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 diff --git a/compat.c b/compat.c
  36 --- a/compat.c
  37 +++ b/compat.c
  38 @@ -41,6 +41,7 @@ extern int checksum_seed;
  39  extern int basis_dir_cnt;
  40  extern int prune_empty_dirs;
  41  extern int protocol_version;
  42 +extern int detect_renamed;
  43  extern int protect_args;
  44  extern int preserve_uid;
  45  extern int preserve_gid;
  46 @@ -120,6 +121,7 @@ void set_allow_inc_recurse(void)
  47                 allow_inc_recurse = 0;
  48         else if (!am_sender
  49          && (delete_before || delete_after
  50 +         || detect_renamed
  51           || delay_updates || prune_empty_dirs))
  52                 allow_inc_recurse = 0;
  53         else if (am_server && !local_server
  54 diff --git a/flist.c b/flist.c
  55 --- a/flist.c
  56 +++ b/flist.c
  57 @@ -63,6 +63,7 @@ extern int non_perishable_cnt;
  58  extern int prune_empty_dirs;
  59  extern int copy_links;
  60  extern int copy_unsafe_links;
  61 +extern int detect_renamed;
  62  extern int protocol_version;
  63  extern int sanitize_paths;
  64  extern int munge_symlinks;
  65 @@ -121,6 +122,8 @@ static int64 tmp_dev, tmp_ino;
  66  #endif
  67  static char tmp_sum[MAX_DIGEST_LEN];
  68
  69 +struct file_list the_fattr_list;
  70 +
  71  static char empty_sum[MAX_DIGEST_LEN];
  72  static int flist_count_offset; /* for --delete --progress */
  73  static int dir_count = 0;
  74 @@ -288,6 +291,45 @@ static int is_excluded(const char *fname, int is_dir, int filter_level)
  75         return 0;
  76  }
  77
  78 +static int fattr_compare(struct file_struct **file1, struct file_struct **file2)
  79 +{
  80 +       struct file_struct *f1 = *file1;
  81 +       struct file_struct *f2 = *file2;
  82 +       int64 len1 = F_LENGTH(f1), len2 = F_LENGTH(f2);
  83 +       int diff;
  84 +
  85 +       if (!f1->basename || !S_ISREG(f1->mode) || !len1) {
  86 +               if (!f2->basename || !S_ISREG(f2->mode) || !len2)
  87 +                       return 0;
  88 +               return 1;
  89 +       }
  90 +       if (!f2->basename || !S_ISREG(f2->mode) || !len2)
  91 +               return -1;
  92 +
  93 +       /* Don't use diff for values that are longer than an int. */
  94 +       if (len1 != len2)
  95 +               return len1 < len2 ? -1 : 1;
  96 +
  97 +       if (always_checksum) {
  98 +               diff = u_memcmp(F_SUM(f1), F_SUM(f2), checksum_len);
  99 +               if (diff)
 100 +                       return diff;
 101 +       } else if (f1->modtime != f2->modtime)
 102 +               return f1->modtime < f2->modtime ? -1 : 1;
 103 +
 104 +       diff = u_strcmp(f1->basename, f2->basename);
 105 +       if (diff)
 106 +               return diff;
 107 +
 108 +       if (f1->dirname == f2->dirname)
 109 +               return 0;
 110 +       if (!f1->dirname)
 111 +               return -1;
 112 +       if (!f2->dirname)
 113 +               return 1;
 114 +       return u_strcmp(f1->dirname, f2->dirname);
 115 +}
 116 +
 117  static void send_directory(int f, struct file_list *flist,
 118                            char *fbuf, int len, int flags);
 119
 120 @@ -2344,6 +2386,25 @@ struct file_list *recv_file_list(int f)
 121
 122         flist_sort_and_clean(flist, relative_paths);
 123
 124 +       if (detect_renamed) {
 125 +               int j = flist->used;
 126 +               the_fattr_list.used = j;
 127 +               the_fattr_list.files = new_array(struct file_struct *, j);
 128 +               if (!the_fattr_list.files)
 129 +                       out_of_memory("recv_file_list");
 130 +               memcpy(the_fattr_list.files, flist->files,
 131 +                      j * sizeof (struct file_struct *));
 132 +               qsort(the_fattr_list.files, j,
 133 +                     sizeof the_fattr_list.files[0], (int (*)())fattr_compare);
 134 +               the_fattr_list.low = 0;
