inplace.diff

   1 Patch from Mark Curtis to implement the --inplace option.
   2 Improved by Wayne Davison.
   3
   4 Run these commands after applying the patch:
   5
   6     autoconf
   7     autoheader
   8     ./configure
   9     make
  10
  11 Optional:
  12
  13     make rsync.1
  14
  15 --- configure.in        30 Apr 2004 18:03:33 -0000      1.196
  16 +++ configure.in        3 Jul 2004 00:16:50 -0000
  17 @@ -442,7 +442,7 @@ dnl AC_FUNC_MEMCMP
  18  AC_FUNC_UTIME_NULL
  19  AC_FUNC_ALLOCA
  20  AC_CHECK_FUNCS(waitpid wait4 getcwd strdup strerror chown chmod mknod mkfifo \
  21 -    fchmod fstat strchr readlink link utime utimes strftime mtrace \
  22 +    fchmod fstat strchr readlink link utime utimes strftime mtrace ftruncate \
  23      memmove lchown vsnprintf snprintf asprintf setsid glob strpbrk \
  24      strlcat strlcpy strtol mallinfo getgroups setgroups geteuid getegid)
  25
  26 --- match.c     2 Jul 2004 23:35:30 -0000       1.63
  27 +++ match.c     3 Jul 2004 00:16:50 -0000
  28 @@ -23,6 +23,7 @@ extern int verbose;
  29  extern int am_server;
  30  extern int do_progress;
  31  extern int checksum_seed;
  32 +extern int inplace;
  33
  34  typedef unsigned short tag;
  35
  36 @@ -52,9 +53,23 @@ static size_t *tag_table;
  37  #define gettag2(s1,s2) (((s1) + (s2)) & 0xFFFF)
  38  #define gettag(sum) gettag2((sum)&0xFFFF,(sum)>>16)
  39
  40 +static struct sum_buf *current_sums;
  41 +static int current_s2length;
  42 +
  43  static int compare_targets(struct target *t1,struct target *t2)
  44  {
  45 -       return (int)t1->t - (int)t2->t;
  46 +       struct sum_buf *s1, *s2;
  47 +
  48 +       if (t1->t != t2->t)
  49 +               return t1->t < t2->t ? -1 : 1;
  50 +
  51 +       s1 = &current_sums[t1->i];
  52 +       s2 = &current_sums[t2->i];
  53 +
  54 +       if (s1->sum1 != s2->sum1)
  55 +               return s1->sum1 < s2->sum1 ? -1 : 1;
  56 +
  57 +       return memcmp(s1->sum2, s2->sum2, current_s2length);
  58  }
  59
  60
  61 @@ -74,6 +89,8 @@ static void build_hash_table(struct sum_
  62                 targets[i].t = gettag(s->sums[i].sum1);
  63         }
  64
  65 +       current_sums = s->sums;
  66 +       current_s2length = s->s2length;
  67         qsort(targets,s->count,sizeof(targets[0]),(int (*)())compare_targets);
  68
  69         for (i = 0; i < TABLESIZE; i++)
  70 @@ -192,12 +209,24 @@ static void hash_search(int f,struct sum
  71                         unsigned int l;
  72                         size_t i = targets[j].i;
  73
  74 -                       if (sum != s->sums[i].sum1)
  75 +                       if (sum != s->sums[i].sum1) {
  76 +                               if (done_csum2)
  77 +                                       break;
  78                                 continue;
  79 +                       }
  80
  81                         /* also make sure the two blocks are the same length */
  82                         l = MIN((OFF_T)s->blength, len-offset);
  83 -                       if (l != s->sums[i].len)
  84 +                       if (l != s->sums[i].len) {
  85 +                               if (done_csum2)
  86 +                                       break;
  87 +                               continue;
  88 +                       }
  89 +
  90 +                       /* inplace: ensure chunk's offset is either >= our
  91 +                        * offset or that the data didn't move. */
  92 +                       if (inplace && s->sums[i].offset < offset
  93 +                           && s->sums[i].i >= 0)
  94                                 continue;
  95
  96                         if (verbose > 3)
  97 @@ -215,15 +244,39 @@ static void hash_search(int f,struct sum
  98                                 continue;
  99                         }
 100
 101 -                       /* we've found a match, but now check to see
 102 -                        * if last_i can hint at a better match */
