4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/utils/lustre_rsync.c
34 * Author: Kalpak Shah <Kalpak.Shah@Sun.COM>
35 * Author: Manoj Joseph <Manoj.Joseph@Sun.COM>
39 * - lustre_rsync is a tool for replicating a lustre filesystem.
41 * - The source-fs is a live lustre filesystem. It is not a
42 * snapshot. It is mounted and undergoing changes
44 * - The target-fs is a copy of the source-fs from the past. Let's
45 * call this point, the 'sync point'.
47 * - There is a changelog of all metadata operations that happened on
48 * the filesystem since the 'sync point'.
50 * - lustre_rsync replicates all the operations saved in the changelog
51 * on to the target filesystem to make it identical to the source.
53 * To facilitate replication, the lustre filesystem provides
54 * a) a way to get the current filesystem path of a given FID
55 * b) a way to open files by specifying its FID
57 * The changelog only has a limited amount of information.
58 * tfid - The FID of the target file
59 * pfid - The FID of the parent of the target file (at the time of
61 * sfid - The FID of the source file
62 * spfid - The FID of the parent of the source file
63 * name - The name of the target file (at the time of the operation), the name
64 * of the source file is appended (delimited with '\0') if this
65 * operation involves a source
67 * With just this information, it is not alwasy possible to determine
68 * the file paths for each operation. For instance, if pfid does not
69 * exist on the source-fs (due to a subsequent deletion), its path
70 * cannot be queried. In such cases, lustre_rsync keeps the files in a
71 * special directory ("/.lustrerepl"). Once all the operations in a
72 * changelog are replayed, all the files in this special directory
73 * will get moved to the location as in the source-fs.
75 * Shorthand used: f2p(fid) = fid2path(fid)
77 * The following are the metadata operations of interest.
79 * If tfid is absent on the source-fs, ignore this operation
80 * If pfid is absent on the source-fs [or]
81 * if f2p(pfid) is not present on target-fs [or]
82 * if f2p(pfid)+name != f2p(tfid)
83 * creat .lustrerepl/tfid
84 * track [pfid,tfid,name]
89 * If .lustrerepl/[tfid] is present on the target
90 * rm .lustrerepl/[tfid]
91 * Else if pfid is present on the source-fs,
92 * if f2p(pfid)+name is present,
95 * 3. move (spfid,sname) to (pfid,name)
97 * if spfid is also present, mv (spfid,sname) to (pfid,name)
98 * else mv .lustrerepl/[sfid] to (pfid,name)
99 * Else if pfid is not present,
100 * if spfid is present, mv (spfid,sname) .lustrerepl/[sfid]
101 * If moving out of .lustrerepl
102 * move out all its children in .lustrerepl.
103 * [pfid,tfid,name] tracked from (1) is used for this.
115 #include <sys/stat.h>
116 #include <sys/types.h>
121 #include <sys/xattr.h>
122 #include <linux/types.h>
123 #include <linux/lustre/lustre_ioctl.h>
125 #include <libcfs/util/ioctl.h>
126 #include <libcfs/util/string.h>
127 #include <libcfs/util/parser.h>
128 #include <linux/lnet/libcfs_debug.h>
129 #include <lustre/lustreapi.h>
130 #include "lustre_rsync.h"
132 #define REPLICATE_STATUS_VER 1
133 #define CLEAR_INTERVAL 100
134 #define DEFAULT_RSYNC_THRESHOLD 0xA00000 /* 10 MB */
136 #define TYPE_STR_LEN 16
138 #define DEFAULT_MDT "-MDT0000"
139 #define SPECIAL_DIR ".lustrerepl"
140 #define RSYNC "rsync"
147 /* Information for processing a changelog record. This structure is
148 allocated on the heap instead of allocating large variables on the
153 unsigned int is_extended:1;
154 enum changelog_rec_type type;
155 char tfid[LR_FID_STR_LEN];
156 char pfid[LR_FID_STR_LEN];
157 char sfid[LR_FID_STR_LEN];
158 char spfid[LR_FID_STR_LEN];
159 char sname[NAME_MAX + 1];
160 char name[NAME_MAX + 1];
161 char src[3 * PATH_MAX + 1];
162 char dest[3 * PATH_MAX + 1];
163 char path[PATH_MAX + 1];
164 char savedpath[PATH_MAX + 1];
165 char link[PATH_MAX + 1];
166 char linktmp[PATH_MAX + 1];
167 char cmd[PATH_MAX * 10];
171 /* Variables for querying the xattributes */
178 struct lr_parent_child_list {
179 struct lr_parent_child_log pc_log;
180 struct lr_parent_child_list *pc_next;
183 struct lustre_rsync_status *status;
184 char *statuslog; /* Name of the status log file */
186 int noxattr; /* Flag to turn off replicating xattrs */
187 int noclear; /* Flag to turn off clearing changelogs */
188 int debug; /* Flag to turn debugging information on and off */
189 int verbose; /* Verbose output */
190 long long rec_count; /* No of changelog records that were processed */
193 int use_rsync; /* Flag to turn on use of rsync to copy data */
