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.
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.
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.
22 To use this patch, run these commands for a successful build:
24 patch -p1 <patches/detect-renamed.diff
25 ./configure (optional if already run)
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
37 @@ -51,6 +51,7 @@ extern int implied_dirs;
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 @@ -68,6 +69,8 @@ int checksum_len;
46 dev_t filesystem_dev; /* used to implement -x */
47 unsigned int file_struct_len;
49 +struct file_list the_fattr_list;
51 static char empty_sum[MD4_SUM_LENGTH];
52 static int flist_count_offset;
54 @@ -250,6 +253,44 @@ static mode_t from_wire_mode(int mode)
58 +static int fattr_compare(struct file_struct **file1, struct file_struct **file2)
60 + struct file_struct *f1 = *file1;
61 + struct file_struct *f2 = *file2;
64 + if (!f1->basename || !S_ISREG(f1->mode) || !f1->length) {
65 + if (!f2->basename || !S_ISREG(f2->mode) || !f2->length)
69 + if (!f2->basename || !S_ISREG(f2->mode) || !f2->length)
72 + /* Don't use diff for values that are longer than an int. */
73 + if (f1->length != f2->length)
74 + return f1->length < f2->length ? -1 : 1;
76 + if (always_checksum) {
77 + diff = u_memcmp(f1->u.sum, f2->u.sum, checksum_len);
80 + } else if (f1->modtime != f2->modtime)
81 + return f1->modtime < f2->modtime ? -1 : 1;
83 + diff = u_strcmp(f1->basename, f2->basename);
87 + if (f1->dirname == f2->dirname)
93 + return u_strcmp(f1->dirname, f2->dirname);
96 static void send_directory(int f, struct file_list *flist,
99 @@ -1379,6 +1420,25 @@ struct file_list *recv_file_list(int f)
101 clean_flist(flist, relative_paths, 1);
103 + if (detect_renamed) {
104 + int j = flist->count;
105 + the_fattr_list.count = j;
106 + the_fattr_list.files = new_array(struct file_struct *, j);
107 + if (!the_fattr_list.files)
109 + memcpy(the_fattr_list.files, flist->files,
110 + j * sizeof (struct file_struct *));
111 + qsort(the_fattr_list.files, j,
112 + sizeof the_fattr_list.files[0], (int (*)())fattr_compare);
113 + the_fattr_list.low = 0;
115 + struct file_struct *fp = the_fattr_list.files[j];
116 + if (fp->basename && S_ISREG(fp->mode) && fp->length)
119 + the_fattr_list.high = j;
123 recv_uid_list(f, flist);
127 @@ -76,6 +76,7 @@ extern char *basis_dir[];
128 extern int compare_dest;
129 extern int copy_dest;
130 extern int link_dest;
131 +extern int detect_renamed;
132 extern int whole_file;
133 extern int list_only;
134 extern int new_root_dir;
135 @@ -91,12 +92,15 @@ extern char *backup_dir;
136 extern char *backup_suffix;
137 extern int backup_suffix_len;
138 extern struct file_list *the_file_list;
139 +extern struct file_list the_fattr_list;
140 extern struct filter_list_struct server_filter_list;
142 static int deletion_count = 0; /* used to implement --max-delete */
143 +static int unexplored_dirs = 1;
145 -/* For calling delete_file() */
146 +/* For calling delete_item() and delete_in_dir() */
147 #define DEL_FORCE_RECURSE (1<<1) /* recurse even w/o --force */
148 +#define DEL_NO_DELETIONS (1<<2)
149 #define DEL_TERSE (1<<3)
152 @@ -106,12 +110,120 @@ static int is_backup_file(char *fn)
153 return k > 0 && strcmp(fn+k, backup_suffix) == 0;
156 +/* Search for a regular file that matches either (1) the size & modified
157 + * time (plus the basename, if possible) or (2) the size & checksum. If
158 + * we find an exact match down to the dirname, return -1 because we found
159 + * an up-to-date file in the transfer, not a renamed file. */
160 +static int fattr_find(struct file_struct *f, char *fname, alloc_pool_t pool)