 135 +               while (j-- > 0) {
 136 +                       struct file_struct *fp = the_fattr_list.files[j];
 137 +                       if (fp->basename && S_ISREG(fp->mode) && F_LENGTH(fp))
 138 +                               break;
 139 +               }
 140 +               the_fattr_list.high = j;
 141 +       }
 142 +
 143         if (protocol_version < 30) {
 144                 /* Recv the io_error flag */
 145                 if (ignore_errors)
 146 diff --git a/generator.c b/generator.c
 147 --- a/generator.c
 148 +++ b/generator.c
 149 @@ -80,6 +80,7 @@ extern char *basis_dir[];
 150  extern int compare_dest;
 151  extern int copy_dest;
 152  extern int link_dest;
 153 +extern int detect_renamed;
 154  extern int whole_file;
 155  extern int list_only;
 156  extern int read_batch;
 157 @@ -98,6 +99,7 @@ extern char *backup_suffix;
 158  extern int backup_suffix_len;
 159  extern struct file_list *cur_flist, *first_flist, *dir_flist;
 160  extern struct filter_list_struct daemon_filter_list;
 161 +extern struct file_list the_fattr_list;
 162
 163  int ignore_perishable = 0;
 164  int non_perishable_cnt = 0;
 165 @@ -105,6 +107,7 @@ int maybe_ATTRS_REPORT = 0;
 166
 167  static dev_t dev_zero;
 168  static int deletion_count = 0; /* used to implement --max-delete */
 169 +static int unexplored_dirs = 1;
 170  static int deldelay_size = 0, deldelay_cnt = 0;
 171  static char *deldelay_buf = NULL;
 172  static int deldelay_fd = -1;
 173 @@ -115,7 +118,7 @@ static int need_retouch_dir_times;
 174  static int need_retouch_dir_perms;
 175  static const char *solo_file = NULL;
 176
 177 -/* For calling delete_item() and delete_dir_contents(). */
 178 +/* For calling delete_item(), delete_dir_contents(), and delete_in_dir(). */
 179  #define DEL_NO_UID_WRITE       (1<<0) /* file/dir has our uid w/o write perm */
 180  #define DEL_RECURSE            (1<<1) /* if dir, delete all contents */
 181  #define DEL_DIR_IS_EMPTY       (1<<2) /* internal delete_FUNCTIONS use only */
 182 @@ -124,6 +127,7 @@ static const char *solo_file = NULL;
 183  #define DEL_FOR_SYMLINK        (1<<5) /* making room for a replacement symlink */
 184  #define DEL_FOR_DEVICE         (1<<6) /* making room for a replacement device */
 185  #define DEL_FOR_SPECIAL        (1<<7) /* making room for a replacement special */
 186 +#define DEL_NO_DELETIONS       (1<<9) /* just check for renames w/o deleting */
 187
 188  #define DEL_MAKE_ROOM (DEL_FOR_FILE|DEL_FOR_DIR|DEL_FOR_SYMLINK|DEL_FOR_DEVICE|DEL_FOR_SPECIAL)
 189
 190 @@ -148,11 +152,121 @@ static int is_backup_file(char *fn)
 191         return k > 0 && strcmp(fn+k, backup_suffix) == 0;
 192  }
 193
 194 +/* Search for a regular file that matches either (1) the size & modified
 195 + * time (plus the basename, if possible) or (2) the size & checksum.  If
 196 + * we find an exact match down to the dirname, return -1 because we found
 197 + * an up-to-date file in the transfer, not a renamed file. */
 198 +static int fattr_find(struct file_struct *f, char *fname)
 199 +{
 200 +       int low = the_fattr_list.low, high = the_fattr_list.high;
 201 +       int mid, ok_match = -1, good_match = -1;
 202 +       struct file_struct *fmid;
 203 +       int diff;
 204 +
 205 +       while (low <= high) {
 206 +               mid = (low + high) / 2;
 207 +               fmid = the_fattr_list.files[mid];
 208 +               if (F_LENGTH(fmid) != F_LENGTH(f)) {
 209 +                       if (F_LENGTH(fmid) < F_LENGTH(f))
 210 +                               low = mid + 1;
 211 +                       else
 212 +                               high = mid - 1;
 213 +                       continue;
 214 +               }
 215 +               if (always_checksum) {
 216 +                       /* We use the FLAG_FILE_SENT flag to indicate when we
 217 +                        * have computed the checksum for an entry. */
