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
45 [[ $FAILURE_MODE = SOFT ]] && \
46 log "WARNING: $0 is not functional with FAILURE_MODE = SOFT, bz22797"
50 check_and_setup_lustre
51 rm -rf $DIR/[df][0-9]*
53 max_recov_time=$(max_recovery_time)
55 # the test node needs to be insulated from a lustre failure as much as possible,
56 # so not even loading the lustre modules is ideal.
58 # -- remove hostname from clients list
59 zconf_umount $(hostname) $MOUNT
60 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
61 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $(hostname))
63 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
65 MDTS=$(get_facets MDS)
67 if [ "$SLOW" = "no" ]; then
68 DURATION=${DURATION:-$((60 * 30))}
69 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 5))}
71 DURATION=${DURATION:-$((60 * 60 * 24))}
72 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 10))} # 10 minutes
77 vmstatLOG=${TESTSUITELOG}_$(basename $0 .sh).vmstat
83 for facet in $MDTS ${failed_clients//,/ }; do
87 echo "$facet failed over $val times"
92 # list is comma separated
96 do_nodes $list "node=\\\$(hostname)
98 log=${TESTSUITELOG}_run_${!var}.sh-\\\$node.debug
99 if [ -e \\\$log ] ; then
100 echo Node \\\$node debug log:
105 summary_and_cleanup () {
111 # Having not empty END_RUN_FILE means the failed loads only
112 if [ -s $END_RUN_FILE ]; then
113 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
116 read END_RUN_NODE < $END_RUN_FILE
118 # a client load will end (i.e. fail) if it finds
119 # the end run file. that does not mean that that client load
120 # actually failed though. the first node in the END_RUN_NODE is
121 # the one we are really interested in.
122 if [ -n "$END_RUN_NODE" ]; then
123 var=$(client_var_name $END_RUN_NODE)_load
124 echo "Client load failed on node $END_RUN_NODE"
126 echo "client $END_RUN_NODE load stdout and debug files :
127 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}
128 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug"
134 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
135 echo "$0" >> $END_RUN_FILE
137 [ $rc -eq 0 ] || result=FAIL
139 log "Duration: $DURATION
140 Server failover period: $SERVER_FAILOVER_PERIOD seconds
141 Exited after: $ELAPSED seconds
142 Number of failovers before exit:
144 Status: $result: rc=$rc"
146 # stop the vmstats on the OSTs
147 if [ "$VMSTAT" ]; then
148 do_nodes $(comma_list $(osts_nodes)) "test -f /tmp/vmstat.pid && \
149 { kill -s TERM \$(cat /tmp/vmstat.pid); rm -f /tmp/vmstat.pid; \
150 gzip -f9 $vmstatLOG-\$(hostname); }"
153 # make sure the client loads die
154 do_nodes $NODES_TO_USE "set -x; test -f $LOAD_PID_FILE && \
155 { kill -s TERM \$(cat $LOAD_PID_FILE) || true; }"
157 # and free up the pdshes that started them, if any are still around
158 if [ -n "$CLIENT_LOAD_PIDS" ]; then
159 kill $CLIENT_LOAD_PIDS || true
161 kill -9 $CLIENT_LOAD_PIDS || true
164 if [ $rc -ne 0 ]; then
165 print_logs $NODES_TO_USE
166 # we are interested in only on failed clients and servers
167 local failedclients=$(cat $END_RUN_FILE | grep -v $0)
168 # FIXME: need ostfailover-s nodes also for FLAVOR=OST
169 local product=$(gather_logs $(comma_list $(osts_nodes) \
170 $mds_HOST $mdsfailover_HOST $failedclients))
171 echo logs files $product
174 [ $rc -eq 0 ] && zconf_mount $(hostname) $MOUNT
182 log "-----============= $0 starting =============-----"
184 trap summary_and_cleanup EXIT # INT
189 if [ "$VMSTAT" ]; then
190 do_nodes $(comma_list $(osts_nodes)) "vmstat 1 > $vmstatLOG-\$(hostname) 2>/dev/null </dev/null & echo \$! > /tmp/vmstat.pid"
193 # Start client loads.
