         int type;
1162         struct revindex_entry *revidx;
1163         struct object_id oid;
1164         revidx = find_pack_revindex(p, obj_offset);
1165         if (!revidx)
1166                 return OBJ_BAD;
1167         nth_packed_object_oid(&oid, p, revidx->nr);
1168         mark_bad_packed_object(p, oid.hash);
1169         type = oid_object_info(r, &oid, NULL);
1170         if (type <= OBJ_NONE)
1171                 return OBJ_BAD;
1172         return type;
1173 }
1174
1175 #define POI_STACK_PREALLOC 64
1176
1177 static enum object_type packed_to_object_type(struct repository *r,
1178                                               struct packed_git *p,
1179                                               off_t obj_offset,
1180                                               enum object_type type,
1181                                               struct pack_window **w_curs,
1182                                               off_t curpos)
1183 {
1184         off_t small_poi_stack[POI_STACK_PREALLOC];
1185         off_t *poi_stack = small_poi_stack;
1186         int poi_stack_nr = 0, poi_stack_alloc = POI_STACK_PREALLOC;
1187
1188         while (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1189                 off_t base_offset;
1190                 unsigned long size;
1191                 /* Push the object we're going to leave behind */
1192                 if (poi_stack_nr >= poi_stack_alloc && poi_stack == small_poi_stack) {
1193                         poi_stack_alloc = alloc_nr(poi_stack_nr);
1194                         ALLOC_ARRAY(poi_stack, poi_stack_alloc);
1195                         memcpy(poi_stack, small_poi_stack, sizeof(off_t)*poi_stack_nr);
1196                 } else {
1197                         ALLOC_GROW(poi_stack, poi_stack_nr+1, poi_stack_alloc);
1198                 }
1199                 poi_stack[poi_stack_nr++] = obj_offset;
1200                 /* If parsing the base offset fails, just unwind */
1201                 base_offset = get_delta_base(p, w_curs, &curpos, type, obj_offset);
1202                 if (!base_offset)
1203                         goto unwind;
1204                 curpos = obj_offset = base_offset;
1205                 type = unpack_object_header(p, w_curs, &curpos, &size);
1206                 if (type <= OBJ_NONE) {
1207                         /* If getting the base itself fails, we first
1208                          * retry the base, otherwise unwind */
1209                         type = retry_bad_packed_offset(r, p, base_offset);
1210                         if (type > OBJ_NONE)
1211                                 goto out;
1212                         goto unwind;
1213                 }
1214         }
1215
1216         switch (type) {
1217         case OBJ_BAD:
1218         case OBJ_COMMIT:
1219         case OBJ_TREE:
1220         case OBJ_BLOB:
1221         case OBJ_TAG:
1222                 break;
1223         default:
1224                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1225                       type, (uintmax_t)obj_offset, p->pack_name);
1226                 type = OBJ_BAD;
1227         }
1228
1229 out:
1230         if (poi_stack != small_poi_stack)
1231                 free(poi_stack);
1232         return type;
1233
1234 unwind:
1235         while (poi_stack_nr) {
1236                 obj_offset = poi_stack[--poi_stack_nr];
1237                 type = retry_bad_packed_offset(r, p, obj_offset);
1238                 if (type > OBJ_NONE)
1239                         goto out;
1240         }
1241         type = OBJ_BAD;
1242         goto out;
1243 }
1244
1245 static struct hashmap delta_base_cache;
1246 static size_t delta_base_cached;
1247
1248 static LIST_HEAD(delta_base_cache_lru);
1249
1250 struct delta_base_cache_key {
1251         struct packed_git *p;
1252         off_t base_offset;
1253 };
1254
1255 struct delta_base_cache_entry {
1256         struct hashmap hash;
1257         struct delta_base_cache_key key;
1258         struct list_head lru;
1259         void *data;
1260         unsigned long size;
1261         enum object_type type;
1262 };
1263
1264 static unsigned int pack_entry_hash(struct packed_git *p, off_t base_offset)
1265 {
1266         unsigned int hash;
1267
1268         hash = (unsigned int)(intptr_t)p + (unsigned int)base_offset;
1269         hash += (hash >> 8) + (hash >> 16);
1270         return hash;
1271 }
1272
1273 static struct delta_base_cache_entry *
1274 get_delta_base_cache_entry(struct packed_git *p, off_t base_offset)
1275 {
1276         struct hashmap_entry entry;
1277         struct delta_base_cache_key key;
1278