 218 +                       if (!(f->flags & FLAG_FILE_SENT)) {
 219 +                               if (fmid->modtime == f->modtime
 220 +                                && f_name_cmp(fmid, f) == 0)
 221 +                                       return -1; /* assume we can't help */
 222 +                               file_checksum(fname, F_SUM(f), F_LENGTH(f));
 223 +                               f->flags |= FLAG_FILE_SENT;
 224 +                       }
 225 +                       diff = u_memcmp(F_SUM(fmid), F_SUM(f), checksum_len);
 226 +                       if (diff) {
 227 +                               if (diff < 0)
 228 +                                       low = mid + 1;
 229 +                               else
 230 +                                       high = mid - 1;
 231 +                               continue;
 232 +                       }
 233 +               } else {
 234 +                       if (fmid->modtime != f->modtime) {
 235 +                               if (fmid->modtime < f->modtime)
 236 +                                       low = mid + 1;
 237 +                               else
 238 +                                       high = mid - 1;
 239 +                               continue;
 240 +                       }
 241 +               }
 242 +               ok_match = mid;
 243 +               diff = u_strcmp(fmid->basename, f->basename);
 244 +               if (diff == 0) {
 245 +                       good_match = mid;
 246 +                       if (fmid->dirname == f->dirname)
 247 +                               return -1; /* file is up-to-date */
 248 +                       if (!fmid->dirname) {
 249 +                               low = mid + 1;
 250 +                               continue;
 251 +                       }
 252 +                       if (!f->dirname) {
 253 +                               high = mid - 1;
 254 +                               continue;
 255 +                       }
 256 +                       diff = u_strcmp(fmid->dirname, f->dirname);
 257 +                       if (diff == 0)
 258 +                               return -1; /* file is up-to-date */
 259 +               }
 260 +               if (diff < 0)
 261 +                       low = mid + 1;
 262 +               else
 263 +                       high = mid - 1;
 264 +       }
 265 +
 266 +       return good_match >= 0 ? good_match : ok_match;
 267 +}
 268 +
 269 +static void look_for_rename(struct file_struct *file, char *fname)
 270 +{
 271 +       struct file_struct *fp;
 272 +       char *partialptr, *fn;
 273 +       STRUCT_STAT st;
 274 +       int ndx;
 275 +
 276 +       if (!partial_dir || (ndx = fattr_find(file, fname)) < 0)
 277 +               return;
 278 +
 279 +       fp = the_fattr_list.files[ndx];
 280 +       fn = f_name(fp, NULL);
 281 +       /* We don't provide an alternate-basis file if there is a basis file. */
 282 +       if (link_stat(fn, &st, 0) == 0)
 283 +               return;
 284 +
 285 +       if (!dry_run) {
 286 +               if ((partialptr = partial_dir_fname(fn)) == NULL
 287 +                || !handle_partial_dir(partialptr, PDIR_CREATE))
 288 +                       return;
 289 +               /* We only use the file if we can hard-link it into our tmp dir. */
 290 +               if (link(fname, partialptr) != 0) {
 291 +                       if (errno != EEXIST)
 292 +                               handle_partial_dir(partialptr, PDIR_DELETE);
 293 +                       return;
 294 +               }
 295 +       }
 296 +
 297 +       /* I think this falls into the -vv category with "%s is uptodate", etc. */
 298 +       if (verbose > 1)
 299 +               rprintf(FINFO, "found renamed: %s => %s\n", fname, fn);
 300 +}
 301 +
 302  /* Delete a file or directory.  If DEL_RECURSE is set in the flags, this will
 303   * delete recursively.
