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 { FAIL_ON_ERROR=true skip_env "$0 Empty SHARED_DIRECTORY" && exit 0; }
37 check_shared_dir $SHARED_DIRECTORY ||
38 error "$SHARED_DIRECTORY isn't a shared directory"
40 [ -n "$CLIENTS" ] || \
41 { FAIL_ON_ERROR=true skip_env "$0 Need two or more remote clients" && exit 0; }
43 [ $CLIENTCOUNT -ge 3 ] || \
44 { FAIL_ON_ERROR=true skip_env "$0 Need two or more remote clients, have $((CLIENTCOUNT - 1))" && exit 0; }
46 END_RUN_FILE=${END_RUN_FILE:-$SHARED_DIRECTORY/end_run_file}
47 LOAD_PID_FILE=${LOAD_PID_FILE:-$TMP/client-load.pid}
49 remote_mds_nodsh && skip "remote MDS with nodsh" && exit 0
51 [[ $FAILURE_MODE = SOFT ]] && \
52 log "WARNING: $0 is not functional with FAILURE_MODE = SOFT, bz22797"
56 check_and_setup_lustre
57 rm -rf $DIR/[df][0-9]*
59 max_recov_time=$(max_recovery_time)
61 # the test node needs to be insulated from a lustre failure as much as possible,
62 # so not even loading the lustre modules is ideal.
64 # -- remove hostname from clients list
65 zconf_umount $(hostname) $MOUNT
66 NODES_TO_USE=${NODES_TO_USE:-$CLIENTS}
67 NODES_TO_USE=$(exclude_items_from_list $NODES_TO_USE $(hostname))
69 check_progs_installed $NODES_TO_USE ${CLIENT_LOADS[@]}
71 MDTS=$(get_facets MDS)
73 if [ "$SLOW" = "no" ]; then
74 DURATION=${DURATION:-$((60 * 30))}
75 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 5))}
77 DURATION=${DURATION:-$((60 * 60 * 24))}
78 SERVER_FAILOVER_PERIOD=${SERVER_FAILOVER_PERIOD:-$((60 * 10))} # 10 minutes
83 vmstatLOG=${TESTSUITELOG}_$(basename $0 .sh).vmstat
89 for facet in $MDTS ${failed_clients//,/ }; do
93 echo "$facet failed over $val times"
98 # list is comma separated
102 do_nodes $list "node=\\\$(hostname)
104 log=${TESTSUITELOG}_run_${!var}.sh-\\\$node.debug
105 if [ -e \\\$log ] ; then
106 echo Node \\\$node debug log:
111 summary_and_cleanup () {
117 # Having not empty END_RUN_FILE means the failed loads only
118 if [ -s $END_RUN_FILE ]; then
119 echo "Found the END_RUN_FILE file: $END_RUN_FILE"
122 read END_RUN_NODE < $END_RUN_FILE
124 # a client load will end (i.e. fail) if it finds
125 # the end run file. that does not mean that that client load
126 # actually failed though. the first node in the END_RUN_NODE is
127 # the one we are really interested in.
128 if [ -n "$END_RUN_NODE" ]; then
129 var=$(node_var_name $END_RUN_NODE)_load
130 echo "Client load failed on node $END_RUN_NODE"
132 echo "client $END_RUN_NODE load stdout and debug files :
133 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}
134 ${TESTSUITELOG}_run_${!var}.sh-${END_RUN_NODE}.debug"
140 echo $(date +'%F %H:%M:%S') Terminating clients loads ...
141 echo "$0" >> $END_RUN_FILE
143 [ $rc -eq 0 ] || result=FAIL
145 log "Duration: $DURATION
146 Server failover period: $SERVER_FAILOVER_PERIOD seconds
147 Exited after: $ELAPSED seconds
148 Number of failovers before exit:
150 Status: $result: rc=$rc"
152 # stop the vmstats on the OSTs
153 if [ "$VMSTAT" ]; then
154 do_nodes $(comma_list $(osts_nodes)) "test -f /tmp/vmstat.pid && \
155 { kill -s TERM \$(cat /tmp/vmstat.pid); rm -f /tmp/vmstat.pid; \
156 gzip -f9 $vmstatLOG-\$(hostname); }"
159 # make sure the client loads die
160 do_nodes $NODES_TO_USE "set -x; test -f $LOAD_PID_FILE && \
161 { kill -s TERM \$(cat $LOAD_PID_FILE) || true; }"
163 # and free up the pdshes that started them, if any are still around
164 if [ -n "$CLIENT_LOAD_PIDS" ]; then
165 kill $CLIENT_LOAD_PIDS || true
167 kill -9 $CLIENT_LOAD_PIDS || true
170 if [ $rc -ne 0 ]; then
171 print_logs $NODES_TO_USE
172 # we are interested in only on failed clients and servers
173 local failedclients=$(cat $END_RUN_FILE | grep -v $0)
174 # FIXME: need ostfailover-s nodes also for FLAVOR=OST
175 local product=$(gather_logs $(comma_list $(osts_nodes) \
176 $(mdts_nodes) $mdsfailover_HOST $failedclients) 1)
177 echo logs files $product
180 [ $rc -eq 0 ] && zconf_mount $(hostname) $MOUNT
188 log "-----============= $0 starting =============-----"
190 trap summary_and_cleanup EXIT # INT
195 if [ "$VMSTAT" ]; then
196 do_nodes $(comma_list $(osts_nodes)) "vmstat 1 > $vmstatLOG-\$(hostname) 2>/dev/null </dev/null & echo \$! > /tmp/vmstat.pid"
199 # Start client loads.