194 start_client_loads $NODES_TO_USE
196 echo clients load pids:
197 if ! do_nodesv $NODES_TO_USE "cat $LOAD_PID_FILE"; then
204 MINSLEEP=${MINSLEEP:-120}
205 REQFAIL_PERCENT=${REQFAIL_PERCENT:-3} # bug17839 comment 62
206 REQFAIL=${REQFAIL:-$(( DURATION / SERVER_FAILOVER_PERIOD * REQFAIL_PERCENT / 100))}
210 # This is used for FAIL_CLIENT only
212 while [ $ELAPSED -lt $DURATION -a ! -e $END_RUN_FILE ]; do
214 # In order to perform the
215 # expected number of failovers, we need to account the following :
216 # 1) the time that has elapsed during the client load checking
217 # 2) time takes for failover
219 it_time_start=$(date +%s)
221 FAIL_CLIENT=$(get_random_entry $NODES_TO_USE)
222 client_var=$(client_var_name $FAIL_CLIENT)_nums
224 # store the list of failed clients
225 # lists are comma separated
226 failed_clients=$(expand_list $failed_clients $FAIL_CLIENT)
228 SERVERFACET=$(get_random_entry $MDTS)
229 var=${SERVERFACET}_nums
231 # Check that our client loads are still running. If any have died,
232 # that means they have died outside of recovery, which is unacceptable.
234 log "==== Checking the clients loads BEFORE failover -- failure NOT OK \
235 ELAPSED=$ELAPSED DURATION=$DURATION PERIOD=$SERVER_FAILOVER_PERIOD"
237 if ! check_client_loads $NODES_TO_USE; then
241 log "FAIL CLIENT $FAIL_CLIENT ... "
242 shutdown_client $FAIL_CLIENT
244 log "Starting failover on $SERVERFACET"
246 facet_failover "$SERVERFACET" || exit 1
247 if ! wait_recovery_complete $SERVERFACET $((TIMEOUT * 10)); then
248 echo "$SERVERFACET recovery is not completed!"
252 boot_node $FAIL_CLIENT
253 echo "Reintegrating $FAIL_CLIENT"
254 zconf_mount $FAIL_CLIENT $MOUNT || exit $?
256 # Increment the number of failovers
259 val=$((${!client_var} + 1))
260 eval $client_var=$val
262 # load script on failed clients could create END_RUN_FILE
263 # We shuold remove it and ignore the failure if this
264 # file contains the failed client only.
265 # We can not use ERRORS_OK when start all loads at the start of this script
266 # because the application errors allowed for random failed client only, but
267 # not for all clients.
268 if [ -e $END_RUN_FILE ]; then
269 read END_RUN_NODE < $END_RUN_FILE
270 [[ $END_RUN_NODE = $FAIL_CLIENT ]] &&
271 rm -f $END_RUN_FILE || exit 13
274 restart_client_loads $FAIL_CLIENT $ERRORS_OK || exit $?
276 # Check that not failed clients loads are still running.
277 # No application failures should occur on clients that was not failed.
279 log "==== Checking the clients loads AFTER failed client reintegrated -- failure NOT OK"
280 if ! ERRORS_OK= check_client_loads $(exclude_items_from_list $NODES_TO_USE $FAIL_CLIENT); then
281 log "Client load failed. Exiting"
285 CURRENT_TS=$(date +%s)
286 ELAPSED=$((CURRENT_TS - START_TS))
288 sleep=$((SERVER_FAILOVER_PERIOD-(CURRENT_TS - it_time_start)))
290 # keep count the number of itterations when
291 # time spend to failover and two client loads check exceeded
292 # the value ( SERVER_FAILOVER_PERIOD - MINSLEEP )
293 if [ $sleep -lt $MINSLEEP ]; then
294 reqfail=$((reqfail +1))
295 log "WARNING: failover, client reintegration and check_client_loads time exceeded SERVER_FAILOVER_PERIOD - MINSLEEP !
296 Failed to load the filesystem with I/O for a minimum period of $MINSLEEP $reqfail times ( REQFAIL=$REQFAIL ).
297 This iteration, the load was only applied for sleep=$sleep seconds.
298 Estimated max recovery time : $max_recov_time
299 Probably the hardware is taking excessively long to boot.
300 Try to increase SERVER_FAILOVER_PERIOD (current is $SERVER_FAILOVER_PERIOD), bug 20918"
301 [ $reqfail -gt $REQFAIL ] && exit 6
304 log " Number of failovers:
305 $(numfailovers) and counting..."
307 if [ $((ELAPSED + sleep)) -gt $DURATION ]; then
311 if [ $sleep -gt 0 ]; then
312 echo "sleeping $sleep seconds ... "