7 # This test needs to be run on the client
10 LUSTRE=${LUSTRE:-$(cd $(dirname $0)/..; echo $PWD)}
13 . $LUSTRE/tests/test-framework.sh
15 . ${CONFIG:=$LUSTRE/tests/cfg/$NAME.sh}
16 CHECK_GRANT=${CHECK_GRANT:-"yes"}
17 GRANT_CHECK_LIST=${GRANT_CHECK_LIST:-""}
19 remote_mds_nodsh && log "SKIP: remote MDS with nodsh" && exit 0
22 # bug number: 17466 15962
23 ALWAYS_EXCEPT="61d $REPLAY_SINGLE_EXCEPT"
25 if [ "$FAILURE_MODE" = "HARD" ] && mixed_ost_devs; then
26 CONFIG_EXCEPTIONS="0b 42 47 61a 61c"
27 echo -n "Several ost services on one ost node are used with FAILURE_MODE=$FAILURE_MODE. "
28 echo "Except the tests: $CONFIG_EXCEPTIONS"
29 ALWAYS_EXCEPT="$ALWAYS_EXCEPT $CONFIG_EXCEPTIONS"
32 # 63 min 7 min AT AT AT AT"
33 [ "$SLOW" = "no" ] && EXCEPT_SLOW="1 2 3 4 6 12 16 44a 44b 65 66 67 68"
37 check_and_setup_lustre
42 rm -rf $DIR/[df][0-9]*
44 test_0a() { # was test_0
47 replay_barrier $SINGLEMDS
51 run_test 0a "empty replay"
54 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
56 # this test attempts to trigger a race in the precreation code,
57 # and must run before any other objects are created on the filesystem
59 createmany -o $DIR/$tfile 20 || return 1
60 unlinkmany $DIR/$tfile 20 || return 2
62 run_test 0b "ensure object created after recover exists. (3284)"
68 lctl set_param -n seq.cli-srv-$mds-mdc-*.width=$width
74 lctl get_param -n seq.cli-srv-$mds-mdc-*.width
77 # This test should pass for single-mds and multi-mds configs.
78 # But for different configurations it tests different things.
82 # (1) fld_create replay should happen;
84 # (2) fld_create replay should not return -EEXISTS, if it does
85 # this means sequence manager recovery code is buggy and allocated
86 # same sequence two times after recovery.
90 # (1) fld_create replay may not happen, because its home MDS is
91 # MDS2 which is not involved to revovery;
93 # (2) as fld_create does not happen on MDS1, it does not make any
96 local label=`mdsdevlabel 1`
97 [ -z "$label" ] && echo "No label for mds1" && return 1
99 replay_barrier $SINGLEMDS
100 local sw=`seq_get_width $label`
102 # make seq manager switch to next sequence each
103 # time as new fid is needed.
104 seq_set_width $label 1
106 # make sure that fld has created at least one new
108 touch $DIR/$tfile || return 2
109 seq_set_width $label $sw
111 # fail $SINGLEMDS and start recovery, replay RPCs, etc.
114 # wait for recovery finish
118 # flush fld cache and dentry cache to make it lookup
119 # created entry instead of revalidating existent one
121 zconf_mount `hostname` $MOUNT
123 # issue lookup which should call fld lookup which
124 # should fail if client did not replay fld create
125 # correctly and server has no fld entry
126 touch $DIR/$tfile || return 3
127 rm $DIR/$tfile || return 4
129 run_test 0c "fld create"
132 replay_barrier $SINGLEMDS
135 $CHECKSTAT -t file $DIR/$tfile || return 1
138 run_test 1 "simple create"
141 replay_barrier $SINGLEMDS
144 $CHECKSTAT -t file $DIR/$tfile || return 1
151 replay_barrier $SINGLEMDS
154 $CHECKSTAT -t file $DIR/$tfile || return 1
160 local file=$DIR/$tfile
161 replay_barrier $SINGLEMDS
163 openfile -f O_DIRECTORY $file
165 $CHECKSTAT -t file $file || return 2
168 run_test 3a "replay failed open(O_DIRECTORY)"
171 replay_barrier $SINGLEMDS
172 #define OBD_FAIL_MDS_OPEN_PACK | OBD_FAIL_ONCE
173 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000114"
175 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
177 $CHECKSTAT -t file $DIR/$tfile && return 2
180 run_test 3b "replay failed open -ENOMEM"
183 replay_barrier $SINGLEMDS
184 #define OBD_FAIL_MDS_ALLOC_OBDO | OBD_FAIL_ONCE
185 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000128"
187 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
190 $CHECKSTAT -t file $DIR/$tfile && return 2
193 run_test 3c "replay failed open -ENOMEM"
195 test_4a() { # was test_4
196 replay_barrier $SINGLEMDS
197 for i in `seq 10`; do
198 echo "tag-$i" > $DIR/$tfile-$i
201 for i in `seq 10`; do
202 grep -q "tag-$i" $DIR/$tfile-$i || error "$tfile-$i"
205 run_test 4a "|x| 10 open(O_CREAT)s"
208 replay_barrier $SINGLEMDS
211 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
213 run_test 4b "|x| rm 10 files"
215 # The idea is to get past the first block of precreated files on both
216 # osts, and then replay.