 304   *
 305   * Note that fbuf must point to a MAXPATHLEN buffer if the mode indicates it's
 306   * a directory! (The buffer is used for recursion, but returned unchanged.)
 307 + *
 308 + * Also note: --detect-rename may use this routine with DEL_NO_DELETIONS set!
 309   */
 310  static enum delret delete_item(char *fbuf, uint16 mode, uint16 flags)
 311  {
 312 @@ -186,6 +300,8 @@ static enum delret delete_item(char *fbuf, uint16 mode, uint16 flags)
 313                         goto check_ret;
 314                 /* OK: try to delete the directory. */
 315         }
 316 +       if (flags & DEL_NO_DELETIONS)
 317 +               return DR_SUCCESS;
 318
 319         if (!(flags & DEL_MAKE_ROOM) && max_delete >= 0 && ++deletion_count > max_delete)
 320                 return DR_AT_LIMIT;
 321 @@ -241,6 +357,8 @@ static enum delret delete_item(char *fbuf, uint16 mode, uint16 flags)
 322   * its contents, otherwise just checks for content.  Returns DR_SUCCESS or
 323   * DR_NOT_EMPTY.  Note that fname must point to a MAXPATHLEN buffer!  (The
 324   * buffer is used for recursion, but returned unchanged.)
 325 + *
 326 + * Note: --detect-rename may use this routine with DEL_NO_DELETIONS set!
 327   */
 328  static enum delret delete_dir_contents(char *fname, uint16 flags)
 329  {
 330 @@ -260,7 +378,9 @@ static enum delret delete_dir_contents(char *fname, uint16 flags)
 331         save_filters = push_local_filters(fname, dlen);
 332
 333         non_perishable_cnt = 0;
 334 +       file_extra_cnt += SUM_EXTRA_CNT;
 335         dirlist = get_dirlist(fname, dlen, 0);
 336 +       file_extra_cnt -= SUM_EXTRA_CNT;
 337         ret = non_perishable_cnt ? DR_NOT_EMPTY : DR_SUCCESS;
 338
 339         if (!dirlist->used)
 340 @@ -300,7 +420,8 @@ static enum delret delete_dir_contents(char *fname, uint16 flags)
 341                 if (S_ISDIR(fp->mode)) {
 342                         if (delete_dir_contents(fname, flags | DEL_RECURSE) != DR_SUCCESS)
 343                                 ret = DR_NOT_EMPTY;
 344 -               }
 345 +               } else if (detect_renamed && S_ISREG(fp->mode))
 346 +                       look_for_rename(fp, fname);
 347                 if (delete_item(fname, fp->mode, flags) != DR_SUCCESS)
 348                         ret = DR_NOT_EMPTY;
 349         }
 350 @@ -465,13 +586,18 @@ static void do_delayed_deletions(char *delbuf)
 351   * all the --delete-WHEN options.  Note that the fbuf pointer must point to a
 352   * MAXPATHLEN buffer with the name of the directory in it (the functions we
 353   * call will append names onto the end, but the old dir value will be restored
 354 - * on exit). */
 355 -static void delete_in_dir(char *fbuf, struct file_struct *file, dev_t *fs_dev)
 356 + * on exit).
 357 + *
 358 + * Note:  --detect-rename may use this routine with DEL_NO_DELETIONS set!