1279         if (!delta_base_cache.cmpfn)
1280                 return NULL;
1281
1282         hashmap_entry_init(&entry, pack_entry_hash(p, base_offset));
1283         key.p = p;
1284         key.base_offset = base_offset;
1285         return hashmap_get(&delta_base_cache, &entry, &key);
1286 }
1287
1288 static int delta_base_cache_key_eq(const struct delta_base_cache_key *a,
1289                                    const struct delta_base_cache_key *b)
1290 {
1291         return a->p == b->p && a->base_offset == b->base_offset;
1292 }
1293
1294 static int delta_base_cache_hash_cmp(const void *unused_cmp_data,
1295                                      const void *va, const void *vb,
1296                                      const void *vkey)
1297 {
1298         const struct delta_base_cache_entry *a = va, *b = vb;
1299         const struct delta_base_cache_key *key = vkey;
1300         if (key)
1301                 return !delta_base_cache_key_eq(&a->key, key);
1302         else
1303                 return !delta_base_cache_key_eq(&a->key, &b->key);
1304 }
1305
1306 static int in_delta_base_cache(struct packed_git *p, off_t base_offset)
1307 {
1308         return !!get_delta_base_cache_entry(p, base_offset);
1309 }
1310
1311 /*
1312  * Remove the entry from the cache, but do _not_ free the associated
1313  * entry data. The caller takes ownership of the "data" buffer, and
1314  * should copy out any fields it wants before detaching.
1315  */
1316 static void detach_delta_base_cache_entry(struct delta_base_cache_entry *ent)
1317 {
1318         hashmap_remove(&delta_base_cache, ent, &ent->key);
1319         list_del(&ent->lru);
1320         delta_base_cached -= ent->size;
1321         free(ent);
1322 }
1323
1324 static void *cache_or_unpack_entry(struct repository *r, struct packed_git *p,
1325                                    off_t base_offset, unsigned long *base_size,
1326                                    enum object_type *type)
1327 {
1328         struct delta_base_cache_entry *ent;
1329
1330         ent = get_delta_base_cache_entry(p, base_offset);
1331         if (!ent)
1332                 return unpack_entry(r, p, base_offset, type, base_size);
1333
1334         if (type)
1335                 *type = ent->type;
1336         if (base_size)
1337                 *base_size = ent->size;
1338         return xmemdupz(ent->data, ent->size);
1339 }
1340
1341 static inline void release_delta_base_cache(struct delta_base_cache_entry *ent)
1342 {
1343         free(ent->data);
1344         detach_delta_base_cache_entry(ent);
1345 }
1346
1347 void clear_delta_base_cache(void)
1348 {
1349         struct list_head *lru, *tmp;
1350         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1351                 struct delta_base_cache_entry *entry =
1352                         list_entry(lru, struct delta_base_cache_entry, lru);
1353                 release_delta_base_cache(entry);
1354         }
1355 }
1356
1357 static void add_delta_base_cache(struct packed_git *p, off_t base_offset,
1358         void *base, unsigned long base_size, enum object_type type)
1359 {
1360         struct delta_base_cache_entry *ent = xmalloc(sizeof(*ent));
1361         struct list_head *lru, *tmp;
1362
1363         delta_base_cached += base_size;
1364
1365         list_for_each_safe(lru, tmp, &delta_base_cache_lru) {
1366                 struct delta_base_cache_entry *f =
1367                         list_entry(lru, struct delta_base_cache_entry, lru);
1368                 if (delta_base_cached <= delta_base_cache_limit)
1369                         break;
1370                 release_delta_base_cache(f);
1371         }
1372
1373         ent->key.p = p;
1374         ent->key.base_offset = base_offset;
1375         ent->type = type;
1376         ent->data = base;
1377         ent->size = base_size;
1378         list_add_tail(&ent->lru, &delta_base_cache_lru);
1379
1380         if (!delta_base_cache.cmpfn)
1381                 hashmap_init(&delta_base_cache, delta_base_cache_hash_cmp, NULL, 0);
1382         hashmap_entry_init(ent, pack_entry_hash(p, base_offset));
1383         hashmap_add(&delta_base_cache, ent);
1384 }
1385
1386 int packed_object_info(struct repository *r, struct packed_git *p,
1387                        off_t obj_offset, struct object_info *oi)
1388 {
1389         struct pack_window *w_curs = NULL;
1390         unsigned long size;
1391         off_t curpos = obj_offset;
1392         enum object_type type;
1393
1394         /*
1395          * We always get the representation type, but only convert it to
1396          * a "real" type later if the caller is interested.