218 replay_barrier $SINGLEMDS
219 for i in `seq 220`; do
220 echo "tag-$i" > $DIR/$tfile-$i
223 for i in `seq 220`; do
224 grep -q "tag-$i" $DIR/$tfile-$i || error "$tfile-$i"
228 # waiting for commitment of removal
230 run_test 5 "|x| 220 open(O_CREAT)"
233 test_6a() { # was test_6
235 replay_barrier $SINGLEMDS
236 mcreate $DIR/$tdir/$tfile
238 $CHECKSTAT -t dir $DIR/$tdir || return 1
239 $CHECKSTAT -t file $DIR/$tdir/$tfile || return 2
241 # waiting for log process thread
243 run_test 6a "mkdir + contained create"
247 replay_barrier $SINGLEMDS
250 $CHECKSTAT -t dir $DIR/$tdir && return 1 || true
252 run_test 6b "|X| rmdir"
256 replay_barrier $SINGLEMDS
257 mcreate $DIR/$tdir/$tfile
259 $CHECKSTAT -t dir $DIR/$tdir || return 1
260 $CHECKSTAT -t file $DIR/$tdir/$tfile || return 2
263 run_test 7 "mkdir |X| contained create"
266 # make sure no side-effect from previous test.
268 replay_barrier $SINGLEMDS
269 multiop_bg_pause $DIR/$tfile mo_c || return 4
273 $CHECKSTAT -t file $DIR/$tfile || return 1
274 kill -USR1 $MULTIPID || return 2
275 wait $MULTIPID || return 3
278 run_test 8 "creat open |X| close"
281 replay_barrier $SINGLEMDS
283 local old_inum=`ls -i $DIR/$tfile | awk '{print $1}'`
285 local new_inum=`ls -i $DIR/$tfile | awk '{print $1}'`
287 echo " old_inum == $old_inum, new_inum == $new_inum"
288 if [ $old_inum -eq $new_inum ] ;
290 echo " old_inum and new_inum match"
292 echo "!!!! old_inum and new_inum NOT match"
297 run_test 9 "|X| create (same inum/gen)"
301 replay_barrier $SINGLEMDS
302 mv $DIR/$tfile $DIR/$tfile-2
305 $CHECKSTAT $DIR/$tfile && return 1
306 $CHECKSTAT $DIR/$tfile-2 ||return 2
310 run_test 10 "create |X| rename unlink"
314 echo "old" > $DIR/$tfile
315 mv $DIR/$tfile $DIR/$tfile-2
316 replay_barrier $SINGLEMDS
317 echo "new" > $DIR/$tfile
319 grep old $DIR/$tfile-2
321 grep new $DIR/$tfile || return 1
322 grep old $DIR/$tfile-2 || return 2
324 run_test 11 "create open write rename |X| create-old-name read"
328 multiop_bg_pause $DIR/$tfile o_tSc || return 3
331 replay_barrier $SINGLEMDS
333 wait $pid || return 1
336 [ -e $DIR/$tfile ] && return 2
339 run_test 12 "open, unlink |X| close"
342 # 1777 - replay open after committed chmod that would make
343 # a regular open a failure
346 multiop_bg_pause $DIR/$tfile O_wc || return 3
349 $CHECKSTAT -p 0 $DIR/$tfile
350 replay_barrier $SINGLEMDS
353 wait $pid || return 1
355 $CHECKSTAT -s 1 -p 0 $DIR/$tfile || return 2
356 rm $DIR/$tfile || return 4
359 run_test 13 "open chmod 0 |x| write close"
362 multiop_bg_pause $DIR/$tfile O_tSc || return 4
365 replay_barrier $SINGLEMDS
366 kill -USR1 $pid || return 1
367 wait $pid || return 2
370 [ -e $DIR/$tfile ] && return 3
373 run_test 14 "open(O_CREAT), unlink |X| close"
376 multiop_bg_pause $DIR/$tfile O_tSc || return 5
379 replay_barrier $SINGLEMDS
380 touch $DIR/g11 || return 1
382 wait $pid || return 2
385 [ -e $DIR/$tfile ] && return 3
386 touch $DIR/h11 || return 4
389 run_test 15 "open(O_CREAT), unlink |X| touch new, close"
393 replay_barrier $SINGLEMDS
396 mcreate $DIR/$tfile-2
398 [ -e $DIR/$tfile ] && return 1
399 [ -e $DIR/$tfile-2 ] || return 2
400 munlink $DIR/$tfile-2 || return 3
402 run_test 16 "|X| open(O_CREAT), unlink, touch new, unlink new"
405 replay_barrier $SINGLEMDS
406 multiop_bg_pause $DIR/$tfile O_c || return 4
409 kill -USR1 $pid || return 1
410 wait $pid || return 2
411 $CHECKSTAT -t file $DIR/$tfile || return 3
414 run_test 17 "|X| open(O_CREAT), |replay| close"
417 replay_barrier $SINGLEMDS