 359 + */
 360 +static void delete_in_dir(char *fbuf, struct file_struct *file, dev_t *fs_dev,
 361 +                         int del_flags)
 362  {
 363         static int already_warned = 0;
 364         struct file_list *dirlist;
 365 -       char delbuf[MAXPATHLEN];
 366 -       int dlen, i;
 367 +       char *p, delbuf[MAXPATHLEN];
 368 +       unsigned remainder;
 369 +       int dlen, i, restore_dot = 0;
 370         int save_uid_ndx = uid_ndx;
 371
 372         if (!fbuf) {
 373 @@ -486,17 +612,22 @@ static void delete_in_dir(char *fbuf, struct file_struct *file, dev_t *fs_dev)
 374                 maybe_send_keepalive();
 375
 376         if (io_error && !ignore_errors) {
 377 -               if (already_warned)
 378 +               if (!already_warned) {
 379 +                       rprintf(FINFO,
 380 +                           "IO error encountered -- skipping file deletion\n");
 381 +                       already_warned = 1;
 382 +               }
 383 +               if (!detect_renamed)
 384                         return;
 385 -               rprintf(FINFO,
 386 -                       "IO error encountered -- skipping file deletion\n");
 387 -               already_warned = 1;
 388 -               return;
 389 +               del_flags |= DEL_NO_DELETIONS;
 390         }
 391
 392         dlen = strlen(fbuf);
 393         change_local_filter_dir(fbuf, dlen, F_DEPTH(file));
 394
 395 +       if (detect_renamed)
 396 +               unexplored_dirs--;
 397 +
 398         if (one_file_system) {
 399                 if (file->flags & FLAG_TOP_DIR)
 400                         filesystem_dev = *fs_dev;
 401 @@ -509,6 +640,14 @@ static void delete_in_dir(char *fbuf, struct file_struct *file, dev_t *fs_dev)
 402
 403         dirlist = get_dirlist(fbuf, dlen, 0);
 404
 405 +       p = fbuf + dlen;
 406 +       if (dlen == 1 && *fbuf == '.') {
 407 +               restore_dot = 1;
 408 +               p = fbuf;
 409 +       } else if (dlen != 1 || *fbuf != '/')
 410 +               *p++ = '/';
 411 +       remainder = MAXPATHLEN - (p - fbuf);
 412 +
 413         /* If an item in dirlist is not found in flist, delete it
 414          * from the filesystem. */
 415         for (i = dirlist->used; i--; ) {
 416 @@ -521,6 +660,10 @@ static void delete_in_dir(char *fbuf, struct file_struct *file, dev_t *fs_dev)
 417                                         f_name(fp, NULL));
 418                         continue;
 419                 }
 420 +               if (detect_renamed && S_ISREG(fp->mode)) {
 421 +                       strlcpy(p, fp->basename, remainder);
 422 +                       look_for_rename(fp, fbuf);
 423 +               }
 424                 /* Here we want to match regardless of file type.  Replacement
 425                  * of a file with one of another type is handled separately by
 426                  * a delete_item call with a DEL_MAKE_ROOM flag. */
 427 @@ -529,14 +672,19 @@ static void delete_in_dir(char *fbuf, struct file_struct *file, dev_t *fs_dev)
 428                         if (!(fp->mode & S_IWUSR) && !am_root && (uid_t)F_OWNER(fp) == our_uid)
 429                                 flags |= DEL_NO_UID_WRITE;
 430                         f_name(fp, delbuf);
 431 -                       if (delete_during == 2) {
 432 -                               if (!remember_delete(fp, delbuf, flags))
 433 +                       if (delete_during == 2 && !(del_flags & DEL_NO_DELETIONS)) {
 434 +                               if (!remember_delete(fp, delbuf, del_flags | flags))
 435                                         break;
 436                         } else
 437 -                               delete_item(delbuf, fp->mode, flags);
 438 -               }
 439 +                               delete_item(delbuf, fp->mode, del_flags | flags);
 440 +               } else if (detect_renamed && S_ISDIR(fp->mode))
 441 +                       unexplored_dirs++;
