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 mixed_ost_devs && CONFIG_EXCEPTIONS="0 2 4 5 6 8" && \
17 echo -n "Several ost services on one ost node are used with FAILURE_MODE=$FAILURE_MODE. " && \
18 echo "Except the tests: $CONFIG_EXCEPTIONS" && \
19 ALWAYS_EXCEPT="$ALWAYS_EXCEPT $CONFIG_EXCEPTIONS"
23 [ "$SLOW" = "no" ] && EXCEPT_SLOW=""
26 CLEANUP=${CLEANUP:-""}
30 SINGLECLIENT=${SINGLECLIENT:-$HOSTNAME}
31 LIVE_CLIENT=${LIVE_CLIENT:-$SINGLECLIENT}
32 FAIL_CLIENTS=${FAIL_CLIENTS:-$RCLIENTS}
34 assert_env mds_HOST MDS_MKFS_OPTS MDSDEV
35 assert_env ost_HOST OST_MKFS_OPTS OSTCOUNT
36 assert_env LIVE_CLIENT FSNAME
38 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
39 remote_ost_nodsh && skip "remote OST with nodsh" && exit 0
41 # FAIL_CLIENTS list should not contain the LIVE_CLIENT
42 FAIL_CLIENTS=$(echo " $FAIL_CLIENTS " | sed -re "s/\s+$LIVE_CLIENT\s+/ /g")
45 TESTDIR=$DIR/d0.$(basename $0 .sh)
48 # fail clients round robin
50 # list of failable clients
51 FAIL_LIST=($FAIL_CLIENTS)
52 FAIL_NUM=${#FAIL_LIST[*]}
55 DOWN_NUM=0 # number of nodes currently down
57 # set next client to fail
59 FAIL_CLIENT=${FAIL_LIST[$FAIL_NEXT]}
60 FAIL_NEXT=$(( (FAIL_NEXT+1) % FAIL_NUM ))
61 echo "fail $FAIL_CLIENT, next is $FAIL_NEXT"
67 log "Request clients to fail: ${num}. Num of clients to fail: ${FAIL_NUM}, already failed: $DOWN_NUM"
68 if [ -z "$num" ] || [ "$num" -gt $((FAIL_NUM - DOWN_NUM)) ]; then
69 num=$((FAIL_NUM - DOWN_NUM))
72 if [ -z "$num" ] || [ "$num" -le 0 ]; then
73 log "No clients failed!"
79 for i in `seq $num`; do
82 DOWN_CLIENTS="$DOWN_CLIENTS $client"
83 shutdown_client $client
86 echo "down clients: $DOWN_CLIENTS"
88 for client in $DOWN_CLIENTS; do
91 DOWN_NUM=`echo $DOWN_CLIENTS | wc -w`
95 reintegrate_clients() {
96 for client in $DOWN_CLIENTS; do
98 echo "Restarting $client"
99 zconf_mount $client $MOUNT || return 1
106 start ost$1 `ostdevname $1` $OST_MOUNT_OPTS
113 for c in $LIVE_CLIENT $FAIL_CLIENTS; do
114 if echo $DOWN_CLIENTS | grep -q $c; then continue; fi
115 $PDSH $c touch $TESTDIR/${c}_$file || return 1
121 for c in $LIVE_CLIENT $FAIL_CLIENTS; do
122 $PDSH $c rm $TESTDIR/${c}_$file
127 for c in $LIVE_CLIENT $FAIL_CLIENTS; do
128 echo "$c mkdir $TESTDIR/$c"
129 $PDSH $c "mkdir $TESTDIR/$c && ls -l $TESTDIR/$c"
134 for c in $LIVE_CLIENT $FAIL_CLIENTS; do
135 echo "rmdir $TESTDIR/$c"
136 $PDSH $LIVE_CLIENT "rmdir $TESTDIR/$c"
140 clients_recover_osts() {
142 # do_node $CLIENTS "$LCTL "'--device %OSC_`hostname`_'"${facet}_svc_MNT_client_facet recover"
145 check_and_setup_lustre
150 # 9 Different Failure Modes Combinations
151 echo "Starting Test 17 at `date`"
155 echo "Waiting for df pid: $DFPID"
156 wait $DFPID || { echo "df returned $?" && return 1; }
158 for i in $(seq $OSTCOUNT) ; do
159 facet_failover ost$i || return 4
160 echo "Waiting for df pid: $DFPID"
161 wait $DFPID || { echo "df returned $?" && return 3; }
165 run_test 0 "Fail all nodes, independently"
167 ############### First Failure Mode ###############
169 echo "Don't do a MDS - MDS Failure Case"
170 echo "This makes no sense"
172 run_test 1 "MDS/MDS failure"
173 ###################################################
175 ############### Second Failure Mode ###############
177 echo "Verify Lustre filesystem is up and running"
178 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
185 # prepare for MDS failover
195 echo "Reintegrating OST"
198 start_ost 1 || return 2
201 start mds $MDSDEV $MDS_MOUNT_OPTS || return $?
