3 # lfs_migrate: a simple tool to copy and check files.
5 # To avoid allocating objects on one or more OSTs, they should be
6 # deactivated on the MDS via "lctl --device {device_number} deactivate",
7 # where {device_number} is from the output of "lctl dl" on the MDS.
9 # To guard against corruption, the file is compared after migration
10 # to verify the copy is correct and the file has not been modified.
11 # This is not a protection against the file being open by another
12 # process, but it would catch the worst cases of in-use files, but
13 # to be 100% safe the administrator needs to ensure this is safe.
16 OPT_RSYNC=${LFS_MIGRATE_RSYNC_MODE:-false}
19 RSYNC_WITH_HLINKS=false
20 LFS_MIGRATE_TMP=${TMPDIR:-/tmp}
21 MIGRATED_SET="$(mktemp ${LFS_MIGRATE_TMP}/lfs_migrate.links.XXXXXX)"
23 REMOVE_FID='s/^\[[0-9a-fx:]*\] //'
30 echo "$old_fid $path" >> "$MIGRATED_SET"
36 sed -e "$REMOVE_FID" $MIGRATED_SET | grep -q "^$path$"
42 grep "^\\$old_fid" "$MIGRATED_SET" | head -n 1 |
48 usage: lfs_migrate [--dry-run|-n] [--help|-h] [--no-rsync|--rsync] [--quiet|-q]
49 [--auto-stripe|-A [-C <cap>]
50 [--min-free|-M <min_free>] [--max-free|-X <max_free>]]
51 [--stripe-count|-c <stripe_count>]
52 [--stripe-size|-S <stripe_size>]
54 [--restripe|-R] [--skip|-s] [--verbose|-v] [--yes|-y] [-0]
56 -A restripe file using an automatically selected stripe count,
57 uses stripe_count = sqrt(size_in_GB) + 1
59 restripe file using the specified <stripe_count>
60 -C <cap> when -A is set, limit the migrated file to use on each OST
61 at most 1/<cap> of the available space of the smallest OST
62 -D do not use direct I/O to copy file contents
63 -h show this usage message
65 when -A is set, an OST must contain more available space than
66 <min_free> KB in order for it to be considered available for
68 --no-rsync do not fall back to rsync mode even if lfs migrate fails
69 -n only print the names of files to be migrated
70 -q run quietly (don't print filenames or status)
71 --rsync force rsync mode instead of using lfs migrate
72 -R restripe file using default directory striping
73 -s skip file data comparison after migrate
75 restripe file using the specified stripe size
76 -v show verbose debug messages
78 when -A is set, limit the amount of space on each OST that
79 can be considered available for the migration to
81 -y answer 'y' to usage question
82 -0 input file names on stdin are separated by a null character
84 Options '-A', '-c', and '-R' are mutually exclusive.
85 Options '-C', '-M', and '-X' are ignored if '-A' is not set.
87 The --rsync and --no-rsync options may not be specified at the same time.
89 If a directory is an argument, all files in the directory are migrated.
90 If no file/directory is given, the file list is read from standard input.
92 Any arguments that are not explicitly recognized by the script are passed
93 through to the 'lfs migrate' utility.
96 lfs_migrate /mnt/lustre/dir
97 lfs_migrate -p newpool /mnt/lustre/dir
98 lfs find /test -O test-OST0004 -size +4G | lfs_migrate -y
104 rm -f "$MIGRATED_SET"
105 [ -n "$NEWNAME" ] && rm -f "$NEWNAME"
129 # Examine any long options and arguments. getopts does not support long
130 # options, so they must be stripped out and classified as either options
131 # for the script, or passed through to "lfs migrate".