162 + int low = the_fattr_list.low, high = the_fattr_list.high;
163 + int mid, ok_match = -1, good_match = -1;
164 + struct file_struct *fmid;
167 + while (low <= high) {
168 + mid = (low + high) / 2;
169 + fmid = the_fattr_list.files[mid];
170 + if (fmid->length != f->length) {
171 + if (fmid->length < f->length)
177 + if (always_checksum) {
179 + if (fmid->modtime == f->modtime
180 + && f_name_cmp(fmid, f) == 0)
181 + return -1; /* assume we can't help */
182 + f->u.sum = pool_alloc(pool, MD4_SUM_LENGTH,
184 + file_checksum(fname, f->u.sum, f->length);
186 + diff = u_memcmp(fmid->u.sum, f->u.sum, checksum_len);
195 + if (fmid->modtime != f->modtime) {
196 + if (fmid->modtime < f->modtime)
204 + diff = u_strcmp(fmid->basename, f->basename);
207 + if (fmid->dirname == f->dirname)
208 + return -1; /* file is up-to-date */
209 + if (!fmid->dirname) {
217 + diff = u_strcmp(fmid->dirname, f->dirname);
219 + return -1; /* file is up-to-date */
227 + return good_match >= 0 ? good_match : ok_match;
230 +static void look_for_rename(struct file_struct *file, char *fname,
233 + struct file_struct *fp;
234 + char *partialptr, *fn;
238 + if ((ndx = fattr_find(file, fname, pool)) < 0)
241 + fp = the_fattr_list.files[ndx];
242 + fn = f_name(fp, NULL);
243 + /* We don't provide an alternate-basis file if there is a basis file. */
244 + if (link_stat(fn, &st, 0) == 0)
246 + if ((partialptr = partial_dir_fname(fn)) == NULL
247 + || !handle_partial_dir(partialptr, PDIR_CREATE))
250 + /* We only use the file if we can hard-link it into our tmp dir. */
251 + if (link(fname, partialptr) == 0) {
253 + rprintf(FINFO, "found renamed: %s => %s\n",
254 + fname, partialptr);
259 + if (errno != EEXIST)
260 + handle_partial_dir(partialptr, PDIR_DELETE);
263 /* Delete a file or directory. If DEL_FORCE_RECURSE is set in the flags, or if
264 * force_delete is set, this will delete recursively.
266 * Note that fname must point to a MAXPATHLEN buffer if the mode indicates it's
267 * a directory! (The buffer is used for recursion, but returned unchanged.)
269 + * Also Note: --detect-rename may use this routine with DEL_NO_DELETIONS set!
271 static int delete_item(char *fname, int mode, int flags)
273 @@ -122,6 +234,8 @@ static int delete_item(char *fname, int
276 if (!S_ISDIR(mode)) {
277 + if (flags & DEL_NO_DELETIONS)
279 if (max_delete && ++deletion_count > max_delete)
281 if (make_backups && (backup_dir || !is_backup_file(fname)))
282 @@ -144,6 +258,7 @@ static int delete_item(char *fname, int
284 zap_dir = flags & DEL_FORCE_RECURSE || force_delete;
285 if ((max_delete && ++deletion_count > max_delete)
286 + || flags & DEL_NO_DELETIONS
287 || (dry_run && zap_dir)) {
290 @@ -186,6 +301,8 @@ static int delete_item(char *fname, int
293 strlcpy(p, fp->basename, remainder);
294 + if (detect_renamed && S_ISREG(fp->mode))
295 + look_for_rename(fp, fname, dirlist->file_pool);
296 delete_item(fname, fp->mode, flags & ~DEL_TERSE);
299 @@ -194,7 +311,8 @@ static int delete_item(char *fname, int
301 pop_local_filters(save_filters);
303 - if (max_delete && ++deletion_count > max_delete)
304 + if (flags & DEL_NO_DELETIONS
305 + || (max_delete && ++deletion_count > max_delete))
308 if (do_rmdir(fname) == 0) {
309 @@ -214,15 +332,19 @@ static int delete_item(char *fname, int
310 * all the --delete-WHEN options. Note that the fbuf pointer must point to a
311 * MAXPATHLEN buffer with the name of the directory in it (the functions we
312 * call will append names onto the end, but the old dir value will be restored
316 + * Note: --detect-rename may use this routine with DEL_NO_DELETIONS set!
