2 # Test multiple failures, AKA Test 17
6 LUSTRE=${LUSTRE:-`dirname $0`/..}
7 . $LUSTRE/tests/test-framework.sh
11 . ${CONFIG:=$LUSTRE/tests/cfg/$NAME.sh}
13 ALWAYS_EXCEPT="10 $INSANITY_EXCEPT"
15 if [ "$FAILURE_MODE" = "HARD" ]; then
16 skip_env "$0: is not functional with FAILURE_MODE = HARD, please use recovery-double-scale, bz20407"
21 [ "$SLOW" = "no" ] && EXCEPT_SLOW=""
24 CLEANUP=${CLEANUP:-""}
28 SINGLECLIENT=${SINGLECLIENT:-$HOSTNAME}
29 LIVE_CLIENT=${LIVE_CLIENT:-$SINGLECLIENT}
30 FAIL_CLIENTS=${FAIL_CLIENTS:-$RCLIENTS}
32 assert_env mds_HOST MDS_MKFS_OPTS MDSDEV
33 assert_env ost_HOST OST_MKFS_OPTS OSTCOUNT
34 assert_env LIVE_CLIENT FSNAME
36 require_dsh_mds || exit 0
37 require_dsh_ost || exit 0
39 # FAIL_CLIENTS list should not contain the LIVE_CLIENT
40 FAIL_CLIENTS=$(echo " $FAIL_CLIENTS " | sed -re "s/\s+$LIVE_CLIENT\s+/ /g")
43 TESTDIR=$DIR/d0.$(basename $0 .sh)
46 # fail clients round robin
48 # list of failable clients
49 FAIL_LIST=($FAIL_CLIENTS)
50 FAIL_NUM=${#FAIL_LIST[*]}
53 DOWN_NUM=0 # number of nodes currently down
55 # set next client to fail
57 FAIL_CLIENT=${FAIL_LIST[$FAIL_NEXT]}
58 FAIL_NEXT=$(( (FAIL_NEXT+1) % FAIL_NUM ))
59 echo "fail $FAIL_CLIENT, next is $FAIL_NEXT"
65 log "Request clients to fail: ${num}. Num of clients to fail: ${FAIL_NUM}, already failed: $DOWN_NUM"
66 if [ -z "$num" ] || [ "$num" -gt $((FAIL_NUM - DOWN_NUM)) ]; then
67 num=$((FAIL_NUM - DOWN_NUM))
70 if [ -z "$num" ] || [ "$num" -le 0 ]; then
71 log "No clients failed!"
77 for i in `seq $num`; do
80 DOWN_CLIENTS="$DOWN_CLIENTS $client"
81 shutdown_client $client
84 echo "down clients: $DOWN_CLIENTS"
86 for client in $DOWN_CLIENTS; do
89 DOWN_NUM=`echo $DOWN_CLIENTS | wc -w`
93 reintegrate_clients() {
94 for client in $DOWN_CLIENTS; do
96 echo "Restarting $client"
97 zconf_mount $client $MOUNT || return 1
104 start ost$1 `ostdevname $1` $OST_MOUNT_OPTS
111 for c in $LIVE_CLIENT $FAIL_CLIENTS; do
112 if echo $DOWN_CLIENTS | grep -q $c; then continue; fi
113 $PDSH $c touch $TESTDIR/${c}_$file || return 1
119 for c in $LIVE_CLIENT $FAIL_CLIENTS; do
120 $PDSH $c rm $TESTDIR/${c}_$file
125 for c in $LIVE_CLIENT $FAIL_CLIENTS; do
126 echo "$c mkdir $TESTDIR/$c"
127 $PDSH $c "mkdir $TESTDIR/$c && ls -l $TESTDIR/$c"
132 for c in $LIVE_CLIENT $FAIL_CLIENTS; do
133 echo "rmdir $TESTDIR/$c"
134 $PDSH $LIVE_CLIENT "rmdir $TESTDIR/$c"
138 clients_recover_osts() {
140 # do_node $CLIENTS "$LCTL "'--device %OSC_`hostname`_'"${facet}_svc_MNT_client_facet recover"
143 check_and_setup_lustre
148 # 9 Different Failure Modes Combinations
149 echo "Starting Test 17 at `date`"
153 echo "Waiting for df pid: $DFPID"
154 wait $DFPID || { echo "df returned $?" && return 1; }
156 for i in $(seq $OSTCOUNT) ; do
157 facet_failover ost$i || return 4
158 echo "Waiting for df pid: $DFPID"
159 wait $DFPID || { echo "df returned $?" && return 3; }
163 run_test 0 "Fail all nodes, independently"
165 ############### First Failure Mode ###############
167 echo "Don't do a MDS - MDS Failure Case"
168 echo "This makes no sense"
170 run_test 1 "MDS/MDS failure"
171 ###################################################
173 ############### Second Failure Mode ###############
175 echo "Verify Lustre filesystem is up and running"
176 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
183 # prepare for MDS failover
193 echo "Reintegrating OST"
196 start_ost 1 || return 2
199 start mds $MDSDEV $MDS_MOUNT_OPTS || return $?