200 start_client_loads $NODES_TO_USE
202 echo clients load pids:
203 if ! do_nodesv $NODES_TO_USE "cat $LOAD_PID_FILE"; then
210 MINSLEEP=${MINSLEEP:-120}
211 REQFAIL_PERCENT=${REQFAIL_PERCENT:-3} # bug17839 comment 62
212 REQFAIL=${REQFAIL:-$(( DURATION / SERVER_FAILOVER_PERIOD * REQFAIL_PERCENT / 100))}
216 # This is used for FAIL_CLIENT only
218 while [ $ELAPSED -lt $DURATION -a ! -e $END_RUN_FILE ]; do
220 # In order to perform the
221 # expected number of failovers, we need to account the following :
222 # 1) the time that has elapsed during the client load checking
223 # 2) time takes for failover
225 it_time_start=$(date +%s)
227 FAIL_CLIENT=$(get_random_entry $NODES_TO_USE)
228 client_var=$(node_var_name $FAIL_CLIENT)_nums
230 # store the list of failed clients
231 # lists are comma separated
232 failed_clients=$(expand_list $failed_clients $FAIL_CLIENT)
234 SERVERFACET=$(get_random_entry $MDTS)
235 var=${SERVERFACET}_nums
237 # Check that our client loads are still running. If any have died,
238 # that means they have died outside of recovery, which is unacceptable.
240 log "==== Checking the clients loads BEFORE failover -- failure NOT OK \
241 ELAPSED=$ELAPSED DURATION=$DURATION PERIOD=$SERVER_FAILOVER_PERIOD"
243 if ! check_client_loads $NODES_TO_USE; then
247 log "FAIL CLIENT $FAIL_CLIENT ... "
248 shutdown_client $FAIL_CLIENT
250 log "Starting failover on $SERVERFACET"
252 facet_failover "$SERVERFACET" || exit 1
253 if ! wait_recovery_complete $SERVERFACET ; then
254 echo "$SERVERFACET recovery is not completed!"
258 boot_node $FAIL_CLIENT
259 echo "Reintegrating $FAIL_CLIENT"
260 zconf_mount $FAIL_CLIENT $MOUNT || exit $?
262 # Increment the number of failovers
265 val=$((${!client_var} + 1))
266 eval $client_var=$val
268 # load script on failed clients could create END_RUN_FILE
269 # We shuold remove it and ignore the failure if this
270 # file contains the failed client only.
271 # We can not use ERRORS_OK when start all loads at the start of this script
272 # because the application errors allowed for random failed client only, but
273 # not for all clients.
274 if [ -e $END_RUN_FILE ]; then
275 read END_RUN_NODE < $END_RUN_FILE
276 [[ $END_RUN_NODE = $FAIL_CLIENT ]] &&
277 rm -f $END_RUN_FILE || exit 13
280 restart_client_loads $FAIL_CLIENT $ERRORS_OK || exit $?
282 # Check that not failed clients loads are still running.
283 # No application failures should occur on clients that was not failed.
285 log "==== Checking the clients loads AFTER failed client reintegrated -- failure NOT OK"
286 if ! ERRORS_OK= check_client_loads $(exclude_items_from_list $NODES_TO_USE $FAIL_CLIENT); then
287 log "Client load failed. Exiting"
291 CURRENT_TS=$(date +%s)
292 ELAPSED=$((CURRENT_TS - START_TS))
293 sleep=$((SERVER_FAILOVER_PERIOD-(CURRENT_TS - it_time_start)))
295 # keep count the number of itterations when
296 # time spend to failover and two client loads check exceeded
297 # the value ( SERVER_FAILOVER_PERIOD - MINSLEEP )
298 if [ $sleep -lt $MINSLEEP ]; then
299 reqfail=$((reqfail +1))
300 log "WARNING: failover, client reintegration and check_client_loads time exceeded SERVER_FAILOVER_PERIOD - MINSLEEP !
301 Failed to load the filesystem with I/O for a minimum period of $MINSLEEP $reqfail times ( REQFAIL=$REQFAIL ).
302 This iteration, the load was only applied for sleep=$sleep seconds.
303 Estimated max recovery time : $max_recov_time
304 Probably the hardware is taking excessively long to boot.
305 Try to increase SERVER_FAILOVER_PERIOD (current is $SERVER_FAILOVER_PERIOD), bug 20918"
306 [ $reqfail -gt $REQFAIL ] && exit 6
309 log " Number of failovers:
310 $(numfailovers) and counting..."
312 if [ $((ELAPSED + sleep)) -ge $DURATION ]; then
316 if [ $sleep -gt 0 ]; then
317 echo "sleeping $sleep seconds ... "