3 # client failure does not affect other clients
5 # Start load on clients (each client works on it's own directory).
6 # At defined (5-10 minutes) interval fail one random client and then fail mds.
7 # Reintegrate failed client after recovery completed,
8 # application errors are allowed for that client but not on other clients.
9 # 10 minute intervals and verify that no application errors occur.
11 # Test runs one of CLIENT_LOAD progs on remote clients.
13 LUSTRE=${LUSTRE:-`dirname $0`/..}
15 CLEANUP=${CLEANUP:-""}
16 . $LUSTRE/tests/test-framework.sh
20 . ${CONFIG:=$LUSTRE/tests/cfg/$NAME.sh}
22 TESTSUITELOG=${TESTSUITELOG:-$TMP/$(basename $0 .sh)}
23 DEBUGLOG=$TESTSUITELOG.debug
28 echo "--- env ---" >&2
30 echo "--- env ---" >&2
33 [ "$SHARED_DIRECTORY" ] || \
34 { skip "$0: Empty SHARED_DIRECTORY" && exit 0; }
36 [ -n "$CLIENTS" ] || { skip "$0 Need two or more remote clients" && exit 0; }
37 [ $CLIENTCOUNT -ge 3 ] || \
38 { skip "$0 Need two or more clients, have $CLIENTCOUNT" && exit 0; }
40 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
41 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
43 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
47 check_and_setup_lustre
48 rm -rf $DIR/[df][0-9]*
50 max_recov_time=$(max_recovery_time)
52 # the test node needs to be insulated from a lustre failure as much as possible,
53 # so not even loading the lustre modules is ideal.
55 # -- remove hostname from clients list
56 zconf_umount $(hostname) $MOUNT
57 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
58 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $(hostname))
60 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
62 MDTS=$(get_facets MDS)
64 if [ "$SLOW" = "no" ]; then
65 DURATION=${DURATION:-$((60 * 30))}
66 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 5))}
68 DURATION=${DURATION:-$((60 * 60 * 24))}
69 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 10))} # 10 minutes
74 vmstatLOG=${TESTSUITELOG}_$(basename $0 .sh).vmstat
80 for facet in $MDTS ${failed_clients//,/ }; do
84 echo "$facet failed over $val times"
89 # list is comma separated
93 do_nodes $list "node=\\\$(hostname)
95 log=${TESTSUITELOG}_run_${!var}.sh-\\\$node.debug
96 if [ -e \\\$log ] ; then
97 echo Node \\\$node debug log:
102 summary_and_cleanup () {
108 # Having not empty END_RUN_FILE means the failed loads only
109 if [ -s $END_RUN_FILE ]; then
110 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
113 read END_RUN_NODE < $END_RUN_FILE
115 # a client load will end (i.e. fail) if it finds
116 # the end run file. that does not mean that that client load
117 # actually failed though. the first node in the END_RUN_NODE is
118 # the one we are really interested in.
119 if [ -n "$END_RUN_NODE" ]; then
120 var=$(client_var_name $END_RUN_NODE)_load
121 echo "Client load failed on node $END_RUN_NODE"
123 echo "client $END_RUN_NODE load stdout and debug files :
124 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}
125 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug"
131 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
132 echo "$0" >> $END_RUN_FILE
134 [ $rc -eq 0 ] || result=FAIL
136 log "Duration: $DURATION
137 Server failover period: $SERVER_FAILOVER_PERIOD seconds
138 Exited after: $ELAPSED seconds
139 Number of failovers before exit:
141 Status: $result: rc=$rc"
143 # stop the vmstats on the OSTs
144 if [ "$VMSTAT" ]; then
145 do_nodes $(comma_list $(osts_nodes)) "test -f /tmp/vmstat.pid && \
146 { kill -s TERM \$(cat /tmp/vmstat.pid); rm -f /tmp/vmstat.pid; \
147 gzip -f9 $vmstatLOG-\$(hostname); }"
150 # make sure the client loads die
151 do_nodes $NODES_TO_USE "set -x; test -f $LOAD_PID_FILE && \
152 { kill -s TERM \$(cat $LOAD_PID_FILE) || true; }"
154 # and free up the pdshes that started them, if any are still around
155 if [ -n "$CLIENT_LOAD_PIDS" ]; then
156 kill $CLIENT_LOAD_PIDS || true
158 kill -9 $CLIENT_LOAD_PIDS || true
161 if [ $rc -ne 0 ]; then
162 print_logs $NODES_TO_USE
163 # we are interested in only on failed clients and servers
164 local failedclients=$(cat $END_RUN_FILE | grep -v $0)
165 # FIXME: need ostfailover-s nodes also for FLAVOR=OST
166 local product=$(gather_logs $(comma_list $(osts_nodes) \
167 $mds_HOST $mdsfailover_HOST $failedclients))
168 echo logs files $product
171 [ $rc -eq 0 ] && zconf_mount $(hostname) $MOUNT
179 log "-----============= $0 starting =============-----"
181 trap summary_and_cleanup EXIT # INT
186 if [ "$VMSTAT" ]; then
187 do_nodes $(comma_list $(osts_nodes)) "vmstat 1 > $vmstatLOG-\$(hostname) 2>/dev/null </dev/null & echo \$! > /tmp/vmstat.pid"
190 # Start client loads.