132 while [ -n "$*" ]; do
136 -l|--link) ;; # maintained backward compatibility for now
137 -n|--dry-run) OPT_DRYRUN=true; OPT_YES=true
138 echo "$PROG: -n deprecated, use --dry-run or --non-block" 1>&2;;
140 -R|--restripe) OPT_RESTRIPE=true;;
141 -s|--skip) OPT_CHECK=false;;
142 -v|--verbose) OPT_DEBUG=true; ECHO=echo;;
143 -y|--yes) OPT_YES=true;;
145 -b|--block|--non-block|--non-direct|--no-verify)
146 # Always pass non-layout options to 'lfs migrate'
147 OPT_PASSTHROUGH+=("$arg");;
148 --rsync) OPT_RSYNC=true;;
149 --no-rsync) OPT_NO_RSYNC=true;;
150 --copy|--yaml|--file)
151 # these options have files as arguments, pass both through
152 OPT_LAYOUT+="$arg $2"; shift;;
153 --auto-stripe|-A) OPT_AUTOSTRIPE=true;;
154 -C) OPT_CAP="$2"; shift;;
155 -D) LFS_OPT_DIRECTIO="-D";;
156 -M|--min-free) OPT_MINFREE="$2"; shift;;
157 -X|--max-free) OPT_MAXFREE="$2"; shift;;
158 -c|--stripe-count) OPT_STRIPE_COUNT="$2"; shift;;
159 -S|--stripe-size) OPT_STRIPE_SIZE="$2"; shift;;
160 *) # Pass other non-file layout options to 'lfs migrate'
161 [ -e "$arg" ] && OPT_FILE+="$arg " && break || OPT_LAYOUT+="$arg "
166 if $OPT_RESTRIPE && [ -n "$OPT_LAYOUT" ]; then
167 echo "$PROG: Options $OPT_LAYOUT cannot be used with the -R option" 1>&2
169 elif $OPT_RESTRIPE && [[ "$OPT_STRIPE_COUNT" || "$OPT_STRIPE_SIZE" ]]; then
170 echo "$(basename $0): Options -c <stripe_count> and -S <stripe_size> "\
171 "may not be specified at the same time as the -R option." 1>&2
173 elif $OPT_AUTOSTRIPE && [ -n "$OPT_STRIPE_COUNT" ]; then
175 echo "$(basename $0) error: The -c <stripe_count> option may not" 1>&2
176 echo "be specified at the same time as the -A option." 1>&2
178 elif $OPT_AUTOSTRIPE && $OPT_RESTRIPE; then
180 echo "$(basename $0) error: The -A option may not be specified at" 1>&2
181 echo "the same time as the -R option." 1>&2
185 if $OPT_RSYNC && $OPT_NO_RSYNC; then
186 echo "$PROG: Options --rsync and --no-rsync may not be" \
187 "specified at the same time." 1>&2
193 echo "lfs_migrate is currently NOT SAFE for moving in-use files." 1>&2
194 echo "Use it only when you are sure migrated files are unused." 1>&2
196 echo "If emptying an OST that is active on the MDS, new files may" 1>&2
197 echo "use it. To stop allocating any new objects on OSTNNNN run:" 1>&2
198 echo " lctl set_param osp.<fsname>-OSTNNNN*.max_create_count=0'" 1>&2
199 echo "on each MDS using the OST(s) being emptied." 1>&2
200 echo -n "Continue? (y/n) "
202 [ "$CHECK" != "y" -a "$CHECK" != "yes" ] && exit 1
205 # if rsync has --xattr support, then try to copy the xattrs.
206 $RSYNC --help 2>&1 | grep -q xattr && RSYNC_OPTS="$RSYNC_OPTS -X"
207 $RSYNC --help 2>&1 | grep -q acls && RSYNC_OPTS="$RSYNC_OPTS -A"
208 # If rsync copies lustre xattrs in the future, then we can skip lfs (bug 22189)
209 strings $(which $RSYNC) 2>&1 | grep -q lustre && LFS=:
211 # rsync creates its temporary files with lenient permissions, even if
212 # permissions on the original files are more strict. Tighten umask here
213 # to avoid the brief window where unprivileged users might be able to
214 # access the temporary file.
217 # Use stripe count = sqrt(size_in_GB) + 1, but cap object size per OST.
218 function calc_stripe()
223 local filegb=$((filekb / 1048576))
225 local ost_max_count=0
227 # Files up to 1GB will have 1 stripe if they fit within the object max
228 if [[ $filegb -lt 1 && "$obj_max_kb" && $filekb -le $obj_max_kb ]]; then
229 echo 1 "$obj_max_kb" && return
232 stripe_count=$(bc <<< "scale=0; 1 + sqrt($filegb)" 2> /dev/null) ||
233 { echo "cannot auto calculate stripe count" >&2; return; }
235 if [ -z "$obj_max_kb" ]; then
236 local ost_min_kb=$((1 << 62))
238 # Calculate cap on object size at 1% of smallest OST
239 # but only include OSTs that have 256MB+ available space
240 while IFS='' read avail; do
241 [[ "$OPT_MAXFREE" && $avail -gt $OPT_MAXFREE ]] &&
243 if [ $avail -ge $OPT_MINFREE ]; then
244 ost_max_count=$((ost_max_count + 1))
245 if [ $avail -lt $ost_min_kb ]; then
249 done < <($LFS df $OLDNAME | awk '/OST/ { print $4 }')
250 # Once this script supports pools, the lfs df command above
251 # should also include the -p <pool> option to restrict the
252 # listed OSTs to the correct pool.