1397          */
1398         if (oi->contentp) {
1399                 *oi->contentp = cache_or_unpack_entry(r, p, obj_offset, oi->sizep,
1400                                                       &type);
1401                 if (!*oi->contentp)
1402                         type = OBJ_BAD;
1403         } else {
1404                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1405         }
1406
1407         if (!oi->contentp && oi->sizep) {
1408                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1409                         off_t tmp_pos = curpos;
1410                         off_t base_offset = get_delta_base(p, &w_curs, &tmp_pos,
1411                                                            type, obj_offset);
1412                         if (!base_offset) {
1413                                 type = OBJ_BAD;
1414                                 goto out;
1415                         }
1416                         *oi->sizep = get_size_from_delta(p, &w_curs, tmp_pos);
1417                         if (*oi->sizep == 0) {
1418                                 type = OBJ_BAD;
1419                                 goto out;
1420                         }
1421                 } else {
1422                         *oi->sizep = size;
1423                 }
1424         }
1425
1426         if (oi->disk_sizep) {
1427                 struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1428                 *oi->disk_sizep = revidx[1].offset - obj_offset;
1429         }
1430
1431         if (oi->typep || oi->type_name) {
1432                 enum object_type ptot;
1433                 ptot = packed_to_object_type(r, p, obj_offset,
1434                                              type, &w_curs, curpos);
1435                 if (oi->typep)
1436                         *oi->typep = ptot;
1437                 if (oi->type_name) {
1438                         const char *tn = type_name(ptot);
1439                         if (tn)
1440                                 strbuf_addstr(oi->type_name, tn);
1441                 }
1442                 if (ptot < 0) {
1443                         type = OBJ_BAD;
1444                         goto out;
1445                 }
1446         }
1447
1448         if (oi->delta_base_sha1) {
1449                 if (type == OBJ_OFS_DELTA || type == OBJ_REF_DELTA) {
1450                         const unsigned char *base;
1451
1452                         base = get_delta_base_sha1(p, &w_curs, curpos,
1453                                                    type, obj_offset);
1454                         if (!base) {
1455                                 type = OBJ_BAD;
1456                                 goto out;
1457                         }
1458
1459                         hashcpy(oi->delta_base_sha1, base);
1460                 } else
1461                         hashclr(oi->delta_base_sha1);
1462         }
1463
1464         oi->whence = in_delta_base_cache(p, obj_offset) ? OI_DBCACHED :
1465                                                           OI_PACKED;
1466
1467 out:
1468         unuse_pack(&w_curs);
1469         return type;
1470 }
1471
1472 static void *unpack_compressed_entry(struct packed_git *p,
1473                                     struct pack_window **w_curs,
1474                                     off_t curpos,
1475                                     unsigned long size)
1476 {
1477         int st;
1478         git_zstream stream;
1479         unsigned char *buffer, *in;
1480
1481         buffer = xmallocz_gently(size);
1482         if (!buffer)
1483                 return NULL;
1484         memset(&stream, 0, sizeof(stream));
1485         stream.next_out = buffer;
1486         stream.avail_out = size + 1;
1487
1488         git_inflate_init(&stream);
1489         do {
1490                 in = use_pack(p, w_curs, curpos, &stream.avail_in);
1491                 stream.next_in = in;
1492                 st = git_inflate(&stream, Z_FINISH);
1493                 if (!stream.avail_out)
1494                         break; /* the payload is larger than it should be */
1495                 curpos += stream.next_in - in;
1496         } while (st == Z_OK || st == Z_BUF_ERROR);
1497         git_inflate_end(&stream);
1498         if ((st != Z_STREAM_END) || stream.total_out != size) {
1499                 free(buffer);
1500                 return NULL;
1501         }
1502
