3 # All pairwise combinations of node failures.
6 # Author: Chris Cooper <ccooper@clusterfs.com>
8 # Script fails pair of nodes:
9 # -- in parallel by default
10 # -- in series if SERIAL is set
12 LUSTRE=${LUSTRE:-`dirname $0`/..}
14 CLEANUP=${CLEANUP:-""}
15 . $LUSTRE/tests/test-framework.sh
19 . ${CONFIG:=$LUSTRE/tests/cfg/$NAME.sh}
20 TESTSUITELOG=${TESTSUITELOG:-$TMP/$(basename $0 .sh)}
21 DEBUGLOG=$TESTSUITELOG.debug
26 echo "--- env ---" >&2
28 echo "--- env ---" >&2
31 [ "$SHARED_DIRECTORY" ] || \
32 { skip "$0: Empty SHARED_DIRECTORY" && exit 0; }
34 [ -n "$CLIENTS" ] || { skip "$0 Need two or more remote clients" && exit 0; }
35 [ $CLIENTCOUNT -ge 3 ] || \
36 { skip "$0 Need two or more remote clients, have $CLIENTCOUNT" && exit 0; }
38 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
39 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
41 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
42 remote_ost_nodsh && skip "remote OST with nodsh" && exit 0
44 check_timeout || exit 1
46 [[ $FAILURE_MODE = SOFT ]] && \
47 log "WARNING: $0 is not functional with FAILURE_MODE = SOFT, bz22797"
51 check_and_setup_lustre
52 rm -rf $DIR/[df][0-9]*
54 # the test node needs to be insulated from a lustre failure as much as possible,
55 # so not even loading the lustre modules is ideal.
57 # -- remove hostname from clients list
58 zconf_umount $(hostname) $MOUNT
59 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
60 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $(hostname))
62 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
64 MDTS=$(get_facets MDS)
65 OSTS=$(get_facets OST)
69 reboot_recover_node () {
70 # item var contains a pair of clients if nodetype=clients
71 # I would prefer to have a list here
74 local timeout=$($LCTL get_param -n timeout)
76 # MDS, OST item contains the facet
78 MDS|OST ) facet_failover $item
79 [ "$SERIAL" ] && wait_recovery_complete $item $((timeout * 4)) || true
81 clients) for c in ${item//,/ }; do
84 echo "Reintegrating $c"
85 # one client fails; need dk logs from this client only
86 zconf_mount $c $MOUNT || NODES="$c $(facet_host mds) $(osts_nodes)" error_exit "zconf_mount failed"
88 start_client_loads $item
91 # don't use error (), the logs from all nodes not needed
92 * ) echo "reboot_recover_node: nodetype=$nodetype. Must be one of 'MDS', 'OST', or 'clients'."
99 local excluded=${2:-""}
105 clients) list=$NODES_TO_USE
108 # don't use error (), the logs from all nodes not needed
109 * ) echo "Invalid type=$type. Must be one of 'MDS', 'OST', or 'clients'."
113 [ "$excluded" ] && list=$(exclude_items_from_list $list $excluded)
115 if [ ! "$(echo $list)" ]; then
120 item=$(get_random_entry $list)
121 if [ "$type" = clients ] ; then
122 item="$item $(get_random_entry $(exclude_items_from_list $list $item))"
123 item=$(comma_list $item)
130 # for the two nodetypes specified, chooses a random node(s) from each
131 # class, reboots the nodes sequentially, and then restarts lustre on
138 local client_nodes=""
145 ==== START === $title "
147 item1=$(get_item_type $type1)
149 { echo "type1=$type1 item1 is empty" && return 0; }
150 item2=$(get_item_type $type2 $item1)
152 { echo "type1=$type1 item1=$item1 type2=$type2 item2=$item2 is empty" && return 0; }
154 # Check that our client loads are still running. If any have died,
155 # that means they have died outside of recovery, which is unacceptable.
156 log "==== Checking the clients loads BEFORE failover -- failure NOT OK"
158 # FIXME. need print summary on exit
159 if ! check_client_loads $NODES_TO_USE; then
163 log "Done checking client loads. Failing type1=$type1 item1=$item1 ... "
165 reboot_recover_node $item1 $type1
167 # Hendrix test17 description:
168 # Introduce a failure, wait at
169 # least 5 minutes (for recovery),
171 # failure, and wait another 5
174 # reboot_recover_node waits recovery in according to
176 # We have a "double failures" if SERIAL is not set,
177 # do not need a sleep between failures for "double failures"
179 log " Failing type2=$type2 item2=$item2 ... "
180 reboot_recover_node $item2 $type2
182 # Client loads are allowed to die while in recovery, so we just
184 log "==== Checking the clients loads AFTER failovers -- ERRORS_OK=$ERRORS_OK"
185 restart_client_loads $NODES_TO_USE $ERRORS_OK || return $?