 103 +                       /* We've found a match, but now check to see if last_i
 104 +                        * can hint at a better match.  If inplace is enabled,
 105 +                        * the best possible match is one with an identical
 106 +                        * offset, so we prefer that over a last_i+1 match. */
 107 +                       if (inplace) {
 108 +                               for (; j < s->count && targets[j].t == t; j++) {
 109 +                                       size_t i2 = targets[j].i;
 110 +                                       if (s->sums[i2].offset != offset)
 111 +                                               continue;
 112 +                                       if (i2 != i) {
 113 +                                               if (sum != s->sums[i2].sum1)
 114 +                                                       break;
 115 +                                               if (memcmp(sum2, s->sums[i2].sum2,
 116 +                                                          s->s2length) != 0)
 117 +                                                       break;
 118 +                                               i = i2;
 119 +                                       }
 120 +                                       /* Use this as a flag to indicate that
 121 +                                        * this chunk was at the same offset on
 122 +                                        * both the sender and the receiver. */
 123 +                                       s->sums[i].i = -1;
 124 +                                       goto set_last_i;
 125 +                               }
 126 +                       }
 127                         if (i != last_i + 1 && last_i + 1 < s->count
 128 +                           && (!inplace || s->sums[last_i+1].offset >= offset || s->sums[last_i+1].i < 0)
 129                             && sum == s->sums[last_i+1].sum1
 130                             && memcmp(sum2, s->sums[last_i+1].sum2, s->s2length) == 0) {
 131                                 /* we've found an adjacent match - the RLL coder
 132                                  * will be happy */
 133                                 i = last_i + 1;
 134                         }
 135 +                   set_last_i:
 136                         last_i = i;
 137
 138                         matched(f,s,buf,offset,i);
 139 --- options.c   20 Jun 2004 19:47:05 -0000      1.157
 140 +++ options.c   3 Jul 2004 00:16:50 -0000
 141 @@ -94,6 +94,7 @@ int ignore_errors = 0;
 142  int modify_window = 0;
 143  int blocking_io = -1;
 144  int checksum_seed = 0;
 145 +int inplace = 0;
 146  unsigned int block_size = 0;
 147
 148
 149 @@ -149,6 +150,7 @@ char *bind_address;
 150  static void print_rsync_version(enum logcode f)
 151  {
 152         char const *got_socketpair = "no ";
 153 +       char const *have_inplace = "no ";
 154         char const *hardlinks = "no ";
 155         char const *links = "no ";
 156         char const *ipv6 = "no ";
 157 @@ -158,6 +160,10 @@ static void print_rsync_version(enum log
 158         got_socketpair = "";
 159  #endif
 160
 161 +#if HAVE_FTRUNCATE
 162 +       have_inplace = "";
 163 +#endif
 164 +
 165  #if SUPPORT_HARD_LINKS
 166         hardlinks = "";
 167  #endif
 168 @@ -183,8 +189,8 @@ static void print_rsync_version(enum log
 169         /* Note that this field may not have type ino_t.  It depends
 170          * on the complicated interaction between largefile feature
 171          * macros. */
 172 -       rprintf(f, "              %sIPv6, %d-bit system inums, %d-bit internal inums\n",
 173 -               ipv6,
 174 +       rprintf(f, "              %sinplace, %sIPv6, %d-bit system inums, %d-bit internal inums\n",
 175 +               have_inplace, ipv6,
 176                 (int) (sizeof dumstat->st_ino * 8),
 177                 (int) (sizeof (uint64) * 8));
 178  #ifdef MAINTAINER_MODE
 179 @@ -234,6 +240,7 @@ void usage(enum logcode F)