1503         /* versions of zlib can clobber unconsumed portion of outbuf */
1504         buffer[size] = '\0';
1505
1506         return buffer;
1507 }
1508
1509 static void write_pack_access_log(struct packed_git *p, off_t obj_offset)
1510 {
1511         static struct trace_key pack_access = TRACE_KEY_INIT(PACK_ACCESS);
1512         trace_printf_key(&pack_access, "%s %"PRIuMAX"\n",
1513                          p->pack_name, (uintmax_t)obj_offset);
1514 }
1515
1516 int do_check_packed_object_crc;
1517
1518 #define UNPACK_ENTRY_STACK_PREALLOC 64
1519 struct unpack_entry_stack_ent {
1520         off_t obj_offset;
1521         off_t curpos;
1522         unsigned long size;
1523 };
1524
1525 static void *read_object(struct repository *r,
1526                          const struct object_id *oid,
1527                          enum object_type *type,
1528                          unsigned long *size)
1529 {
1530         struct object_info oi = OBJECT_INFO_INIT;
1531         void *content;
1532         oi.typep = type;
1533         oi.sizep = size;
1534         oi.contentp = &content;
1535
1536         if (oid_object_info_extended(r, oid, &oi, 0) < 0)
1537                 return NULL;
1538         return content;
1539 }
1540
1541 void *unpack_entry(struct repository *r, struct packed_git *p, off_t obj_offset,
1542                    enum object_type *final_type, unsigned long *final_size)
1543 {
1544         struct pack_window *w_curs = NULL;
1545         off_t curpos = obj_offset;
1546         void *data = NULL;
1547         unsigned long size;
1548         enum object_type type;
1549         struct unpack_entry_stack_ent small_delta_stack[UNPACK_ENTRY_STACK_PREALLOC];
1550         struct unpack_entry_stack_ent *delta_stack = small_delta_stack;
1551         int delta_stack_nr = 0, delta_stack_alloc = UNPACK_ENTRY_STACK_PREALLOC;
1552         int base_from_cache = 0;
1553
1554         write_pack_access_log(p, obj_offset);
1555
1556         /* PHASE 1: drill down to the innermost base object */
1557         for (;;) {
1558                 off_t base_offset;
1559                 int i;
1560                 struct delta_base_cache_entry *ent;
1561
1562                 ent = get_delta_base_cache_entry(p, curpos);
1563                 if (ent) {
1564                         type = ent->type;
1565                         data = ent->data;
1566                         size = ent->size;
1567                         detach_delta_base_cache_entry(ent);
1568                         base_from_cache = 1;
1569                         break;
1570                 }
1571
1572                 if (do_check_packed_object_crc && p->index_version > 1) {
1573                         struct revindex_entry *revidx = find_pack_revindex(p, obj_offset);
1574                         off_t len = revidx[1].offset - obj_offset;
1575                         if (check_pack_crc(p, &w_curs, obj_offset, len, revidx->nr)) {
1576                                 struct object_id oid;
1577                                 nth_packed_object_oid(&oid, p, revidx->nr);
1578                                 error("bad packed object CRC for %s",
1579                                       oid_to_hex(&oid));
1580                                 mark_bad_packed_object(p, oid.hash);
1581                                 data = NULL;
1582                                 goto out;
1583                         }
1584                 }
1585
1586                 type = unpack_object_header(p, &w_curs, &curpos, &size);
1587                 if (type != OBJ_OFS_DELTA && type != OBJ_REF_DELTA)
1588                         break;
1589
1590                 base_offset = get_delta_base(p, &w_curs, &curpos, type, obj_offset);
1591                 if (!base_offset) {
1592                         error("failed to validate delta base reference "
1593                               "at offset %"PRIuMAX" from %s",
1594                               (uintmax_t)curpos, p->pack_name);
1595                         /* bail to phase 2, in hopes of recovery */
1596                         data = NULL;
1597                         break;
1598                 }
1599
1600                 /* push object, proceed to base */
1601                 if (delta_stack_nr >= delta_stack_alloc
1602                     && delta_stack == small_delta_stack) {
1603                         delta_stack_alloc = alloc_nr(delta_stack_nr);