418 multiop_bg_pause $DIR/$tfile O_tSc || return 8
421 touch $DIR/$tfile-2 || return 1
422 echo "pid: $pid will close"
424 wait $pid || return 2
427 [ -e $DIR/$tfile ] && return 3
428 [ -e $DIR/$tfile-2 ] || return 4
429 # this touch frequently fails
430 touch $DIR/$tfile-3 || return 5
431 munlink $DIR/$tfile-2 || return 6
432 munlink $DIR/$tfile-3 || return 7
435 run_test 18 "|X| open(O_CREAT), unlink, touch new, close, touch, unlink"
437 # bug 1855 (a simpler form of test_11 above)
439 replay_barrier $SINGLEMDS
441 echo "old" > $DIR/$tfile
442 mv $DIR/$tfile $DIR/$tfile-2
443 grep old $DIR/$tfile-2
445 grep old $DIR/$tfile-2 || return 2
447 run_test 19 "|X| mcreate, open, write, rename "
449 test_20a() { # was test_20
450 replay_barrier $SINGLEMDS
451 multiop_bg_pause $DIR/$tfile O_tSc || return 3
457 wait $pid || return 1
458 [ -e $DIR/$tfile ] && return 2
461 run_test 20a "|X| open(O_CREAT), unlink, replay, close (test mds_cleanup_orphans)"
463 test_20b() { # bug 10480
464 BEFOREUSED=`df -P $DIR | tail -1 | awk '{ print $3 }'`
466 dd if=/dev/zero of=$DIR/$tfile bs=4k count=10000 &
468 while [ ! -e $DIR/$tfile ] ; do
469 usleep 60 # give dd a chance to start
472 lfs getstripe $DIR/$tfile || return 1
473 rm -f $DIR/$tfile || return 2 # make it an orphan
475 df -P $DIR || df -P $DIR || true # reconnect
477 fail $SINGLEMDS # start orphan recovery
478 df -P $DIR || df -P $DIR || true # reconnect
479 wait_recovery_complete $SINGLEMDS || error "MDS recovery not done"
481 # just because recovery is done doesn't mean we've finished
482 # orphan cleanup. Wait for llogs to get synchronized.
483 echo waiting for orphan cleanup...
485 local -a sync=($(do_facet ost "$LCTL get_param obdfilter.*.mds_sync" | awk -F= ' {print $2}'))
487 for ((i=0; i<${#sync[@]}; i++)); do
488 [ ${sync[$i]} -eq 0 ] && continue
489 # there is a not finished MDS-OST synchronization
493 [ ${con} -eq 1 ] && break
497 # let the statfs cache to get old enough.
500 AFTERUSED=`df -P $DIR | tail -1 | awk '{ print $3 }'`
501 log "before $BEFOREUSED, after $AFTERUSED"
502 [ $AFTERUSED -gt $((BEFOREUSED + 20)) ] && \
503 error "after $AFTERUSED > before $BEFOREUSED"
506 run_test 20b "write, unlink, eviction, replay, (test mds_cleanup_orphans)"
508 test_20c() { # bug 10480
509 multiop_bg_pause $DIR/$tfile Ow_c || return 1
516 df -P $DIR || df -P $DIR || true # reconnect
519 test -s $DIR/$tfile || error "File was truncated"
521 wait $pid || return 1
524 run_test 20c "check that client eviction does not affect file content"
527 replay_barrier $SINGLEMDS
528 multiop_bg_pause $DIR/$tfile O_tSc || return 5
531 touch $DIR/g11 || return 1
535 wait $pid || return 2
536 [ -e $DIR/$tfile ] && return 3
537 touch $DIR/h11 || return 4
540 run_test 21 "|X| open(O_CREAT), unlink touch new, replay, close (test mds_cleanup_orphans)"
543 multiop_bg_pause $DIR/$tfile O_tSc || return 3
546 replay_barrier $SINGLEMDS
551 wait $pid || return 1
552 [ -e $DIR/$tfile ] && return 2
555 run_test 22 "open(O_CREAT), |X| unlink, replay, close (test mds_cleanup_orphans)"
558 multiop_bg_pause $DIR/$tfile O_tSc || return 5
561 replay_barrier $SINGLEMDS
563 touch $DIR/g11 || return 1
567 wait $pid || return 2
568 [ -e $DIR/$tfile ] && return 3
569 touch $DIR/h11 || return 4
572 run_test 23 "open(O_CREAT), |X| unlink touch new, replay, close (test mds_cleanup_orphans)"
575 multiop_bg_pause $DIR/$tfile O_tSc || return 3
578 replay_barrier $SINGLEMDS
582 wait $pid || return 1
583 [ -e $DIR/$tfile ] && return 2