 180    rprintf(F,"     --backup-dir            make backups into this directory\n");
 181    rprintf(F,"     --suffix=SUFFIX         backup suffix (default %s w/o --backup-dir)\n",BACKUP_SUFFIX);
 182    rprintf(F," -u, --update                update only (don't overwrite newer files)\n");
 183 +  rprintf(F,"     --inplace               update the destination file inplace (see man page)\n");
 184    rprintf(F," -K, --keep-dirlinks         treat symlinked dir on receiver as dir\n");
 185    rprintf(F," -l, --links                 copy symlinks as symlinks\n");
 186    rprintf(F," -L, --copy-links            copy the referent of all symlinks\n");
 187 @@ -341,6 +348,7 @@ static struct poptOption long_options[]
 188    {"sparse",          'S', POPT_ARG_NONE,   &sparse_files, 0, 0, 0 },
 189    {"cvs-exclude",     'C', POPT_ARG_NONE,   &cvs_exclude, 0, 0, 0 },
 190    {"update",          'u', POPT_ARG_NONE,   &update_only, 0, 0, 0 },
 191 +  {"inplace",          0,  POPT_ARG_NONE,   &inplace, 0, 0, 0 },
 192    {"keep-dirlinks",   'K', POPT_ARG_NONE,   &keep_dirlinks, 0, 0, 0 },
 193    {"links",           'l', POPT_ARG_NONE,   &preserve_links, 0, 0, 0 },
 194    {"copy-links",      'L', POPT_ARG_NONE,   &copy_links, 0, 0, 0 },
 195 @@ -739,6 +747,18 @@ int parse_arguments(int *argc, const cha
 196                         bwlimit_writemax = 512;
 197         }
 198
 199 +       if (inplace) {
 200 +#if HAVE_FTRUNCATE
 201 +               if (keep_partial)
 202 +                       keep_partial = 0;
 203 +#else
 204 +               snprintf(err_buf, sizeof err_buf,
 205 +                        "inplace is not supported on this %s\n",
 206 +                        am_server ? "server" : "client");
 207 +               return 0;
 208 +#endif
 209 +       }
 210 +
 211         if (files_from) {
 212                 char *colon;
 213                 if (*argc != 2 && !(am_server && am_sender && *argc == 1)) {
 214 @@ -963,6 +983,9 @@ void server_options(char **args,int *arg
 215         if (opt_ignore_existing && am_sender)
 216                 args[ac++] = "--ignore-existing";
 217
 218 +       if (inplace)
 219 +               args[ac++] = "--inplace";
 220 +
 221         if (tmpdir) {
 222                 args[ac++] = "--temp-dir";
 223                 args[ac++] = tmpdir;
 224 --- receiver.c  2 Jul 2004 18:23:57 -0000       1.86
 225 +++ receiver.c  3 Jul 2004 00:16:50 -0000
 226 @@ -48,6 +48,7 @@ extern int ignore_errors;
 227  extern int orig_umask;
 228  extern int keep_partial;
 229  extern int checksum_seed;
 230 +extern int inplace;
 231
 232  static void delete_one(char *fn, int is_dir)
 233  {
 234 @@ -255,16 +256,30 @@ static int receive_data(int f_in,struct
 235                         sum_update(map,len);
 236                 }
 237
 238 -               if (fd != -1 && write_file(fd, map, len) != (int)len) {
 239 -                       rsyserr(FERROR, errno, "write failed on %s",
 240 -                               full_fname(fname));
 241 -                       exit_cleanup(RERR_FILEIO);
 242 +               if (!inplace || offset != offset2) {
 243 +                       if (fd != -1 && write_file(fd, map, len) != (int)len) {
 244 +                               rsyserr(FERROR, errno, "write failed on %s",
 245 +                                       full_fname(fname));
 246 +                               exit_cleanup(RERR_FILEIO);
 247 +                       }
 248 +               } else {
 249 +                       flush_write_file(fd);
 250 +                       if (do_lseek(fd,(OFF_T)len,SEEK_CUR) != offset+len) {
 251 +                               rprintf(FERROR, "lseek failed on %s: %s, %lli, %lli, %i\n",
 252 +                                       full_fname(fname), strerror(errno), do_lseek(fd,0,SEEK_CUR), (offset+len), i);
 253 +                               exit_cleanup(RERR_FILEIO);
 254 +                       }
 255                 }
 256                 offset += len;