1604                         ALLOC_ARRAY(delta_stack, delta_stack_alloc);
1605                         memcpy(delta_stack, small_delta_stack,
1606                                sizeof(*delta_stack)*delta_stack_nr);
1607                 } else {
1608                         ALLOC_GROW(delta_stack, delta_stack_nr+1, delta_stack_alloc);
1609                 }
1610                 i = delta_stack_nr++;
1611                 delta_stack[i].obj_offset = obj_offset;
1612                 delta_stack[i].curpos = curpos;
1613                 delta_stack[i].size = size;
1614
1615                 curpos = obj_offset = base_offset;
1616         }
1617
1618         /* PHASE 2: handle the base */
1619         switch (type) {
1620         case OBJ_OFS_DELTA:
1621         case OBJ_REF_DELTA:
1622                 if (data)
1623                         BUG("unpack_entry: left loop at a valid delta");
1624                 break;
1625         case OBJ_COMMIT:
1626         case OBJ_TREE:
1627         case OBJ_BLOB:
1628         case OBJ_TAG:
1629                 if (!base_from_cache)
1630                         data = unpack_compressed_entry(p, &w_curs, curpos, size);
1631                 break;
1632         default:
1633                 data = NULL;
1634                 error("unknown object type %i at offset %"PRIuMAX" in %s",
1635                       type, (uintmax_t)obj_offset, p->pack_name);
1636         }
1637
1638         /* PHASE 3: apply deltas in order */
1639
1640         /* invariants:
1641          *   'data' holds the base data, or NULL if there was corruption
1642          */
1643         while (delta_stack_nr) {
1644                 void *delta_data;
1645                 void *base = data;
1646                 void *external_base = NULL;
1647                 unsigned long delta_size, base_size = size;
1648                 int i;
1649
1650                 data = NULL;
1651
1652                 if (base)
1653                         add_delta_base_cache(p, obj_offset, base, base_size, type);
1654
1655                 if (!base) {
1656                         /*
1657                          * We're probably in deep shit, but let's try to fetch
1658                          * the required base anyway from another pack or loose.
1659                          * This is costly but should happen only in the presence
1660                          * of a corrupted pack, and is better than failing outright.
1661                          */
1662                         struct revindex_entry *revidx;
1663                         struct object_id base_oid;
1664                         revidx = find_pack_revindex(p, obj_offset);
1665                         if (revidx) {
1666                                 nth_packed_object_oid(&base_oid, p, revidx->nr);
1667                                 error("failed to read delta base object %s"
1668                                       " at offset %"PRIuMAX" from %s",
1669                                       oid_to_hex(&base_oid), (uintmax_t)obj_offset,
1670                                       p->pack_name);
1671                                 mark_bad_packed_object(p, base_oid.hash);
1672                                 base = read_object(r, &base_oid, &type, &base_size);
1673                                 external_base = base;
1674                         }
1675                 }
1676
1677                 i = --delta_stack_nr;
1678                 obj_offset = delta_stack[i].obj_offset;
1679                 curpos = delta_stack[i].curpos;
1680                 delta_size = delta_stack[i].size;
1681
1682                 if (!base)
1683                         continue;
1684
1685                 delta_data = unpack_compressed_entry(p, &w_curs, curpos, delta_size);
1686
1687                 if (!delta_data) {
1688                         error("failed to unpack compressed delta "
1689                               "at offset %"PRIuMAX" from %s",
1690                               (uintmax_t)curpos, p->pack_name);
1691                         data = NULL;
1692                         free(external_base);
1693                         continue;
1694                 }
1695
1696                 data = patch_delta(base, base_size,
1697                                    delta_data, delta_size,
1698                                    &size);
1699
1700                 /*
1701                  * We could not apply the delta; warn the user, but keep going.
