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 hashcmp(a->sha1, b->sha1);
  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                            unsigned char *sha1)
  48 {
  49         struct sha1file *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         git_SHA_CTX ctx;
  55         uint32_t index_version;
  56
  57         if (nr_objects) {
  58                 sorted_by_sha = objects;
  59                 list = sorted_by_sha;
  60                 last = sorted_by_sha + nr_objects;
  61                 for (i = 0; i < nr_objects; ++i) {
  62                         if (objects[i]->offset > last_obj_offset)
  63                                 last_obj_offset = objects[i]->offset;
  64                 }
  65                 qsort(sorted_by_sha, nr_objects, sizeof(sorted_by_sha[0]),
  66                       sha1_compare);
  67         }
  68         else
  69                 sorted_by_sha = list = last = NULL;
  70
  71         if (opts->flags & WRITE_IDX_VERIFY) {
  72                 assert(index_name);
  73                 f = sha1fd_check(index_name);
  74         } else {
  75                 if (!index_name) {
  76                         static char tmpfile[PATH_MAX];
  77                         fd = odb_mkstemp(tmpfile, sizeof(tmpfile), "pack/tmp_idx_XXXXXX");
  78                         index_name = xstrdup(tmpfile);
  79                 } else {
  80                         unlink(index_name);
  81                         fd = open(index_name, O_CREAT|O_EXCL|O_WRONLY, 0600);
  82                 }
  83                 if (fd < 0)
  84                         die_errno("unable to create '%s'", index_name);
  85                 f = sha1fd(fd, index_name);
  86         }
  87
  88         /* if last object's offset is >= 2^31 we should use index V2 */
  89         index_version = need_large_offset(last_obj_offset, opts) ? 2 : opts->version;
  90
  91         /* index versions 2 and above need a header */
  92         if (index_version >= 2) {
  93                 struct pack_idx_header hdr;
  94                 hdr.idx_signature = htonl(PACK_IDX_SIGNATURE);
  95                 hdr.idx_version = htonl(index_version);
  96                 sha1write(f, &hdr, sizeof(hdr));
  97         }
  98
  99         /*
 100          * Write the first-level table (the list is sorted,
 101          * but we use a 256-entry lookup to be able to avoid
 102          * having to do eight extra binary search iterations).
 103          */
 104         for (i = 0; i < 256; i++) {
 105                 struct pack_idx_entry **next = list;
 106                 while (next < last) {
 107                         struct pack_idx_entry *obj = *next;
 108                         if (obj->sha1[0] != i)
 109                                 break;
 110                         next++;
 111                 }
 112                 array[i] = htonl(next - sorted_by_sha);
 113                 list = next;
 114         }
 115         sha1write(f, array, 256 * 4);
 116
 117         /* compute the SHA1 hash of sorted object names. */
 118         git_SHA1_Init(&ctx);
 119
 120         /*
 121          * Write the actual SHA1 entries..
 122          */
 123         list = sorted_by_sha;
 124         for (i = 0; i < nr_objects; i++) {
 125                 struct pack_idx_entry *obj = *list++;
 126                 if (index_version < 2) {
 127                         uint32_t offset = htonl(obj->offset);
 128                         sha1write(f, &offset, 4);
 129                 }
 130                 sha1write(f, obj->sha1, 20);
 131                 git_SHA1_Update(&ctx, obj->sha1, 20);
 132                 if ((opts->flags & WRITE_IDX_STRICT) &&
 133                     (i && !hashcmp(list[-2]->sha1, obj->sha1)))
 134                         die("The same object %s appears twice in the pack",
 135                             sha1_to_hex(obj->sha1));
 136         }
 137
 138         if (index_version >= 2) {
 139                 unsigned int nr_large_offset = 0;
 140
 141                 /* write the crc32 table */
 142                 list = sorted_by_sha;
 143                 for (i = 0; i < nr_objects; i++) {
 144                         struct pack_idx_entry *obj = *list++;
 145                         uint32_t crc32_val = htonl(obj->crc32);
 146                         sha1write(f, &crc32_val, 4);
 147                 }
 148
 149                 /* write the 32-bit offset table */
 150                 list = sorted_by_sha;
 151                 for (i = 0; i < nr_objects; i++) {
 152                         struct pack_idx_entry *obj = *list++;
 153                         uint32_t offset;
 154
 155                         offset = (need_large_offset(obj->offset, opts)
 156                                   ? (0x80000000 | nr_large_offset++)
 157                                   : obj->offset);