586 run_test 24 "open(O_CREAT), replay, unlink, close (test mds_cleanup_orphans)"
589 multiop_bg_pause $DIR/$tfile O_tSc || return 3
593 replay_barrier $SINGLEMDS
596 wait $pid || return 1
597 [ -e $DIR/$tfile ] && return 2
600 run_test 25 "open(O_CREAT), unlink, replay, close (test mds_cleanup_orphans)"
603 replay_barrier $SINGLEMDS
604 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
606 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
611 wait $pid2 || return 1
615 wait $pid1 || return 2
616 [ -e $DIR/$tfile-1 ] && return 3
617 [ -e $DIR/$tfile-2 ] && return 4
620 run_test 26 "|X| open(O_CREAT), unlink two, close one, replay, close one (test mds_cleanup_orphans)"
623 replay_barrier $SINGLEMDS
624 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
626 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
633 wait $pid1 || return 1
635 wait $pid2 || return 2
636 [ -e $DIR/$tfile-1 ] && return 3
637 [ -e $DIR/$tfile-2 ] && return 4
640 run_test 27 "|X| open(O_CREAT), unlink two, replay, close two (test mds_cleanup_orphans)"
643 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
645 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
647 replay_barrier $SINGLEMDS
651 wait $pid2 || return 1
655 wait $pid1 || return 2
656 [ -e $DIR/$tfile-1 ] && return 3
657 [ -e $DIR/$tfile-2 ] && return 4
660 run_test 28 "open(O_CREAT), |X| unlink two, close one, replay, close one (test mds_cleanup_orphans)"
663 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
665 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
667 replay_barrier $SINGLEMDS
673 wait $pid1 || return 1
675 wait $pid2 || return 2
676 [ -e $DIR/$tfile-1 ] && return 3
677 [ -e $DIR/$tfile-2 ] && return 4
680 run_test 29 "open(O_CREAT), |X| unlink two, replay, close two (test mds_cleanup_orphans)"
683 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
685 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
690 replay_barrier $SINGLEMDS
693 wait $pid1 || return 1
695 wait $pid2 || return 2
696 [ -e $DIR/$tfile-1 ] && return 3
697 [ -e $DIR/$tfile-2 ] && return 4
700 run_test 30 "open(O_CREAT) two, unlink two, replay, close two (test mds_cleanup_orphans)"
703 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
705 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
709 replay_barrier $SINGLEMDS
713 wait $pid1 || return 1
715 wait $pid2 || return 2
716 [ -e $DIR/$tfile-1 ] && return 3
717 [ -e $DIR/$tfile-2 ] && return 4
720 run_test 31 "open(O_CREAT) two, unlink one, |X| unlink one, close two (test mds_cleanup_orphans)"
722 # tests for bug 2104; completion without crashing is success. The close is
723 # stale, but we always return 0 for close, so the app never sees it.
725 multiop_bg_pause $DIR/$tfile O_c || return 2
727 multiop_bg_pause $DIR/$tfile O_c || return 3
730 df $MOUNT || sleep 1 && df $MOUNT || return 1
733 wait $pid1 || return 4
734 wait $pid2 || return 5
737 run_test 32 "close() notices client eviction; close() after client eviction"
739 # Abort recovery before client complete
740 test_33a() { # was test_33
741 replay_barrier $SINGLEMDS
742 createmany -o $DIR/$tfile-%d 100
743 fail_abort $SINGLEMDS
744 # this file should be gone, because the replay was aborted
745 $CHECKSTAT -t file $DIR/$tfile-* && return 3
746 unlinkmany $DIR/$tfile-%d 0 100
749 run_test 33a "abort recovery before client does replay"
751 # Stale FID sequence bug 15962
752 test_33b() { # was test_33a
753 replay_barrier $SINGLEMDS
754 createmany -o $DIR/$tfile-%d 10
755 fail_abort $SINGLEMDS
