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         int i, fd;
  52         uint32_t index_version;
  53
  54         if (nr_objects) {
  55                 sorted_by_sha = objects;
  56                 list = sorted_by_sha;
  57                 last = sorted_by_sha + nr_objects;
  58                 for (i = 0; i < nr_objects; ++i) {
  59                         if (objects[i]->offset > last_obj_offset)
  60                                 last_obj_offset = objects[i]->offset;
  61                 }
  62                 QSORT(sorted_by_sha, nr_objects, sha1_compare);
  63         }
  64         else
  65                 sorted_by_sha = list = last = NULL;
  66
  67         if (opts->flags & WRITE_IDX_VERIFY) {
  68                 assert(index_name);
  69                 f = hashfd_check(index_name);
  70         } else {
  71                 if (!index_name) {
  72                         struct strbuf tmp_file = STRBUF_INIT;
  73                         fd = odb_mkstemp(&tmp_file, "pack/tmp_idx_XXXXXX");
  74                         index_name = strbuf_detach(&tmp_file, NULL);
  75                 } else {
  76                         unlink(index_name);
  77                         fd = open(index_name, O_CREAT|O_EXCL|O_WRONLY, 0600);
  78                         if (fd < 0)
  79                                 die_errno("unable to create '%s'", index_name);
  80                 }
  81                 f = hashfd(fd, index_name);
  82         }
  83
  84         /* if last object's offset is >= 2^31 we should use index V2 */
  85         index_version = need_large_offset(last_obj_offset, opts) ? 2 : opts->version;
  86
  87         /* index versions 2 and above need a header */
  88         if (index_version >= 2) {
  89                 struct pack_idx_header hdr;
  90                 hdr.idx_signature = htonl(PACK_IDX_SIGNATURE);
  91                 hdr.idx_version = htonl(index_version);
  92                 hashwrite(f, &hdr, sizeof(hdr));
  93         }
  94
  95         /*
  96          * Write the first-level table (the list is sorted,
  97          * but we use a 256-entry lookup to be able to avoid
  98          * having to do eight extra binary search iterations).
  99          */
 100         for (i = 0; i < 256; i++) {
 101                 struct pack_idx_entry **next = list;
 102                 while (next < last) {
 103                         struct pack_idx_entry *obj = *next;
 104                         if (obj->oid.hash[0] != i)
 105                                 break;
 106                         next++;
 107                 }
 108                 hashwrite_be32(f, next - sorted_by_sha);
 109                 list = next;
 110         }
 111
 112         /*
 113          * Write the actual SHA1 entries..
 114          */
 115         list = sorted_by_sha;
 116         for (i = 0; i < nr_objects; i++) {
 117                 struct pack_idx_entry *obj = *list++;
 118                 if (index_version < 2)
 119                         hashwrite_be32(f, obj->offset);
 120                 hashwrite(f, obj->oid.hash, the_hash_algo->rawsz);
 121                 if ((opts->flags & WRITE_IDX_STRICT) &&
 122                     (i && oideq(&list[-2]->oid, &obj->oid)))
 123                         die("The same object %s appears twice in the pack",
 124                             oid_to_hex(&obj->oid));
 125         }
 126
 127         if (index_version >= 2) {
 128                 unsigned int nr_large_offset = 0;
 129
 130                 /* write the crc32 table */
 131                 list = sorted_by_sha;
 132                 for (i = 0; i < nr_objects; i++) {
 133                         struct pack_idx_entry *obj = *list++;
 134                         hashwrite_be32(f, obj->crc32);
 135                 }
 136
 137                 /* write the 32-bit offset table */
 138                 list = sorted_by_sha;
 139                 for (i = 0; i < nr_objects; i++) {
 140                         struct pack_idx_entry *obj = *list++;
 141                         uint32_t offset;
 142
 143                         offset = (need_large_offset(obj->offset, opts)
 144                                   ? (0x80000000 | nr_large_offset++)
 145                                   : obj->offset);
 146                         hashwrite_be32(f, offset);
 147                 }
 148
 149                 /* write the large offset table */
 150                 list = sorted_by_sha;
 151                 while (nr_large_offset) {
 152                         struct pack_idx_entry *obj = *list++;
 153                         uint64_t offset = obj->offset;
 154
 155                         if (!need_large_offset(offset, opts))
 156                                 continue;
 157                         hashwrite_be64(f, offset);
 158                         nr_large_offset--;
 159                 }
 160         }
 161
 162         hashwrite(f, sha1, the_hash_algo->rawsz);
 163         finalize_hashfile(f, NULL, CSUM_HASH_IN_STREAM | CSUM_CLOSE |
 164                                     ((opts->flags & WRITE_IDX_VERIFY)
 165                                     ? 0 : CSUM_FSYNC));
 166         return index_name;
 167 }
 168
 169 off_t write_pack_header(struct hashfile *f, uint32_t nr_entries)
 170 {
 171         struct pack_header hdr;
 172
 173         hdr.hdr_signature = htonl(PACK_SIGNATURE);
 174         hdr.hdr_version = htonl(PACK_VERSION);
 175         hdr.hdr_entries = htonl(nr_entries);
 176         hashwrite(f, &hdr, sizeof(hdr));
 177         return sizeof(hdr);
 178 }
 179
 180 /*
 181  * Update pack header with object_count and compute new SHA1 for pack data
 182  * associated to pack_fd, and write that SHA1 at the end.  That new SHA1
 183  * is also returned in new_pack_sha1.
