git.oblomov.eu Git - git/blob - pack-write.c

   1 #include "cache.h"
   2 #include "pack.h"
   3 #include "csum-file.h"
   4
   5 void reset_pack_idx_option(struct pack_idx_option *opts)
   6 {
   7         memset(opts, 0, sizeof(*opts));
   8         opts->version = 2;
   9         opts->off32_limit = 0x7fffffff;
  10 }
  11
  12 static int sha1_compare(const void *_a, const void *_b)
  13 {
  14         struct pack_idx_entry *a = *(struct pack_idx_entry **)_a;
  15         struct pack_idx_entry *b = *(struct pack_idx_entry **)_b;
  16         return oidcmp(&a->oid, &b->oid);
  17 }
  18
  19 static int cmp_uint32(const void *a_, const void *b_)
  20 {
  21         uint32_t a = *((uint32_t *)a_);
  22         uint32_t b = *((uint32_t *)b_);
  23
  24         return (a < b) ? -1 : (a != b);
  25 }
  26
  27 static int need_large_offset(off_t offset, const struct pack_idx_option *opts)
  28 {
  29         uint32_t ofsval;
  30
  31         if ((offset >> 31) || (opts->off32_limit < offset))
  32                 return 1;
  33         if (!opts->anomaly_nr)
  34                 return 0;
  35         ofsval = offset;
  36         return !!bsearch(&ofsval, opts->anomaly, opts->anomaly_nr,
  37                          sizeof(ofsval), cmp_uint32);
  38 }
  39
  40 /*
  41  * The *sha1 contains the pack content SHA1 hash.
  42  * The objects array passed in will be sorted by SHA1 on exit.
  43  */
  44 const char *write_idx_file(const char *index_name, struct pack_idx_entry **objects,
  45                            int nr_objects, const struct pack_idx_option *opts,
  46                            const unsigned char *sha1)
  47 {
  48         struct hashfile *f;
  49         struct pack_idx_entry **sorted_by_sha, **list, **last;
  50         off_t last_obj_offset = 0;
  51         uint32_t array[256];
  52         int i, fd;
  53         uint32_t index_version;
  54
  55         if (nr_objects) {
  56                 sorted_by_sha = objects;
  57                 list = sorted_by_sha;
  58                 last = sorted_by_sha + nr_objects;
  59                 for (i = 0; i < nr_objects; ++i) {
  60                         if (objects[i]->offset > last_obj_offset)
  61                                 last_obj_offset = objects[i]->offset;
  62                 }
  63                 QSORT(sorted_by_sha, nr_objects, sha1_compare);
  64         }
  65         else
  66                 sorted_by_sha = list = last = NULL;
  67
  68         if (opts->flags & WRITE_IDX_VERIFY) {
  69                 assert(index_name);
  70                 f = hashfd_check(index_name);
  71         } else {
  72                 if (!index_name) {
  73                         struct strbuf tmp_file = STRBUF_INIT;
  74                         fd = odb_mkstemp(&tmp_file, "pack/tmp_idx_XXXXXX");
  75                         index_name = strbuf_detach(&tmp_file, NULL);
  76                 } else {
  77                         unlink(index_name);
  78                         fd = open(index_name, O_CREAT|O_EXCL|O_WRONLY, 0600);
  79                         if (fd < 0)
  80                                 die_errno("unable to create '%s'", index_name);
  81                 }
  82                 f = hashfd(fd, index_name);
  83         }
  84
  85         /* if last object's offset is >= 2^31 we should use index V2 */
  86         index_version = need_large_offset(last_obj_offset, opts) ? 2 : opts->version;
  87
  88         /* index versions 2 and above need a header */
  89         if (index_version >= 2) {
  90                 struct pack_idx_header hdr;
  91                 hdr.idx_signature = htonl(PACK_IDX_SIGNATURE);
  92                 hdr.idx_version = htonl(index_version);
  93                 hashwrite(f, &hdr, sizeof(hdr));
  94         }
  95
  96         /*
  97          * Write the first-level table (the list is sorted,
  98          * but we use a 256-entry lookup to be able to avoid
  99          * having to do eight extra binary search iterations).
