detect-renamed.diff

   1 This patch adds the --detect-renamed option which makes rsync notice files
   2 that either (1) match in size & modify-time (plus the basename, if possible)
   3 or (2) match in size & checksum (when --checksum was also specified) and use
   4 each match as an alternate basis file to speed up the transfer.
   5
   6 The algorithm attempts to scan the receiving-side's files in an efficient
   7 manner.  If --delete[-before] is enabled, we'll take advantage of the
   8 pre-transfer delete pass to prepare any alternate-basis-file matches we
   9 might find.  If --delete-before is not enabled, rsync does the rename scan
  10 during the regular file-sending scan (scanning each directory right before
  11 the generator starts updating files from that dir).  In this latter mode,
  12 rsync might delay the updating of a file (if no alternate-basis match was
  13 yet found) until the full scan of the receiving side is complete, at which
  14 point any delayed files are processed.
  15
  16 I chose to hard-link the alternate-basis files into a ".~tmp~" subdir that
  17 takes advantage of rsync's pre-existing partial-dir logic.  This uses less
  18 memory than trying to keep track of the matches internally, and also allows
  19 any deletions or file-updates to occur normally without interfering with
  20 these alternate-basis discoveries.
  21
  22 To use this patch, run these commands for a successful build:
  23
  24     patch -p1 <patches/detect-renamed.diff
  25     ./configure                                 (optional if already run)
  26     make
  27
  28 TODO:
  29
  30   We need to never return a match from fattr_find() that has a basis
  31   file.  This will ensure that we don't try to give a renamed file to
  32   a file that can't use it, while missing out on giving it to a file
  33   that could use it.
  34
  35 --- old/flist.c
  36 +++ new/flist.c
  37 @@ -56,6 +56,7 @@ extern int non_perishable_cnt;
  38  extern int prune_empty_dirs;
  39  extern int copy_links;
  40  extern int copy_unsafe_links;
  41 +extern int detect_renamed;
  42  extern int protocol_version;
  43  extern int sanitize_paths;
  44  extern struct stats stats;
  45 @@ -86,6 +87,8 @@ static int64 tmp_dev, tmp_ino;
  46  #endif
  47  static char tmp_sum[MD4_SUM_LENGTH];
  48
  49 +struct file_list the_fattr_list;
  50 +
  51  static char empty_sum[MD4_SUM_LENGTH];
  52  static int flist_count_offset; /* for --delete --progress */
  53
  54 @@ -268,6 +271,45 @@ static mode_t from_wire_mode(int mode)
  55         return mode;
  56  }
  57
  58 +static int fattr_compare(struct file_struct **file1, struct file_struct **file2)
  59 +{
  60 +       struct file_struct *f1 = *file1;
  61 +       struct file_struct *f2 = *file2;
  62 +       int64 len1 = F_LENGTH(f1), len2 = F_LENGTH(f2);
  63 +       int diff;
  64 +
  65 +       if (!f1->basename || !S_ISREG(f1->mode) || !len1) {
  66 +               if (!f2->basename || !S_ISREG(f2->mode) || !len2)
  67 +                       return 0;
  68 +               return 1;
  69 +       }
  70 +       if (!f2->basename || !S_ISREG(f2->mode) || !len2)
  71 +               return -1;
  72 +
  73 +       /* Don't use diff for values that are longer than an int. */
  74 +       if (len1 != len2)
  75 +               return len1 < len2 ? -1 : 1;
  76 +
  77 +       if (always_checksum) {
  78 +               diff = u_memcmp(F_SUM(f1), F_SUM(f2), checksum_len);
  79 +               if (diff)
  80 +                       return diff;
  81 +       } else if (f1->modtime != f2->modtime)
  82 +               return f1->modtime < f2->modtime ? -1 : 1;
  83 +
  84 +       diff = u_strcmp(f1->basename, f2->basename);
  85 +       if (diff)
  86 +               return diff;
  87 +
  88 +       if (f1->dirname == f2->dirname)
  89 +               return 0;
  90 +       if (!f1->dirname)
  91 +               return -1;
  92 +       if (!f2->dirname)
  93 +               return 1;
  94 +       return u_strcmp(f1->dirname, f2->dirname);
  95 +}
  96 +
  97  static void send_directory(int f, struct file_list *flist, int ndx,