318 static void delete_in_dir(struct file_list *flist, char *fbuf,
319 - struct file_struct *file, STRUCT_STAT *stp)
320 + struct file_struct *file, STRUCT_STAT *stp, int flags)
322 static int min_depth = MAXPATHLEN, cur_depth = -1;
323 static void *filt_array[MAXPATHLEN/2+1];
324 static int already_warned = 0;
325 struct file_list *dirlist;
326 - char delbuf[MAXPATHLEN];
327 + char *p, delbuf[MAXPATHLEN];
328 + unsigned remainder;
332 @@ -236,6 +358,8 @@ static void delete_in_dir(struct file_li
334 rprintf(FINFO, "delete_in_dir(%s)\n", fbuf);
336 + flags |= DEL_FORCE_RECURSE;
339 maybe_send_keepalive();
341 @@ -243,12 +367,14 @@ static void delete_in_dir(struct file_li
342 return; /* Impossible... */
344 if (io_error && !(lp_ignore_errors(module_id) || ignore_errors)) {
345 - if (already_warned)
346 + if (!already_warned) {
348 + "IO error encountered -- skipping file deletion\n");
349 + already_warned = 1;
351 + if (!detect_renamed)
354 - "IO error encountered -- skipping file deletion\n");
355 - already_warned = 1;
357 + flags |= DEL_NO_DELETIONS;
360 while (cur_depth >= file->dir.depth && cur_depth >= min_depth)
361 @@ -259,6 +385,9 @@ static void delete_in_dir(struct file_li
363 filt_array[cur_depth] = push_local_filters(fbuf, dlen);
365 + if (detect_renamed)
368 if (one_file_system) {
369 if (file->flags & FLAG_TOP_DIR)
370 filesystem_dev = stp->st_dev;
371 @@ -268,18 +397,30 @@ static void delete_in_dir(struct file_li
373 dirlist = get_dirlist(fbuf, dlen, 0);
376 + if (dlen != 1 || *fbuf != '/')
378 + remainder = MAXPATHLEN - (p - fbuf);
380 /* If an item in dirlist is not found in flist, delete it
381 * from the filesystem. */
382 for (i = dirlist->count; i--; ) {
383 struct file_struct *fp = dirlist->files[i];
384 if (!fp->basename || fp->flags & FLAG_MOUNT_POINT)
386 + if (detect_renamed && S_ISREG(fp->mode)) {
387 + strlcpy(p, fp->basename, remainder);
388 + look_for_rename(fp, fbuf, dirlist->file_pool);
390 if (flist_find(flist, fp) < 0) {
392 - delete_item(delbuf, fp->mode, DEL_FORCE_RECURSE);
394 + delete_item(delbuf, fp->mode, flags);
395 + } else if (detect_renamed && S_ISDIR(fp->mode))
404 @@ -309,9 +450,9 @@ static void do_delete_pass(struct file_l
405 || !S_ISDIR(st.st_mode))
408 - delete_in_dir(flist, fbuf, file, &st);
409 + delete_in_dir(flist, fbuf, file, &st, 0);
411 - delete_in_dir(NULL, NULL, NULL, NULL);
412 + delete_in_dir(NULL, NULL, NULL, NULL, 0);
414 if (do_progress && !am_server)
415 rprintf(FINFO, " \r");
416 @@ -771,6 +912,7 @@ static int try_dests_non(struct file_str
420 +static struct bitbag *delayed_bits = NULL;
421 static int phase = 0;
423 /* Acts on the_file_list->file's ndx'th item, whose name is fname. If a dir,
424 @@ -944,8 +1086,12 @@ static void recv_generator(char *fname,
425 && verbose && code != FNONE && f_out != -1)
426 rprintf(code, "%s/\n", fname);
427 if (delete_during && f_out != -1 && !phase && dry_run < 2
428 - && (file->flags & FLAG_DEL_HERE))
429 - delete_in_dir(the_file_list, fname, file, &st);
