This patch adds the --detect-renamed option which makes rsync notice files that either (1) match in size & modify-time (plus the basename, if possible) or (2) match in size & checksum (when --checksum was also specified) and use each match as an alternate basis file to speed up the transfer. The algorithm attempts to scan the receiving-side's files in an efficient manner. If --delete[-before] is enabled, we'll take advantage of the pre-transfer delete pass to prepare any alternate-basis-file matches we might find. If --delete-before is not enabled, rsync does the rename scan during the regular file-sending scan (scanning each directory right before the generator starts updating files from that dir). In this latter mode, rsync might delay the updating of a file (if no alternate-basis match was yet found) until the full scan of the receiving side is complete, at which point any delayed files are processed. I chose to hard-link the alternate-basis files into a ".~tmp~" subdir that takes advantage of rsync's pre-existing partial-dir logic. This uses less memory than trying to keep track of the matches internally, and also allows any deletions or file-updates to occur normally without interfering with these alternate-basis discoveries. To use this patch, run these commands for a successful build: patch -p1 basename || !S_ISREG(f1->mode) || !len1) { + if (!f2->basename || !S_ISREG(f2->mode) || !len2) + return 0; + return 1; + } + if (!f2->basename || !S_ISREG(f2->mode) || !len2) + return -1; + + /* Don't use diff for values that are longer than an int. */ + if (len1 != len2) + return len1 < len2 ? -1 : 1; + + if (always_checksum) { + diff = u_memcmp(F_SUM(f1), F_SUM(f2), checksum_len); + if (diff) + return diff; + } else if (f1->modtime != f2->modtime) + return f1->modtime < f2->modtime ? -1 : 1; + + diff = u_strcmp(f1->basename, f2->basename); + if (diff) + return diff; + + if (f1->dirname == f2->dirname) + return 0; + if (!f1->dirname) + return -1; + if (!f2->dirname) + return 1; + return u_strcmp(f1->dirname, f2->dirname); +} + static void send_directory(int f, struct file_list *flist, char *fbuf, int len, int flags); @@ -2344,6 +2386,25 @@ struct file_list *recv_file_list(int f) flist_sort_and_clean(flist, relative_paths); + if (detect_renamed) { + int j = flist->used; + the_fattr_list.used = j; + the_fattr_list.files = new_array(struct file_struct *, j); + if (!the_fattr_list.files) + out_of_memory("recv_file_list"); + memcpy(the_fattr_list.files, flist->files, + j * sizeof (struct file_struct *)); + qsort(the_fattr_list.files, j, + sizeof the_fattr_list.files[0], (int (*)())fattr_compare); + the_fattr_list.low = 0; + while (j-- > 0) { + struct file_struct *fp = the_fattr_list.files[j]; + if (fp->basename && S_ISREG(fp->mode) && F_LENGTH(fp)) + break; + } + the_fattr_list.high = j; + } + if (protocol_version < 30) { /* Recv the io_error flag */ if (ignore_errors) diff --git a/generator.c b/generator.c --- a/generator.c +++ b/generator.c @@ -80,6 +80,7 @@ extern char *basis_dir[]; extern int compare_dest; extern int copy_dest; extern int link_dest; +extern int detect_renamed; extern int whole_file; extern int list_only; extern int read_batch; @@ -98,6 +99,7 @@ extern char *backup_suffix; extern int backup_suffix_len; extern struct file_list *cur_flist, *first_flist, *dir_flist; extern struct filter_list_struct daemon_filter_list; +extern struct file_list