186 log "Done checking / re-Starting client loads. PASS"
190 summary_and_cleanup () {
194 # Having not empty END_RUN_FILE means the failed loads only
195 if [ -s $END_RUN_FILE ]; then
196 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
199 read END_RUN_NODE < $END_RUN_FILE
201 # a client load will end (i.e. fail) if it finds
202 # the end run file. that does not mean that that client load
203 # actually failed though. the first node in the END_RUN_NODE is
204 # the one we are really interested in.
205 if [ -n "$END_RUN_NODE" ]; then
206 var=$(client_var_name $END_RUN_NODE)_load
207 echo "Client load failed on node $END_RUN_NODE"
209 echo "client $END_RUN_NODE load debug output :"
210 local logfile=${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug
211 do_node ${END_RUN_NODE} "set -x; [ -e $logfile ] && cat $logfile " || true
216 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
217 echo "$0" >> $END_RUN_FILE
219 [ $rc -eq 0 ] || result=FAIL
222 Server failover period: $FAILOVER_PERIOD seconds
223 Exited after: $ELAPSED seconds
224 Status: $result: rc=$rc"
226 # make sure the client loads die
227 do_nodes $NODES_TO_USE "set -x; test -f $TMP/client-load.pid && \
228 { kill -s TERM \$(cat $TMP/client-load.pid) || true; }"
230 # and free up the pdshes that started them, if any are still around
231 if [ -n "$CLIENT_LOAD_PIDS" ]; then
232 kill $CLIENT_LOAD_PIDS || true
234 kill -9 $CLIENT_LOAD_PIDS || true
237 if [ $rc -ne 0 ]; then
238 # we are interested in only on failed clients and servers
239 local failedclients=$(cat $END_RUN_FILE | grep -v $0)
240 # FIXME: need ostfailover-s nodes also for FLAVOR=OST
241 local product=$(gather_logs $(comma_list $(osts_nodes) \
242 $mds_HOST $mdsfailover_HOST $failedclients))
243 echo logs files $product
246 [ $rc -eq 0 ] && zconf_mount $(hostname) $MOUNT
250 trap summary_and_cleanup EXIT TERM INT
255 log "-----============= $0 starting =============-----"
261 # Set SERIAL to serialize the failure through a recovery of the first failure.
265 [ "$SERIAL" ] && ERRORS_OK=""
267 FAILOVER_PERIOD=${FAILOVER_PERIOD:-$((60*5))} # 5 minutes
269 # Start client loads.
270 start_client_loads $NODES_TO_USE
271 echo clients load pids:
272 if ! do_nodesv $NODES_TO_USE "cat $TMP/client-load.pid"; then
276 # FIXME: Do we want to have an initial sleep period where the clients
277 # just run before introducing a failure?
278 sleep $FAILOVER_PERIOD
281 failover_pair MDS OST "test 1: failover MDS, then OST =========="
282 sleep $FAILOVER_PERIOD
285 failover_pair MDS clients "test 2: failover MDS, then 2 clients ===="
286 sleep $FAILOVER_PERIOD
289 # No test 3 for 1.8.x lustre version
292 if [ $OSTCOUNT -gt 1 ]; then
293 failover_pair OST OST "test 4: failover OST, then another OST =="
294 sleep $FAILOVER_PERIOD
296 skip "$0 : $OSTCOUNT < 2 OSTs, test 4 skipped"
300 failover_pair OST clients "test 5: failover OST, then 2 clients ===="
301 sleep $FAILOVER_PERIOD
304 failover_pair OST MDS "test 6: failover OST, then MDS =========="
305 sleep $FAILOVER_PERIOD
308 failover_pair clients MDS "test 7: failover 2 clients, then MDS ===="
309 sleep $FAILOVER_PERIOD
312 #failover_pair clients OST "test 8: failover 2 clients, then OST ===="
313 sleep $FAILOVER_PERIOD
316 if [ $CLIENTCOUNT -ge 5 ]; then
317 failover_pair clients clients "test 9: failover 2 clients, then 2 different clients =="
318 sleep $FAILOVER_PERIOD
320 log "==== Checking the clients loads AFTER all failovers -- failure NOT OK"
321 if ! check_client_loads $NODES_TO_USE; then
322 log "Client load failed after failover. Exiting"
326 CURRENT_TS=$(date +%s)
327 ELAPSED=$((CURRENT_TS - START_TS))
329 log "Completed successfully in $ELAPSED seconds"