 442         }
 443
 444 +       if (restore_dot)
 445 +               fbuf[0] = '.';
 446 +       fbuf[dlen] = '\0';
 447 +
 448         flist_free(dirlist);
 449
 450         if (!save_uid_ndx) {
 451 @@ -574,9 +722,9 @@ static void do_delete_pass(void)
 452                  || !S_ISDIR(st.st_mode))
 453                         continue;
 454
 455 -               delete_in_dir(fbuf, file, &st.st_dev);
 456 +               delete_in_dir(fbuf, file, &st.st_dev, 0);
 457         }
 458 -       delete_in_dir(NULL, NULL, &dev_zero);
 459 +       delete_in_dir(NULL, NULL, &dev_zero, 0);
 460
 461         if (do_progress && !am_server)
 462                 rprintf(FINFO, "                    \r");
 463 @@ -1229,6 +1377,7 @@ static void list_file_entry(struct file_struct *f)
 464         }
 465  }
 466
 467 +static struct bitbag *delayed_bits = NULL;
 468  static int phase = 0;
 469  static int dflt_perms;
 470
 471 @@ -1505,9 +1654,12 @@ static void recv_generator(char *fname, struct file_struct *file, int ndx,
 472                 }
 473                 else if (delete_during && f_out != -1 && !phase
 474                     && !(file->flags & FLAG_MISSING_DIR)) {
 475 -                       if (file->flags & FLAG_CONTENT_DIR)
 476 -                               delete_in_dir(fname, file, &real_sx.st.st_dev);
 477 -                       else
 478 +                       if (file->flags & FLAG_CONTENT_DIR) {
 479 +                               if (detect_renamed && real_ret != 0)
 480 +                                       unexplored_dirs++;
 481 +                               delete_in_dir(fname, file, &real_sx.st.st_dev,
 482 +                                             delete_during < 0 ? DEL_NO_DELETIONS : 0);
 483 +                       } else
 484                                 change_local_filter_dir(fname, strlen(fname), F_DEPTH(file));
 485                 }
 486                 goto cleanup;
 487 @@ -1785,8 +1937,14 @@ static void recv_generator(char *fname, struct file_struct *file, int ndx,
 488                         goto cleanup;
 489                 }
 490  #endif
 491 -               if (stat_errno == ENOENT)
 492 +               if (stat_errno == ENOENT) {
 493 +                       if (detect_renamed && unexplored_dirs > 0
 494 +                        && F_LENGTH(file)) {
 495 +                               bitbag_set_bit(delayed_bits, ndx);
 496 +                               return;
 497 +                       }
 498                         goto notify_others;
 499 +               }
 500                 rsyserr(FERROR_XFER, stat_errno, "recv_generator: failed to stat %s",
 501                         full_fname(fname));
 502                 goto cleanup;
 503 @@ -2187,6 +2345,12 @@ void generate_files(int f_out, const char *local_name)
 504         if (verbose > 2)
 505                 rprintf(FINFO, "generator starting pid=%ld\n", (long)getpid());
 506
 507 +       if (detect_renamed) {
 508 +               delayed_bits = bitbag_create(cur_flist->used);
 509 +               if (!delete_before && !delete_during)
 510 +                       delete_during = -1;
 511 +       }
 512 +
 513         if (delete_before && !solo_file && cur_flist->used > 0)
 514                 do_delete_pass();
 515         if (delete_during == 2) {
 516 @@ -2197,7 +2361,7 @@ void generate_files(int f_out, const char *local_name)
 517         }
 518         do_progress = 0;
 519
 520 -       if (append_mode > 0 || whole_file < 0)
 521 +       if (append_mode > 0 || detect_renamed || whole_file < 0)
 522                 whole_file = 0;
 523         if (verbose >= 2) {
 524                 rprintf(FINFO, "delta-transmission %s\n",
 525 @@ -2239,7 +2403,7 @@ void generate_files(int f_out, const char *local_name)
 526                                                 dirdev = MAKEDEV(DEV_MAJOR(devp), DEV_MINOR(devp));
 527                                         } else