194 long long rsync_threshold = DEFAULT_RSYNC_THRESHOLD;
195 int quit; /* Flag to stop processing the changelog; set on the
196 receipt of a signal */
197 int abort_on_err = 0;
199 char rsync[PATH_MAX];
200 char rsync_ver[PATH_MAX];
201 struct lr_parent_child_list *parents;
205 /* Command line options */
206 struct option long_opts[] = {
207 { .val = 'l', .name = "statuslog", .has_arg = required_argument },
208 { .val = 'm', .name = "mdt", .has_arg = required_argument },
209 { .val = 's', .name = "source", .has_arg = required_argument },
210 { .val = 't', .name = "target", .has_arg = required_argument },
211 { .val = 'u', .name = "user", .has_arg = required_argument },
212 { .val = 'v', .name = "verbose", .has_arg = no_argument },
213 { .val = 'x', .name = "xattr", .has_arg = required_argument },
214 { .val = 'z', .name = "dry-run", .has_arg = no_argument },
215 /* Undocumented options follow */
216 { .val = 'a', .name = "abort-on-err", .has_arg = no_argument },
217 { .val = 'c', .name = "cl-clear", .has_arg = required_argument },
218 { .val = 'd', .name = "debug", .has_arg = required_argument },
219 { .val = 'D', .name = "debuglog", .has_arg = required_argument },
220 { .val = 'n', .name = "start-recno", .has_arg = required_argument },
221 { .val = 'r', .name = "use-rsync", .has_arg = no_argument },
222 { .val = 'y', .name = "rsync-threshold",
223 .has_arg = required_argument },
226 /* Command line usage */
229 fprintf(stderr, "\tlustre_rsync -s <lustre_root_path> -t <target_path> "
230 "-m <mdt> -r <user id> -l <status log>\n"
231 "lustre_rsync can also pick up parameters from a "
232 "status log created earlier.\n"
233 "\tlustre_rsync -l <log_file>\n"
235 "\t--xattr <yes|no> replicate EAs\n"
236 "\t--abort-on-err abort at first err\n"
238 "\t--dry-run don't write anything\n");
241 #define DEBUG_ENTRY(info) \
242 lr_debug(D_TRACE, "***** Start %lld %s (%d) %s %s %s *****\n", \
243 (info)->recno, changelog_type2str((info)->type), \
244 (info)->type, (info)->tfid, (info)->pfid, (info)->name);
246 #define DEBUG_EXIT(info, rc) \
247 lr_debug(D_TRACE, "##### End %lld %s (%d) %s %s %s rc=%d #####\n", \
248 (info)->recno, changelog_type2str((info)->type), \
249 (info)->type, (info)->tfid, (info)->pfid, (info)->name, rc);
251 /* Print debug information. This is controlled by the value of the
252 global variable 'debug' */
253 void lr_debug(int level, const char *fmt, ...)
261 if (debug_log != NULL)
262 vfprintf(debug_log, fmt, ap);
264 vfprintf(stdout, fmt, ap);
269 void * lr_grow_buf(void *buf, int size)
273 ptr = realloc(buf, size);
280 /* Use rsync to replicate file data */
281 int lr_rsync_data(struct lr_info *info)
283 struct stat st_src, st_dest;
286 lr_debug(DTRACE, "Syncing data%s\n", info->tfid);
288 rc = stat(info->src, &st_src);
290 fprintf(stderr, "Error: Unable to stat src=%s %s\n",
291 info->src, info->name);
297 rc = stat(info->dest, &st_dest);
299 fprintf(stderr, "Error: Unable to stat dest=%s\n",
304 if (st_src.st_mtime != st_dest.st_mtime ||
305 st_src.st_size != st_dest.st_size) {
306 /* XXX spawning off an rsync for every data sync and
307 * waiting synchronously is bad for performance.
308 * librsync could possibly used here. But it does not
309 * seem to be of production grade. Multi-threaded
310 * replication is also to be considered.
314 if (snprintf(info->cmd, sizeof(info->cmd), "%s --inplace %s %s",
315 rsync, info->src, info->dest) >= sizeof(info->cmd)) {
319 lr_debug(DTRACE, "\t%s %s\n", info->cmd, info->tfid);
320 status = system(info->cmd);
323 } else if (WIFEXITED(status)) {
324 status = WEXITSTATUS(status);
327 else if (status == 23 || status == 24)
328 /* Error due to vanished source files;
334 lr_debug(DINFO, "rsync %s exited with %d %d\n",
335 info->src, status, rc);
340 lr_debug(DTRACE, "Not syncing %s and %s %s\n", info->src,
341 info->dest, info->tfid);
347 int lr_copy_data(struct lr_info *info)
357 fd_src = open(info->src, O_RDONLY);
360 if (fstat(fd_src, &st_src) == -1 ||
361 stat(info->dest, &st_dest) == -1)
364 if (st_src.st_mtime == st_dest.st_mtime &&
365 st_src.st_size == st_dest.st_size)
368 if (st_src.st_size > rsync_threshold && rsync[0] != '\0') {
369 /* It is more efficient to use rsync to replicate
370 large files. Any file larger than rsync_threshold
371 is handed off to rsync. */
372 lr_debug(DTRACE, "Using rsync to replicate %s\n", info->tfid);
373 rc = lr_rsync_data(info);
377 fd_dest = open(info->dest, O_WRONLY | O_TRUNC, st_src.st_mode);
382 bufsize = st_dest.st_blksize;
384 if (info->bufsize < bufsize) {
386 info->buf = lr_grow_buf(info->buf, bufsize);