  98                            char *fbuf, int len, int flags);
  99
 100 @@ -1711,6 +1753,25 @@ struct file_list *recv_file_list(int f)
 101
 102         clean_flist(flist, relative_paths, 1);
 103
 104 +       if (detect_renamed) {
 105 +               int j = flist->count;
 106 +               the_fattr_list.count = j;
 107 +               the_fattr_list.files = new_array(struct file_struct *, j);
 108 +               if (!the_fattr_list.files)
 109 +                       out_of_memory("recv_file_list");
 110 +               memcpy(the_fattr_list.files, flist->files,
 111 +                      j * sizeof (struct file_struct *));
 112 +               qsort(the_fattr_list.files, j,
 113 +                     sizeof the_fattr_list.files[0], (int (*)())fattr_compare);
 114 +               the_fattr_list.low = 0;
 115 +               while (j-- > 0) {
 116 +                       struct file_struct *fp = the_fattr_list.files[j];
 117 +                       if (fp->basename && S_ISREG(fp->mode) && F_LENGTH(fp))
 118 +                               break;
 119 +               }
 120 +               the_fattr_list.high = j;
 121 +       }
 122 +
 123         if (inc_recurse) {
 124                 qsort(dir_flist->files + dstart, dir_flist->count - dstart,
 125                       sizeof dir_flist->files[0], (int (*)())file_compare);
 126 --- old/generator.c
 127 +++ new/generator.c
 128 @@ -79,6 +79,7 @@ extern char *basis_dir[];
 129  extern int compare_dest;
 130  extern int copy_dest;
 131  extern int link_dest;
 132 +extern int detect_renamed;
 133  extern int whole_file;
 134  extern int list_only;
 135  extern int new_root_dir;
 136 @@ -95,6 +96,7 @@ extern char *backup_suffix;
 137  extern int backup_suffix_len;
 138  extern struct file_list *cur_flist, *first_flist, *dir_flist;
 139  extern struct filter_list_struct server_filter_list;
 140 +extern struct file_list the_fattr_list;
 141
 142  int ignore_perishable = 0;
 143  int non_perishable_cnt = 0;
 144 @@ -102,6 +104,7 @@ int maybe_ATTRS_REPORT = 0;
 145
 146  static dev_t dev_zero;
 147  static int deletion_count = 0; /* used to implement --max-delete */
 148 +static int unexplored_dirs = 1;
 149  static int deldelay_size = 0, deldelay_cnt = 0;
 150  static char *deldelay_buf = NULL;
 151  static int deldelay_fd = -1;
 152 @@ -110,7 +113,8 @@ static int dir_tweaking;
 153  static int need_retouch_dir_times;
 154  static const char *solo_file = NULL;
 155
 156 -/* For calling delete_item() and delete_dir_contents(). */
 157 +/* For calling delete_item(), delete_dir_contents(), and delete_in_dir(). */
 158 +#define DEL_NO_DELETIONS       (1<<0)
 159  #define DEL_RECURSE            (1<<1) /* recurse */
 160  #define DEL_DIR_IS_EMPTY       (1<<2) /* internal delete_FUNCTIONS use only */
 161
 162 @@ -132,11 +136,120 @@ static int is_backup_file(char *fn)
 163         return k > 0 && strcmp(fn+k, backup_suffix) == 0;
 164  }
 165
 166 +/* Search for a regular file that matches either (1) the size & modified
 167 + * time (plus the basename, if possible) or (2) the size & checksum.  If
 168 + * we find an exact match down to the dirname, return -1 because we found
 169 + * an up-to-date file in the transfer, not a renamed file. */
 170 +static int fattr_find(struct file_struct *f, char *fname)
 171 +{
 172 +       int low = the_fattr_list.low, high = the_fattr_list.high;
 173 +       int mid, ok_match = -1, good_match = -1;
 174 +       struct file_struct *fmid;
 175 +       int diff;
 176 +
 177 +       while (low <= high) {
 178 +               mid = (low + high) / 2;
 179 +               fmid = the_fattr_list.files[mid];
 180 +               if (F_LENGTH(fmid) != F_LENGTH(f)) {
 181 +                       if (F_LENGTH(fmid) < F_LENGTH(f))
 182 +                               low = mid + 1;
 183 +                       else
 184 +                               high = mid - 1;
 185 +                       continue;
 186 +               }
 187 +               if (always_checksum) {
 188 +                       /* We use the FLAG_SENT flag to indicate when we