 100          */
 101         for (i = 0; i < 256; i++) {
 102                 struct pack_idx_entry **next = list;
 103                 while (next < last) {
 104                         struct pack_idx_entry *obj = *next;
 105                         if (obj->oid.hash[0] != i)
 106                                 break;
 107                         next++;
 108                 }
 109                 array[i] = htonl(next - sorted_by_sha);
 110                 list = next;
 111         }
 112         hashwrite(f, array, 256 * 4);
 113
 114         /*
 115          * Write the actual SHA1 entries..
 116          */
 117         list = sorted_by_sha;
 118         for (i = 0; i < nr_objects; i++) {
 119                 struct pack_idx_entry *obj = *list++;
 120                 if (index_version < 2)
 121                         hashwrite_be32(f, obj->offset);
 122                 hashwrite(f, obj->oid.hash, the_hash_algo->rawsz);
 123                 if ((opts->flags & WRITE_IDX_STRICT) &&
 124                     (i && oideq(&list[-2]->oid, &obj->oid)))
 125                         die("The same object %s appears twice in the pack",
 126                             oid_to_hex(&obj->oid));
 127         }
 128
 129         if (index_version >= 2) {
 130                 unsigned int nr_large_offset = 0;
 131
 132                 /* write the crc32 table */
 133                 list = sorted_by_sha;
 134                 for (i = 0; i < nr_objects; i++) {
 135                         struct pack_idx_entry *obj = *list++;
 136                         hashwrite_be32(f, obj->crc32);
 137                 }
 138
 139                 /* write the 32-bit offset table */
 140                 list = sorted_by_sha;
 141                 for (i = 0; i < nr_objects; i++) {
 142                         struct pack_idx_entry *obj = *list++;
 143                         uint32_t offset;
 144
 145                         offset = (need_large_offset(obj->offset, opts)
 146                                   ? (0x80000000 | nr_large_offset++)
 147                                   : obj->offset);
 148                         hashwrite_be32(f, offset);
 149                 }
 150
 151                 /* write the large offset table */
 152                 list = sorted_by_sha;
 153                 while (nr_large_offset) {
 154                         struct pack_idx_entry *obj = *list++;
 155                         uint64_t offset = obj->offset;
 156                         uint32_t split[2];
 157
 158                         if (!need_large_offset(offset, opts))
 159                                 continue;
 160                         split[0] = htonl(offset >> 32);
 161                         split[1] = htonl(offset & 0xffffffff);
 162                         hashwrite(f, split, 8);
 163                         nr_large_offset--;
 164                 }
 165         }
 166
 167         hashwrite(f, sha1, the_hash_algo->rawsz);
 168         finalize_hashfile(f, NULL, CSUM_HASH_IN_STREAM | CSUM_CLOSE |
 169                                     ((opts->flags & WRITE_IDX_VERIFY)
 170                                     ? 0 : CSUM_FSYNC));
 171         return index_name;
 172 }
 173
 174 off_t write_pack_header(struct hashfile *f, uint32_t nr_entries)
 175 {
 176         struct pack_header hdr;
 177
 178         hdr.hdr_signature = htonl(PACK_SIGNATURE);
 179         hdr.hdr_version = htonl(PACK_VERSION);
 180         hdr.hdr_entries = htonl(nr_entries);
 181         hashwrite(f, &hdr, sizeof(hdr));
 182         return sizeof(hdr);
 183 }
 184
 185 /*
 186  * Update pack header with object_count and compute new SHA1 for pack data
 187  * associated to pack_fd, and write that SHA1 at the end.  That new SHA1
 188  * is also returned in new_pack_sha1.
 189  *
 190  * If partial_pack_sha1 is non null, then the SHA1 of the existing pack
 191  * (without the header update) is computed and validated against the
 192  * one provided in partial_pack_sha1.  The validation is performed at
 193  * partial_pack_offset bytes in the pack file.  The SHA1 of the remaining
 194  * data (i.e. from partial_pack_offset to the end) is then computed and
 195  * returned in partial_pack_sha1.