387 if (info->buf == NULL) {
391 info->bufsize = bufsize;
399 rsize = read(fd_src, buf, bufsize);
409 wsize = write(fd_dest, buf, rsize);
429 /* Copy data from source to destination */
430 int lr_sync_data(struct lr_info *info)
433 return lr_rsync_data(info);
435 return lr_copy_data(info);
438 /* Copy all attributes from file src to file dest */
439 int lr_copy_attr(char *src, char *dest)
444 if (stat(src, &st) == -1 ||
445 chmod(dest, st.st_mode) == -1 ||
446 chown(dest, st.st_uid, st.st_gid) == -1)
449 time.actime = st.st_atime;
450 time.modtime = st.st_mtime;
451 if (utime(dest, &time) == -1)
456 /* Copy all xattrs from file info->src to info->dest */
457 int lr_copy_xattr(struct lr_info *info)
459 size_t size = info->xsize;
468 rc = llistxattr(info->src, info->xlist, size);
469 lr_debug(DTRACE, "llistxattr(%s,%p) returned %d, errno=%d\n",
470 info->src, info->xlist, rc, errno);
471 if ((rc > 0 && info->xlist == NULL) || errno == ERANGE) {
472 size = rc > PATH_MAX ? rc : PATH_MAX;
473 info->xlist = lr_grow_buf(info->xlist, size);
474 if (info->xlist == NULL)
477 rc = llistxattr(info->src, info->xlist, size);
478 lr_debug(DTRACE, "llistxattr %s returned %d, errno=%d\n",
479 info->src, rc, errno);
486 while (start < len) {
488 rc = lgetxattr(info->src, info->xlist + start,
490 if (info->xvalue == NULL || errno == ERANGE) {
491 size = rc > PATH_MAX ? rc : PATH_MAX;
492 info->xvalue = lr_grow_buf(info->xvalue, size);
493 if (info->xvalue == NULL)
496 rc = lgetxattr(info->src, info->xlist + start,
499 lr_debug(DTRACE, "\t(%s,%d) rc=%p\n", info->xlist + start,
503 rc = lsetxattr(info->dest, info->xlist + start,
504 info->xvalue, size, 0);
505 lr_debug(DTRACE, "\tlsetxattr(), rc=%d, errno=%d\n",
508 if (errno != ENOTSUP) {
509 fprintf(stderr, "Error replicating "
510 " xattr for %s: %d\n",
517 start += strlen(info->xlist + start) + 1;
520 lr_debug(DINFO, "setxattr: %s %s\n", info->src, info->dest);
525 /* Retrieve the filesystem path for a given FID and a given
526 linkno. The path is returned in info->path */
527 int lr_get_path_ln(struct lr_info *info, char *fidstr, int linkno)
529 long long recno = -1;
532 rc = llapi_fid2path(status->ls_source, fidstr, info->path,
533 PATH_MAX, &recno, &linkno);
534 if (rc < 0 && rc != -ENOENT) {
535 fprintf(stderr, "fid2path error: (%s, %s) %d %s\n",
536 status->ls_source, fidstr, -rc, strerror(errno = -rc));
542 /* Retrieve the filesystem path for a given FID. The path is returned
544 int lr_get_path(struct lr_info *info, char *fidstr)
546 return lr_get_path_ln(info, fidstr, 0);
549 /* Generate the path for opening by FID */
550 void lr_get_FID_PATH(char *mntpt, char *fidstr, char *buf, int bufsize)
552 /* Open-by-FID path is <mntpt>/.lustre/fid/[SEQ:OID:VER] */
553 snprintf(buf, bufsize, "%s/%s/fid/%s", mntpt, dot_lustre_name,
558 /* Read the symlink information into 'info->link' */
559 int lr_get_symlink(struct lr_info *info)
564 lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
565 rc = readlink(info->src, info->linktmp, PATH_MAX);
567 info->linktmp[rc] = '\0';
570 lr_debug(DTRACE, "symlink: readlink returned %s\n", info->linktmp);
572 if (strncmp(info->linktmp, status->ls_source,
573 strlen(status->ls_source)) == 0) {
574 /* Strip source fs path and replace with target fs path. */
575 link = info->linktmp + strlen(status->ls_source);
576 snprintf(info->src, sizeof(info->src), "%s%s",
577 status->ls_targets[info->target_no], link);
580 link = info->linktmp;
582 strlcpy(info->link, link, sizeof(info->link));
587 /* Create file/directory/device file/symlink. */
588 int lr_mkfile(struct lr_info *info)
594 lr_debug(DINFO, "mkfile(%d) %s \n", info->type, info->dest);
595 if (info->type == CL_MKDIR) {
596 rc = mkdir(info->dest, 0777);
597 } else if (info->type == CL_SOFTLINK) {
598 lr_get_symlink(info);
599 rc = symlink(info->link, info->dest);
600 } else if (info->type == CL_MKNOD) {
601 lr_get_FID_PATH(status->ls_source, info->tfid,
602 info->src, PATH_MAX);
603 rc = stat(info->src, &st);
610 rc = mknod(info->dest, st.st_mode, st.st_rdev);
612 rc = mknod(info->dest, S_IFREG | 0777, 0);
622 /* Sync data and attributes */
623 if (info->type == CL_CREATE || info->type == CL_MKDIR) {
624 lr_debug(DTRACE, "Syncing data and attributes %s\n",
626 (void) lr_copy_xattr(info);
627 if (info->type == CL_CREATE)
628 rc = lr_sync_data(info);
630 rc = lr_copy_attr(info->src, info->dest);
633 /* Source file has disappeared. Not an error. */
636 lr_debug(DTRACE, "Not syncing data and attributes %s\n",
643 int lr_add_pc(const char *pfid, const char *tfid, const char *name)
