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 @@ -55,6 +55,7 @@ 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 @@ -80,6 +81,8 @@ static int64 tmp_dev, tmp_ino;
47 static char tmp_sum[MD4_SUM_LENGTH];
49 +struct file_list the_fattr_list;
51 static char empty_sum[MD4_SUM_LENGTH];
52 static int flist_count_offset; /* for --delete --progress */
54 @@ -262,6 +265,45 @@ 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;
62 + int64 len1 = F_LENGTH(f1), len2 = F_LENGTH(f2);
65 + if (!f1->basename || !S_ISREG(f1->mode) || !len1) {
66 + if (!f2->basename || !S_ISREG(f2->mode) || !len2)
70 + if (!f2->basename || !S_ISREG(f2->mode) || !len2)
73 + /* Don't use diff for values that are longer than an int. */
75 + return len1 < len2 ? -1 : 1;
77 + if (always_checksum) {
78 + diff = u_memcmp(F_SUM(f1), F_SUM(f2), checksum_len);
81 + } else if (f1->modtime != f2->modtime)
82 + return f1->modtime < f2->modtime ? -1 : 1;
84 + diff = u_strcmp(f1->basename, f2->basename);
88 + if (f1->dirname == f2->dirname)
94 + return u_strcmp(f1->dirname, f2->dirname);
97 static void send_directory(int f, struct file_list *flist,
100 @@ -1501,6 +1543,25 @@ struct file_list *recv_file_list(int f)
102 clean_flist(flist, relative_paths, 1);
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;
116 + struct file_struct *fp = the_fattr_list.files[j];
117 + if (fp->basename && S_ISREG(fp->mode) && F_LENGTH(fp))
120 + the_fattr_list.high = j;
124 recv_uid_list(f, flist);
128 @@ -76,6 +76,7 @@ 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 @@ -91,6 +92,7 @@ extern char *backup_dir;
137 extern char *backup_suffix;
138 extern int backup_suffix_len;
139 extern struct file_list *the_file_list;
140 +extern struct file_list the_fattr_list;
141 extern struct filter_list_struct server_filter_list;
143 int ignore_perishable = 0;
144 @@ -98,12 +100,14 @@ int non_perishable_cnt = 0;
145 int maybe_ATTRS_REPORT = 0;
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 static BOOL solo_file = 0;
154 -/* For calling delete_item() and delete_dir_contents(). */
155 +/* For calling delete_item(), delete_dir_contents(), and delete_in_dir(). */
156 +#define DEL_NO_DELETIONS (1<<0)
157 #define DEL_RECURSE (1<<1) /* recurse */
158 #define DEL_DIR_IS_EMPTY (1<<2) /* internal delete_FUNCTIONS use only */
160 @@ -125,11 +129,120 @@ static int is_backup_file(char *fn)
161 return k > 0 && strcmp(fn+k, backup_suffix) == 0;
164 +/* Search for a regular file that matches either (1) the size & modified
165 + * time (plus the basename, if possible) or (2) the size & checksum. If
166 + * we find an exact match down to the dirname, return -1 because we found
167 + * an up-to-date file in the transfer, not a renamed file. */
168 +static int fattr_find(struct file_struct *f, char *fname)
170 + int low = the_fattr_list.low, high = the_fattr_list.high;
171 + int mid, ok_match = -1, good_match = -1;
172 + struct file_struct *fmid;
175 + while (low <= high) {
176 + mid = (low + high) / 2;
177 + fmid = the_fattr_list.files[mid];
178 + if (F_LENGTH(fmid) != F_LENGTH(f)) {
179 + if (F_LENGTH(fmid) < F_LENGTH(f))
185 + if (always_checksum) {
186 + /* We use the FLAG_SENT flag to indicate when we
187 + * have computed the checksum for an entry. */
188 + if (!(f->flags & FLAG_SENT)) {
189 + if (fmid->modtime == f->modtime
190 + && f_name_cmp(fmid, f) == 0)
191 + return -1; /* assume we can't help */
