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  * On entry *sha1 contains the pack content SHA1 hash, on exit it is
  42  * the SHA1 hash of sorted object names. The objects array passed in
  43  * will be sorted by SHA1 on exit.
  44  */
  45 const char *write_idx_file(const char *index_name, struct pack_idx_entry **objects,
  46                            int nr_objects, const struct pack_idx_option *opts,
  47                            const unsigned char *sha1)
  48 {
  49         struct hashfile *f;
  50         struct pack_idx_entry **sorted_by_sha, **list, **last;
  51         off_t last_obj_offset = 0;
  52         uint32_t array[256];
  53         int i, fd;
  54         uint32_t index_version;
  55
  56         if (nr_objects) {
  57                 sorted_by_sha = objects;
  58                 list = sorted_by_sha;
  59                 last = sorted_by_sha + nr_objects;
  60                 for (i = 0; i < nr_objects; ++i) {
  61                         if (objects[i]->offset > last_obj_offset)
  62                                 last_obj_offset = objects[i]->offset;
  63                 }
  64                 QSORT(sorted_by_sha, nr_objects, sha1_compare);
  65         }
  66         else
  67                 sorted_by_sha = list = last = NULL;
  68
  69         if (opts->flags & WRITE_IDX_VERIFY) {
  70                 assert(index_name);
  71                 f = hashfd_check(index_name);
  72         } else {
  73                 if (!index_name) {
  74                         struct strbuf tmp_file = STRBUF_INIT;
  75                         fd = odb_mkstemp(&tmp_file, "pack/tmp_idx_XXXXXX");
  76                         index_name = strbuf_detach(&tmp_file, NULL);
  77                 } else {
  78                         unlink(index_name);
  79                         fd = open(index_name, O_CREAT|O_EXCL|O_WRONLY, 0600);
  80                         if (fd < 0)
  81                                 die_errno("unable to create '%s'", index_name);
  82                 }
  83                 f = hashfd(fd, index_name);
  84         }
  85
  86         /* if last object's offset is >= 2^31 we should use index V2 */
  87         index_version = need_large_offset(last_obj_offset, opts) ? 2 : opts->version;
  88
  89         /* index versions 2 and above need a header */
  90         if (index_version >= 2) {
  91                 struct pack_idx_header hdr;
  92                 hdr.idx_signature = htonl(PACK_IDX_SIGNATURE);
  93                 hdr.idx_version = htonl(index_version);
  94                 hashwrite(f, &hdr, sizeof(hdr));
  95         }
  96
  97         /*
  98          * Write the first-level table (the list is sorted,
  99          * but we use a 256-entry lookup to be able to avoid
 100          * having to do eight extra binary search iterations).
 101          */
 102         for (i = 0; i < 256; i++) {
 103                 struct pack_idx_entry **next = list;
 104                 while (next < last) {
 105                         struct pack_idx_entry *obj = *next;
 106                         if (obj->oid.hash[0] != i)
 107                                 break;
 108                         next++;
 109                 }
 110                 array[i] = htonl(next - sorted_by_sha);
 111                 list = next;
 112         }
 113         hashwrite(f, array, 256 * 4);
 114
 115         /*
 116          * Write the actual SHA1 entries..
 117          */
 118         list = sorted_by_sha;
 119         for (i = 0; i < nr_objects; i++) {
 120                 struct pack_idx_entry *obj = *list++;
 121                 if (index_version < 2) {
 122                         uint32_t offset = htonl(obj->offset);
 123                         hashwrite(f, &offset, 4);
 124                 }
 125                 hashwrite(f, obj->oid.hash, the_hash_algo->rawsz);
 126                 if ((opts->flags & WRITE_IDX_STRICT) &&
 127                     (i && !oidcmp(&list[-2]->oid, &obj->oid)))
 128                         die("The same object %s appears twice in the pack",
 129                             oid_to_hex(&obj->oid));
 130         }
 131
 132         if (index_version >= 2) {
 133                 unsigned int nr_large_offset = 0;
 134
 135                 /* write the crc32 table */
 136                 list = sorted_by_sha;
 137                 for (i = 0; i < nr_objects; i++) {
 138                         struct pack_idx_entry *obj = *list++;
 139                         uint32_t crc32_val = htonl(obj->crc32);
 140                         hashwrite(f, &crc32_val, 4);
 141                 }
 142
 143                 /* write the 32-bit offset table */
 144                 list = sorted_by_sha;
 145                 for (i = 0; i < nr_objects; i++) {
 146                         struct pack_idx_entry *obj = *list++;
 147                         uint32_t offset;
 148
 149                         offset = (need_large_offset(obj->offset, opts)
 150                                   ? (0x80000000 | nr_large_offset++)
 151                                   : obj->offset);
 152                         offset = htonl(offset);
 153                         hashwrite(f, &offset, 4);
 154                 }
 155
 156                 /* write the large offset table */