645 struct lr_parent_child_list *p;
648 p = calloc(1, sizeof(*p));
651 len = strlcpy(p->pc_log.pcl_pfid, pfid, sizeof(p->pc_log.pcl_pfid));
652 if (len >= sizeof(p->pc_log.pcl_pfid))
654 len = strlcpy(p->pc_log.pcl_tfid, tfid, sizeof(p->pc_log.pcl_tfid));
655 if (len >= sizeof(p->pc_log.pcl_tfid))
657 len = strlcpy(p->pc_log.pcl_name, name, sizeof(p->pc_log.pcl_name));
658 if (len >= sizeof(p->pc_log.pcl_name))
661 p->pc_next = parents;
670 void lr_cascade_move(const char *fid, const char *dest, struct lr_info *info)
672 struct lr_parent_child_list *curr, *prev;
673 char d[4 * PATH_MAX + 1];
676 prev = curr = parents;
678 if (strcmp(curr->pc_log.pcl_pfid, fid) == 0) {
679 if (snprintf(d, sizeof(d), "%s/%s", dest,
680 curr->pc_log.pcl_name) >= sizeof(d)) {
681 fprintf(stderr, "Buffer truncated\n");
684 if (snprintf(info->src, sizeof(info->src), "%s/%s/%s",
685 status->ls_targets[info->target_no],
686 SPECIAL_DIR, curr->pc_log.pcl_tfid) >=
689 rc = rename(info->src, d);
691 fprintf(stderr, "Error renaming file "
693 info->src, d, errno);
697 parents = curr->pc_next;
699 prev->pc_next = curr->pc_next;
700 lr_cascade_move(curr->pc_log.pcl_tfid, d, info);
702 prev = curr = parents;
706 curr = curr->pc_next;
711 /* remove [info->spfid, info->sfid] from parents */
712 int lr_remove_pc(const char *pfid, const char *tfid)
714 struct lr_parent_child_list *curr, *prev;
716 for (prev = curr = parents; curr; prev = curr, curr = curr->pc_next) {
717 if (strcmp(curr->pc_log.pcl_pfid, pfid) == 0 &&
718 strcmp(curr->pc_log.pcl_tfid, tfid) == 0) {
720 parents = curr->pc_next;
722 prev->pc_next = curr->pc_next;
730 /* Create file under SPECIAL_DIR with its tfid as its name. */
731 int lr_mk_special(struct lr_info *info)
735 snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
736 status->ls_targets[info->target_no], SPECIAL_DIR,
739 rc = lr_mkfile(info);
743 rc = lr_add_pc(info->pfid, info->tfid, info->name);
747 /* Remove a file or directory */
748 int lr_rmfile(struct lr_info *info)
752 if (info->type == CL_RMDIR)
753 rc = rmdir(info->dest);
755 rc = unlink(info->dest);
761 /* Recursively remove directory and its contents */
762 int lr_rm_recursive(struct lr_info *info)
766 snprintf(info->cmd, sizeof(info->cmd), "rm -rf %s",
768 rc = system(info->cmd);
775 /* Remove a file under SPECIAL_DIR with its tfid as its name. */
776 int lr_rm_special(struct lr_info *info)
780 snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
781 status->ls_targets[info->target_no], SPECIAL_DIR,
783 rc = lr_rmfile(info);
786 lr_debug(DINFO, "remove: %s; rc=%d, errno=%d\n",
787 info->dest, rc, errno);
791 /* Replicate file and directory create events */
792 int lr_create(struct lr_info *info)
799 /* Is target FID present on the source? */
800 rc = lr_get_path(info, info->tfid);
802 /* Source file has disappeared. Not an error. */
803 lr_debug(DINFO, "create: tfid %s not found on"
804 "source-fs\n", info->tfid);
809 strcpy(info->savedpath, info->path);
811 /* Is parent FID present on the source */
812 rc = lr_get_path(info, info->pfid);
814 lr_debug(DINFO, "create: pfid %s not found on source-fs\n",
821 /* Is f2p(pfid)+name != f2p(tfid)? If not the file has moved. */
822 len = strlen(info->path);
823 if (len == 1 && info->path[0] == '/')
824 snprintf(info->dest, sizeof(info->dest), "%s", info->name);
825 else if (len - 1 > 0 && info->path[len - 1] == '/')
826 snprintf(info->dest, sizeof(info->dest), "%s%s", info->path,
829 snprintf(info->dest, sizeof(info->dest), "%s/%s", info->path,
832 lr_debug(DTRACE, "dest = %s; savedpath = %s\n", info->dest,
834 if (strncmp(info->dest, info->savedpath, PATH_MAX) != 0) {
835 lr_debug(DTRACE, "create: file moved (%s). %s != %s\n",
836 info->tfid, info->dest, info->savedpath);
840 /* Is f2p(pfid) present on the target? If not, the parent has
843 snprintf(info->dest, sizeof(info->dest), "%s/%s",
844 status->ls_targets[0], info->path);
845 if (access(info->dest, F_OK) != 0) {
846 lr_debug(DTRACE, "create: parent %s not found\n",
851 for (info->target_no = 0; info->target_no < status->ls_num_targets;
853 snprintf(info->dest, sizeof(info->dest), "%s/%s",
854 status->ls_targets[info->target_no], info->savedpath);
855 lr_get_FID_PATH(status->ls_source, info->tfid, info->src,
859 rc1 = lr_mkfile(info);
860 if (mkspecial || rc1 == -ENOENT) {
861 rc1 = lr_mk_special(info);
869 /* Replicate a file remove (rmdir/unlink) operation */
870 int lr_remove(struct lr_info *info)
875 for (info->target_no = 0; info->target_no < status->ls_num_targets;
878 rc1 = lr_rm_special(info);
882 rc1 = lr_get_path(info, info->pfid);