192 + file_checksum(fname, (char*)F_SUM(f), F_LENGTH(f));
193 + f->flags |= FLAG_SENT;
195 + diff = u_memcmp(F_SUM(fmid), F_SUM(f), checksum_len);
204 + if (fmid->modtime != f->modtime) {
205 + if (fmid->modtime < f->modtime)
213 + diff = u_strcmp(fmid->basename, f->basename);
216 + if (fmid->dirname == f->dirname)
217 + return -1; /* file is up-to-date */
218 + if (!fmid->dirname) {
226 + diff = u_strcmp(fmid->dirname, f->dirname);
228 + return -1; /* file is up-to-date */
236 + return good_match >= 0 ? good_match : ok_match;
239 +static void look_for_rename(struct file_struct *file, char *fname)
241 + struct file_struct *fp;
242 + char *partialptr, *fn;
246 + if ((ndx = fattr_find(file, fname)) < 0)
249 + fp = the_fattr_list.files[ndx];
250 + fn = f_name(fp, NULL);
251 + /* We don't provide an alternate-basis file if there is a basis file. */
252 + if (link_stat(fn, &st, 0) == 0)
254 + if ((partialptr = partial_dir_fname(fn)) == NULL
255 + || !handle_partial_dir(partialptr, PDIR_CREATE))
258 + /* We only use the file if we can hard-link it into our tmp dir. */
259 + if (link(fname, partialptr) == 0) {
261 + rprintf(FINFO, "found renamed: %s => %s\n",
262 + fname, partialptr);
267 + if (errno != EEXIST)
268 + handle_partial_dir(partialptr, PDIR_DELETE);
271 /* Delete a file or directory. If DEL_RECURSE is set in the flags, this will
272 * delete recursively.
274 * Note that fbuf must point to a MAXPATHLEN buffer if the mode indicates it's
275 * a directory! (The buffer is used for recursion, but returned unchanged.)
277 + * Also note: --detect-rename may use this routine with DEL_NO_DELETIONS set!
279 static enum delret delete_item(char *fbuf, int mode, char *replace, int flags)
281 @@ -151,6 +264,8 @@ static enum delret delete_item(char *fbu
283 /* OK: try to delete the directory. */
285 + if (flags & DEL_NO_DELETIONS)
288 if (!replace && max_delete >= 0 && ++deletion_count > max_delete)
290 @@ -197,6 +312,8 @@ static enum delret delete_item(char *fbu
291 * its contents, otherwise just checks for content. Returns DR_SUCCESS or
292 * DR_NOT_EMPTY. Note that fname must point to a MAXPATHLEN buffer! (The
293 * buffer is used for recursion, but returned unchanged.)
295 + * Note: --detect-rename may use this routine with DEL_NO_DELETIONS set!
297 static enum delret delete_dir_contents(char *fname, int flags)
299 @@ -216,7 +333,9 @@ static enum delret delete_dir_contents(c
300 save_filters = push_local_filters(fname, dlen);
302 non_perishable_cnt = 0;
303 + file_extra_cnt += SUM_EXTRA_CNT;
304 dirlist = get_dirlist(fname, dlen, 0);
305 + file_extra_cnt -= SUM_EXTRA_CNT;
306 ret = non_perishable_cnt ? DR_NOT_EMPTY : DR_SUCCESS;
309 @@ -253,6 +372,8 @@ static enum delret delete_dir_contents(c
310 if (S_ISDIR(fp->mode)
311 && delete_dir_contents(fname, flags | DEL_RECURSE) != DR_SUCCESS)
313 + if (detect_renamed && S_ISREG(fp->mode))
314 + look_for_rename(fp, fname);
315 if (delete_item(fname, fp->mode, NULL, flags) != DR_SUCCESS)
318 @@ -405,15 +526,19 @@ static void do_delayed_deletions(char *d
319 * all the --delete-WHEN options. Note that the fbuf pointer must point to a
320 * MAXPATHLEN buffer with the name of the directory in it (the functions we
321 * call will append names onto the end, but the old dir value will be restored
325 + * Note: --detect-rename may use this routine with DEL_NO_DELETIONS set!