 189 +                        * have computed the checksum for an entry. */
 190 +                       if (!(f->flags & FLAG_SENT)) {
 191 +                               if (fmid->modtime == f->modtime
 192 +                                && f_name_cmp(fmid, f) == 0)
 193 +                                       return -1; /* assume we can't help */
 194 +                               file_checksum(fname, (char*)F_SUM(f), F_LENGTH(f));
 195 +                               f->flags |= FLAG_SENT;
 196 +                       }
 197 +                       diff = u_memcmp(F_SUM(fmid), F_SUM(f), checksum_len);
 198 +                       if (diff) {
 199 +                               if (diff < 0)
 200 +                                       low = mid + 1;
 201 +                               else
 202 +                                       high = mid - 1;
 203 +                               continue;
 204 +                       }
 205 +               } else {
 206 +                       if (fmid->modtime != f->modtime) {
 207 +                               if (fmid->modtime < f->modtime)
 208 +                                       low = mid + 1;
 209 +                               else
 210 +                                       high = mid - 1;
 211 +                               continue;
 212 +                       }
 213 +               }
 214 +               ok_match = mid;
 215 +               diff = u_strcmp(fmid->basename, f->basename);
 216 +               if (diff == 0) {
 217 +                       good_match = mid;
 218 +                       if (fmid->dirname == f->dirname)
 219 +                               return -1; /* file is up-to-date */
 220 +                       if (!fmid->dirname) {
 221 +                               low = mid + 1;
 222 +                               continue;
 223 +                       }
 224 +                       if (!f->dirname) {
 225 +                               high = mid - 1;
 226 +                               continue;
 227 +                       }
 228 +                       diff = u_strcmp(fmid->dirname, f->dirname);
 229 +                       if (diff == 0)
 230 +                               return -1; /* file is up-to-date */
 231 +               }
 232 +               if (diff < 0)
 233 +                       low = mid + 1;
 234 +               else
 235 +                       high = mid - 1;
 236 +       }
 237 +
 238 +       return good_match >= 0 ? good_match : ok_match;
 239 +}
 240 +
 241 +static void look_for_rename(struct file_struct *file, char *fname)
 242 +{
 243 +       struct file_struct *fp;
 244 +       char *partialptr, *fn;
 245 +       STRUCT_STAT st;
 246 +       int ndx;
 247 +
 248 +       if ((ndx = fattr_find(file, fname)) < 0)
 249 +               return;
 250 +
 251 +       fp = the_fattr_list.files[ndx];
 252 +       fn = f_name(fp, NULL);
 253 +       /* We don't provide an alternate-basis file if there is a basis file. */
 254 +       if (link_stat(fn, &st, 0) == 0)
 255 +               return;
 256 +       if ((partialptr = partial_dir_fname(fn)) == NULL
 257 +        || !handle_partial_dir(partialptr, PDIR_CREATE))
 258 +               return;
 259 +
 260 +       /* We only use the file if we can hard-link it into our tmp dir. */
 261 +       if (link(fname, partialptr) == 0) {
 262 +               if (verbose > 2) {
 263 +                       rprintf(FINFO, "found renamed: %s => %s\n",
 264 +                               fname, partialptr);
 265 +               }
 266 +               return;
 267 +       }
 268 +
 269 +       if (errno != EEXIST)
 270 +               handle_partial_dir(partialptr, PDIR_DELETE);
 271 +}
 272 +
 273  /* Delete a file or directory.  If DEL_RECURSE is set in the flags, this will
 274   * delete recursively.
 275   *
 276   * Note that fbuf must point to a MAXPATHLEN buffer if the mode indicates it's
 277   * a directory! (The buffer is used for recursion, but returned unchanged.)
 278 + *
 279 + * Also note: --detect-rename may use this routine with DEL_NO_DELETIONS set!