 196  *
 197  * Note that new_pack_sha1 is updated last, so both new_pack_sha1 and
 198  * partial_pack_sha1 can refer to the same buffer if the caller is not
 199  * interested in the resulting SHA1 of pack data above partial_pack_offset.
 200  */
 201 void fixup_pack_header_footer(int pack_fd,
 202                          unsigned char *new_pack_hash,
 203                          const char *pack_name,
 204                          uint32_t object_count,
 205                          unsigned char *partial_pack_hash,
 206                          off_t partial_pack_offset)
 207 {
 208         int aligned_sz, buf_sz = 8 * 1024;
 209         git_hash_ctx old_hash_ctx, new_hash_ctx;
 210         struct pack_header hdr;
 211         char *buf;
 212         ssize_t read_result;
 213
 214         the_hash_algo->init_fn(&old_hash_ctx);
 215         the_hash_algo->init_fn(&new_hash_ctx);
 216
 217         if (lseek(pack_fd, 0, SEEK_SET) != 0)
 218                 die_errno("Failed seeking to start of '%s'", pack_name);
 219         read_result = read_in_full(pack_fd, &hdr, sizeof(hdr));
 220         if (read_result < 0)
 221                 die_errno("Unable to reread header of '%s'", pack_name);
 222         else if (read_result != sizeof(hdr))
 223                 die_errno("Unexpected short read for header of '%s'",
 224                           pack_name);
 225         if (lseek(pack_fd, 0, SEEK_SET) != 0)
 226                 die_errno("Failed seeking to start of '%s'", pack_name);
 227         the_hash_algo->update_fn(&old_hash_ctx, &hdr, sizeof(hdr));
 228         hdr.hdr_entries = htonl(object_count);
 229         the_hash_algo->update_fn(&new_hash_ctx, &hdr, sizeof(hdr));
 230         write_or_die(pack_fd, &hdr, sizeof(hdr));
 231         partial_pack_offset -= sizeof(hdr);
 232
 233         buf = xmalloc(buf_sz);
 234         aligned_sz = buf_sz - sizeof(hdr);
 235         for (;;) {
 236                 ssize_t m, n;
 237                 m = (partial_pack_hash && partial_pack_offset < aligned_sz) ?
 238                         partial_pack_offset : aligned_sz;
 239                 n = xread(pack_fd, buf, m);
 240                 if (!n)
 241                         break;
 242                 if (n < 0)
 243                         die_errno("Failed to checksum '%s'", pack_name);
 244                 the_hash_algo->update_fn(&new_hash_ctx, buf, n);
 245
 246                 aligned_sz -= n;
 247                 if (!aligned_sz)
 248                         aligned_sz = buf_sz;
 249
 250                 if (!partial_pack_hash)
 251                         continue;
 252
 253                 the_hash_algo->update_fn(&old_hash_ctx, buf, n);
 254                 partial_pack_offset -= n;
 255                 if (partial_pack_offset == 0) {
 256                         unsigned char hash[GIT_MAX_RAWSZ];
 257                         the_hash_algo->final_fn(hash, &old_hash_ctx);
 258                         if (!hasheq(hash, partial_pack_hash))
 259                                 die("Unexpected checksum for %s "
 260                                     "(disk corruption?)", pack_name);
 261                         /*
 262                          * Now let's compute the SHA1 of the remainder of the
 263                          * pack, which also means making partial_pack_offset
 264                          * big enough not to matter anymore.