254 if [ $ost_max_count -eq 0 ]; then
255 echo "no OSTs with sufficient available space" >&2
259 if (( ost_min_kb == (1 << 62) )); then
260 echo "warning: unable to determine minimum OST size, " \
261 "object size not capped" >&2
263 echo "$stripe_count" "$obj_max_kb"
267 obj_max_kb=$((ost_min_kb / $OPT_CAP))
268 elif [ $obj_max_kb -eq 0 ]; then
269 echo "warning: unable to determine minimum OST size " \
270 "from previous migrate, object size not capped" >&2
271 echo "$stripe_count" "$obj_max_kb"
275 # If disk usage would exceed the cap, increase the number of stripes.
276 # Round up to the nearest MB to ensure file will fit.
277 (( filekb > stripe_count * obj_max_kb )) &&
278 stripe_count=$(((filekb + obj_max_kb - 1024) / obj_max_kb))
280 # Limit the count to the number of eligible OSTs
281 if [ "$stripe_count" -gt $ost_max_count ]; then
282 echo "$ost_max_count" "$obj_max_kb"
284 echo "$stripe_count" "$obj_max_kb"
289 while IFS='' read -d '' OLDNAME; do
291 local stripe_size="$OPT_STRIPE_SIZE"
292 local stripe_count="$OPT_STRIPE_COUNT"
293 local parent_count=""
299 $ECHO -n "$OLDNAME: "
301 # avoid duplicate stat if possible
302 local nlink_type=($(LANG=C stat -c "%h %F %s" "$OLDNAME" \
305 # skip non-regular files, since they don't have any objects
306 # and there is no point in trying to migrate them.
307 if [ "${nlink_type[1]}" != "regular" ]; then
308 echo -e "\r$OLDNAME: not a regular file, skipped" 1>&2
312 # working out write perms is hard, let the shell do it
313 if [ ! -w "$OLDNAME" ]; then
314 echo -e "\r$OLDNAME: no write permission, skipped" 1>&2
318 if $OPT_DRYRUN && ! $OPT_DEBUG; then
319 $ECHO "dry run, skipped"
323 # xattrs use absolute file paths, so ensure provided path is
324 # also absolute so that the names can be compared
325 local oldname_absolute=$(readlink -f "$OLDNAME")
326 if [ -z "$oldname_absolute" ]; then
327 echo -e "\r$OLDNAME: cannot resolve full path, skipped" 1>&2
330 OLDNAME=$oldname_absolute
332 # In the future, the path2fid and fid2path calls below
333 # should be replaced with a single call to
334 # "lfs path2links" once that command is available. The logic
335 # for detecting unlisted hard links could then be removed.
336 local fid=$($LFS path2fid "$OLDNAME" 2> /dev/null)
337 if [ $? -ne 0 ]; then
338 echo -e "\r$OLDNAME: cannot determine FID; skipping; " \
339 "is this a Lustre file system?" 1>&2
343 if [[ ${nlink_type[0]} -gt 1 ]] || $RSYNC_WITH_HLINKS; then
344 # don't migrate a hard link if it was already migrated
345 if path_in_set "$OLDNAME"; then
346 $ECHO "already migrated via another hard link"
350 # There is limited space available in the xattrs
351 # to store all of the hard links for a file, so it's
352 # possible that $OLDNAME is part of a link set but is
353 # not listed in xattrs and therefore not listed as
355 local migrated=$(old_fid_in_set "$fid")
356 if [ -n "$migrated" ]; then
357 $ECHO "already migrated via another hard link"
359 # Only the rsync case has to relink.