 257         }
 258
 259         flush_write_file(fd);
 260
 261 +#ifdef HAVE_FTRUNCATE
 262 +       if (inplace)
 263 +               ftruncate(fd, offset);
 264 +#endif
 265 +
 266         if (do_progress)
 267                 end_progress(total_size);
 268
 269 @@ -414,44 +429,59 @@ int recv_files(int f_in,struct file_list
 270                 } else
 271                         mapbuf = NULL;
 272
 273 -               if (!get_tmpname(fnametmp,fname)) {
 274 -                       if (mapbuf)
 275 -                               unmap_file(mapbuf);
 276 -                       if (fd1 != -1)
 277 -                               close(fd1);
 278 -                       continue;
 279 -               }
 280 +               /* We now check to see if we are writing file "inplace" */
 281 +               if (inplace)  {
 282 +                       fd2 = do_open(fnamecmp, O_WRONLY|O_CREAT, 0);
 283 +                       if (fd2 == -1) {
 284 +                               rsyserr(FERROR, errno, "open %s failed",
 285 +                                       full_fname(fnamecmp));
 286 +                               receive_data(f_in,mapbuf,-1,NULL,file->length);
 287 +                               if (mapbuf)
 288 +                                       unmap_file(mapbuf);
 289 +                               if (fd1 != -1)
 290 +                                       close(fd1);
 291 +                               continue;
 292 +                       }
 293 +               } else {
 294 +                       if (!get_tmpname(fnametmp,fname)) {
 295 +                               if (mapbuf)
 296 +                                       unmap_file(mapbuf);
 297 +                               if (fd1 != -1)
 298 +                                       close(fd1);
 299 +                               continue;
 300 +                       }
 301
 302 -               strlcpy(template, fnametmp, sizeof template);
 303 +                       strlcpy(template, fnametmp, sizeof template);
 304
 305 -               /* we initially set the perms without the
 306 -                * setuid/setgid bits to ensure that there is no race
 307 -                * condition. They are then correctly updated after
 308 -                * the lchown. Thanks to snabb@epipe.fi for pointing
 309 -                * this out.  We also set it initially without group
 310 -                * access because of a similar race condition. */
 311 -               fd2 = do_mkstemp(fnametmp, file->mode & INITACCESSPERMS);
 312 -
 313 -               /* in most cases parent directories will already exist
 314 -                * because their information should have been previously
 315 -                * transferred, but that may not be the case with -R */
 316 -               if (fd2 == -1 && relative_paths && errno == ENOENT &&
 317 -                   create_directory_path(fnametmp, orig_umask) == 0) {
 318 -                       strlcpy(fnametmp, template, sizeof fnametmp);
 319 +                       /* we initially set the perms without the
 320 +                        * setuid/setgid bits to ensure that there is no race
 321 +                        * condition. They are then correctly updated after
 322 +                        * the lchown. Thanks to snabb@epipe.fi for pointing
 323 +                        * this out.  We also set it initially without group
 324 +                        * access because of a similar race condition. */
 325                         fd2 = do_mkstemp(fnametmp, file->mode & INITACCESSPERMS);
 326 -               }
 327 -               if (fd2 == -1) {
 328 -                       rsyserr(FERROR, errno, "mkstemp %s failed",
 329 -                               full_fname(fnametmp));
 330 -                       receive_data(f_in,mapbuf,-1,NULL,file->length);
 331 -                       if (mapbuf)
 332 -                               unmap_file(mapbuf);
 333 -                       if (fd1 != -1)
