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/$NAME.sh}
23 TESTSUITELOG=${TESTSUITELOG:-$TMP/$(basename $0 .sh)}
24 DEBUGLOG=$TESTSUITELOG.debug
29 echo "--- env ---" >&2
31 echo "--- env ---" >&2
34 [ "$SHARED_DIRECTORY" ] || \
35 { skip "$0: Empty SHARED_DIRECTORY" && exit 0; }
37 [ -n "$CLIENTS" ] || { skip "$0 Need two or more remote clients" && exit 0; }
38 [ $CLIENTCOUNT -ge 3 ] || \
39 { skip "$0 Need two or more clients, have $CLIENTCOUNT" && exit 0; }
41 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
42 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
44 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
46 [[ $FAILURE_MODE = SOFT ]] && \
47 log "WARNING: $0 is not functional with FAILURE_MODE = SOFT, bz22797"
51 check_and_setup_lustre
52 rm -rf $DIR/[df][0-9]*
54 max_recov_time=$(max_recovery_time)
56 # the test node needs to be insulated from a lustre failure as much as possible,
57 # so not even loading the lustre modules is ideal.
59 # -- remove hostname from clients list
60 zconf_umount $(hostname) $MOUNT
61 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
62 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $(hostname))
64 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
66 MDTS=$(get_facets MDS)
68 if [ "$SLOW" = "no" ]; then
69 DURATION=${DURATION:-$((60 * 30))}
70 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 5))}
72 DURATION=${DURATION:-$((60 * 60 * 24))}
73 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 10))} # 10 minutes
78 vmstatLOG=${TESTSUITELOG}_$(basename $0 .sh).vmstat
84 for facet in $MDTS ${failed_clients//,/ }; do
88 echo "$facet failed over $val times"
93 # list is comma separated
97 do_nodes $list "node=\\\$(hostname)
99 log=${TESTSUITELOG}_run_${!var}.sh-\\\$node.debug
100 if [ -e \\\$log ] ; then
101 echo Node \\\$node debug log:
106 summary_and_cleanup () {
112 # Having not empty END_RUN_FILE means the failed loads only
113 if [ -s $END_RUN_FILE ]; then
114 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
117 read END_RUN_NODE < $END_RUN_FILE
119 # a client load will end (i.e. fail) if it finds
120 # the end run file. that does not mean that that client load
121 # actually failed though. the first node in the END_RUN_NODE is
122 # the one we are really interested in.
123 if [ -n "$END_RUN_NODE" ]; then
124 var=$(client_var_name $END_RUN_NODE)_load
125 echo "Client load failed on node $END_RUN_NODE"
127 echo "client $END_RUN_NODE load stdout and debug files :
128 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}
129 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug"
135 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
136 echo "$0" >> $END_RUN_FILE
138 [ $rc -eq 0 ] || result=FAIL
140 log "Duration: $DURATION
141 Server failover period: $SERVER_FAILOVER_PERIOD seconds
142 Exited after: $ELAPSED seconds
143 Number of failovers before exit:
145 Status: $result: rc=$rc"
147 # stop the vmstats on the OSTs
148 if [ "$VMSTAT" ]; then
149 do_nodes $(comma_list $(osts_nodes)) "test -f /tmp/vmstat.pid && \
150 { kill -s TERM \$(cat /tmp/vmstat.pid); rm -f /tmp/vmstat.pid; \
151 gzip -f9 $vmstatLOG-\$(hostname); }"
154 # make sure the client loads die
155 do_nodes $NODES_TO_USE "set -x; test -f $LOAD_PID_FILE && \
156 { kill -s TERM \$(cat $LOAD_PID_FILE) || true; }"
158 # and free up the pdshes that started them, if any are still around
159 if [ -n "$CLIENT_LOAD_PIDS" ]; then
160 kill $CLIENT_LOAD_PIDS || true
162 kill -9 $CLIENT_LOAD_PIDS || true
165 if [ $rc -ne 0 ]; then
166 print_logs $NODES_TO_USE
167 # we are interested in only on failed clients and servers
168 local failedclients=$(cat $END_RUN_FILE | grep -v $0)
169 # FIXME: need ostfailover-s nodes also for FLAVOR=OST
170 local product=$(gather_logs $(comma_list $(osts_nodes) \
171 $(mdts_nodes) $mdsfailover_HOST $failedclients))
172 echo logs files $product
175 [ $rc -eq 0 ] && zconf_mount $(hostname) $MOUNT
183 log "-----============= $0 starting =============-----"
185 trap summary_and_cleanup EXIT # INT
190 if [ "$VMSTAT" ]; then
191 do_nodes $(comma_list $(osts_nodes)) "vmstat 1 > $vmstatLOG-\$(hostname) 2>/dev/null </dev/null & echo \$! > /tmp/vmstat.pid"
194 # Start client loads.