756 unlinkmany $DIR/$tfile-%d 0 10
757 # recreate shouldn't fail
758 createmany -o $DIR/$tfile-%d 10 || return 3
759 unlinkmany $DIR/$tfile-%d 0 10
762 run_test 33b "fid shouldn't be reused after abort recovery"
765 multiop_bg_pause $DIR/$tfile O_c || return 2
769 replay_barrier $SINGLEMDS
770 fail_abort $SINGLEMDS
772 wait $pid || return 3
773 [ -e $DIR/$tfile ] && return 1
777 run_test 34 "abort recovery before client does replay (test mds_cleanup_orphans)"
779 # bug 2278 - generate one orphan on OST, then destroy it during recovery from llog
783 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
784 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000119"
789 # give a chance to remove from MDS
790 fail_abort $SINGLEMDS
791 $CHECKSTAT -t file $DIR/$tfile && return 1 || true
793 run_test 35 "test recovery from llog for unlink op"
795 # b=2432 resent cancel after replay uses wrong cookie,
796 # so don't resend cancels
798 replay_barrier $SINGLEMDS
800 checkstat $DIR/$tfile
801 facet_failover $SINGLEMDS
803 if dmesg | grep "unknown lock cookie"; then
804 echo "cancel after replay failed"
808 run_test 36 "don't resend cancel"
811 # directory orphans can't be unlinked from PENDING directory
813 rmdir $DIR/$tfile 2>/dev/null
814 multiop_bg_pause $DIR/$tfile dD_c || return 2
818 replay_barrier $SINGLEMDS
819 # clear the dmesg buffer so we only see errors from this recovery
821 fail_abort $SINGLEMDS
823 dmesg | grep "mds_unlink_orphan.*error .* unlinking orphan" && return 1
824 wait $pid || return 3
828 run_test 37 "abort recovery before client does replay (test mds_cleanup_orphans for directories)"
831 createmany -o $DIR/$tfile-%d 800
832 unlinkmany $DIR/$tfile-%d 0 400
833 replay_barrier $SINGLEMDS
835 unlinkmany $DIR/$tfile-%d 400 400
837 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
839 run_test 38 "test recovery from unlink llog (test llog_gen_rec) "
841 test_39() { # bug 4176
842 createmany -o $DIR/$tfile-%d 800
843 replay_barrier $SINGLEMDS
844 unlinkmany $DIR/$tfile-%d 0 400
846 unlinkmany $DIR/$tfile-%d 400 400
848 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
850 run_test 39 "test recovery from unlink llog (test llog_gen_rec) "
853 lctl get_param -n osc.*.stats | awk -vwrites=0 '/ost_write/ { writes += $2 } END { print writes; }'
858 $LCTL mark multiop $MOUNT/$tfile OS_c
859 multiop $MOUNT/$tfile OS_c &
861 writeme -s $MOUNT/${tfile}-2 &
864 facet_failover $SINGLEMDS
865 #define OBD_FAIL_MDS_CONNECT_NET 0x117
866 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000117"
868 stat1=`count_ost_writes`
870 stat2=`count_ost_writes`
871 echo "$stat1, $stat2"
872 if [ $stat1 -lt $stat2 ]; then
873 echo "writes continuing during recovery"
876 echo "writes not continuing during recovery, bug 2477"
879 echo "waiting for writeme $WRITE_PID"
883 echo "waiting for multiop $PID"
884 wait $PID || return 2
885 do_facet client munlink $MOUNT/$tfile || return 3
886 do_facet client munlink $MOUNT/${tfile}-2 || return 3
889 run_test 40 "cause recovery in ptlrpc, ensure IO continues"
893 # make sure that a read to one osc doesn't try to double-unlock its page just
894 # because another osc is invalid. trigger_group_io used to mistakenly return
895 # an error if any oscs were invalid even after having successfully put rpcs
896 # on valid oscs. This was fatal if the caller was ll_readpage who unlocked
897 # the page, guarnateeing that the unlock from the RPC completion would
898 # assert on trying to unlock the unlocked page.