 334 -                               close(fd1);
 335 -                       continue;
 336 -               }
 337
 338 -               cleanup_set(fnametmp, fname, file, mapbuf, fd1, fd2);
 339 +                       /* in most cases parent directories will already exist
 340 +                        * because their information should have been previously
 341 +                        * transferred, but that may not be the case with -R */
 342 +                       if (fd2 == -1 && relative_paths && errno == ENOENT
 343 +                           && create_directory_path(fnametmp, orig_umask) == 0) {
 344 +                               strlcpy(fnametmp, template, sizeof fnametmp);
 345 +                               fd2 = do_mkstemp(fnametmp, file->mode & INITACCESSPERMS);
 346 +                       }
 347 +                       if (fd2 == -1) {
 348 +                               rsyserr(FERROR, errno, "mkstemp %s failed",
 349 +                                       full_fname(fnametmp));
 350 +                               receive_data(f_in,mapbuf,-1,NULL,file->length);
 351 +                               if (mapbuf)
 352 +                                       unmap_file(mapbuf);
 353 +                               if (fd1 != -1)
 354 +                                       close(fd1);
 355 +                               continue;
 356 +                       }
 357 +
 358 +                       cleanup_set(fnametmp, fname, file, mapbuf, fd1, fd2);
 359 +               }
 360
 361                 if (!am_server && verbose)
 362                         rprintf(FINFO, "%s\n", fname);
 363 --- rsync.c     2 Jul 2004 18:13:53 -0000       1.142
 364 +++ rsync.c     3 Jul 2004 00:16:51 -0000
 365 @@ -34,6 +34,7 @@ extern int force_delete;
 366  extern int recurse;
 367  extern int make_backups;
 368  extern char *backup_dir;
 369 +extern int inplace;
 370
 371
 372  /*
 373 @@ -239,6 +240,13 @@ void finish_transfer(char *fname, char *
 374         if (make_backups && !make_backup(fname))
 375                 return;
 376
 377 +       if (inplace) {
 378 +               if (verbose > 2)
 379 +                       rprintf(FINFO, "finishing %s\n", fname);
 380 +               set_perms(fname, file, NULL, 0);
 381 +               return;
 382 +       }
 383 +
 384         /* move tmp file over real file */
 385         if (verbose > 2)
 386                 rprintf(FINFO, "renaming %s to %s\n", fnametmp, fname);
 387 --- rsync.yo    5 Jun 2004 16:16:30 -0000       1.171
 388 +++ rsync.yo    3 Jul 2004 00:16:52 -0000
 389 @@ -289,6 +289,7 @@ verb(
 390       --backup-dir            make backups into this directory
 391       --suffix=SUFFIX         backup suffix (default ~ w/o --backup-dir)
 392   -u, --update                update only (don't overwrite newer files)
 393 +     --inplace               update the destination file inplace
 394   -K, --keep-dirlinks         treat symlinked dir on receiver as dir
 395   -l, --links                 copy symlinks as symlinks
 396   -L, --copy-links            copy the referent of all symlinks
 397 @@ -484,6 +485,17 @@ dit(bf(-K, --keep-dirlinks)) On the rece
 398  pointing to a directory, it will be treated as matching a directory
 399  from the sender.
 400
 401 +dit(bf(--inplace)) This causes rsync not to create a new copy of the file
 402 +and then move it into place.  Instead rsync will overwrite the existing
 403 +file, meaning that the rsync algorithm can't extract the full ammount of
 404 +network reduction it might otherwise.
 405 +
 406 +This option is useful for transfer of large files with block based changes
 407 +and also on systems that are disk bound not network bound.
 408 +
 409 +WARNING: If the transfer is interrupted, you will have an inconsistent file
 410 +and the transfer should be run again.
 411 +
 412  dit(bf(-l, --links)) When symlinks are encountered, recreate the
 413  symlink on the destination.
 414