1702                  * Our failure will be noticed either in the next iteration of
1703                  * the loop, or if this is the final delta, in the caller when
1704                  * we return NULL. Those code paths will take care of making
1705                  * a more explicit warning and retrying with another copy of
1706                  * the object.
1707                  */
1708                 if (!data)
1709                         error("failed to apply delta");
1710
1711                 free(delta_data);
1712                 free(external_base);
1713         }
1714
1715         if (final_type)
1716                 *final_type = type;
1717         if (final_size)
1718                 *final_size = size;
1719
1720 out:
1721         unuse_pack(&w_curs);
1722
1723         if (delta_stack != small_delta_stack)
1724                 free(delta_stack);
1725
1726         return data;
1727 }
1728
1729 int bsearch_pack(const struct object_id *oid, const struct packed_git *p, uint32_t *result)
1730 {
1731         const unsigned char *index_fanout = p->index_data;
1732         const unsigned char *index_lookup;
1733         const unsigned int hashsz = the_hash_algo->rawsz;
1734         int index_lookup_width;
1735
1736         if (!index_fanout)
1737                 BUG("bsearch_pack called without a valid pack-index");
1738
1739         index_lookup = index_fanout + 4 * 256;
1740         if (p->index_version == 1) {
1741                 index_lookup_width = hashsz + 4;
1742                 index_lookup += 4;
1743         } else {
1744                 index_lookup_width = hashsz;
1745                 index_fanout += 8;
1746                 index_lookup += 8;
1747         }
1748
1749         return bsearch_hash(oid->hash, (const uint32_t*)index_fanout,
1750                             index_lookup, index_lookup_width, result);
1751 }
1752
1753 const unsigned char *nth_packed_object_sha1(struct packed_git *p,
1754                                             uint32_t n)
1755 {
1756         const unsigned char *index = p->index_data;
1757         const unsigned int hashsz = the_hash_algo->rawsz;
1758         if (!index) {
1759                 if (open_pack_index(p))
1760                         return NULL;
1761                 index = p->index_data;
1762         }
1763         if (n >= p->num_objects)
1764                 return NULL;
1765         index += 4 * 256;
1766         if (p->index_version == 1) {
1767                 return index + (hashsz + 4) * n + 4;
1768         } else {
1769                 index += 8;
1770                 return index + hashsz * n;
1771         }
1772 }
1773
1774 const struct object_id *nth_packed_object_oid(struct object_id *oid,
1775                                               struct packed_git *p,
1776                                               uint32_t n)
1777 {
1778         const unsigned char *hash = nth_packed_object_sha1(p, n);
1779         if (!hash)
1780                 return NULL;
1781         hashcpy(oid->hash, hash);
1782         return oid;
1783 }
1784
1785 void check_pack_index_ptr(const struct packed_git *p, const void *vptr)
1786 {
1787         const unsigned char *ptr = vptr;
1788         const unsigned char *start = p->index_data;
1789         const unsigned char *end = start + p->index_size;
1790         if (ptr < start)
1791                 die(_("offset before start of pack index for %s (corrupt index?)"),
1792                     p->pack_name);
1793         /* No need to check for underflow; .idx files must be at least 8 bytes */
1794         if (ptr >= end - 8)