327 static void delete_in_dir(struct file_list *flist, char *fbuf,
328 - struct file_struct *file, STRUCT_STAT *stp)
329 + struct file_struct *file, STRUCT_STAT *stp, int flags)
331 static int min_depth = MAXPATHLEN, cur_depth = -1;
332 static void *filt_array[MAXPATHLEN/2+1];
333 static int already_warned = 0;
334 struct file_list *dirlist;
335 - char delbuf[MAXPATHLEN];
336 + char *p, delbuf[MAXPATHLEN];
337 + unsigned remainder;
341 @@ -427,6 +552,8 @@ static void delete_in_dir(struct file_li
343 rprintf(FINFO, "delete_in_dir(%s)\n", fbuf);
345 + flags |= DEL_RECURSE;
348 maybe_send_keepalive();
350 @@ -434,12 +561,14 @@ static void delete_in_dir(struct file_li
351 return; /* Impossible... */
353 if (io_error && !ignore_errors) {
354 - if (already_warned)
355 + if (!already_warned) {
357 + "IO error encountered -- skipping file deletion\n");
358 + already_warned = 1;
360 + if (!detect_renamed)
363 - "IO error encountered -- skipping file deletion\n");
364 - already_warned = 1;
366 + flags |= DEL_NO_DELETIONS;
369 while (cur_depth >= F_DEPTH(file) && cur_depth >= min_depth)
370 @@ -450,6 +579,9 @@ static void delete_in_dir(struct file_li
372 filt_array[cur_depth] = push_local_filters(fbuf, dlen);
374 + if (detect_renamed)
377 if (one_file_system) {
378 if (file->flags & FLAG_TOP_DIR)
379 filesystem_dev = stp->st_dev;
380 @@ -459,6 +591,11 @@ static void delete_in_dir(struct file_li
382 dirlist = get_dirlist(fbuf, dlen, 0);
385 + if (dlen != 1 || *fbuf != '/')
387 + remainder = MAXPATHLEN - (p - fbuf);
389 /* If an item in dirlist is not found in flist, delete it
390 * from the filesystem. */
391 for (i = dirlist->count; i--; ) {
392 @@ -471,16 +608,23 @@ static void delete_in_dir(struct file_li
396 + if (detect_renamed && S_ISREG(fp->mode)) {
397 + strlcpy(p, fp->basename, remainder);
398 + look_for_rename(fp, fbuf);
400 if (flist_find(flist, fp) < 0) {
402 - if (delete_during == 2) {
403 + if (delete_during == 2 && !(flags & DEL_NO_DELETIONS)) {
404 if (!remember_delete(fp, delbuf))
407 - delete_item(delbuf, fp->mode, NULL, DEL_RECURSE);
409 + delete_item(delbuf, fp->mode, NULL, flags);
410 + } else if (detect_renamed && S_ISDIR(fp->mode))
419 @@ -510,9 +654,9 @@ static void do_delete_pass(struct file_l
420 || !S_ISDIR(st.st_mode))
423 - delete_in_dir(flist, fbuf, file, &st);
424 + delete_in_dir(flist, fbuf, file, &st, 0);
426 - delete_in_dir(NULL, NULL, NULL, NULL);
427 + delete_in_dir(NULL, NULL, NULL, NULL, 0);
429 if (do_progress && !am_server)
430 rprintf(FINFO, " \r");
431 @@ -1047,6 +1191,7 @@ static int try_dests_non(struct file_str
435 +static struct bitbag *delayed_bits = NULL;
436 static int phase = 0;
438 /* Acts on the_file_list->file's ndx'th item, whose name is fname. If a dir,
439 @@ -1232,8 +1377,12 @@ static void recv_generator(char *fname,
440 if (real_ret != 0 && one_file_system)
441 real_st.st_dev = filesystem_dev;
442 if (delete_during && f_out != -1 && !phase && dry_run < 2
443 - && (file->flags & FLAG_XFER_DIR))
444 - delete_in_dir(the_file_list, fname, file, &real_st);