430 + && (file->flags & FLAG_DEL_HERE)) {
431 + if (detect_renamed && statret != 0)
433 + delete_in_dir(the_file_list, fname, file, &st,
434 + delete_during < 0 ? DEL_NO_DELETIONS : 0);
439 @@ -1201,8 +1347,14 @@ static void recv_generator(char *fname,
440 && hard_link_check(file, ndx, fname, statret, &st,
441 itemizing, code, HL_SKIP))
443 - if (stat_errno == ENOENT)
444 + if (stat_errno == ENOENT) {
445 + if (detect_renamed && unexplored_dirs > 0
447 + bitbag_set_bit(delayed_bits, ndx);
452 rsyserr(FERROR, stat_errno, "recv_generator: failed to stat %s",
455 @@ -1388,11 +1540,17 @@ void generate_files(int f_out, struct fi
456 (long)getpid(), flist->count);
459 + if (detect_renamed) {
460 + delayed_bits = bitbag_create(flist->count);
461 + if (!delete_before && !delete_during)
462 + delete_during = -1;
465 if (delete_before && !local_name && flist->count > 0)
466 do_delete_pass(flist);
469 - if (append_mode || whole_file < 0)
470 + if (append_mode || detect_renamed || whole_file < 0)
473 rprintf(FINFO, "delta-transmission %s\n",
474 @@ -1447,7 +1605,23 @@ void generate_files(int f_out, struct fi
476 recv_generator(NULL, NULL, 0, 0, 0, code, -1);
478 - delete_in_dir(NULL, NULL, NULL, NULL);
479 + delete_in_dir(NULL, NULL, NULL, NULL, 0);
481 + if (detect_renamed) {
482 + if (delete_during < 0)
484 + detect_renamed = 0;
486 + for (i = -1; (i = bitbag_next_bit(delayed_bits, i)) >= 0; ) {
487 + struct file_struct *file = flist->files[i];
489 + strlcpy(fbuf, local_name, sizeof fbuf);
491 + f_name(file, fbuf);
492 + recv_generator(fbuf, file, i, itemizing,
493 + maybe_ATTRS_REPORT, code, f_out);
498 csum_length = SUM_LENGTH;
501 @@ -78,6 +78,7 @@ int am_generator = 0;
502 int am_starting_up = 1;
503 int relative_paths = -1;
504 int implied_dirs = 1;
505 +int detect_renamed = 0;
507 int allow_8bit_chars = 0;
508 int force_delete = 0;
509 @@ -346,6 +347,7 @@ void usage(enum logcode F)
510 rprintf(F," --modify-window=NUM compare mod-times with reduced accuracy\n");
511 rprintf(F," -T, --temp-dir=DIR create temporary files in directory DIR\n");
512 rprintf(F," -y, --fuzzy find similar file for basis if no dest file\n");
513 + rprintf(F," --detect-renamed try to find renamed files to speed up the transfer\n");
514 rprintf(F," --compare-dest=DIR also compare destination files relative to DIR\n");
515 rprintf(F," --copy-dest=DIR ... and include copies of unchanged files\n");
516 rprintf(F," --link-dest=DIR hardlink to files in DIR when unchanged\n");
517 @@ -499,6 +501,7 @@ static struct poptOption long_options[]
518 {"compare-dest", 0, POPT_ARG_STRING, 0, OPT_COMPARE_DEST, 0, 0 },
519 {"copy-dest", 0, POPT_ARG_STRING, 0, OPT_COPY_DEST, 0, 0 },
520 {"link-dest", 0, POPT_ARG_STRING, 0, OPT_LINK_DEST, 0, 0 },
521 + {"detect-renamed", 0, POPT_ARG_NONE, &detect_renamed, 0, 0, 0 },
522 {"fuzzy", 'y', POPT_ARG_NONE, &fuzzy_basis, 0, 0, 0 },
523 {"compress", 'z', POPT_ARG_NONE, 0, 'z', 0, 0 },