1795                 die(_("offset beyond end of pack index for %s (truncated index?)"),
1796                     p->pack_name);
1797 }
1798
1799 off_t nth_packed_object_offset(const struct packed_git *p, uint32_t n)
1800 {
1801         const unsigned char *index = p->index_data;
1802         const unsigned int hashsz = the_hash_algo->rawsz;
1803         index += 4 * 256;
1804         if (p->index_version == 1) {
1805                 return ntohl(*((uint32_t *)(index + (hashsz + 4) * n)));
1806         } else {
1807                 uint32_t off;
1808                 index += 8 + p->num_objects * (hashsz + 4);
1809                 off = ntohl(*((uint32_t *)(index + 4 * n)));
1810                 if (!(off & 0x80000000))
1811                         return off;
1812                 index += p->num_objects * 4 + (off & 0x7fffffff) * 8;
1813                 check_pack_index_ptr(p, index);
1814                 return get_be64(index);
1815         }
1816 }
1817
1818 off_t find_pack_entry_one(const unsigned char *sha1,
1819                                   struct packed_git *p)
1820 {
1821         const unsigned char *index = p->index_data;
1822         struct object_id oid;
1823         uint32_t result;
1824
1825         if (!index) {
1826                 if (open_pack_index(p))
1827                         return 0;
1828         }
1829
1830         hashcpy(oid.hash, sha1);
1831         if (bsearch_pack(&oid, p, &result))
1832                 return nth_packed_object_offset(p, result);
1833         return 0;
1834 }
1835
1836 int is_pack_valid(struct packed_git *p)
1837 {
1838         /* An already open pack is known to be valid. */
1839         if (p->pack_fd != -1)
1840                 return 1;
1841
1842         /* If the pack has one window completely covering the
1843          * file size, the pack is known to be valid even if
1844          * the descriptor is not currently open.
1845          */
1846         if (p->windows) {
1847                 struct pack_window *w = p->windows;
1848
1849                 if (!w->offset && w->len == p->pack_size)
1850                         return 1;
1851         }
1852
1853         /* Force the pack to open to prove its valid. */
1854         return !open_packed_git(p);
1855 }
1856
1857 struct packed_git *find_sha1_pack(const unsigned char *sha1,
1858                                   struct packed_git *packs)
1859 {
1860         struct packed_git *p;
1861
1862         for (p = packs; p; p = p->next) {
1863                 if (find_pack_entry_one(sha1, p))
1864                         return p;
1865         }
1866         return NULL;
1867
1868 }
1869
1870 static int fill_pack_entry(const struct object_id *oid,
1871                            struct pack_entry *e,
1872                            struct packed_git *p)
1873 {
1874         off_t offset;
1875
1876         if (p->num_bad_objects) {
1877                 unsigned i;
1878                 for (i = 0; i < p->num_bad_objects; i++)
1879                         if (!hashcmp(oid->hash,
1880                                      p->bad_object_sha1 + the_hash_algo->rawsz * i))
1881                                 return 0;
1882         }
1883
1884         offset = find_pack_entry_one(oid->hash, p);
1885         if (!offset)
1886                 return 0;
1887
1888         /*
1889          * We are about to tell the caller where they can locate the
1890          * requested object.  We better make sure the packfile is
1891          * still here and can be accessed before supplying that
1892          * answer, as it may have been deleted since the index was
1893          * loaded!