 528                                                 dirdev = MAKEDEV(0, 0);
 529 -                                       delete_in_dir(fbuf, fp, &dirdev);
 530 +                                       delete_in_dir(fbuf, fp, &dirdev, 0);
 531                                 } else
 532                                         change_local_filter_dir(fbuf, strlen(fbuf), F_DEPTH(fp));
 533                         }
 534 @@ -2289,7 +2453,21 @@ void generate_files(int f_out, const char *local_name)
 535                 write_ndx(f_out, NDX_DONE);
 536
 537         if (delete_during)
 538 -               delete_in_dir(NULL, NULL, &dev_zero);
 539 +               delete_in_dir(NULL, NULL, &dev_zero, 0);
 540 +       if (detect_renamed) {
 541 +               if (delete_during < 0)
 542 +                       delete_during = 0;
 543 +               detect_renamed = 0;
 544 +
 545 +               for (i = -1; (i = bitbag_next_bit(delayed_bits, i)) >= 0; ) {
 546 +                       struct file_struct *file = cur_flist->files[i];
 547 +                       if (local_name)
 548 +                               strlcpy(fbuf, local_name, sizeof fbuf);
 549 +                       else
 550 +                               f_name(file, fbuf);
 551 +                       recv_generator(fbuf, file, i, itemizing, code, f_out);
 552 +               }
 553 +       }
 554         phase++;
 555         if (verbose > 2)
 556                 rprintf(FINFO, "generate_files phase=%d\n", phase);
 557 diff --git a/options.c b/options.c
 558 --- a/options.c
 559 +++ b/options.c
 560 @@ -82,6 +82,7 @@ int am_generator = 0;
 561  int am_starting_up = 1;
 562  int relative_paths = -1;
 563  int implied_dirs = 1;
 564 +int detect_renamed = 0;
 565  int numeric_ids = 0;
 566  int allow_8bit_chars = 0;
 567  int force_delete = 0;
 568 @@ -392,6 +393,7 @@ void usage(enum logcode F)
 569    rprintf(F,"     --modify-window=NUM     compare mod-times with reduced accuracy\n");
 570    rprintf(F," -T, --temp-dir=DIR          create temporary files in directory DIR\n");
 571    rprintf(F," -y, --fuzzy                 find similar file for basis if no dest file\n");
 572 +  rprintf(F,"     --detect-renamed        try to find renamed files to speed up the transfer\n");
 573    rprintf(F,"     --compare-dest=DIR      also compare destination files relative to DIR\n");
 574    rprintf(F,"     --copy-dest=DIR         ... and include copies of unchanged files\n");
 575    rprintf(F,"     --link-dest=DIR         hardlink to files in DIR when unchanged\n");
 576 @@ -579,6 +581,7 @@ static struct poptOption long_options[] = {
 577    {"compare-dest",     0,  POPT_ARG_STRING, 0, OPT_COMPARE_DEST, 0, 0 },
 578    {"copy-dest",        0,  POPT_ARG_STRING, 0, OPT_COPY_DEST, 0, 0 },
 579    {"link-dest",        0,  POPT_ARG_STRING, 0, OPT_LINK_DEST, 0, 0 },
 580 +  {"detect-renamed",   0,  POPT_ARG_NONE,   &detect_renamed, 0, 0, 0 },
 581    {"fuzzy",           'y', POPT_ARG_VAL,    &fuzzy_basis, 1, 0, 0 },
 582    {"no-fuzzy",         0,  POPT_ARG_VAL,    &fuzzy_basis, 0, 0, 0 },
 583    {"no-y",             0,  POPT_ARG_VAL,    &fuzzy_basis, 0, 0, 0 },
 584 @@ -1591,7 +1594,7 @@ int parse_arguments(int *argc_p, const char ***argv_p)
 585                 inplace = 1;
 586         }
 587
 588 -       if (delay_updates && !partial_dir)
 589 +       if ((delay_updates || detect_renamed) && !partial_dir)
 590                 partial_dir = tmp_partialdir;
 591
 592         if (inplace) {
 593 @@ -1600,6 +1603,7 @@ int parse_arguments(int *argc_p, const char ***argv_p)
 594                         snprintf(err_buf, sizeof err_buf,
 595                                  "--%s cannot be used with --%s\n",
 596                                  append_mode ? "append" : "inplace",
 597 +                                detect_renamed ? "detect-renamed" :
 598                                  delay_updates ? "delay-updates" : "partial-dir");