205 clients_recover_osts ost1
206 echo "Verify reintegration"
207 clients_up || return 1
210 run_test 2 "Second Failure Mode: MDS/OST `date`"
211 ###################################################
214 ############### Third Failure Mode ###############
217 echo "Verify Lustre filesystem is up and running"
218 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
222 wait $DFPID || echo df failed: $?
225 echo "Test Lustre stability after MDS failover"
229 echo "Failing 2 CLIENTS"
233 echo "Test Lustre stability after CLIENT failure"
237 echo "Reintegrating CLIENTS"
238 reintegrate_clients || return 1
240 clients_up || return 3
242 run_test 3 "Thirdb Failure Mode: MDS/CLIENT `date`"
243 ###################################################
245 ############### Fourth Failure Mode ###############
247 echo "Fourth Failure Mode: OST/MDS `date`"
253 echo "Test Lustre stability after OST failure"
262 # prepare for MDS failover
271 echo "Reintegrating OST"
277 start mds $MDSDEV $MDS_MOUNT_OPTS
282 clients_recover_osts ost1
283 echo "Test Lustre stability after MDS failover"
284 clients_up || return 1
286 run_test 4 "Fourth Failure Mode: OST/MDS `date`"
287 ###################################################
289 ############### Fifth Failure Mode ###############
291 [ $OSTCOUNT -lt 2 ] && skip_env "$OSTCOUNT < 2, not enough OSTs" && return 0
293 echo "Fifth Failure Mode: OST/OST `date`"
296 echo "Verify Lustre filesystem is up and running"
297 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
306 echo "Test Lustre stability after OST failure"
316 echo "Test Lustre stability after OST failure"
322 echo "Reintegrating OSTs"
328 clients_recover_osts ost1
329 clients_recover_osts ost2
334 clients_up || return 2
336 run_test 5 "Fifth Failure Mode: OST/OST `date`"
337 ###################################################
339 ############### Sixth Failure Mode ###############
341 echo "Sixth Failure Mode: OST/CLIENT `date`"
344 echo "Verify Lustre filesystem is up and running"
345 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
348 client_touch testfile || return 2
355 echo "Test Lustre stability after OST failure"
362 echo "Failing CLIENTs"
366 echo "Test Lustre stability after CLIENTs failure"
373 echo "Reintegrating OST/CLIENTs"
376 reintegrate_clients || return 1
379 wait_remote_prog df $((TIMEOUT * 3 + 10))
383 echo "Verifying mount"
384 [ -z "$(mounted_lustre_filesystems)" ] && return 3
387 run_test 6 "Sixth Failure Mode: OST/CLIENT `date`"
388 ###################################################
391 ############### Seventh Failure Mode ###############
393 echo "Seventh Failure Mode: CLIENT/MDS `date`"
396 echo "Verify Lustre filesystem is up and running"
397 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
400 client_touch testfile || return 1
403 echo "Part 1: Failing CLIENT"
407 echo "Test Lustre stability after CLIENTs failure"
409 $PDSH $LIVE_CLIENT "ls -l $TESTDIR"
410 $PDSH $LIVE_CLIENT "rm -f $TESTDIR/*_testfile"
413 echo "Wait 1 minutes"