900 [ $OSTCOUNT -lt 2 ] && \
901 skip "skipping test 41: we don't have a second OST to test with" && \
904 local f=$MOUNT/$tfile
905 # make sure the start of the file is ost1
906 lfs setstripe $f -s $((128 * 1024)) -i 0
907 do_facet client dd if=/dev/zero of=$f bs=4k count=1 || return 3
909 # fail ost2 and read from ost1
910 local osc2dev=`do_facet $SINGLEMDS "lctl get_param -n devices | grep ${ost2_svc}-osc-MDT0000" | awk '{print $1}'`
911 [ -z "$osc2dev" ] && echo "OST: $ost2_svc" && lctl get_param -n devices && return 4
912 do_facet $SINGLEMDS $LCTL --device $osc2dev deactivate || return 1
913 do_facet client dd if=$f of=/dev/null bs=4k count=1 || return 3
914 do_facet $SINGLEMDS $LCTL --device $osc2dev activate || return 2
917 run_test 41 "read from a valid osc while other oscs are invalid"
919 # test MDS recovery after ost failure
921 blocks=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
922 createmany -o $DIR/$tfile-%d 800
924 unlinkmany $DIR/$tfile-%d 0 400
926 lctl set_param debug=-1
929 # osc is evicted, fs is smaller (but only with failout OSTs (bug 7287)
930 #blocks_after=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
931 #[ $blocks_after -lt $blocks ] || return 1
932 echo wait for MDS to timeout and recover
933 sleep $((TIMEOUT * 2))
935 unlinkmany $DIR/$tfile-%d 400 400
936 $CHECKSTAT -t file $DIR/$tfile-* && return 2 || true
938 run_test 42 "recovery after ost failure"
940 # timeout in MDS/OST recovery RPC will LBUG MDS
941 test_43() { # bug 2530
942 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
944 replay_barrier $SINGLEMDS
946 # OBD_FAIL_OST_CREATE_NET 0x204
947 do_facet ost1 "lctl set_param fail_loc=0x80000204"
950 do_facet ost1 "lctl set_param fail_loc=0"
954 run_test 43 "mds osc import failure during recovery; don't LBUG"
956 test_44a() { # was test_44
959 mdcdev=`lctl get_param -n devices | awk '/MDT0000-mdc-/ {print $1}'`
960 [ "$mdcdev" ] || exit 2
962 # adaptive timeouts slow this way down
963 if at_is_enabled; then
964 at_max_saved=$(at_max_get mds)
968 for i in `seq 1 10`; do
969 echo "$i of 10 ($(date +%s))"
970 do_facet $SINGLEMDS "lctl get_param -n mdt.*.mdt.timeouts | grep service"
971 #define OBD_FAIL_TGT_CONN_RACE 0x701
972 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000701"
973 $LCTL --device $mdcdev recover
976 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
977 [ $at_max_saved -ne 0 ] && at_max_set $at_max_saved mds
980 run_test 44a "race in target handle connect"
983 mdcdev=`lctl get_param -n devices | awk '/MDT0000-mdc-/ {print $1}'`
984 [ "$mdcdev" ] || exit 2
985 for i in `seq 1 10`; do
986 echo "$i of 10 ($(date +%s))"
987 do_facet $SINGLEMDS "lctl get_param -n mdt.*.mdt.timeouts | grep service"
988 #define OBD_FAIL_TGT_DELAY_RECONNECT 0x704
989 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000704"
990 $LCTL --device $mdcdev recover
993 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
996 run_test 44b "race in target handle connect"
998 # Handle failed close
1000 mdcdev=`lctl get_param -n devices | awk '/MDT0000-mdc-/ {print $1}'`
1001 [ "$mdcdev" ] || exit 2
1002 $LCTL --device $mdcdev recover
1004 multiop_bg_pause $DIR/$tfile O_c || return 1
1007 # This will cause the CLOSE to fail before even
1008 # allocating a reply buffer
1009 $LCTL --device $mdcdev deactivate || return 4
1013 wait $pid || return 1
1015 $LCTL --device $mdcdev activate || return 5
1018 $CHECKSTAT -t file $DIR/$tfile || return 2
1021 run_test 45 "Handle failed close"
1025 drop_reply "touch $DIR/$tfile"
1027 # ironically, the previous test, 45, will cause a real forced close,
1028 # so just look for one for this test
1029 dmesg | grep -i "force closing client file handle for $tfile" && return 1
1032 run_test 46 "Don't leak file handle after open resend (3325)"
1034 test_47() { # bug 2824
1035 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1037 # create some files to make sure precreate has been done on all
1038 # OSTs. (just in case this test is run independently)
1039 createmany -o $DIR/$tfile 20 || return 1
1041 # OBD_FAIL_OST_CREATE_NET 0x204
1043 do_facet ost1 "lctl set_param fail_loc=0x80000204"
1044 df $MOUNT || return 2
1046 # let the MDS discover the OST failure, attempt to recover, fail
1047 # and recover again.