 157                 list = sorted_by_sha;
 158                 while (nr_large_offset) {
 159                         struct pack_idx_entry *obj = *list++;
 160                         uint64_t offset = obj->offset;
 161                         uint32_t split[2];
 162
 163                         if (!need_large_offset(offset, opts))
 164                                 continue;
 165                         split[0] = htonl(offset >> 32);
 166                         split[1] = htonl(offset & 0xffffffff);
 167                         hashwrite(f, split, 8);
 168                         nr_large_offset--;
 169                 }
 170         }
 171
 172         hashwrite(f, sha1, the_hash_algo->rawsz);
 173         hashclose(f, NULL, ((opts->flags & WRITE_IDX_VERIFY)
 174                             ? CSUM_CLOSE : CSUM_FSYNC));
 175         return index_name;
 176 }
 177
 178 off_t write_pack_header(struct hashfile *f, uint32_t nr_entries)
 179 {
 180         struct pack_header hdr;
 181
 182         hdr.hdr_signature = htonl(PACK_SIGNATURE);
 183         hdr.hdr_version = htonl(PACK_VERSION);
 184         hdr.hdr_entries = htonl(nr_entries);
 185         hashwrite(f, &hdr, sizeof(hdr));
 186         return sizeof(hdr);
 187 }
 188
 189 /*
 190  * Update pack header with object_count and compute new SHA1 for pack data
 191  * associated to pack_fd, and write that SHA1 at the end.  That new SHA1
 192  * is also returned in new_pack_sha1.
 193  *
 194  * If partial_pack_sha1 is non null, then the SHA1 of the existing pack
 195  * (without the header update) is computed and validated against the
 196  * one provided in partial_pack_sha1.  The validation is performed at
 197  * partial_pack_offset bytes in the pack file.  The SHA1 of the remaining
 198  * data (i.e. from partial_pack_offset to the end) is then computed and
 199  * returned in partial_pack_sha1.
 200  *
 201  * Note that new_pack_sha1 is updated last, so both new_pack_sha1 and
 202  * partial_pack_sha1 can refer to the same buffer if the caller is not
 203  * interested in the resulting SHA1 of pack data above partial_pack_offset.
 204  */
 205 void fixup_pack_header_footer(int pack_fd,
 206                          unsigned char *new_pack_hash,
 207                          const char *pack_name,
 208                          uint32_t object_count,
 209                          unsigned char *partial_pack_hash,
 210                          off_t partial_pack_offset)
 211 {
 212         int aligned_sz, buf_sz = 8 * 1024;
 213         git_hash_ctx old_hash_ctx, new_hash_ctx;
 214         struct pack_header hdr;
 215         char *buf;
 216         ssize_t read_result;
 217
 218         the_hash_algo->init_fn(&old_hash_ctx);
 219         the_hash_algo->init_fn(&new_hash_ctx);
 220
 221         if (lseek(pack_fd, 0, SEEK_SET) != 0)
 222                 die_errno("Failed seeking to start of '%s'", pack_name);
 223         read_result = read_in_full(pack_fd, &hdr, sizeof(hdr));
 224         if (read_result < 0)
 225                 die_errno("Unable to reread header of '%s'", pack_name);
 226         else if (read_result != sizeof(hdr))
 227                 die_errno("Unexpected short read for header of '%s'",
 228                           pack_name);
 229         if (lseek(pack_fd, 0, SEEK_SET) != 0)
 230                 die_errno("Failed seeking to start of '%s'", pack_name);
 231         the_hash_algo->update_fn(&old_hash_ctx, &hdr, sizeof(hdr));
 232         hdr.hdr_entries = htonl(object_count);
 233         the_hash_algo->update_fn(&new_hash_ctx, &hdr, sizeof(hdr));
 234         write_or_die(pack_fd, &hdr, sizeof(hdr));
 235         partial_pack_offset -= sizeof(hdr);
 236
 237         buf = xmalloc(buf_sz);
 238         aligned_sz = buf_sz - sizeof(hdr);
 239         for (;;) {
 240                 ssize_t m, n;
 241                 m = (partial_pack_hash && partial_pack_offset < aligned_sz) ?
 242                         partial_pack_offset : aligned_sz;
 243                 n = xread(pack_fd, buf, m);
 244                 if (!n)
 245                         break;
 246                 if (n < 0)
 247                         die_errno("Failed to checksum '%s'", pack_name);
 248                 the_hash_algo->update_fn(&new_hash_ctx, buf, n);
 249
 250                 aligned_sz -= n;
 251                 if (!aligned_sz)
 252                         aligned_sz = buf_sz;
 253
 254                 if (!partial_pack_hash)
 255                         continue;
 256
 257                 the_hash_algo->update_fn(&old_hash_ctx, buf, n);
 258                 partial_pack_offset -= n;
 259                 if (partial_pack_offset == 0) {
 260                         unsigned char hash[GIT_MAX_RAWSZ];
 261                         the_hash_algo->final_fn(hash, &old_hash_ctx);
 262                         if (hashcmp(hash, partial_pack_hash) != 0)
 263                                 die("Unexpected checksum for %s "
 264                                     "(disk corruption?)", pack_name);
 265                         /*
 266                          * Now let's compute the SHA1 of the remainder of the
 267                          * pack, which also means making partial_pack_offset
 268                          * big enough not to matter anymore.