 280   */
 281  static enum delret delete_item(char *fbuf, int mode, char *replace, int flags)
 282  {
 283 @@ -158,6 +271,8 @@ static enum delret delete_item(char *fbu
 284                         goto check_ret;
 285                 /* OK: try to delete the directory. */
 286         }
 287 +       if (flags & DEL_NO_DELETIONS)
 288 +               return DR_SUCCESS;
 289
 290         if (!replace && max_delete >= 0 && ++deletion_count > max_delete)
 291                 return DR_AT_LIMIT;
 292 @@ -204,6 +319,8 @@ static enum delret delete_item(char *fbu
 293   * its contents, otherwise just checks for content.  Returns DR_SUCCESS or
 294   * DR_NOT_EMPTY.  Note that fname must point to a MAXPATHLEN buffer!  (The
 295   * buffer is used for recursion, but returned unchanged.)
 296 + *
 297 + * Note: --detect-rename may use this routine with DEL_NO_DELETIONS set!
 298   */
 299  static enum delret delete_dir_contents(char *fname, int flags)
 300  {
 301 @@ -223,7 +340,9 @@ static enum delret delete_dir_contents(c
 302         save_filters = push_local_filters(fname, dlen);
 303
 304         non_perishable_cnt = 0;
 305 +       file_extra_cnt += SUM_EXTRA_CNT;
 306         dirlist = get_dirlist(fname, dlen, 0);
 307 +       file_extra_cnt -= SUM_EXTRA_CNT;
 308         ret = non_perishable_cnt ? DR_NOT_EMPTY : DR_SUCCESS;
 309
 310         if (!dirlist->count)
 311 @@ -260,6 +379,8 @@ static enum delret delete_dir_contents(c
 312                 if (S_ISDIR(fp->mode)
 313                  && delete_dir_contents(fname, flags | DEL_RECURSE) != DR_SUCCESS)
 314                         ret = DR_NOT_EMPTY;
 315 +               if (detect_renamed && S_ISREG(fp->mode))
 316 +                       look_for_rename(fp, fname);
 317                 if (delete_item(fname, fp->mode, NULL, flags) != DR_SUCCESS)
 318                         ret = DR_NOT_EMPTY;
 319         }
 320 @@ -412,13 +533,17 @@ static void do_delayed_deletions(char *d
 321   * all the --delete-WHEN options.  Note that the fbuf pointer must point to a
 322   * MAXPATHLEN buffer with the name of the directory in it (the functions we
 323   * call will append names onto the end, but the old dir value will be restored
 324 - * on exit). */
 325 + * on exit).
 326 + *
 327 + * Note:  --detect-rename may use this routine with DEL_NO_DELETIONS set!
 328 + */
 329  static void delete_in_dir(struct file_list *flist, char *fbuf,
 330 -                         struct file_struct *file, dev_t *fs_dev)
 331 +                         struct file_struct *file, dev_t *fs_dev, int flags)
 332  {
 333         static int already_warned = 0;
 334         struct file_list *dirlist;
 335 -       char delbuf[MAXPATHLEN];
 336 +       char *p, delbuf[MAXPATHLEN];
 337 +       unsigned remainder;
 338         int dlen, i;
 339
 340         if (!flist) {
 341 @@ -429,21 +554,28 @@ static void delete_in_dir(struct file_li
 342         if (verbose > 2)
 343                 rprintf(FINFO, "delete_in_dir(%s)\n", fbuf);
 344
 345 +       flags |= DEL_RECURSE;
 346 +
 347         if (allowed_lull)
 348                 maybe_send_keepalive();
 349
 350         if (io_error && !ignore_errors) {
 351 -               if (already_warned)
 352 +               if (!already_warned) {
 353 +                       rprintf(FINFO,
 354 +                           "IO error encountered -- skipping file deletion\n");
 355 +                       already_warned = 1;
 356 +               }
 357 +               if (!detect_renamed)
 358                         return;
 359 -               rprintf(FINFO,
 360 -                       "IO error encountered -- skipping file deletion\n");
 361 -               already_warned = 1;
 362 -               return;
 363 +               flags |= DEL_NO_DELETIONS;
 364         }
 365
 366         dlen = strlen(fbuf);
 367         change_local_filter_dir(fbuf, dlen, F_DEPTH(file));
 368
 369 +       if (detect_renamed)
 370 +               unexplored_dirs--;
 371 +
 372         if (one_file_system) {
 373                 if (file->flags & FLAG_TOP_DIR)
 374                         filesystem_dev = *fs_dev;