445 + && (file->flags & FLAG_XFER_DIR)) {
446 + if (detect_renamed && real_ret != 0)
448 + delete_in_dir(the_file_list, fname, file, &real_st,
449 + delete_during < 0 ? DEL_NO_DELETIONS : 0);
454 @@ -1497,8 +1646,14 @@ static void recv_generator(char *fname,
455 if (preserve_hard_links && F_HLINK_NOT_LAST(file))
458 - if (stat_errno == ENOENT)
459 + if (stat_errno == ENOENT) {
460 + if (detect_renamed && unexplored_dirs > 0
461 + && F_LENGTH(file)) {
462 + bitbag_set_bit(delayed_bits, ndx);
467 rsyserr(FERROR, stat_errno, "recv_generator: failed to stat %s",
470 @@ -1688,6 +1843,12 @@ void generate_files(int f_out, struct fi
471 (long)getpid(), flist->count);
474 + if (detect_renamed) {
475 + delayed_bits = bitbag_create(flist->count);
476 + if (!delete_before && !delete_during)
477 + delete_during = -1;
480 if (delete_before && !local_name && flist->count > 0)
481 do_delete_pass(flist);
482 if (delete_during == 2) {
483 @@ -1698,7 +1859,7 @@ void generate_files(int f_out, struct fi
487 - if (append_mode || whole_file < 0)
488 + if (append_mode || detect_renamed || whole_file < 0)
491 rprintf(FINFO, "delta-transmission %s\n",
492 @@ -1754,7 +1915,22 @@ void generate_files(int f_out, struct fi
494 recv_generator(NULL, NULL, 0, 0, code, -1);
496 - delete_in_dir(NULL, NULL, NULL, NULL);
497 + delete_in_dir(NULL, NULL, NULL, NULL, 0);
499 + if (detect_renamed) {
500 + if (delete_during < 0)
502 + detect_renamed = 0;
504 + for (i = -1; (i = bitbag_next_bit(delayed_bits, i)) >= 0; ) {
505 + struct file_struct *file = flist->files[i];
507 + strlcpy(fbuf, local_name, sizeof fbuf);
509 + f_name(file, fbuf);
510 + recv_generator(fbuf, file, i, itemizing, code, f_out);
515 csum_length = SUM_LENGTH;
518 @@ -78,6 +78,7 @@ int am_generator = 0;
519 int am_starting_up = 1;
520 int relative_paths = -1;
521 int implied_dirs = 1;
522 +int detect_renamed = 0;
524 int allow_8bit_chars = 0;
525 int force_delete = 0;
526 @@ -342,6 +343,7 @@ void usage(enum logcode F)
527 rprintf(F," --modify-window=NUM compare mod-times with reduced accuracy\n");
528 rprintf(F," -T, --temp-dir=DIR create temporary files in directory DIR\n");
529 rprintf(F," -y, --fuzzy find similar file for basis if no dest file\n");
530 + rprintf(F," --detect-renamed try to find renamed files to speed up the transfer\n");
531 rprintf(F," --compare-dest=DIR also compare destination files relative to DIR\n");
532 rprintf(F," --copy-dest=DIR ... and include copies of unchanged files\n");
533 rprintf(F," --link-dest=DIR hardlink to files in DIR when unchanged\n");
534 @@ -496,6 +498,7 @@ static struct poptOption long_options[]
535 {"compare-dest", 0, POPT_ARG_STRING, 0, OPT_COMPARE_DEST, 0, 0 },
536 {"copy-dest", 0, POPT_ARG_STRING, 0, OPT_COPY_DEST, 0, 0 },
537 {"link-dest", 0, POPT_ARG_STRING, 0, OPT_LINK_DEST, 0, 0 },
538 + {"detect-renamed", 0, POPT_ARG_NONE, &detect_renamed, 0, 0, 0 },
539 {"fuzzy", 'y', POPT_ARG_NONE, &fuzzy_basis, 0, 0, 0 },
540 {"compress", 'z', POPT_ARG_NONE, 0, 'z', 0, 0 },