 184  *
 185  * If partial_pack_sha1 is non null, then the SHA1 of the existing pack
 186  * (without the header update) is computed and validated against the
 187  * one provided in partial_pack_sha1.  The validation is performed at
 188  * partial_pack_offset bytes in the pack file.  The SHA1 of the remaining
 189  * data (i.e. from partial_pack_offset to the end) is then computed and
 190  * returned in partial_pack_sha1.
 191  *
 192  * Note that new_pack_sha1 is updated last, so both new_pack_sha1 and
 193  * partial_pack_sha1 can refer to the same buffer if the caller is not
 194  * interested in the resulting SHA1 of pack data above partial_pack_offset.
 195  */
 196 void fixup_pack_header_footer(int pack_fd,
 197                          unsigned char *new_pack_hash,
 198                          const char *pack_name,
 199                          uint32_t object_count,
 200                          unsigned char *partial_pack_hash,
 201                          off_t partial_pack_offset)
 202 {
 203         int aligned_sz, buf_sz = 8 * 1024;
 204         git_hash_ctx old_hash_ctx, new_hash_ctx;
 205         struct pack_header hdr;
 206         char *buf;
 207         ssize_t read_result;
 208
 209         the_hash_algo->init_fn(&old_hash_ctx);
 210         the_hash_algo->init_fn(&new_hash_ctx);
 211
 212         if (lseek(pack_fd, 0, SEEK_SET) != 0)
 213                 die_errno("Failed seeking to start of '%s'", pack_name);
 214         read_result = read_in_full(pack_fd, &hdr, sizeof(hdr));
 215         if (read_result < 0)
 216                 die_errno("Unable to reread header of '%s'", pack_name);
 217         else if (read_result != sizeof(hdr))
 218                 die_errno("Unexpected short read for header of '%s'",
 219                           pack_name);
 220         if (lseek(pack_fd, 0, SEEK_SET) != 0)
 221                 die_errno("Failed seeking to start of '%s'", pack_name);
 222         the_hash_algo->update_fn(&old_hash_ctx, &hdr, sizeof(hdr));
 223         hdr.hdr_entries = htonl(object_count);
 224         the_hash_algo->update_fn(&new_hash_ctx, &hdr, sizeof(hdr));
 225         write_or_die(pack_fd, &hdr, sizeof(hdr));
 226         partial_pack_offset -= sizeof(hdr);
 227
 228         buf = xmalloc(buf_sz);
 229         aligned_sz = buf_sz - sizeof(hdr);
 230         for (;;) {
 231                 ssize_t m, n;
 232                 m = (partial_pack_hash && partial_pack_offset < aligned_sz) ?
 233                         partial_pack_offset : aligned_sz;
 234                 n = xread(pack_fd, buf, m);
 235                 if (!n)
 236                         break;
 237                 if (n < 0)
 238                         die_errno("Failed to checksum '%s'", pack_name);
 239                 the_hash_algo->update_fn(&new_hash_ctx, buf, n);
 240
 241                 aligned_sz -= n;
 242                 if (!aligned_sz)
 243                         aligned_sz = buf_sz;
 244
 245                 if (!partial_pack_hash)
 246                         continue;
 247
 248                 the_hash_algo->update_fn(&old_hash_ctx, buf, n);
 249                 partial_pack_offset -= n;
 250                 if (partial_pack_offset == 0) {
 251                         unsigned char hash[GIT_MAX_RAWSZ];
 252                         the_hash_algo->final_fn(hash, &old_hash_ctx);
 253                         if (!hasheq(hash, partial_pack_hash))
 254                                 die("Unexpected checksum for %s "
 255                                     "(disk corruption?)", pack_name);
 256                         /*
 257                          * Now let's compute the SHA1 of the remainder of the
 258                          * pack, which also means making partial_pack_offset
 259                          * big enough not to matter anymore.