 375 @@ -453,6 +585,11 @@ static void delete_in_dir(struct file_li
 376
 377         dirlist = get_dirlist(fbuf, dlen, 0);
 378
 379 +       p = fbuf + dlen;
 380 +       if (dlen != 1 || *fbuf != '/')
 381 +               *p++ = '/';
 382 +       remainder = MAXPATHLEN - (p - fbuf);
 383 +
 384         /* If an item in dirlist is not found in flist, delete it
 385          * from the filesystem. */
 386         for (i = dirlist->count; i--; ) {
 387 @@ -465,16 +602,23 @@ static void delete_in_dir(struct file_li
 388                                         f_name(fp, NULL));
 389                         continue;
 390                 }
 391 +               if (detect_renamed && S_ISREG(fp->mode)) {
 392 +                       strlcpy(p, fp->basename, remainder);
 393 +                       look_for_rename(fp, fbuf);
 394 +               }
 395                 if (flist_find(flist, fp) < 0) {
 396                         f_name(fp, delbuf);
 397 -                       if (delete_during == 2) {
 398 +                       if (delete_during == 2 && !(flags & DEL_NO_DELETIONS)) {
 399                                 if (!remember_delete(fp, delbuf))
 400                                         break;
 401                         } else
 402 -                               delete_item(delbuf, fp->mode, NULL, DEL_RECURSE);
 403 -               }
 404 +                               delete_item(delbuf, fp->mode, NULL, flags);
 405 +               } else if (detect_renamed && S_ISDIR(fp->mode))
 406 +                       unexplored_dirs++;
 407         }
 408
 409 +       fbuf[dlen] = '\0';
 410 +
 411         flist_free(dirlist);
 412  }
 413
 414 @@ -504,9 +648,9 @@ static void do_delete_pass(struct file_l
 415                  || !S_ISDIR(st.st_mode))
 416                         continue;
 417
 418 -               delete_in_dir(flist, fbuf, file, &st.st_dev);
 419 +               delete_in_dir(flist, fbuf, file, &st.st_dev, 0);
 420         }
 421 -       delete_in_dir(NULL, NULL, NULL, &dev_zero);
 422 +       delete_in_dir(NULL, NULL, NULL, &dev_zero, 0);
 423
 424         if (do_progress && !am_server)
 425                 rprintf(FINFO, "                    \r");
 426 @@ -1023,6 +1167,7 @@ static int try_dests_non(struct file_str
 427         return j;
 428  }
 429
 430 +static struct bitbag *delayed_bits = NULL;
 431  static int phase = 0;
 432
 433  /* Acts on cur_flist->file's ndx'th item, whose name is fname.  If a dir,
 434 @@ -1214,8 +1359,12 @@ static void recv_generator(char *fname,
 435                         }
 436                 }
 437                 else if (delete_during && f_out != -1 && !phase && dry_run < 2
 438 -                   && (file->flags & FLAG_XFER_DIR))
 439 -                       delete_in_dir(cur_flist, fname, file, &real_st.st_dev);
 440 +                   && (file->flags & FLAG_XFER_DIR)) {
 441 +                       if (detect_renamed && real_ret != 0)
 442 +                               unexplored_dirs++;
 443 +                       delete_in_dir(cur_flist, fname, file, &real_st.st_dev,
 444 +                                     delete_during < 0 ? DEL_NO_DELETIONS : 0);
 445 +               }
 446                 return;
 447         }
 448
 449 @@ -1479,8 +1628,14 @@ static void recv_generator(char *fname,
 450                 if (preserve_hard_links && F_HLINK_NOT_LAST(file))
 451                         return;
 452  #endif
 453 -               if (stat_errno == ENOENT)
 454 +               if (stat_errno == ENOENT) {
 455 +                       if (detect_renamed && unexplored_dirs > 0
 456 +                        && F_LENGTH(file)) {
 457 +                               bitbag_set_bit(delayed_bits, ndx);
 458 +                               return;
 459 +                       }
 460                         goto notify_others;
 461 +               }
 462                 rsyserr(FERROR, stat_errno, "recv_generator: failed to stat %s",
 463                         full_fname(fname));
 464                 return;
 465 @@ -1775,6 +1930,12 @@ void generate_files(int f_out, const cha