203 clients_recover_osts ost1
204 echo "Verify reintegration"
205 clients_up || return 1
208 run_test 2 "Second Failure Mode: MDS/OST `date`"
209 ###################################################
212 ############### Third Failure Mode ###############
215 echo "Verify Lustre filesystem is up and running"
216 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
220 wait $DFPID || echo df failed: $?
223 echo "Test Lustre stability after MDS failover"
227 echo "Failing 2 CLIENTS"
231 echo "Test Lustre stability after CLIENT failure"
235 echo "Reintegrating CLIENTS"
236 reintegrate_clients || return 1
238 clients_up || return 3
240 run_test 3 "Thirdb Failure Mode: MDS/CLIENT `date`"
241 ###################################################
243 ############### Fourth Failure Mode ###############
245 echo "Fourth Failure Mode: OST/MDS `date`"
251 echo "Test Lustre stability after OST failure"
260 # prepare for MDS failover
269 echo "Reintegrating OST"
275 start mds $MDSDEV $MDS_MOUNT_OPTS
280 clients_recover_osts ost1
281 echo "Test Lustre stability after MDS failover"
282 clients_up || return 1
284 run_test 4 "Fourth Failure Mode: OST/MDS `date`"
285 ###################################################
287 ############### Fifth Failure Mode ###############
289 [ $OSTCOUNT -lt 2 ] && skip_env "$OSTCOUNT < 2, not enough OSTs" && return 0
291 echo "Fifth Failure Mode: OST/OST `date`"
294 echo "Verify Lustre filesystem is up and running"
295 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
304 echo "Test Lustre stability after OST failure"
314 echo "Test Lustre stability after OST failure"
320 echo "Reintegrating OSTs"
326 clients_recover_osts ost1
327 clients_recover_osts ost2
332 clients_up || return 2
334 run_test 5 "Fifth Failure Mode: OST/OST `date`"
335 ###################################################
337 ############### Sixth Failure Mode ###############
339 echo "Sixth Failure Mode: OST/CLIENT `date`"
342 echo "Verify Lustre filesystem is up and running"
343 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
346 client_touch testfile || return 2
353 echo "Test Lustre stability after OST failure"
360 echo "Failing CLIENTs"
364 echo "Test Lustre stability after CLIENTs failure"
371 echo "Reintegrating OST/CLIENTs"
374 reintegrate_clients || return 1
377 wait_remote_prog df $((TIMEOUT * 3 + 10))
381 echo "Verifying mount"
382 [ -z "$(mounted_lustre_filesystems)" ] && return 3
385 run_test 6 "Sixth Failure Mode: OST/CLIENT `date`"
386 ###################################################
389 ############### Seventh Failure Mode ###############
391 echo "Seventh Failure Mode: CLIENT/MDS `date`"
394 echo "Verify Lustre filesystem is up and running"
395 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
398 client_touch testfile || return 1
401 echo "Part 1: Failing CLIENT"
405 echo "Test Lustre stability after CLIENTs failure"
407 $PDSH $LIVE_CLIENT "ls -l $TESTDIR"
408 $PDSH $LIVE_CLIENT "rm -f $TESTDIR/*_testfile"