the_fattr_list; int ignore_perishable = 0; int non_perishable_cnt = 0; @@ -105,6 +107,7 @@ int maybe_ATTRS_REPORT = 0; static dev_t dev_zero; static int deletion_count = 0; /* used to implement --max-delete */ +static int unexplored_dirs = 1; static int deldelay_size = 0, deldelay_cnt = 0; static char *deldelay_buf = NULL; static int deldelay_fd = -1; @@ -115,7 +118,7 @@ static int need_retouch_dir_times; static int need_retouch_dir_perms; static const char *solo_file = NULL; -/* For calling delete_item() and delete_dir_contents(). */ +/* For calling delete_item(), delete_dir_contents(), and delete_in_dir(). */ #define DEL_NO_UID_WRITE (1<<0) /* file/dir has our uid w/o write perm */ #define DEL_RECURSE (1<<1) /* if dir, delete all contents */ #define DEL_DIR_IS_EMPTY (1<<2) /* internal delete_FUNCTIONS use only */ @@ -124,6 +127,7 @@ static const char *solo_file = NULL; #define DEL_FOR_SYMLINK (1<<5) /* making room for a replacement symlink */ #define DEL_FOR_DEVICE (1<<6) /* making room for a replacement device */ #define DEL_FOR_SPECIAL (1<<7) /* making room for a replacement special */ +#define DEL_NO_DELETIONS (1<<9) /* just check for renames w/o deleting */ #define DEL_MAKE_ROOM (DEL_FOR_FILE|DEL_FOR_DIR|DEL_FOR_SYMLINK|DEL_FOR_DEVICE|DEL_FOR_SPECIAL) @@ -148,11 +152,121 @@ static int is_backup_file(char *fn) return k > 0 && strcmp(fn+k, backup_suffix) == 0; } +/* Search for a regular file that matches either (1) the size & modified + * time (plus the basename, if possible) or (2) the size & checksum. If + * we find an exact match down to the dirname, return -1 because we found + * an up-to-date file in the transfer, not a renamed file. */ +static int fattr_find(struct file_struct *f, char *fname) +{ + int low = the_fattr_list.low, high = the_fattr_list.high; + int mid, ok_match = -1, good_match = -1; + struct file_struct *fmid; + int diff; + + while (low <= high) { + mid = (low + high) / 2; + fmid = the_fattr_list.files[mid]; + if (F_LENGTH(fmid) != F_LENGTH(f)) { + if (F_LENGTH(fmid) < F_LENGTH(f)) + low = mid + 1; + else + high = mid - 1; + continue; + } + if (always_checksum) { + /* We use the FLAG_FILE_SENT flag to indicate when we + * have computed the checksum for an entry. */ + if (!(f->flags & FLAG_FILE_SENT)) { + if (fmid->modtime == f->modtime + && f_name_cmp(fmid, f) == 0) + return -1; /* assume we can't help */ + file_checksum(fname, F_SUM(f), F_LENGTH(f)); + f->flags |= FLAG_FILE_SENT; + } + diff = u_memcmp(F_SUM(fmid), F_SUM(f), checksum_len); + if (diff) { + if (diff < 0) + low = mid + 1; + else + high = mid - 1; + continue; + } + } else { + if (fmid->modtime != f->modtime) { + if (fmid->modtime < f->modtime) + low = mid + 1; + else + high = mid - 1; + continue; + } + } + ok_match = mid; + diff = u_strcmp(fmid->basename, f->basename); + if (diff == 0) { + good_match = mid; + if (fmid->dirname == f->dirname) + return -1; /* file is up-to-date */ + if (!fmid->dirname) { + low = mid + 1; + continue; + } + if (!f->dirname) { + high = mid - 1; + continue; + } + diff = u_strcmp(fmid->dirname, f->dirname); + if (diff == 0) + return -1; /* file is up-to-date */ + } + if (diff < 0) + low = mid + 1; + else + high = mid - 1; + } + + return good_match >= 0 ? good_match : ok_match; +} + +static void look_for_rename(struct file_struct *file, char *fname) +{ + struct file_struct *fp; + char *partialptr, *fn; + STRUCT_STAT st; + int ndx; + + if (!partial_dir || (ndx = fattr_find(file, fname)) < 0) + return; + + fp = the_fattr_list.files[ndx]; + fn = f_name(fp, NULL); + /* We don't provide an alternate-basis file if there is a basis file. */ + if (link_stat(fn, &st, 0) == 0) + return; + + if (!dry_run) { + if ((partialptr = partial_dir_fname(fn)) == NULL + || !handle_partial_dir(partialptr, PDIR_CREATE)) + return; + /* We only use the file if we can hard-link it into our tmp dir. */ + if (link(fname, partialptr) != 0) { + if (errno != EEXIST) + handle_partial_dir(partialptr, PDIR_DELETE); + return; + } + } + + /* I think this falls into the -vv category with "%s is uptodate", etc. */ + if (verbose > 1) + rprintf(FINFO, "found renamed: %s => %s\n", fname, fn); +} + /* Delete a file or directory. If DEL_RECURSE is set in the flags, this will * delete recursively. * * Note that fbuf must point to a MAXPATHLEN buffer if the mode indicates it's * a directory! (The buffer is used for recursion, but returned unchanged.) + * + * Also note: --detect-rename may use this routine with DEL_NO_DELETIONS set! */ static enum delret delete_item(char *fbuf, uint16 mode, uint16 flags) { @@ -186,6 +300,8 @@ static enum delret delete_item(char *fbuf, uint16 mode, uint16 flags) goto check_ret; /* OK: try to delete the directory. */ } + if (flags & DEL_NO_DELETIONS) + return DR_SUCCESS; if (!(flags & DEL_MAKE_ROOM) && max_delete >= 0 && ++deletion_count > max_delete) return DR_AT_LIMIT; @@ -241,6 +357,8 @@ static enum delret delete_item(char *fbuf, uint16 mode, uint16 flags) * its contents, otherwise just checks for content. Returns DR_SUCCESS or * DR_NOT_EMPTY. Note that fname must point to a MAXPATHLEN buffer! (The * buffer is used for recursion, but returned unchanged.) + * + * Note: --detect-rename may use this routine with DEL_NO_DELETIONS set! */ static enum delret delete_dir_contents(char *fname, uint16 flags) { @@ -260,7 +378,9 @@ static enum delret delete_dir_contents(char *fname, uint16 flags) save_filters = push_local_filters(fname, dlen); non_perishable_cnt = 0; + file_extra_cnt += SUM_EXTRA_CNT; dirlist = get_dirlist(fname, dlen, 0); + file_extra_cnt -= SUM_EXTRA_CNT; ret = non_perishable_cnt ? DR_NOT_EMPTY : DR_SUCCESS; if (!dirlist->used) @@ -300,7 +420,8 @@ static enum delret delete_dir_contents(char *fname, uint16 flags) if (S_ISDIR(fp->mode)) { if (delete_dir_contents(fname, flags | DEL_RECURSE) != DR_SUCCESS) ret = DR_NOT_EMPTY; - } + } else if (detect_renamed && S_ISREG(fp->mode)) + look_for_rename(fp, fname); if (delete_item(fname, fp->mode, flags) != DR_SUCCESS) ret = DR_NOT_EMPTY; } @@ -465,13 +586,18 @@ static void do_delayed_deletions(char *delbuf) * all the --delete-WHEN options. Note that the fbuf pointer must point to a * MAXPATHLEN buffer with the name of the directory in it (the functions we * call will append names onto the end, but the old dir value will be restored - * on exit). */ -static void delete_in_dir(char *fbuf, struct file_struct *file, dev_t *fs_dev) + * on exit). + * + * Note: --detect-rename may use this routine with DEL_NO_DELETIONS set! + */ +static void delete_in_dir(char *fbuf, struct file_struct *file, dev_t *fs_dev, + int del_flags) { static int already_warned = 0; struct file_list *dirlist; - char delbuf[MAXPATHLEN]; - int dlen, i; + char *p, delbuf[MAXPATHLEN]; + unsigned remainder; + int dlen, i, restore_dot = 0; int save_uid_ndx = uid_ndx; if (!fbuf) { @@ -486,17 +612,22 @@ static void delete_in_dir(char *fbuf, struct file_struct *file, dev_t *fs_dev) maybe_send_keepalive(); if (io_error && !ignore_errors) { - if (already_warned) + if (!already_warned) { + rprintf(FINFO, + "IO error encountered -- skipping file deletion\n"); + already_warned = 1; + } + if (!detect_renamed) return; - rprintf(FINFO, - "IO error encountered -- skipping file deletion\n"); - already_warned = 1; - return; + del_flags |= DEL_NO_DELETIONS; } dlen = strlen(fbuf); change_local_filter_dir(fbuf, dlen, F_DEPTH(file)); + if (detect_renamed) + unexplored_dirs--; + if (one_file_system) { if (file->flags & FLAG_TOP_DIR) filesystem_dev = *fs_dev; @@ -509,6 +640,14 @@ static void delete_in_dir(char *fbuf, struct file_struct *file, dev_t *fs_dev) dirlist = get_dirlist(fbuf, dlen, 0); + p = fbuf + dlen; + if (dlen == 1 && *fbuf == '.') { + restore_dot = 1; + p = fbuf; + } else if (dlen != 1 || *fbuf != '/') + *p++ = '/'; + remainder = MAXPATHLEN - (p - fbuf); + /* If an item in dirlist is not found in flist, delete it * from the filesystem. */ for (i = dirlist->used; i--; ) { @@ -521,6 +660,10 @@ static void delete_in_dir(char *fbuf, struct file_struct *file, dev_t *fs_dev) f_name(fp, NULL)); continue; } + if (detect_renamed && S_ISREG(fp->mode)) { + strlcpy(p, fp->basename, remainder); + look_for_rename(fp, fbuf); + } /* Here we want to match regardless of file type. Replacement * of a file with one of another type is handled separately by * a delete_item call with a DEL_MAKE_ROOM flag. */ @@ -529,14 +672,19 @@ static void delete_in_dir(char *fbuf, struct file_struct *file, dev_t *fs_dev) if (!(fp->mode & S_IWUSR) && !am_root && (uid_t)F_OWNER(fp) == our_uid) flags |= DEL_NO_UID_WRITE; f_name(fp, delbuf); - if (delete_during == 2) { - if (!remember_delete(fp, delbuf, flags)) + if (delete_during == 2 && !(del_flags & DEL_NO_DELETIONS)) { + if (!remember_delete(fp, delbuf, del_flags | flags)) break; } else - delete_item(delbuf, fp->mode, flags); - } + delete_item(delbuf, fp->mode, del_flags | flags); + } else if (detect_renamed && S_ISDIR(fp->mode)) + unexplored_dirs++; } + if (restore_dot) + fbuf[0] = '.'; + fbuf[dlen] = '\0'; + flist_free(dirlist); if (!save_uid_ndx) { @@ -574,9 +722,9 @@ static void do_delete_pass(void) || !S_ISDIR(st.st_mode)) continue; - delete_in_dir(fbuf, file, &st.st_dev); + delete_in_dir(fbuf, file, &st.st_dev, 0); } - delete_in_dir(NULL, NULL, &dev_zero); + delete_in_dir(NULL, NULL, &dev_zero, 0); if (do_progress && !am_server) rprintf(FINFO, " \r"); @@ -1229,6 +1377,7 @@ static void list_file_entry(struct file_struct *f) } } +static struct bitbag *delayed_bits = NULL; static int phase = 0; static int dflt_perms; @@ -1505,9 +1654,12 @@ static void recv_generator(char *fname, struct file_struct *file, int ndx, } else if (delete_during && f_out != -1 && !phase && !(file->flags & FLAG_MISSING_DIR)) { - if (file->flags & FLAG_CONTENT_DIR) - delete_in_dir(fname, file, &real_sx.st.st_dev); - else + if (file->flags & FLAG_CONTENT_DIR) { + if (detect_renamed && real_ret != 0) + unexplored_dirs++; + delete_in_dir(fname, file, &real_sx.st.st_dev, + delete_during < 0 ? DEL_NO_DELETIONS : 0); + } else change_local_filter_dir(fname, strlen(fname), F_DEPTH(file)); } goto cleanup; @@ -1785,8 +1937,14 @@ static void recv_generator(char *fname, struct file_struct *file, int ndx, goto cleanup; } #endif - if (stat_errno == ENOENT) + if (stat_errno == ENOENT) { + if (detect_renamed && unexplored_dirs > 0 + && F_LENGTH(file)) { + bitbag_set_bit(delayed_bits, ndx); + return; + } goto notify_others; + } rsyserr(FERROR_XFER, stat_errno, "recv_generator: failed to stat %s", full_fname(fname)); goto cleanup; @@ -2187,6 +2345,12 @@ void generate_files(int f_out, const char *local_name) if (verbose > 2) rprintf(FINFO, "generator starting pid=%ld\n", (long)getpid()); + if (detect_renamed) { + delayed_bits = bitbag_create(cur_flist->used); + if (!delete_before && !delete_during) + delete_during = -1; + } + if (delete_before && !solo_file && cur_flist->used > 0) do_delete_pass(); if (delete_during == 2) { @@ -2197,7 +2361,7 @@ void generate_files(int f_out, const char *local_name) } do_progress = 0; - if (append_mode > 0 || whole_file < 0) + if (append_mode > 0 || detect_renamed || whole_file < 0) whole_file = 0; if (verbose >= 2) { rprintf(FINFO, "delta-transmission %s\n", @@ -2239,7 +2403,7 @@ void generate_files(int f_out, const char *local_name) dirdev = MAKEDEV(DEV_MAJOR(devp), DEV_MINOR(devp)); } else dirdev = MAKEDEV(0, 0); - delete_in_dir(fbuf, fp, &dirdev); + delete_in_dir(fbuf, fp, &dirdev, 0); } else change_local_filter_dir(fbuf, strlen(fbuf), F_DEPTH(fp)); } @@ -2289,7 +2453,21 @@ void generate_files(int f_out, const char *local_name) write_ndx(f_out, NDX_DONE); if (delete_during) - delete_in_dir(NULL, NULL, &dev_zero); + delete_in_dir(NULL, NULL, &dev_zero, 0); + if (detect_renamed) { + if (delete_during < 0) + delete_during = 0; + detect_renamed = 0; + + for (i = -1; (i = bitbag_next_bit(delayed_bits, i)) >= 0; ) { + struct file_struct *file = cur_flist->files[i]; + if (local_name) + strlcpy(fbuf, local_name, sizeof fbuf); + else + f_name(file, fbuf); + recv_generator(fbuf, file, i, itemizing, code, f_out); + } + } phase++; if (verbose > 2) rprintf(FINFO, "generate_files phase=%d\n", phase); diff --git a/options.c b/options.c --- a/options.c +++ b/options.c @@ -82,6 +82,7 @@ int am_generator = 0; int am_starting_up = 1; int relative_paths = -1; int implied_dirs = 1; +int detect_renamed = 0; int numeric_ids = 0; int allow_8bit_chars = 0; int force_delete = 0; @@ -392,6 +393,7 @@ void usage(enum logcode F) rprintf(F," --modify-window=NUM compare mod-times with reduced accuracy\n"); rprintf(F," -T, --temp-dir=DIR create temporary files in directory DIR\n"); rprintf(F," -y, --fuzzy find similar file for basis if no dest file\n"); + rprintf(F," --detect-renamed try to find renamed files to speed up the transfer\n"); rprintf(F," --compare-dest=DIR also compare destination files relative to DIR\n"); rprintf(F," --copy-dest=DIR ... and include copies of unchanged files\n"); rprintf(F," --link-dest=DIR hardlink to files in DIR when unchanged\n"); @@ -579,6 +581,7 @@ static struct poptOption long_options[] = { {"compare-dest", 0, POPT_ARG_STRING, 0, OPT_COMPARE_DEST, 0, 0 }, {"copy-dest", 0, POPT_ARG_STRING, 0, OPT_COPY_DEST, 0, 0 }, {"link-dest", 0, POPT_ARG_STRING, 0, OPT_LINK_DEST, 0, 0 }, + {"detect-renamed", 0, POPT_ARG_NONE, &detect_renamed, 0, 0, 0 }, {"fuzzy", 'y', POPT_ARG_VAL, &fuzzy_basis, 1, 0, 0 }, {"no-fuzzy", 0, POPT_ARG_VAL, &fuzzy_basis, 0, 0, 0 }, {"no-y", 0, POPT_ARG_VAL, &fuzzy_basis, 0, 0, 0 }, @@ -1591,7 +1594,7 @@ int parse_arguments(int *argc_p, const char ***argv_p) inplace = 1; } - if (delay_updates && !partial_dir) + if ((delay_updates || detect_renamed) && !partial_dir) partial_dir = tmp_partialdir; if (inplace) { @@ -1600,6 +1603,7 @@ int parse_arguments(int *argc_p, const char ***argv_p) snprintf(err_buf, sizeof err_buf, "--%s cannot be used with --%s\n", append_mode ? "append" : "inplace", + detect_renamed ? "detect-renamed" : delay_updates ? "delay-updates" : "partial-dir"); return 0; } @@ -1961,6 +1965,8 @@ void server_options(char **args, int *argc_p) args[ac++] = "--super"; if (size_only) args[ac++] = "--size-only"; + if (detect_renamed) + args[ac++] = "--detect-renamed"; } else { if (skip_compress) { if (asprintf(&arg, "--skip-compress=%s", skip_compress) < 0) diff --git a/rsync.yo b/rsync.yo --- a/rsync.yo +++ b/rsync.yo @@ -389,6 +389,7 @@ to the detailed description below for a complete description. verb( --modify-window=NUM compare mod-times with reduced accuracy -T, --temp-dir=DIR create temporary files in directory DIR -y, --fuzzy find similar file for basis if no dest file + --detect-renamed try to find renamed files to speed the xfer --compare-dest=DIR also compare received files relative to DIR --copy-dest=DIR ... and include copies of unchanged files --link-dest=DIR hardlink to files in DIR when unchanged @@ -1491,6 +1492,21 @@ Note that the use of the bf(--delete) option might get rid of any potential fuzzy-match files, so either use bf(--delete-after) or specify some filename exclusions if you need to prevent this. +dit(bf(--detect-renamed)) With this option, for each new source file +(call it em(src/S)), rsync looks for a file em(dest/D) anywhere in the +destination that passes the quick check with em(src/S). If such a em(dest/D) +is found, rsync uses it as an alternate basis for transferring em(S). The +idea is that if em(src/S) was renamed from em(src/D) (as opposed to em(src/S) +passing the quick check with em(dest/D) by coincidence), the delta-transfer +algorithm will find that all the data matches between em(src/S) and em(dest/D), +and the transfer will be really fast. + +By default, alternate-basis files are hard-linked into a directory named +".~tmp~" in each file's destination directory, but if you've specified +the bf(--partial-dir) option, that directory will be used instead. These +potential alternate-basis files will be removed as the transfer progresses. +This option conflicts with bf(--inplace) and bf(--append). + dit(bf(--compare-dest=DIR)) This option instructs rsync to use em(DIR) on the destination machine as an additional hierarchy to compare destination files against doing transfers (if the files are missing in the destination diff --git a/util.c b/util.c --- a/util.c +++ b/util.c @@ -1168,6 +1168,32 @@ int handle_partial_dir(const char *fname, int create) return 1; } +/* We need to supply our own strcmp function for file list comparisons + * to ensure that signed/unsigned usage is consistent between machines. */ +int u_strcmp(const char *p1, const char *p2) +{ + for ( ; *p1; p1++, p2++) { + if (*p1 != *p2) + break; + } + + return (int)*(uchar*)p1 - (int)*(uchar*)p2; +} + +/* We need a memcmp function compares unsigned-byte values. */ +int u_memcmp(const void *p1, const void *p2, size_t len) +{ + const uchar *u1 = p1; + const uchar *u2 = p2; + + while (len--) { + if (*u1 != *u2) + return (int)*u1 - (int)*u2; + } + + return 0; +} + /** * Determine if a symlink points outside the current directory tree. * This is considered "unsafe" because e.g. when mirroring somebody