 158                         offset = htonl(offset);
 159                         sha1write(f, &offset, 4);
 160                 }
 161
 162                 /* write the large offset table */
 163                 list = sorted_by_sha;
 164                 while (nr_large_offset) {
 165                         struct pack_idx_entry *obj = *list++;
 166                         uint64_t offset = obj->offset;
 167                         uint32_t split[2];
 168
 169                         if (!need_large_offset(offset, opts))
 170                                 continue;
 171                         split[0] = htonl(offset >> 32);
 172                         split[1] = htonl(offset & 0xffffffff);
 173                         sha1write(f, split, 8);
 174                         nr_large_offset--;
 175                 }
 176         }
 177
 178         sha1write(f, sha1, 20);
 179         sha1close(f, NULL, ((opts->flags & WRITE_IDX_VERIFY)
 180                             ? CSUM_CLOSE : CSUM_FSYNC));
 181         git_SHA1_Final(sha1, &ctx);
 182         return index_name;
 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_sha1,
 203                          const char *pack_name,
 204                          uint32_t object_count,
 205                          unsigned char *partial_pack_sha1,
 206                          off_t partial_pack_offset)
 207 {
 208         int aligned_sz, buf_sz = 8 * 1024;
 209         git_SHA_CTX old_sha1_ctx, new_sha1_ctx;
 210         struct pack_header hdr;
 211         char *buf;
 212
 213         git_SHA1_Init(&old_sha1_ctx);
 214         git_SHA1_Init(&new_sha1_ctx);
 215
 216         if (lseek(pack_fd, 0, SEEK_SET) != 0)
 217                 die_errno("Failed seeking to start of '%s'", pack_name);
 218         if (read_in_full(pack_fd, &hdr, sizeof(hdr)) != sizeof(hdr))
 219                 die_errno("Unable to reread header of '%s'", pack_name);
 220         if (lseek(pack_fd, 0, SEEK_SET) != 0)
 221                 die_errno("Failed seeking to start of '%s'", pack_name);
 222         git_SHA1_Update(&old_sha1_ctx, &hdr, sizeof(hdr));
 223         hdr.hdr_entries = htonl(object_count);
 224         git_SHA1_Update(&new_sha1_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_sha1 && 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                 git_SHA1_Update(&new_sha1_ctx, buf, n);
 240
 241                 aligned_sz -= n;
 242                 if (!aligned_sz)
 243                         aligned_sz = buf_sz;
 244
 245                 if (!partial_pack_sha1)
 246                         continue;
 247
 248                 git_SHA1_Update(&old_sha1_ctx, buf, n);
 249                 partial_pack_offset -= n;
 250                 if (partial_pack_offset == 0) {
 251                         unsigned char sha1[20];
 252                         git_SHA1_Final(sha1, &old_sha1_ctx);
 253                         if (hashcmp(sha1, partial_pack_sha1) != 0)
 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                         git_SHA1_Init(&old_sha1_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_sha1)
 269                 git_SHA1_Final(partial_pack_sha1, &old_sha1_ctx);
 270         git_SHA1_Final(new_pack_sha1, &new_sha1_ctx);
 271         write_or_die(pack_fd, new_pack_sha1, 20);
 272         fsync_or_die(pack_fd, pack_name);
 273 }
 274
 275 char *index_pack_lockfile(int ip_out)
 276 {
 277         char packname[46];
 278
 279         /*
 280          * The first thing we expect from index-pack's output
 281          * is "pack\t%40s\n" or "keep\t%40s\n" (46 bytes) where
 282          * %40s is the newly created pack SHA1 name.  In the "keep"
 283          * case, we need it to remove the corresponding .keep file
 284          * later on.  If we don't get that then tough luck with it.
 285          */
 286         if (read_in_full(ip_out, packname, 46) == 46 && packname[45] == '\n' &&
 287             memcmp(packname, "keep\t", 5) == 0) {
 288                 char path[PATH_MAX];
 289                 packname[45] = 0;
 290                 snprintf(path, sizeof(path), "%s/pack/pack-%s.keep",
 291                          get_object_directory(), packname + 5);
 292                 return xstrdup(path);
 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(enum object_type type, uintmax_t size, unsigned char *hdr)
 305 {
 306         int n = 1;
 307         unsigned char c;
 308
 309         if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
 310                 die("bad type %d", type);
 311
 312         c = (type << 4) | (size & 15);
 313         size >>= 4;
 314         while (size) {
 315                 *hdr++ = c | 0x80;
 316                 c = size & 0x7f;
 317                 size >>= 7;
 318                 n++;
 319         }
 320         *hdr = c;
 321         return n;
 322 }