883 if (rc1 == -ENOENT) {
884 lr_debug(DINFO, "remove: pfid %s not found\n",
892 snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
893 status->ls_targets[info->target_no], info->path,
896 rc1 = lr_rmfile(info);
897 lr_debug(DINFO, "remove: %s; rc1=%d, errno=%d\n",
898 info->dest, rc1, errno);
899 if (rc1 == -ENOTEMPTY)
900 rc1 = lr_rm_recursive(info);
910 /* Replicate a rename/move operation. */
911 int lr_move(struct lr_info *info)
917 int special_dest = 0;
918 char srcpath[PATH_MAX + 1] = "";
920 assert(info->is_extended);
922 rc_src = lr_get_path(info, info->spfid);
923 if (rc_src < 0 && rc_src != -ENOENT)
925 memcpy(srcpath, info->path, strlen(info->path));
927 rc_dest = lr_get_path(info, info->pfid);
928 if (rc_dest < 0 && rc_dest != -ENOENT)
931 for (info->target_no = 0; info->target_no < status->ls_num_targets;
935 snprintf(info->dest, sizeof(info->dest), "%s/%s",
936 status->ls_targets[info->target_no],
938 if (access(info->dest, F_OK) != 0) {
941 snprintf(info->dest, sizeof(info->dest),
943 status->ls_targets[info->target_no],
944 info->path, info->name);
946 lr_debug(DINFO, "dest path %s rc_dest=%d\n", info->dest,
949 if (rc_dest == -ENOENT) {
950 snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
951 status->ls_targets[info->target_no],
952 SPECIAL_DIR, info->sfid);
954 lr_debug(DINFO, "special dest %s\n", info->dest);
958 snprintf(info->src, sizeof(info->src), "%s/%s/%s",
959 status->ls_targets[info->target_no],
960 srcpath, info->sname);
961 lr_debug(DINFO, "src path %s rc_src=%d\n", info->src,
964 if (rc_src == -ENOENT || (access(info->src, F_OK) != 0 &&
966 snprintf(info->src, sizeof(info->src), "%s/%s/%s",
967 status->ls_targets[info->target_no],
968 SPECIAL_DIR, info->sfid);
970 lr_debug(DINFO, "special src %s\n", info->src);
975 if (strcmp(info->src, info->dest) != 0) {
976 rc1 = rename(info->src, info->dest);
979 lr_debug(DINFO, "rename returns %d\n", rc1);
983 rc1 = lr_remove_pc(info->spfid, info->sfid);
986 lr_cascade_move(info->sfid, info->dest, info);
988 rc1 = lr_add_pc(info->pfid, info->sfid, info->name);
990 lr_debug(DINFO, "move: %s [to] %s rc1=%d, errno=%d\n",
991 info->src, info->dest, rc1, errno);
998 /* Replicate a hard link */
999 int lr_link(struct lr_info *info)
1006 lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
1007 rc = stat(info->src, &st);
1011 for (info->target_no = 0; info->target_no < status->ls_num_targets;
1012 info->target_no++) {
1019 * The changelog record has the new parent directory FID and
1020 * name of the target file. So info->dest can be constructed
1021 * by getting the path of the new parent directory and
1022 * appending the target file name.
1024 rc1 = lr_get_path(info, info->pfid);
1025 lr_debug(rc1 ? 0 : DTRACE, "\tparent fid2path %s, %s, rc=%d\n",
1026 info->path, info->name, rc1);
1029 snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
1030 status->ls_targets[info->target_no],
1031 info->path, info->name);
1032 lr_debug(DINFO, "link destination is %s\n", info->dest);
1035 /* Search through the hardlinks to get the src */
1036 for (i = 0; i < st.st_nlink && info->src[0] == 0; i++) {
1039 rc1 = lr_get_path_ln(info, info->tfid, i);
1040 lr_debug(rc1 ? 0:DTRACE, "\tfid2path %s, %s, %d rc=%d\n",
1041 info->path, info->name, i, rc1);
1046 * Compare the path of target FID with info->dest
1047 * to find out info->src.
1049 len = sizeof(status->ls_targets[info->target_no]) +
1051 char srcpath[len + 1];
1053 snprintf(srcpath, sizeof(srcpath), "%s/%s",
1054 status->ls_targets[info->target_no],
1057 if (strcmp(srcpath, info->dest) != 0) {
1058 strlcpy(info->src, srcpath, sizeof(info->src));
1059 lr_debug(DINFO, "link source is %s\n",
1069 if (info->src[0] == 0)
1070 snprintf(info->src, sizeof(info->src), "%s/%s/%s",
1071 status->ls_targets[info->target_no],
1072 SPECIAL_DIR, info->tfid);
1073 else if (info->dest[0] == 0)
1074 snprintf(info->dest, sizeof(info->dest), "%s/%s/%s",
1075 status->ls_targets[info->target_no],
1076 SPECIAL_DIR, info->tfid);
1078 rc1 = link(info->src, info->dest);
1079 lr_debug(DINFO, "link: %s [to] %s; rc1=%d %s\n",
1080 info->src, info->dest, rc1,
1081 strerror(rc1 ? errno : 0));
1089 /* Replicate file attributes */
1090 int lr_setattr(struct lr_info *info)
1095 lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
1097 rc = lr_get_path(info, info->tfid);
1099 lr_debug(DINFO, "setattr: %s not present on source-fs\n",
1104 for (info->target_no = 0; info->target_no < status->ls_num_targets;
1105 info->target_no++) {
1107 snprintf(info->dest, sizeof(info->dest), "%s/%s",
1108 status->ls_targets[info->target_no], info->path);