 265                          */
 266                         the_hash_algo->init_fn(&old_hash_ctx);
 267                         partial_pack_offset = ~partial_pack_offset;
 268                         partial_pack_offset -= MSB(partial_pack_offset, 1);
 269                 }
 270         }
 271         free(buf);
 272
 273         if (partial_pack_hash)
 274                 the_hash_algo->final_fn(partial_pack_hash, &old_hash_ctx);
 275         the_hash_algo->final_fn(new_pack_hash, &new_hash_ctx);
 276         write_or_die(pack_fd, new_pack_hash, the_hash_algo->rawsz);
 277         fsync_or_die(pack_fd, pack_name);
 278 }
 279
 280 char *index_pack_lockfile(int ip_out)
 281 {
 282         char packname[GIT_MAX_HEXSZ + 6];
 283         const int len = the_hash_algo->hexsz + 6;
 284
 285         /*
 286          * The first thing we expect from index-pack's output
 287          * is "pack\t%40s\n" or "keep\t%40s\n" (46 bytes) where
 288          * %40s is the newly created pack SHA1 name.  In the "keep"
 289          * case, we need it to remove the corresponding .keep file
 290          * later on.  If we don't get that then tough luck with it.
 291          */
 292         if (read_in_full(ip_out, packname, len) == len && packname[len-1] == '\n') {
 293                 const char *name;
 294                 packname[len-1] = 0;
 295                 if (skip_prefix(packname, "keep\t", &name))
 296                         return xstrfmt("%s/pack/pack-%s.keep",
 297                                        get_object_directory(), name);
 298         }
 299         return NULL;
 300 }
 301
 302 /*
 303  * The per-object header is a pretty dense thing, which is
 304  *  - first byte: low four bits are "size", then three bits of "type",
 305  *    and the high bit is "size continues".
 306  *  - each byte afterwards: low seven bits are size continuation,
 307  *    with the high bit being "size continues"
 308  */
 309 int encode_in_pack_object_header(unsigned char *hdr, int hdr_len,
 310                                  enum object_type type, uintmax_t size)
 311 {
 312         int n = 1;
 313         unsigned char c;
 314
 315         if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
 316                 die("bad type %d", type);
 317
 318         c = (type << 4) | (size & 15);
 319         size >>= 4;
 320         while (size) {
 321                 if (n == hdr_len)
 322                         die("object size is too enormous to format");
 323                 *hdr++ = c | 0x80;
 324                 c = size & 0x7f;
 325                 size >>= 7;
 326                 n++;
 327         }
 328         *hdr = c;
 329         return n;
 330 }
 331
 332 struct hashfile *create_tmp_packfile(char **pack_tmp_name)
 333 {
 334         struct strbuf tmpname = STRBUF_INIT;
 335         int fd;
 336
 337         fd = odb_mkstemp(&tmpname, "pack/tmp_pack_XXXXXX");
 338         *pack_tmp_name = strbuf_detach(&tmpname, NULL);
 339         return hashfd(fd, *pack_tmp_name);
 340 }
 341
 342 void finish_tmp_packfile(struct strbuf *name_buffer,
 343                          const char *pack_tmp_name,
 344                          struct pack_idx_entry **written_list,
 345                          uint32_t nr_written,
 346                          struct pack_idx_option *pack_idx_opts,
 347                          unsigned char hash[])
 348 {
 349         const char *idx_tmp_name;
 350         int basename_len = name_buffer->len;
 351
 352         if (adjust_shared_perm(pack_tmp_name))
 353                 die_errno("unable to make temporary pack file readable");
 354
 355         idx_tmp_name = write_idx_file(NULL, written_list, nr_written,
 356                                       pack_idx_opts, hash);
 357         if (adjust_shared_perm(idx_tmp_name))
 358                 die_errno("unable to make temporary index file readable");
 359
 360         strbuf_addf(name_buffer, "%s.pack", hash_to_hex(hash));
 361
 362         if (rename(pack_tmp_name, name_buffer->buf))
 363                 die_errno("unable to rename temporary pack file");
 364
 365         strbuf_setlen(name_buffer, basename_len);
 366
 367         strbuf_addf(name_buffer, "%s.idx", hash_to_hex(hash));
 368         if (rename(idx_tmp_name, name_buffer->buf))
 369                 die_errno("unable to rename temporary index file");
 370
 371         strbuf_setlen(name_buffer, basename_len);
 372
 373         free((void *)idx_tmp_name);
 374 }