 260                          */
 261                         the_hash_algo->init_fn(&old_hash_ctx);
 262                         partial_pack_offset = ~partial_pack_offset;
 263                         partial_pack_offset -= MSB(partial_pack_offset, 1);
 264                 }
 265         }
 266         free(buf);
 267
 268         if (partial_pack_hash)
 269                 the_hash_algo->final_fn(partial_pack_hash, &old_hash_ctx);
 270         the_hash_algo->final_fn(new_pack_hash, &new_hash_ctx);
 271         write_or_die(pack_fd, new_pack_hash, the_hash_algo->rawsz);
 272         fsync_or_die(pack_fd, pack_name);
 273 }
 274
 275 char *index_pack_lockfile(int ip_out)
 276 {
 277         char packname[GIT_MAX_HEXSZ + 6];
 278         const int len = the_hash_algo->hexsz + 6;
 279
 280         /*
 281          * The first thing we expect from index-pack's output
 282          * is "pack\t%40s\n" or "keep\t%40s\n" (46 bytes) where
 283          * %40s is the newly created pack SHA1 name.  In the "keep"
 284          * case, we need it to remove the corresponding .keep file
 285          * later on.  If we don't get that then tough luck with it.
 286          */
 287         if (read_in_full(ip_out, packname, len) == len && packname[len-1] == '\n') {
 288                 const char *name;
 289                 packname[len-1] = 0;
 290                 if (skip_prefix(packname, "keep\t", &name))
 291                         return xstrfmt("%s/pack/pack-%s.keep",
 292                                        get_object_directory(), name);
 293         }
 294         return NULL;
 295 }
 296
 297 /*
 298  * The per-object header is a pretty dense thing, which is
 299  *  - first byte: low four bits are "size", then three bits of "type",
 300  *    and the high bit is "size continues".
 301  *  - each byte afterwards: low seven bits are size continuation,
 302  *    with the high bit being "size continues"
 303  */
 304 int encode_in_pack_object_header(unsigned char *hdr, int hdr_len,
 305                                  enum object_type type, uintmax_t size)
 306 {
 307         int n = 1;
 308         unsigned char c;
 309
 310         if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
 311                 die("bad type %d", type);
 312
 313         c = (type << 4) | (size & 15);
 314         size >>= 4;
 315         while (size) {
 316                 if (n == hdr_len)
 317                         die("object size is too enormous to format");
 318                 *hdr++ = c | 0x80;
 319                 c = size & 0x7f;
 320                 size >>= 7;
 321                 n++;
 322         }
 323         *hdr = c;
 324         return n;
 325 }
 326
 327 struct hashfile *create_tmp_packfile(char **pack_tmp_name)
 328 {
 329         struct strbuf tmpname = STRBUF_INIT;
 330         int fd;
 331
 332         fd = odb_mkstemp(&tmpname, "pack/tmp_pack_XXXXXX");
 333         *pack_tmp_name = strbuf_detach(&tmpname, NULL);
 334         return hashfd(fd, *pack_tmp_name);
 335 }
 336
 337 void finish_tmp_packfile(struct strbuf *name_buffer,
 338                          const char *pack_tmp_name,
 339                          struct pack_idx_entry **written_list,
 340                          uint32_t nr_written,
 341                          struct pack_idx_option *pack_idx_opts,
 342                          unsigned char hash[])
 343 {
 344         const char *idx_tmp_name;
 345         int basename_len = name_buffer->len;
 346
 347         if (adjust_shared_perm(pack_tmp_name))
 348                 die_errno("unable to make temporary pack file readable");
 349
 350         idx_tmp_name = write_idx_file(NULL, written_list, nr_written,
 351                                       pack_idx_opts, hash);
 352         if (adjust_shared_perm(idx_tmp_name))
 353                 die_errno("unable to make temporary index file readable");
 354
 355         strbuf_addf(name_buffer, "%s.pack", hash_to_hex(hash));
 356
 357         if (rename(pack_tmp_name, name_buffer->buf))
 358                 die_errno("unable to rename temporary pack file");
 359
 360         strbuf_setlen(name_buffer, basename_len);
 361
 362         strbuf_addf(name_buffer, "%s.idx", hash_to_hex(hash));
 363         if (rename(idx_tmp_name, name_buffer->buf))
 364                 die_errno("unable to rename temporary index file");
 365
 366         strbuf_setlen(name_buffer, basename_len);
 367
 368         free((void *)idx_tmp_name);
 369 }