1109 lr_debug(DINFO, "setattr: %s %s %s", info->src, info->dest,
1112 rc1 = lr_sync_data(info);
1114 rc1 = lr_copy_attr(info->src, info->dest);
1121 /* Replicate xattrs */
1122 int lr_setxattr(struct lr_info *info)
1126 lr_get_FID_PATH(status->ls_source, info->tfid, info->src, PATH_MAX);
1128 rc = lr_get_path(info, info->tfid);
1130 lr_debug(DINFO, "setxattr: %s not present on source-fs\n",
1135 for (info->target_no = 0; info->target_no < status->ls_num_targets;
1136 info->target_no++) {
1138 snprintf(info->dest, sizeof(info->dest), "%s/%s",
1139 status->ls_targets[info->target_no], info->path);
1140 lr_debug(DINFO, "setxattr: %s %s %s\n", info->src, info->dest,
1143 rc1 = lr_copy_xattr(info);
1151 /* Parse a line of changelog entry */
1152 int lr_parse_line(void *priv, struct lr_info *info)
1154 struct changelog_rec *rec;
1155 struct changelog_ext_rename *rnm;
1157 size_t copylen = sizeof(info->name);
1159 if (llapi_changelog_recv(priv, &rec) != 0)
1162 info->is_extended = !!(rec->cr_flags & CLF_RENAME);
1163 info->recno = rec->cr_index;
1164 info->type = rec->cr_type;
1165 snprintf(info->tfid, sizeof(info->tfid), DFID, PFID(&rec->cr_tfid));
1166 snprintf(info->pfid, sizeof(info->pfid), DFID, PFID(&rec->cr_pfid));
1168 namelen = strnlen(changelog_rec_name(rec), rec->cr_namelen);
1169 if (copylen > namelen + 1)
1170 copylen = namelen + 1;
1171 strlcpy(info->name, changelog_rec_name(rec), copylen);
1173 /* Don't use rnm if CLF_RENAME isn't set */
1174 rnm = changelog_rec_rename(rec);
1175 if (rec->cr_flags & CLF_RENAME && !fid_is_zero(&rnm->cr_sfid)) {
1176 copylen = sizeof(info->sname);
1178 snprintf(info->sfid, sizeof(info->sfid), DFID,
1179 PFID(&rnm->cr_sfid));
1180 snprintf(info->spfid, sizeof(info->spfid), DFID,
1181 PFID(&rnm->cr_spfid));
1182 namelen = changelog_rec_snamelen(rec);
1183 if (copylen > namelen + 1)
1184 copylen = namelen + 1;
1185 strlcpy(info->sname, changelog_rec_sname(rec), copylen);
1188 printf("Rec %lld: %d %s %s\n", info->recno, info->type,
1189 info->name, info->sname);
1192 printf("Rec %lld: %d %s\n", info->recno, info->type,
1196 llapi_changelog_free(&rec);
1202 /* Initialize the replication parameters */
1203 int lr_init_status()
1205 size_t size = sizeof(struct lustre_rsync_status) + PATH_MAX + 1;
1209 status = calloc(size, 1);
1212 status->ls_version = REPLICATE_STATUS_VER;
1213 status->ls_size = size;
1214 status->ls_last_recno = -1;
1218 /* Make a backup of the statuslog */
1219 void lr_backup_log()
1221 char backupfile[PATH_MAX];
1225 snprintf(backupfile, sizeof(backupfile), "%s.old", statuslog);
1226 (void) rename(statuslog, backupfile);
1232 /* Save replication parameters to a statuslog. */
1237 size_t write_size = status->ls_size;
1238 struct lr_parent_child_list *curr;
1241 if (statuslog == NULL)
1246 fd = open(statuslog, O_WRONLY | O_CREAT | O_SYNC,
1249 fprintf(stderr, "Error opening log file for writing (%s)\n",
1254 size = write(fd, status, write_size);
1255 if (size != write_size) {
1256 fprintf(stderr, "Error writing to log file (%s) %d\n",
1262 for (curr = parents; curr; curr = curr->pc_next) {
1263 size = write(fd, &curr->pc_log, sizeof(curr->pc_log));
1264 if (size != sizeof(curr->pc_log)) {
1265 fprintf(stderr, "Error writing to log file (%s) %d\n",
1275 /* Read statuslog and populate the replication parameters. Command
1276 * line parameters take precedence over parameters in the log file.*/
1279 struct lr_parent_child_list *tmp;
1280 struct lr_parent_child_log rec;
1281 struct lustre_rsync_status *s;
1284 size_t read_size = sizeof(struct lustre_rsync_status) + PATH_MAX + 1;
1287 if (statuslog == NULL)
1290 s = calloc(1, read_size);
1296 fd = open(statuslog, O_RDONLY);
1302 size = read(fd, s, read_size);
1303 if (size != read_size) {
1308 if (read_size < s->ls_size) {
1309 read_size = s->ls_size;
1310 s = lr_grow_buf(s, read_size);
1316 if (lseek(fd, 0, SEEK_SET) == -1) {
1321 size = read(fd, s, read_size);
1322 if (size != read_size) {
1328 while (read(fd, &rec, sizeof(rec)) != 0) {
1329 tmp = calloc(1, sizeof(*tmp));
1336 tmp->pc_next = parents;
1340 /* copy uninitialized fields to status */
1341 if (status->ls_num_targets == 0) {
1342 if (status->ls_size != s->ls_size) {
1343 status = lr_grow_buf(status, s->ls_size);
1344 if (status == NULL) {
1349 status->ls_size = s->ls_size;
1351 status->ls_num_targets = s->ls_num_targets;
1352 memcpy(status->ls_targets, s->ls_targets,
1353 (PATH_MAX + 1) * s->ls_num_targets);
1355 if (status->ls_last_recno == -1)
1356 status->ls_last_recno = s->ls_last_recno;
1358 if (status->ls_registration[0] == '\0')