195 start_client_loads $NODES_TO_USE
197 echo clients load pids:
198 if ! do_nodesv $NODES_TO_USE "cat $LOAD_PID_FILE"; then
205 MINSLEEP=${MINSLEEP:-120}
206 REQFAIL_PERCENT=${REQFAIL_PERCENT:-3} # bug17839 comment 62
207 REQFAIL=${REQFAIL:-$(( DURATION / SERVER_FAILOVER_PERIOD * REQFAIL_PERCENT / 100))}
211 # This is used for FAIL_CLIENT only
213 while [ $ELAPSED -lt $DURATION -a ! -e $END_RUN_FILE ]; do
215 # In order to perform the
216 # expected number of failovers, we need to account the following :
217 # 1) the time that has elapsed during the client load checking
218 # 2) time takes for failover
220 it_time_start=$(date +%s)
222 FAIL_CLIENT=$(get_random_entry $NODES_TO_USE)
223 client_var=$(client_var_name $FAIL_CLIENT)_nums
225 # store the list of failed clients
226 # lists are comma separated
227 failed_clients=$(expand_list $failed_clients $FAIL_CLIENT)
229 SERVERFACET=$(get_random_entry $MDTS)
230 var=${SERVERFACET}_nums
232 # Check that our client loads are still running. If any have died,
233 # that means they have died outside of recovery, which is unacceptable.
235 log "==== Checking the clients loads BEFORE failover -- failure NOT OK \
236 ELAPSED=$ELAPSED DURATION=$DURATION PERIOD=$SERVER_FAILOVER_PERIOD"
238 if ! check_client_loads $NODES_TO_USE; then
242 log "FAIL CLIENT $FAIL_CLIENT ... "
243 shutdown_client $FAIL_CLIENT
245 log "Starting failover on $SERVERFACET"
247 facet_failover "$SERVERFACET" || exit 1
248 if ! wait_recovery_complete $SERVERFACET $((TIMEOUT * 10)); then
249 echo "$SERVERFACET recovery is not completed!"
253 boot_node $FAIL_CLIENT
254 echo "Reintegrating $FAIL_CLIENT"
255 zconf_mount $FAIL_CLIENT $MOUNT || exit $?
257 # Increment the number of failovers
260 val=$((${!client_var} + 1))
261 eval $client_var=$val
263 # load script on failed clients could create END_RUN_FILE
264 # We shuold remove it and ignore the failure if this
265 # file contains the failed client only.
266 # We can not use ERRORS_OK when start all loads at the start of this script
267 # because the application errors allowed for random failed client only, but
268 # not for all clients.
269 if [ -e $END_RUN_FILE ]; then
270 read END_RUN_NODE < $END_RUN_FILE
271 [[ $END_RUN_NODE = $FAIL_CLIENT ]] &&
272 rm -f $END_RUN_FILE || exit 13
275 restart_client_loads $FAIL_CLIENT $ERRORS_OK || exit $?
277 # Check that not failed clients loads are still running.
278 # No application failures should occur on clients that was not failed.
280 log "==== Checking the clients loads AFTER failed client reintegrated -- failure NOT OK"
281 if ! ERRORS_OK= check_client_loads $(exclude_items_from_list $NODES_TO_USE $FAIL_CLIENT); then
282 log "Client load failed. Exiting"
286 CURRENT_TS=$(date +%s)
287 ELAPSED=$((CURRENT_TS - START_TS))
289 sleep=$((SERVER_FAILOVER_PERIOD-(CURRENT_TS - it_time_start)))
291 # keep count the number of itterations when
292 # time spend to failover and two client loads check exceeded
293 # the value ( SERVER_FAILOVER_PERIOD - MINSLEEP )
294 if [ $sleep -lt $MINSLEEP ]; then
295 reqfail=$((reqfail +1))
296 log "WARNING: failover, client reintegration and check_client_loads time exceeded SERVER_FAILOVER_PERIOD - MINSLEEP !
297 Failed to load the filesystem with I/O for a minimum period of $MINSLEEP $reqfail times ( REQFAIL=$REQFAIL ).
298 This iteration, the load was only applied for sleep=$sleep seconds.
299 Estimated max recovery time : $max_recov_time
300 Probably the hardware is taking excessively long to boot.
301 Try to increase SERVER_FAILOVER_PERIOD (current is $SERVER_FAILOVER_PERIOD), bug 20918"
302 [ $reqfail -gt $REQFAIL ] && exit 6
305 log " Number of failovers:
306 $(numfailovers) and counting..."
308 if [ $((ELAPSED + sleep)) -gt $DURATION ]; then
312 if [ $sleep -gt 0 ]; then
313 echo "sleeping $sleep seconds ... "