 466         if (verbose > 2)
 467                 rprintf(FINFO, "generator starting pid=%ld\n", (long)getpid());
 468
 469 +       if (detect_renamed) {
 470 +               delayed_bits = bitbag_create(flist->count);
 471 +               if (!delete_before && !delete_during)
 472 +                       delete_during = -1;
 473 +       }
 474 +
 475         if (delete_before && !solo_file && cur_flist->count > 0)
 476                 do_delete_pass(cur_flist);
 477         if (delete_during == 2) {
 478 @@ -1785,7 +1946,7 @@ void generate_files(int f_out, const cha
 479         }
 480         do_progress = 0;
 481
 482 -       if (append_mode > 0 || whole_file < 0)
 483 +       if (append_mode > 0 || detect_renamed || whole_file < 0)
 484                 whole_file = 0;
 485         if (verbose >= 2) {
 486                 rprintf(FINFO, "delta-transmission %s\n",
 487 @@ -1810,7 +1971,7 @@ void generate_files(int f_out, const cha
 488                                         dirdev = MAKEDEV(DEV_MAJOR(devp), DEV_MINOR(devp));
 489                                 } else
 490                                         dirdev = MAKEDEV(0, 0);
 491 -                               delete_in_dir(cur_flist, f_name(fp, fbuf), fp, &dirdev);
 492 +                               delete_in_dir(cur_flist, f_name(fp, fbuf), fp, &dirdev, 0);
 493                         }
 494                 }
 495                 for (i = cur_flist->low; i <= cur_flist->high; i++) {
 496 @@ -1863,7 +2024,21 @@ void generate_files(int f_out, const cha
 497         } while ((cur_flist = cur_flist->next) != NULL);
 498
 499         if (delete_during)
 500 -               delete_in_dir(NULL, NULL, NULL, &dev_zero);
 501 +               delete_in_dir(NULL, NULL, NULL, &dev_zero, 0);
 502 +       if (detect_renamed) {
 503 +               if (delete_during < 0)
 504 +                       delete_during = 0;
 505 +               detect_renamed = 0;
 506 +
 507 +               for (i = -1; (i = bitbag_next_bit(delayed_bits, i)) >= 0; ) {
 508 +                       struct file_struct *file = flist->files[i];
 509 +                       if (local_name)
 510 +                               strlcpy(fbuf, local_name, sizeof fbuf);
 511 +                       else
 512 +                               f_name(file, fbuf);
 513 +                       recv_generator(fbuf, file, i, itemizing, code, f_out);
 514 +               }
 515 +       }
 516         phase++;
 517         if (verbose > 2)
 518                 rprintf(FINFO, "generate_files phase=%d\n", phase);
 519 --- old/options.c
 520 +++ new/options.c
 521 @@ -78,6 +78,7 @@ int am_generator = 0;
 522  int am_starting_up = 1;
 523  int relative_paths = -1;
 524  int implied_dirs = 1;
 525 +int detect_renamed = 0;
 526  int numeric_ids = 0;
 527  int allow_8bit_chars = 0;
 528  int force_delete = 0;
 529 @@ -343,6 +344,7 @@ void usage(enum logcode F)
 530    rprintf(F,"     --modify-window=NUM     compare mod-times with reduced accuracy\n");
 531    rprintf(F," -T, --temp-dir=DIR          create temporary files in directory DIR\n");
 532    rprintf(F," -y, --fuzzy                 find similar file for basis if no dest file\n");
 533 +  rprintf(F,"     --detect-renamed        try to find renamed files to speed up the transfer\n");
 534    rprintf(F,"     --compare-dest=DIR      also compare destination files relative to DIR\n");
 535    rprintf(F,"     --copy-dest=DIR         ... and include copies of unchanged files\n");
 536    rprintf(F,"     --link-dest=DIR         hardlink to files in DIR when unchanged\n");
 537 @@ -497,6 +499,7 @@ static struct poptOption long_options[]
 538    {"compare-dest",     0,  POPT_ARG_STRING, 0, OPT_COMPARE_DEST, 0, 0 },
 539    {"copy-dest",        0,  POPT_ARG_STRING, 0, OPT_COPY_DEST, 0, 0 },
 540    {"link-dest",        0,  POPT_ARG_STRING, 0, OPT_LINK_DEST, 0, 0 },
 541 +  {"detect-renamed",   0,  POPT_ARG_NONE,   &detect_renamed, 0, 0, 0 },
 542    {"fuzzy",           'y', POPT_ARG_NONE,   &fuzzy_basis, 0, 0, 0 },
 543    {"compress",        'z', POPT_ARG_NONE,   0, 'z', 0, 0 },