1048 sleep $((3 * TIMEOUT))
1050 # Without 2824, this createmany would hang
1051 createmany -o $DIR/$tfile 20 || return 3
1052 unlinkmany $DIR/$tfile 20 || return 4
1054 do_facet ost1 "lctl set_param fail_loc=0"
1057 run_test 47 "MDS->OSC failure during precreate cleanup (2824)"
1060 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1061 [ "$OSTCOUNT" -lt "2" ] && skip "$OSTCOUNT < 2 OSTs -- skipping" && return
1063 replay_barrier $SINGLEMDS
1064 createmany -o $DIR/$tfile 20 || return 1
1065 # OBD_FAIL_OST_EROFS 0x216
1066 facet_failover $SINGLEMDS
1067 do_facet ost1 "lctl set_param fail_loc=0x80000216"
1068 df $MOUNT || return 2
1070 createmany -o $DIR/$tfile 20 20 || return 2
1071 unlinkmany $DIR/$tfile 40 || return 3
1074 run_test 48 "MDS->OSC failure during precreate cleanup (2824)"
1077 local oscdev=`do_facet $SINGLEMDS lctl get_param -n devices | grep ${ost1_svc}-osc-MDT0000 | awk '{print $1}'`
1078 [ "$oscdev" ] || return 1
1079 do_facet $SINGLEMDS $LCTL --device $oscdev recover || return 2
1080 do_facet $SINGLEMDS $LCTL --device $oscdev recover || return 3
1081 # give the mds_lov_sync threads a chance to run
1084 run_test 50 "Double OSC recovery, don't LASSERT (3812)"
1086 # b3764 timed out lock replay
1089 cancel_lru_locks mdc
1091 multiop $DIR/$tfile s || return 1
1092 replay_barrier $SINGLEMDS
1093 #define OBD_FAIL_LDLM_REPLY 0x30c
1094 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000030c"
1095 fail $SINGLEMDS || return 2
1096 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1098 $CHECKSTAT -t file $DIR/$tfile-* && return 3 || true
1100 run_test 52 "time out lock replay (3764)"
1102 # bug 3462 - simultaneous MDC requests
1104 mkdir -p $DIR/${tdir}-1
1105 mkdir -p $DIR/${tdir}-2
1106 multiop $DIR/${tdir}-1/f O_c &
1108 # give multiop a change to open
1111 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1112 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1113 kill -USR1 $close_pid
1114 cancel_lru_locks mdc # force the close
1115 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1117 mcreate $DIR/${tdir}-2/f || return 1
1119 # close should still be here
1120 [ -d /proc/$close_pid ] || return 2
1122 replay_barrier_nodf $SINGLEMDS
1124 wait $close_pid || return 3
1126 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1127 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1128 rm -rf $DIR/${tdir}-*
1130 run_test 53a "|X| close request while two MDC requests in flight"
1133 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1135 mkdir -p $DIR/${tdir}-1
1136 mkdir -p $DIR/${tdir}-2
1137 multiop $DIR/${tdir}-1/f O_c &
1140 #define OBD_FAIL_MDS_REINT_NET 0x107
1141 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1142 mcreate $DIR/${tdir}-2/f &
1146 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1147 kill -USR1 $close_pid
1148 cancel_lru_locks mdc # force the close
1149 wait $close_pid || return 1
1150 # open should still be here
1151 [ -d /proc/$open_pid ] || return 2
1153 replay_barrier_nodf $SINGLEMDS
1155 wait $open_pid || return 3
1157 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1158 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1159 rm -rf $DIR/${tdir}-*
1161 run_test 53b "|X| open request while two MDC requests in flight"
1164 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1166 mkdir -p $DIR/${tdir}-1
1167 mkdir -p $DIR/${tdir}-2
1168 multiop $DIR/${tdir}-1/f O_c &
1171 #define OBD_FAIL_MDS_REINT_NET 0x107
1172 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1173 mcreate $DIR/${tdir}-2/f &
1177 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1178 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1179 kill -USR1 $close_pid
1180 cancel_lru_locks mdc # force the close
1182 replay_barrier_nodf $SINGLEMDS
1183 fail_nodf $SINGLEMDS
1184 wait $open_pid || return 1
1186 # close should be gone
1187 [ -d /proc/$close_pid ] && return 2
1188 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1190 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1191 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1192 rm -rf $DIR/${tdir}-*
1194 run_test 53c "|X| open request and close request while two MDC requests in flight"
1197 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1199 mkdir -p $DIR/${tdir}-1
1200 mkdir -p $DIR/${tdir}-2
1201 multiop $DIR/${tdir}-1/f O_c &
1203 # give multiop a chance to open
1206 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1207 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1208 kill -USR1 $close_pid
1209 cancel_lru_locks mdc # force the close
1210 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1211 mcreate $DIR/${tdir}-2/f || return 1
1213 # close should still be here
1214 [ -d /proc/$close_pid ] || return 2
1216 wait $close_pid || return 3
1218 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1219 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1220 rm -rf $DIR/${tdir}-*
1222 run_test 53d "|X| close reply while two MDC requests in flight"
1225 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1227 mkdir -p $DIR/${tdir}-1
1228 mkdir -p $DIR/${tdir}-2
1229 multiop $DIR/${tdir}-1/f O_c &
1232 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1233 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1234 mcreate $DIR/${tdir}-2/f &
1238 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1239 kill -USR1 $close_pid