411 echo "Wait 1 minutes"
415 echo "Verify Lustre filesystem is up and running"
416 [ -z "$(mounted_lustre_filesystems)" ] && return 2
425 echo "Test Lustre stability after MDS failover"
426 wait $DFPID || echo "df on down clients fails " || return 1
427 $PDSH $LIVE_CLIENT "ls -l $TESTDIR"
428 $PDSH $LIVE_CLIENT "rm -f $TESTDIR/*_testfile"
431 echo "Reintegrating CLIENTs"
432 reintegrate_clients || return 2
436 echo "wait 1 minutes"
439 run_test 7 "Seventh Failure Mode: CLIENT/MDS `date`"
440 ###################################################
443 ############### Eighth Failure Mode ###############
445 echo "Eighth Failure Mode: CLIENT/OST `date`"
448 echo "Verify Lustre filesystem is up and running"
449 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
452 client_touch testfile
455 echo "Failing CLIENTs"
459 echo "Test Lustre stability after CLIENTs failure"
461 $PDSH $LIVE_CLIENT "ls -l $TESTDIR"
462 $PDSH $LIVE_CLIENT "rm -f $TESTDIR/*_testfile"
465 echo "Wait 1 minutes"
469 echo "Verify Lustre filesystem is up and running"
470 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
473 client_touch testfile
481 echo "Test Lustre stability after OST failure"
485 #non-failout hangs forever here
486 #$PDSH $LIVE_CLIENT "ls -l $TESTDIR"
487 #$PDSH $LIVE_CLIENT "rm -f $TESTDIR/*_testfile"
490 echo "Reintegrating CLIENTs/OST"
491 reintegrate_clients || return 3
495 clients_up || return 1
496 client_touch testfile2 || return 2
499 echo "Wait 1 minutes"
502 run_test 8 "Eighth Failure Mode: CLIENT/OST `date`"
503 ###################################################
506 ############### Ninth Failure Mode ###############
511 echo "Verify Lustre filesystem is up and running"
512 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
515 client_touch testfile || return 1
518 echo "Failing CLIENTs"
522 echo "Test Lustre stability after CLIENTs failure"
524 $PDSH $LIVE_CLIENT "ls -l $TESTDIR" || return 1
525 $PDSH $LIVE_CLIENT "rm -f $TESTDIR/*_testfile" || return 2
528 echo "Wait 1 minutes"
532 echo "Verify Lustre filesystem is up and running"
533 $PDSH $LIVE_CLIENT "grep -e $MOUNT /proc/mounts" || return 3
534 client_up $LIVE_CLIENT
535 client_touch testfile || return 4
538 echo "Failing CLIENTs"
542 echo "Test Lustre stability after CLIENTs failure"
544 $PDSH $LIVE_CLIENT "ls -l $TESTDIR" || return 5
545 $PDSH $LIVE_CLIENT "rm -f $TESTDIR/*_testfile" || return 6
548 echo "Reintegrating CLIENTs/CLIENTs"
549 reintegrate_clients || return 7
553 echo "Wait 1 minutes"
556 run_test 9 "Ninth Failure Mode: CLIENT/CLIENT `date`"
557 ###################################################
560 #Run availability after all failures
561 DURATION=${DURATION:-$((2 * 60 * 60))} # 6 hours default
562 LOADTEST=${LOADTEST:-metadata-load.py}
563 $PWD/availability.sh $CONFIG $DURATION $CLIENTS || return 1
565 run_test 10 "Running Availability for 6 hours..."
567 complete $(basename $0) $SECONDS
568 check_and_cleanup_lustre