1894          */
1895         if (!is_pack_valid(p))
1896                 return 0;
1897         e->offset = offset;
1898         e->p = p;
1899         return 1;
1900 }
1901
1902 int find_pack_entry(struct repository *r, const struct object_id *oid, struct pack_entry *e)
1903 {
1904         struct list_head *pos;
1905
1906         prepare_packed_git(r);
1907         if (!r->objects->packed_git)
1908                 return 0;
1909
1910         list_for_each(pos, &r->objects->packed_git_mru) {
1911                 struct packed_git *p = list_entry(pos, struct packed_git, mru);
1912                 if (fill_pack_entry(oid, e, p)) {
1913                         list_move(&p->mru, &r->objects->packed_git_mru);
1914                         return 1;
1915                 }
1916         }
1917         return 0;
1918 }
1919
1920 int has_object_pack(const struct object_id *oid)
1921 {
1922         struct pack_entry e;
1923         return find_pack_entry(the_repository, oid, &e);
1924 }
1925
1926 int has_pack_index(const unsigned char *sha1)
1927 {
1928         struct stat st;
1929         if (stat(sha1_pack_index_name(sha1), &st))
1930                 return 0;
1931         return 1;
1932 }
1933
1934 int for_each_object_in_pack(struct packed_git *p, each_packed_object_fn cb, void *data)
1935 {
1936         uint32_t i;
1937         int r = 0;
1938
1939         for (i = 0; i < p->num_objects; i++) {
1940                 struct object_id oid;
1941
1942                 if (!nth_packed_object_oid(&oid, p, i))
1943                         return error("unable to get sha1 of object %u in %s",
1944                                      i, p->pack_name);
1945
1946                 r = cb(&oid, p, i, data);
1947                 if (r)
1948                         break;
1949         }
1950         return r;
1951 }
1952
1953 int for_each_packed_object(each_packed_object_fn cb, void *data, unsigned flags)
1954 {
1955         struct packed_git *p;
1956         int r = 0;
1957         int pack_errors = 0;
1958
1959         prepare_packed_git(the_repository);
1960         for (p = the_repository->objects->packed_git; p; p = p->next) {
1961                 if ((flags & FOR_EACH_OBJECT_LOCAL_ONLY) && !p->pack_local)
1962                         continue;
1963                 if ((flags & FOR_EACH_OBJECT_PROMISOR_ONLY) &&
1964                     !p->pack_promisor)
1965                         continue;
1966                 if (open_pack_index(p)) {
1967                         pack_errors = 1;
1968                         continue;
1969                 }
1970                 r = for_each_object_in_pack(p, cb, data);
1971                 if (r)
1972                         break;
1973         }
1974         return r ? r : pack_errors;
1975 }
1976
1977 static int add_promisor_object(const struct object_id *oid,
1978                                struct packed_git *pack,
1979                                uint32_t pos,
1980                                void *set_)
1981 {
1982         struct oidset *set = set_;
1983         struct object *obj = parse_object(oid);
1984         if (!obj)
1985                 return 1;
1986
1987         oidset_insert(set, oid);
1988
1989         /*
1990          * If this is a tree, commit, or tag, the objects it refers
1991          * to are also promisor objects. (Blobs refer to no objects->)
1992          */
1993         if (obj->type == OBJ_TREE) {
1994                 struct tree *tree = (struct tree *)obj;
1995                 struct tree_desc desc;
1996                 struct name_entry entry;
1997                 if (init_tree_desc_gently(&desc, tree->buffer, tree->size))
1998                         /*
1999                          * Error messages are given when packs are
2000                          * verified, so do not print any here.
2001                          */
2002                         return 0;
2003                 while (tree_entry_gently(&desc, &entry))
2004                         oidset_insert(set, entry.oid);
2005         } else if (obj->type == OBJ_COMMIT) {
2006                 struct commit *commit = (struct commit *) obj;
2007                 struct commit_list *parents = commit->parents;
2008
2009                 oidset_insert(set, get_commit_tree_oid(commit));
2010                 for (; parents; parents = parents->next)
2011                         oidset_insert(set, &parents->item->object.oid);
2012         } else if (obj->type == OBJ_TAG) {
2013                 struct tag *tag = (struct tag *) obj;
2014                 oidset_insert(set, &tag->tagged->oid);
2015         }
2016         return 0;
2017 }
2018
2019 int is_promisor_object(const struct object_id *oid)
2020 {
2021         static struct oidset promisor_objects;
2022         static int promisor_objects_prepared;
2023
2024         if (!promisor_objects_prepared) {
2025                 if (repository_format_partial_clone) {
2026                         for_each_packed_object(add_promisor_object,
2027                                                &promisor_objects,
2028                                                FOR_EACH_OBJECT_PROMISOR_ONLY);
2029                 }
2030                 promisor_objects_prepared = 1;
2031         }
2032         return oidset_contains(&promisor_objects, oid);
2033 }