 599                         return 0;
 600                 }
 601 @@ -1961,6 +1965,8 @@ void server_options(char **args, int *argc_p)
 602                         args[ac++] = "--super";
 603                 if (size_only)
 604                         args[ac++] = "--size-only";
 605 +               if (detect_renamed)
 606 +                       args[ac++] = "--detect-renamed";
 607         } else {
 608                 if (skip_compress) {
 609                         if (asprintf(&arg, "--skip-compress=%s", skip_compress) < 0)
 610 diff --git a/rsync.yo b/rsync.yo
 611 --- a/rsync.yo
 612 +++ b/rsync.yo
 613 @@ -389,6 +389,7 @@ to the detailed description below for a complete description.  verb(
 614       --modify-window=NUM     compare mod-times with reduced accuracy
 615   -T, --temp-dir=DIR          create temporary files in directory DIR
 616   -y, --fuzzy                 find similar file for basis if no dest file
 617 +     --detect-renamed        try to find renamed files to speed the xfer
 618       --compare-dest=DIR      also compare received files relative to DIR
 619       --copy-dest=DIR         ... and include copies of unchanged files
 620       --link-dest=DIR         hardlink to files in DIR when unchanged
 621 @@ -1491,6 +1492,21 @@ Note that the use of the bf(--delete) option might get rid of any potential
 622  fuzzy-match files, so either use bf(--delete-after) or specify some
 623  filename exclusions if you need to prevent this.
 624
 625 +dit(bf(--detect-renamed)) With this option, for each new source file
 626 +(call it em(src/S)), rsync looks for a file em(dest/D) anywhere in the
 627 +destination that passes the quick check with em(src/S).  If such a em(dest/D)
 628 +is found, rsync uses it as an alternate basis for transferring em(S).  The
 629 +idea is that if em(src/S) was renamed from em(src/D) (as opposed to em(src/S)
 630 +passing the quick check with em(dest/D) by coincidence), the delta-transfer
 631 +algorithm will find that all the data matches between em(src/S) and em(dest/D),
 632 +and the transfer will be really fast.
 633 +
 634 +By default, alternate-basis files are hard-linked into a directory named
 635 +".~tmp~" in each file's destination directory, but if you've specified
 636 +the bf(--partial-dir) option, that directory will be used instead.  These
 637 +potential alternate-basis files will be removed as the transfer progresses.
 638 +This option conflicts with bf(--inplace) and bf(--append).
 639 +
 640  dit(bf(--compare-dest=DIR)) This option instructs rsync to use em(DIR) on
 641  the destination machine as an additional hierarchy to compare destination
 642  files against doing transfers (if the files are missing in the destination
 643 diff --git a/util.c b/util.c
 644 --- a/util.c
 645 +++ b/util.c
 646 @@ -1168,6 +1168,32 @@ int handle_partial_dir(const char *fname, int create)
 647         return 1;
 648  }
 649
 650 +/* We need to supply our own strcmp function for file list comparisons
 651 + * to ensure that signed/unsigned usage is consistent between machines. */
 652 +int u_strcmp(const char *p1, const char *p2)
 653 +{
 654 +        for ( ; *p1; p1++, p2++) {
 655 +               if (*p1 != *p2)
 656 +                       break;
 657 +       }
 658 +
 659 +       return (int)*(uchar*)p1 - (int)*(uchar*)p2;
 660 +}
 661 +
 662 +/* We need a memcmp function compares unsigned-byte values. */
 663 +int u_memcmp(const void *p1, const void *p2, size_t len)
 664 +{
 665 +       const uchar *u1 = p1;
 666 +       const uchar *u2 = p2;
 667 +
 668 +       while (len--) {
 669 +               if (*u1 != *u2)
 670 +                       return (int)*u1 - (int)*u2;
 671 +       }
 672 +
 673 +       return 0;
 674 +}
 675 +
 676  /**
 677   * Determine if a symlink points outside the current directory tree.
 678   * This is considered "unsafe" because e.g. when mirroring somebody