191 start_client_loads $NODES_TO_USE
193 echo clients load pids:
194 if ! do_nodesv $NODES_TO_USE "cat $LOAD_PID_FILE"; then
201 MINSLEEP=${MINSLEEP:-120}
202 REQFAIL_PERCENT=${REQFAIL_PERCENT:-3} # bug17839 comment 62
203 REQFAIL=${REQFAIL:-$(( DURATION / SERVER_FAILOVER_PERIOD * REQFAIL_PERCENT / 100))}
207 # This is used for FAIL_CLIENT only
209 while [ $ELAPSED -lt $DURATION -a ! -e $END_RUN_FILE ]; do
211 # In order to perform the
212 # expected number of failovers, we need to account the following :
213 # 1) the time that has elapsed during the client load checking
214 # 2) time takes for failover
216 it_time_start=$(date +%s)
218 FAIL_CLIENT=$(get_random_entry $NODES_TO_USE)
219 client_var=$(client_var_name $FAIL_CLIENT)_nums
221 # store the list of failed clients
222 # lists are comma separated
223 failed_clients=$(expand_list $failed_clients $FAIL_CLIENT)
225 SERVERFACET=$(get_random_entry $MDTS)
226 var=${SERVERFACET}_nums
228 # Check that our client loads are still running. If any have died,
229 # that means they have died outside of recovery, which is unacceptable.
231 log "==== Checking the clients loads BEFORE failover -- failure NOT OK \
232 ELAPSED=$ELAPSED DURATION=$DURATION PERIOD=$SERVER_FAILOVER_PERIOD"
234 if ! check_client_loads $NODES_TO_USE; then
238 log "FAIL CLIENT $FAIL_CLIENT ... "
239 shutdown_client $FAIL_CLIENT
241 log "Starting failover on $SERVERFACET"
243 facet_failover "$SERVERFACET" || exit 1
244 if ! wait_recovery_complete $SERVERFACET $((TIMEOUT * 10)); then
245 echo "$SERVERFACET recovery is not completed!"
249 boot_node $FAIL_CLIENT
250 echo "Reintegrating $FAIL_CLIENT"
251 zconf_mount $FAIL_CLIENT $MOUNT || exit $?
253 # Increment the number of failovers
256 val=$((${!client_var} + 1))
257 eval $client_var=$val
259 # load script on failed clients could create END_RUN_FILE
260 # We shuold remove it and ignore the failure if this
261 # file contains the failed client only.
262 # We can not use ERRORS_OK when start all loads at the start of this script
263 # because the application errors allowed for random failed client only, but
264 # not for all clients.
265 if [ -e $END_RUN_FILE ]; then
266 read END_RUN_NODE < $END_RUN_FILE
267 [[ $END_RUN_NODE = $FAIL_CLIENT ]] &&
268 rm -f $END_RUN_FILE || exit 13
271 restart_client_loads $FAIL_CLIENT $ERRORS_OK || exit $?
273 # Check that not failed clients loads are still running.
274 # No application failures should occur on clients that was not failed.
276 log "==== Checking the clients loads AFTER failed client reintegrated -- failure NOT OK"
277 if ! ERRORS_OK= check_client_loads $(exclude_items_from_list $NODES_TO_USE $FAIL_CLIENT); then
278 log "Client load failed. Exiting"
282 CURRENT_TS=$(date +%s)
283 ELAPSED=$((CURRENT_TS - START_TS))
285 sleep=$((SERVER_FAILOVER_PERIOD-(CURRENT_TS - it_time_start)))
287 # keep count the number of itterations when
288 # time spend to failover and two client loads check exceeded
289 # the value ( SERVER_FAILOVER_PERIOD - MINSLEEP )
290 if [ $sleep -lt $MINSLEEP ]; then
291 reqfail=$((reqfail +1))
292 log "WARNING: failover, client reintegration and check_client_loads time exceeded SERVER_FAILOVER_PERIOD - MINSLEEP !
293 Failed to load the filesystem with I/O for a minimum period of $MINSLEEP $reqfail times ( REQFAIL=$REQFAIL ).
294 This iteration, the load was only applied for sleep=$sleep seconds.
295 Estimated max recovery time : $max_recov_time
296 Probably the hardware is taking excessively long to boot.
297 Try to increase SERVER_FAILOVER_PERIOD (current is $SERVER_FAILOVER_PERIOD), bug 20918"
298 [ $reqfail -gt $REQFAIL ] && exit 6
301 log " Number of failovers:
302 $(numfailovers) and counting..."
304 if [ $((ELAPSED + sleep)) -gt $DURATION ]; then
308 if [ $sleep -gt 0 ]; then
309 echo "sleeping $sleep seconds ... "