1359 strlcpy(status->ls_registration, s->ls_registration,
1360 sizeof(status->ls_registration));
1362 if (status->ls_mdt_device[0] == '\0')
1363 strlcpy(status->ls_mdt_device, s->ls_mdt_device,
1364 sizeof(status->ls_mdt_device));
1366 if (status->ls_source_fs[0] == '\0')
1367 strlcpy(status->ls_source_fs, s->ls_source_fs,
1368 sizeof(status->ls_source_fs));
1370 if (status->ls_source[0] == '\0')
1371 strlcpy(status->ls_source, s->ls_source,
1372 sizeof(status->ls_source));
1382 /* Clear changelogs every CLEAR_INTERVAL records or at the end of
1384 int lr_clear_cl(struct lr_info *info, int force)
1386 char mdt_device[LR_NAME_MAXLEN + 1];
1390 if (force || info->recno > status->ls_last_recno + CLEAR_INTERVAL) {
1391 if (info->type == CL_RENAME)
1392 rec = info->recno + 1;
1395 if (!noclear && !dryrun) {
1396 /* llapi_changelog_clear modifies the mdt
1397 * device name so make a copy of it until this
1400 strlcpy(mdt_device, status->ls_mdt_device,
1401 sizeof(mdt_device));
1402 rc = llapi_changelog_clear(mdt_device,
1403 status->ls_registration,
1406 printf("Changelog clear (%s, %s, %lld) "
1407 "returned %d\n", status->ls_mdt_device,
1408 status->ls_registration, rec, rc);
1410 if (!rc && !dryrun) {
1411 status->ls_last_recno = rec;
1420 /* Locate a usable version of rsync. At this point we'll use any
1422 int lr_locate_rsync()
1428 snprintf(rsync, sizeof(rsync), "%s -p %s", TYPE, RSYNC);
1429 fp = popen(rsync, "r");
1433 if (fgets(rsync, sizeof(rsync), fp) == NULL) {
1438 len = strlen(rsync);
1439 if (len > 0 && rsync[len - 1] == '\n')
1440 rsync[len - 1] = '\0';
1443 /* Determine the version of rsync */
1444 snprintf(rsync_ver, sizeof(rsync_ver), "%s --version", rsync);
1445 fp = popen(rsync_ver, "r");
1449 if (fgets(rsync_ver, sizeof(rsync_ver), fp) == NULL) {
1453 len = strlen(rsync_ver);
1454 if (len > 0 && rsync_ver[len - 1] == '\n')
1455 rsync_ver[len - 1] = '\0';
1462 /* Print the replication parameters */
1463 void lr_print_status(struct lr_info *info)
1470 printf("Lustre filesystem: %s\n", status->ls_source_fs);
1471 printf("MDT device: %s\n", status->ls_mdt_device);
1472 printf("Source: %s\n", status->ls_source);
1473 for (i = 0; i < status->ls_num_targets; i++)
1474 printf("Target: %s\n", status->ls_targets[i]);
1475 if (statuslog != NULL)
1476 printf("Statuslog: %s\n", statuslog);
1477 printf("Changelog registration: %s\n", status->ls_registration);
1478 printf("Starting changelog record: %jd\n",
1479 (uintmax_t)status->ls_last_recno);
1481 printf("Replicate xattrs: no\n");
1483 printf("Clear changelog after use: no\n");
1485 printf("Using rsync: %s (%s)\n", rsync, rsync_ver);
1488 void lr_print_failure(struct lr_info *info, int rc)
1490 fprintf(stderr, "Replication of operation failed(%d):"
1491 " %lld %s (%d) %s %s %s\n", rc, info->recno,
1492 changelog_type2str(info->type), info->type, info->tfid,
1493 info->pfid, info->name);
1496 /* Replicate filesystem operations from src_path to target_path */
1499 void *changelog_priv;
1500 struct lr_info *info;
1501 struct lr_info *ext = NULL;
1509 info = calloc(1, sizeof(struct lr_info));
1513 rc = llapi_search_fsname(status->ls_source, status->ls_source_fs);
1515 fprintf(stderr, "Source path is not a valid Lustre client "
1520 if (status->ls_mdt_device[0] == '\0') {
1523 len = snprintf(status->ls_mdt_device,
1524 sizeof(status->ls_mdt_device), "%s%s",
1525 status->ls_source_fs, DEFAULT_MDT);
1526 if (len >= sizeof(status->ls_mdt_device)) {
1532 ext = calloc(1, sizeof(struct lr_info));
1538 for (i = 0, xattr_not_supp = 0; i < status->ls_num_targets; i++) {
1539 snprintf(info->dest, sizeof(info->dest), "%s/%s",
1540 status->ls_targets[i], SPECIAL_DIR);
1541 rc = mkdir(info->dest, 0777);
1542 if (rc == -1 && errno != EEXIST) {
1543 fprintf(stderr, "Error writing to target path %s.\n",
1544 status->ls_targets[i]);
1548 rc = llistxattr(info->src, info->xlist, info->xsize);
1549 if (rc == -1 && errno == ENOTSUP) {
1550 fprintf(stderr, "xattrs not supported on %s\n",
1551 status->ls_targets[i]);
1555 if (xattr_not_supp == status->ls_num_targets)
1556 /* None of the targets support xattrs. */
1559 lr_print_status(info);
1561 /* Open changelogs for consumption*/
1562 rc = llapi_changelog_start(&changelog_priv,
1563 CHANGELOG_FLAG_BLOCK |
1564 CHANGELOG_FLAG_JOBID |
1565 CHANGELOG_FLAG_EXTRA_FLAGS,
1566 status->ls_mdt_device, status->ls_last_recno);
1568 fprintf(stderr, "Error opening changelog file for fs %s.\n",
1569 status->ls_source_fs);
1573 rc = llapi_changelog_set_xflags(changelog_priv,
1574 CHANGELOG_EXTRA_FLAG_UIDGID);