417 echo "Verify Lustre filesystem is up and running"
418 [ -z "$(mounted_lustre_filesystems)" ] && return 2
427 echo "Test Lustre stability after MDS failover"
428 wait $DFPID || echo "df on down clients fails " || return 1
429 $PDSH $LIVE_CLIENT "ls -l $TESTDIR"
430 $PDSH $LIVE_CLIENT "rm -f $TESTDIR/*_testfile"
433 echo "Reintegrating CLIENTs"
434 reintegrate_clients || return 2
438 echo "wait 1 minutes"
441 run_test 7 "Seventh Failure Mode: CLIENT/MDS `date`"
442 ###################################################
445 ############### Eighth Failure Mode ###############
447 echo "Eighth Failure Mode: CLIENT/OST `date`"
450 echo "Verify Lustre filesystem is up and running"
451 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
454 client_touch testfile
457 echo "Failing CLIENTs"
461 echo "Test Lustre stability after CLIENTs failure"
463 $PDSH $LIVE_CLIENT "ls -l $TESTDIR"
464 $PDSH $LIVE_CLIENT "rm -f $TESTDIR/*_testfile"
467 echo "Wait 1 minutes"
471 echo "Verify Lustre filesystem is up and running"
472 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
475 client_touch testfile
483 echo "Test Lustre stability after OST failure"
487 #non-failout hangs forever here
488 #$PDSH $LIVE_CLIENT "ls -l $TESTDIR"
489 #$PDSH $LIVE_CLIENT "rm -f $TESTDIR/*_testfile"
492 echo "Reintegrating CLIENTs/OST"
493 reintegrate_clients || return 3
497 clients_up || return 1
498 client_touch testfile2 || return 2
501 echo "Wait 1 minutes"
504 run_test 8 "Eighth Failure Mode: CLIENT/OST `date`"
505 ###################################################
508 ############### Ninth Failure Mode ###############
513 echo "Verify Lustre filesystem is up and running"
514 [ -z "$(mounted_lustre_filesystems)" ] && error "Lustre is not running"
517 client_touch testfile || return 1
520 echo "Failing CLIENTs"
524 echo "Test Lustre stability after CLIENTs failure"
526 $PDSH $LIVE_CLIENT "ls -l $TESTDIR" || return 1
527 $PDSH $LIVE_CLIENT "rm -f $TESTDIR/*_testfile" || return 2
530 echo "Wait 1 minutes"
534 echo "Verify Lustre filesystem is up and running"
535 $PDSH $LIVE_CLIENT "grep -e $MOUNT /proc/mounts" || return 3
536 client_up $LIVE_CLIENT
537 client_touch testfile || return 4
540 echo "Failing CLIENTs"
544 echo "Test Lustre stability after CLIENTs failure"
546 $PDSH $LIVE_CLIENT "ls -l $TESTDIR" || return 5
547 $PDSH $LIVE_CLIENT "rm -f $TESTDIR/*_testfile" || return 6
550 echo "Reintegrating CLIENTs/CLIENTs"
551 reintegrate_clients || return 7
555 echo "Wait 1 minutes"
558 run_test 9 "Ninth Failure Mode: CLIENT/CLIENT `date`"
559 ###################################################
562 #Run availability after all failures
563 DURATION=${DURATION:-$((2 * 60 * 60))} # 6 hours default
564 LOADTEST=${LOADTEST:-metadata-load.py}
565 $PWD/availability.sh $CONFIG $DURATION $CLIENTS || return 1
567 run_test 10 "Running Availability for 6 hours..."
569 equals_msg `basename $0`: test complete, cleaning up
570 check_and_cleanup_lustre
571 [ -f "$TESTSUITELOG" ] && cat $TESTSUITELOG && grep -q FAIL $TESTSUITELOG && exit 1 || true