524 {"compress-level", 0, POPT_ARG_INT, &def_compress_level, 'z', 0, 0 },
525 @@ -1362,7 +1365,7 @@ int parse_arguments(int *argc, const cha
529 - if (delay_updates && !partial_dir)
530 + if ((delay_updates || detect_renamed) && !partial_dir)
531 partial_dir = tmp_partialdir;
534 @@ -1371,6 +1374,7 @@ int parse_arguments(int *argc, const cha
535 snprintf(err_buf, sizeof err_buf,
536 "--%s cannot be used with --%s\n",
537 append_mode ? "append" : "inplace",
538 + detect_renamed ? "detect-renamed" :
539 delay_updates ? "delay-updates" : "partial-dir");
542 @@ -1674,6 +1678,8 @@ void server_options(char **args,int *arg
543 args[ac++] = "--super";
545 args[ac++] = "--size-only";
546 + if (detect_renamed)
547 + args[ac++] = "--detect-renamed";
550 if (modify_window_set) {
553 @@ -363,6 +363,7 @@ to the detailed description below for a
554 --modify-window=NUM compare mod-times with reduced accuracy
555 -T, --temp-dir=DIR create temporary files in directory DIR
556 -y, --fuzzy find similar file for basis if no dest file
557 + --detect-renamed try to find renamed files to speed the xfer
558 --compare-dest=DIR also compare received files relative to DIR
559 --copy-dest=DIR ... and include copies of unchanged files
560 --link-dest=DIR hardlink to files in DIR when unchanged
561 @@ -1259,6 +1260,15 @@ Note that the use of the bf(--delete) op
562 fuzzy-match files, so either use bf(--delete-after) or specify some
563 filename exclusions if you need to prevent this.
565 +dit(bf(--detect-renamed)) This option tells rsync to scan the receiving
566 +side for files that have been renamed, and to use any that are found as
567 +alternate basis files to help speed up the transfer.
568 +By default, alternate-basis files are hard-linked into a directory named
569 +".~tmp~" in each file's destination directory, but if you've specified
570 +the bf(--partial-dir) option, that directory will be used instead. These
571 +potential alternate-basis files will be removed as the transfer progresses.
572 +This option conflicts with bf(--inplace) and bf(--append).
574 dit(bf(--compare-dest=DIR)) This option instructs rsync to use em(DIR) on
575 the destination machine as an additional hierarchy to compare destination
576 files against doing transfers (if the files are missing in the destination
579 @@ -1025,6 +1025,32 @@ int handle_partial_dir(const char *fname
583 +/* We need to supply our own strcmp function for file list comparisons
584 + * to ensure that signed/unsigned usage is consistent between machines. */
585 +int u_strcmp(const char *p1, const char *p2)
587 + for ( ; *p1; p1++, p2++) {
592 + return (int)*(uchar*)p1 - (int)*(uchar*)p2;
595 +/* We need a memcmp function compares unsigned-byte values. */
596 +int u_memcmp(const void *p1, const void *p2, size_t len)
598 + const uchar *u1 = p1;
599 + const uchar *u2 = p2;
603 + return (int)*u1 - (int)*u2;
610 * Determine if a symlink points outside the current directory tree.
611 * This is considered "unsafe" because e.g. when mirroring somebody