 544    {"compress-level",   0,  POPT_ARG_INT,    &def_compress_level, 'z', 0, 0 },
 545 @@ -1368,7 +1371,7 @@ int parse_arguments(int *argc, const cha
 546                 inplace = 1;
 547         }
 548
 549 -       if (delay_updates && !partial_dir)
 550 +       if ((delay_updates || detect_renamed) && !partial_dir)
 551                 partial_dir = tmp_partialdir;
 552
 553         if (inplace) {
 554 @@ -1377,6 +1380,7 @@ int parse_arguments(int *argc, const cha
 555                         snprintf(err_buf, sizeof err_buf,
 556                                  "--%s cannot be used with --%s\n",
 557                                  append_mode ? "append" : "inplace",
 558 +                                detect_renamed ? "detect-renamed" :
 559                                  delay_updates ? "delay-updates" : "partial-dir");
 560                         return 0;
 561                 }
 562 @@ -1690,6 +1694,8 @@ void server_options(char **args,int *arg
 563                         args[ac++] = "--super";
 564                 if (size_only)
 565                         args[ac++] = "--size-only";
 566 +               if (detect_renamed)
 567 +                       args[ac++] = "--detect-renamed";
 568         }
 569
 570         if (modify_window_set) {
 571 --- old/rsync.yo
 572 +++ new/rsync.yo
 573 @@ -364,6 +364,7 @@ to the detailed description below for a
 574       --modify-window=NUM     compare mod-times with reduced accuracy
 575   -T, --temp-dir=DIR          create temporary files in directory DIR
 576   -y, --fuzzy                 find similar file for basis if no dest file
 577 +     --detect-renamed        try to find renamed files to speed the xfer
 578       --compare-dest=DIR      also compare received files relative to DIR
 579       --copy-dest=DIR         ... and include copies of unchanged files
 580       --link-dest=DIR         hardlink to files in DIR when unchanged
 581 @@ -1296,6 +1297,15 @@ Note that the use of the bf(--delete) op
 582  fuzzy-match files, so either use bf(--delete-after) or specify some
 583  filename exclusions if you need to prevent this.
 584
 585 +dit(bf(--detect-renamed)) This option tells rsync to scan the receiving
 586 +side for files that have been renamed, and to use any that are found as
 587 +alternate basis files to help speed up the transfer.
 588 +By default, alternate-basis files are hard-linked into a directory named
 589 +".~tmp~" in each file's destination directory, but if you've specified
 590 +the bf(--partial-dir) option, that directory will be used instead.  These
 591 +potential alternate-basis files will be removed as the transfer progresses.
 592 +This option conflicts with bf(--inplace) and bf(--append).
 593 +
 594  dit(bf(--compare-dest=DIR)) This option instructs rsync to use em(DIR) on
 595  the destination machine as an additional hierarchy to compare destination
 596  files against doing transfers (if the files are missing in the destination
 597 --- old/util.c
 598 +++ new/util.c
 599 @@ -1027,6 +1027,32 @@ int handle_partial_dir(const char *fname
 600         return 1;
 601  }
 602
 603 +/* We need to supply our own strcmp function for file list comparisons
 604 + * to ensure that signed/unsigned usage is consistent between machines. */
 605 +int u_strcmp(const char *p1, const char *p2)
 606 +{
 607 +        for ( ; *p1; p1++, p2++) {
 608 +               if (*p1 != *p2)
 609 +                       break;
 610 +       }
 611 +
 612 +       return (int)*(uchar*)p1 - (int)*(uchar*)p2;
 613 +}
 614 +
 615 +/* We need a memcmp function compares unsigned-byte values. */
 616 +int u_memcmp(const void *p1, const void *p2, size_t len)
 617 +{
 618 +       const uchar *u1 = p1;
 619 +       const uchar *u2 = p2;
 620 +
 621 +       while (len--) {
 622 +               if (*u1 != *u2)
 623 +                       return (int)*u1 - (int)*u2;
 624 +       }
 625 +
 626 +       return 0;
 627 +}
 628 +
 629  /**
 630   * Determine if a symlink points outside the current directory tree.
 631   * This is considered "unsafe" because e.g. when mirroring somebody