541 {"compress-level", 0, POPT_ARG_INT, &def_compress_level, 'z', 0, 0 },
542 @@ -1355,7 +1358,7 @@ int parse_arguments(int *argc, const cha
546 - if (delay_updates && !partial_dir)
547 + if ((delay_updates || detect_renamed) && !partial_dir)
548 partial_dir = tmp_partialdir;
551 @@ -1364,6 +1367,7 @@ int parse_arguments(int *argc, const cha
552 snprintf(err_buf, sizeof err_buf,
553 "--%s cannot be used with --%s\n",
554 append_mode ? "append" : "inplace",
555 + detect_renamed ? "detect-renamed" :
556 delay_updates ? "delay-updates" : "partial-dir");
559 @@ -1674,6 +1678,8 @@ void server_options(char **args,int *arg
560 args[ac++] = "--super";
562 args[ac++] = "--size-only";
563 + if (detect_renamed)
564 + args[ac++] = "--detect-renamed";
567 if (modify_window_set) {
570 @@ -364,6 +364,7 @@ to the detailed description below for a
571 --modify-window=NUM compare mod-times with reduced accuracy
572 -T, --temp-dir=DIR create temporary files in directory DIR
573 -y, --fuzzy find similar file for basis if no dest file
574 + --detect-renamed try to find renamed files to speed the xfer
575 --compare-dest=DIR also compare received files relative to DIR
576 --copy-dest=DIR ... and include copies of unchanged files
577 --link-dest=DIR hardlink to files in DIR when unchanged
578 @@ -1272,6 +1273,15 @@ Note that the use of the bf(--delete) op
579 fuzzy-match files, so either use bf(--delete-after) or specify some
580 filename exclusions if you need to prevent this.
582 +dit(bf(--detect-renamed)) This option tells rsync to scan the receiving
583 +side for files that have been renamed, and to use any that are found as
584 +alternate basis files to help speed up the transfer.
585 +By default, alternate-basis files are hard-linked into a directory named
586 +".~tmp~" in each file's destination directory, but if you've specified
587 +the bf(--partial-dir) option, that directory will be used instead. These
588 +potential alternate-basis files will be removed as the transfer progresses.
589 +This option conflicts with bf(--inplace) and bf(--append).
591 dit(bf(--compare-dest=DIR)) This option instructs rsync to use em(DIR) on
592 the destination machine as an additional hierarchy to compare destination
593 files against doing transfers (if the files are missing in the destination
596 @@ -1027,6 +1027,32 @@ int handle_partial_dir(const char *fname
600 +/* We need to supply our own strcmp function for file list comparisons
601 + * to ensure that signed/unsigned usage is consistent between machines. */
602 +int u_strcmp(const char *p1, const char *p2)
604 + for ( ; *p1; p1++, p2++) {
609 + return (int)*(uchar*)p1 - (int)*(uchar*)p2;
612 +/* We need a memcmp function compares unsigned-byte values. */
613 +int u_memcmp(const void *p1, const void *p2, size_t len)
615 + const uchar *u1 = p1;
616 + const uchar *u2 = p2;
620 + return (int)*u1 - (int)*u2;
627 * Determine if a symlink points outside the current directory tree.
628 * This is considered "unsafe" because e.g. when mirroring somebody