 269                          */
 270                         the_hash_algo->init_fn(&old_hash_ctx);
 271                         partial_pack_offset = ~partial_pack_offset;
 272                         partial_pack_offset -= MSB(partial_pack_offset, 1);
 273                 }
 274         }
 275         free(buf);
 276
 277         if (partial_pack_hash)
 278                 the_hash_algo->final_fn(partial_pack_hash, &old_hash_ctx);
 279         the_hash_algo->final_fn(new_pack_hash, &new_hash_ctx);
 280         write_or_die(pack_fd, new_pack_hash, the_hash_algo->rawsz);
 281         fsync_or_die(pack_fd, pack_name);
 282 }
 283
 284 char *index_pack_lockfile(int ip_out)
 285 {
 286         char packname[GIT_MAX_HEXSZ + 6];
 287         const int len = the_hash_algo->hexsz + 6;
 288
 289         /*
 290          * The first thing we expect from index-pack's output
 291          * is "pack\t%40s\n" or "keep\t%40s\n" (46 bytes) where
 292          * %40s is the newly created pack SHA1 name.  In the "keep"
 293          * case, we need it to remove the corresponding .keep file
 294          * later on.  If we don't get that then tough luck with it.
 295          */
 296         if (read_in_full(ip_out, packname, len) == len && packname[len-1] == '\n') {
 297                 const char *name;
 298                 packname[len-1] = 0;
 299                 if (skip_prefix(packname, "keep\t", &name))
 300                         return xstrfmt("%s/pack/pack-%s.keep",
 301                                        get_object_directory(), name);
 302         }
 303         return NULL;
 304 }
 305
 306 /*
 307  * The per-object header is a pretty dense thing, which is
 308  *  - first byte: low four bits are "size", then three bits of "type",
 309  *    and the high bit is "size continues".
 310  *  - each byte afterwards: low seven bits are size continuation,
 311  *    with the high bit being "size continues"
 312  */
 313 int encode_in_pack_object_header(unsigned char *hdr, int hdr_len,
 314                                  enum object_type type, uintmax_t size)
 315 {
 316         int n = 1;
 317         unsigned char c;
 318
 319         if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
 320                 die("bad type %d", type);
 321
 322         c = (type << 4) | (size & 15);
 323         size >>= 4;
 324         while (size) {
 325                 if (n == hdr_len)
 326                         die("object size is too enormous to format");
 327                 *hdr++ = c | 0x80;
 328                 c = size & 0x7f;
 329                 size >>= 7;
 330                 n++;
 331         }
 332         *hdr = c;
 333         return n;
 334 }
 335
 336 struct hashfile *create_tmp_packfile(char **pack_tmp_name)
 337 {
 338         struct strbuf tmpname = STRBUF_INIT;
 339         int fd;
 340
 341         fd = odb_mkstemp(&tmpname, "pack/tmp_pack_XXXXXX");
 342         *pack_tmp_name = strbuf_detach(&tmpname, NULL);
 343         return hashfd(fd, *pack_tmp_name);
 344 }
 345
 346 void finish_tmp_packfile(struct strbuf *name_buffer,
 347                          const char *pack_tmp_name,
 348                          struct pack_idx_entry **written_list,
 349                          uint32_t nr_written,
 350                          struct pack_idx_option *pack_idx_opts,
 351                          unsigned char sha1[])
 352 {
 353         const char *idx_tmp_name;
 354         int basename_len = name_buffer->len;
 355
 356         if (adjust_shared_perm(pack_tmp_name))
 357                 die_errno("unable to make temporary pack file readable");
 358
 359         idx_tmp_name = write_idx_file(NULL, written_list, nr_written,
 360                                       pack_idx_opts, sha1);
 361         if (adjust_shared_perm(idx_tmp_name))
 362                 die_errno("unable to make temporary index file readable");
 363
 364         strbuf_addf(name_buffer, "%s.pack", sha1_to_hex(sha1));
 365
 366         if (rename(pack_tmp_name, name_buffer->buf))
 367                 die_errno("unable to rename temporary pack file");
 368
 369         strbuf_setlen(name_buffer, basename_len);
 370
 371         strbuf_addf(name_buffer, "%s.idx", sha1_to_hex(sha1));
 372         if (rename(idx_tmp_name, name_buffer->buf))
 373                 die_errno("unable to rename temporary index file");
 374
 375         strbuf_setlen(name_buffer, basename_len);
 376
 377         free((void *)idx_tmp_name);
 378 }