360 # The lfs migrate case preserves the
361 # inode so the links are already
363 [ "$migrated" != "$OLDNAME" ] &&
364 ln -f "$migrated" "$OLDNAME"
366 add_to_set "$fid" "$OLDNAME"
371 if $OPT_RESTRIPE; then
374 # if rsync copies Lustre xattrs properly in the future
375 # (i.e. before the file data, so that it preserves
376 # striping) then we don't need to do this getstripe
380 stripe_pool=$($LFS getstripe -p "$OLDNAME" 2> /dev/null)
381 mirror_count=$($LFS getstripe -N "$OLDFILE" 2> /dev/null)
383 if $OPT_AUTOSTRIPE; then
384 local filekb=$((${nlink_type[3]} / 1024))
386 read stripe_count OBJ_MAX_KB < <(calc_stripe \
387 "$OLDNAME" "$filekb" "$OBJ_MAX_KB")
388 [ -z "$stripe_count" ] && exit 1
389 [ $stripe_count -lt 1 ] && stripe_count=1
391 [ "$OPT_STRIPE_COUNT" ] && stripe_count=$OPT_STRIPE_COUNT ||
392 stripe_count=$($LFS getstripe -c "$OLDNAME" \
395 [ -z "$stripe_size" ] &&
396 stripe_size=$($LFS getstripe -S "$OLDNAME" 2> /dev/null)
398 [ -z "$stripe_count" -o -z "$stripe_size" ] && UNLINK=""
405 if $OPT_RESTRIPE; then
406 parent_count=$($LFS getstripe -c \
407 $(dirname "$OLDNAME") 2> \
409 parent_size=$($LFS getstripe -S \
410 $(dirname "$OLDNAME") 2> \
412 stripe_pool=$($LFS getstripe --pool \
413 $(dirname "$OLDNAME") 2> \
415 mirror_count=$($LFS getstripe -N \
416 $(dirname "$OLDFILE") 2> \
420 $ECHO -n "stripe_count=${stripe_count:-$parent_count},stripe_size=${stripe_size:-$parent_size}"
421 [ -n "$stripe_pool" ] &&
422 $ECHO -n ",pool=${stripe_pool}"
423 [ -n "$mirror_count" ] &&
424 $ECHO -n ",mirror_count=${mirror_count}"
429 $ECHO " dry run, skipped"
433 [ -n "$stripe_count" ] && stripe_count="-c $stripe_count"
434 [ -n "$stripe_size" ] && stripe_size="-S $stripe_size"
435 [ -n "$stripe_pool" ] && stripe_pool="-p $stripe_pool"
436 [ -n "$mirror_count" ] && mirror_count="-N $mirror_count"
437 layout="$stripe_count $stripe_size $stripe_pool $mirror_count \
440 # detect other hard links and store them on a global
441 # list so we don't re-migrate them
442 local mntpoint=$(df -P "$OLDNAME" |
443 awk 'NR==2 { print $NF; exit }')
444 if [ -z "$mntpoint" ]; then
445 echo -e "\r$OLDNAME: cannot determine mount point; skipped" 1>&2
448 hlinks=$($LFS fid2path "$mntpoint" "$fid" 2> /dev/null)
449 if [ $? -ne 0 ]; then
450 echo -e "\r$OLDNAME: cannot determine hard link paths, skipped" 1>&2
455 # first try to migrate via Lustre tools, then fall back to rsync
456 if ! $OPT_RSYNC; then
457 if $LFS migrate "${OPT_PASSTHROUGH[@]}" $layout \
460 for link in ${hlinks[*]}; do
461 add_to_set "$fid" "$link"
464 elif $OPT_NO_RSYNC; then
465 echo -e "\r$OLDNAME: refusing to fall back to rsync, skipped" 1>&2
468 $ECHO -n "falling back to rsync: "
473 NEWNAME=$(mktemp $UNLINK "$OLDNAME-lfs_migrate.tmp.XXXXXX")
474 if [ $? -ne 0 -o -z "$NEWNAME" ]; then
475 echo -e "\r$OLDNAME: cannot make temp file, skipped" 1>&2
479 if [ "$UNLINK" ]; then
480 if ! $LFS setstripe "${OPT_PASSTHROUGH[@]}" $layout \
482 echo -e "\r$NEWNAME: setstripe failed, exiting" 1>&2
487 # we use --inplace, since we created our own temp file already
488 if ! $RSYNC -a --inplace $RSYNC_OPTS "$OLDNAME" "$NEWNAME";then
489 echo -e "\r$OLDNAME: copy error, exiting" 1>&2
493 if $OPT_CHECK && ! cmp -s "$OLDNAME" "$NEWNAME"; then
494 echo -e "\r$NEWNAME: compare failed, exiting" 1>&2
498 if ! mv "$NEWNAME" "$OLDNAME"; then
499 echo -e "\r$OLDNAME: rename error, exiting" 1>&2
504 for link in ${hlinks[*]}; do
505 if [ "$link" != "$OLDNAME" ]; then
506 ln -f "$OLDNAME" "$link"
508 add_to_set "$fid" "$link"
511 # If the number of hlinks exceeds the space in the xattrs,
512 # when the final path is statted it will have a link count
513 # of 1 (all other links will point to the new inode).
514 # This flag indicates that even paths with a link count of
515 # 1 are potentially part of a link set.
516 [ ${#hlinks[*]} -gt 1 ] && RSYNC_WITH_HLINKS=true
520 if [ "$#" -eq 0 ]; then
524 tr '\n' '\0' | lfs_migrate
529 $LFS find "$1" -type f -print0