1576 fprintf(stderr, "Error setting xflag in changelog for fs %s.\n",
1577 status->ls_source_fs);
1581 while (!quit && lr_parse_line(changelog_priv, info) == 0) {
1583 if (info->type == CL_RENAME && !info->is_extended) {
1584 /* Newer rename operations extends changelog to store
1585 * source file information, but old changelog has
1588 if (lr_parse_line(changelog_priv, ext) != 0)
1590 memcpy(info->sfid, info->tfid, sizeof(info->sfid));
1591 memcpy(info->spfid, info->pfid, sizeof(info->spfid));
1592 memcpy(info->tfid, ext->tfid, sizeof(info->tfid));
1593 memcpy(info->pfid, ext->pfid, sizeof(info->pfid));
1594 strlcpy(info->sname, info->name, sizeof(info->sname));
1595 strlcpy(info->name, ext->name, sizeof(info->name));
1596 info->is_extended = 1;
1604 switch(info->type) {
1609 rc = lr_create(info);
1613 rc = lr_remove(info);
1623 rc = lr_setattr(info);
1626 rc = lr_setxattr(info);
1633 /* Nothing needs to be done for these entries */
1638 DEBUG_EXIT(info, rc);
1639 if (rc && rc != -ENOENT) {
1640 lr_print_failure(info, rc);
1645 lr_clear_cl(info, 0);
1647 bzero(info, sizeof(struct lr_info));
1648 bzero(ext, sizeof(struct lr_info));
1652 llapi_changelog_fini(&changelog_priv);
1654 if (errors || verbose)
1655 printf("Errors: %d\n", errors);
1657 /* Clear changelog records used so far */
1658 lr_clear_cl(info, 1);
1661 printf("lustre_rsync took %ld seconds\n", time(NULL) - start);
1662 printf("Changelog records consumed: %lld\n", rec_count);
1677 termination_handler (int signum)
1679 /* Set a flag for the replicator to gracefully shutdown */
1681 printf("lustre_rsync halting.\n");
1684 int main(int argc, char *argv[])
1690 if ((rc = lr_init_status()) != 0)
1693 while ((rc = getopt_long(argc, argv, "as:t:m:u:l:vx:zc:ry:n:d:D:",
1694 long_opts, NULL)) >= 0) {
1697 /* Assume absolute paths */
1701 /* Assume absolute paths */
1702 strlcpy(status->ls_source, optarg,
1703 sizeof(status->ls_source));
1706 status->ls_num_targets++;
1708 if (numtargets != status->ls_num_targets) {
1709 /* Targets were read from a log
1710 file. The ones specified on the
1711 command line take precedence. The
1712 ones from the log file will be
1714 status->ls_num_targets = numtargets;
1716 newsize = sizeof (struct lustre_rsync_status) +
1717 (status->ls_num_targets * (PATH_MAX + 1));
1718 if (status->ls_size != newsize) {
1719 status->ls_size = newsize;
1720 status = lr_grow_buf(status, newsize);
1724 strlcpy(status->ls_targets[status->ls_num_targets - 1],
1725 optarg, sizeof(status->ls_targets[0]));
1728 strlcpy(status->ls_mdt_device, optarg,
1729 sizeof(status->ls_mdt_device));
1732 strlcpy(status->ls_registration, optarg,
1733 sizeof(status->ls_registration));
1737 (void) lr_read_log();
1743 if (strcmp("no", optarg) == 0) {
1745 } else if (strcmp("yes", optarg) != 0) {
1746 printf("Invalid parameter %s. "
1747 "Specify --xattr=no or --xattr=yes\n",
1756 /* Undocumented option cl-clear */
1757 if (strcmp("no", optarg) == 0) {
1759 } else if (strcmp("yes", optarg) != 0) {
1760 printf("Invalid parameter %s. "
1761 "Specify --cl-clear=no "
1762 "or --cl-clear=yes\n",
1768 /* Undocumented option use-rsync */
1772 /* Undocumented option rsync-threshold */
1773 rsync_threshold = atol(optarg);
1776 /* Undocumented option start-recno */
1777 status->ls_last_recno = atol(optarg);
1780 /* Undocumented option debug */
1781 debug = atoi(optarg);
1782 if (debug < 0 || debug > 2)
1786 /* Undocumented option debug log file */
1787 if (debug_log != NULL)
1789 debug_log = fopen(optarg, "a");
1790 if (debug_log == NULL) {
1791 printf("Cannot open %s for debug log\n",
1797 fprintf(stderr, "error: %s: option '%s' "
1798 "unrecognized.\n", argv[0], argv[optind - 1]);
1804 if (status->ls_last_recno == -1)
1805 status->ls_last_recno = 0;
1806 if (strnlen(status->ls_registration, LR_NAME_MAXLEN) == 0) {
1807 /* No registration ID was passed in. */
1808 printf("Please specify changelog consumer registration id.\n");
1812 if (strnlen(status->ls_source, PATH_MAX) == 0) {
1813 fprintf(stderr, "Please specify the source path.\n");
1817 if (strnlen(status->ls_targets[0], PATH_MAX) == 0) {
1818 fprintf(stderr, "Please specify the target path.\n");
1823 /* This plumbing is needed for some of the ioctls behind
1824 llapi calls to work. */
1825 register_ioc_dev(OBD_DEV_ID, OBD_DEV_PATH);
1827 rc = lr_locate_rsync();
1828 if (use_rsync && rc != 0) {
1829 fprintf(stderr, "Error: unable to locate %s.\n", RSYNC);
1833 signal(SIGINT, termination_handler);
1834 signal(SIGHUP, termination_handler);
1835 signal(SIGTERM, termination_handler);
1837 rc = lr_replicate();
1839 if (debug_log != NULL)