1240 cancel_lru_locks mdc # force the close
1241 wait $close_pid || return 1
1242 # open should still be here
1243 [ -d /proc/$open_pid ] || return 2
1245 replay_barrier_nodf $SINGLEMDS
1247 wait $open_pid || return 3
1249 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1250 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1251 rm -rf $DIR/${tdir}-*
1253 run_test 53e "|X| open reply while two MDC requests in flight"
1256 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1258 mkdir -p $DIR/${tdir}-1
1259 mkdir -p $DIR/${tdir}-2
1260 multiop $DIR/${tdir}-1/f O_c &
1263 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1264 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1265 mcreate $DIR/${tdir}-2/f &
1269 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1270 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1271 kill -USR1 $close_pid
1272 cancel_lru_locks mdc # force the close
1274 replay_barrier_nodf $SINGLEMDS
1275 fail_nodf $SINGLEMDS
1276 wait $open_pid || return 1
1278 # close should be gone
1279 [ -d /proc/$close_pid ] && return 2
1280 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1282 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1283 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1284 rm -rf $DIR/${tdir}-*
1286 run_test 53f "|X| open reply and close reply while two MDC requests in flight"
1289 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1291 mkdir -p $DIR/${tdir}-1
1292 mkdir -p $DIR/${tdir}-2
1293 multiop $DIR/${tdir}-1/f O_c &
1296 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1297 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1298 mcreate $DIR/${tdir}-2/f &
1302 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1303 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1304 kill -USR1 $close_pid
1305 cancel_lru_locks mdc # force the close
1307 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1308 replay_barrier_nodf $SINGLEMDS
1309 fail_nodf $SINGLEMDS
1310 wait $open_pid || return 1
1312 # close should be gone
1313 [ -d /proc/$close_pid ] && return 2
1315 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1316 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1317 rm -rf $DIR/${tdir}-*
1319 run_test 53g "|X| drop open reply and close request while close and open are both in flight"
1322 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1324 mkdir -p $DIR/${tdir}-1
1325 mkdir -p $DIR/${tdir}-2
1326 multiop $DIR/${tdir}-1/f O_c &
1329 #define OBD_FAIL_MDS_REINT_NET 0x107
1330 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1331 mcreate $DIR/${tdir}-2/f &
1335 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1336 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1337 kill -USR1 $close_pid
1338 cancel_lru_locks mdc # force the close
1341 replay_barrier_nodf $SINGLEMDS
1342 fail_nodf $SINGLEMDS
1343 wait $open_pid || return 1
1345 # close should be gone
1346 [ -d /proc/$close_pid ] && return 2
1347 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1349 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1350 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1351 rm -rf $DIR/${tdir}-*
1353 run_test 53h "|X| open request and close reply while two MDC requests in flight"
1355 #b_cray 54 "|X| open request and close reply while two MDC requests in flight"
1357 #b3761 ASSERTION(hash != 0) failed
1359 # OBD_FAIL_MDS_OPEN_CREATE | OBD_FAIL_ONCE
1360 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012b"
1362 # give touch a chance to run
1364 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1368 run_test 55 "let MDS_CHECK_RESENT return the original return code instead of 0"
1370 #b3440 ASSERTION(rec->ur_fid2->id) failed
1372 ln -s foo $DIR/$tfile
1373 replay_barrier $SINGLEMDS
1374 #drop_reply "cat $DIR/$tfile"
1378 run_test 56 "don't replay a symlink open request (3440)"
1380 #recovery one mds-ost setattr from llog
1382 #define OBD_FAIL_MDS_OST_SETATTR 0x12c
1383 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012c"
1385 replay_barrier $SINGLEMDS
1388 $CHECKSTAT -t file $DIR/$tfile || return 1
1389 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1392 run_test 57 "test recovery from llog for setattr op"
1394 #recovery many mds-ost setattr from llog
1397 #define OBD_FAIL_MDS_OST_SETATTR 0x12c
1398 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012c"
1399 createmany -o $DIR/$tdir/$tfile-%d 2500
1400 replay_barrier $SINGLEMDS
1403 $CHECKSTAT -t file $DIR/$tdir/$tfile-* >/dev/null || return 1
1404 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1405 unlinkmany $DIR/$tdir/$tfile-%d 2500
1408 run_test 58a "test recovery from llog for setattr op (test llog_gen_rec)"
1411 mount_client $MOUNT2
1413 touch $DIR/$tdir/$tfile
1414 replay_barrier $SINGLEMDS
1415 setfattr -n trusted.foo -v bar $DIR/$tdir/$tfile
1417 VAL=`getfattr --absolute-names --only-value -n trusted.foo $MOUNT2/$tdir/$tfile`
1418 [ x$VAL = x"bar" ] || return 1
1419 rm -f $DIR/$tdir/$tfile
1421 zconf_umount `hostname` $MOUNT2
1423 run_test 58b "test replay of setxattr op"
1425 test_58c() { # bug 16570
1426 mount_client $MOUNT2
1428 touch $DIR/$tdir/$tfile
1429 drop_request "setfattr -n trusted.foo -v bar $DIR/$tdir/$tfile" || \
1431 VAL=`getfattr --absolute-names --only-value -n trusted.foo $MOUNT2/$tdir/$tfile`
1432 [ x$VAL = x"bar" ] || return 2
1433 drop_reint_reply "setfattr -n trusted.foo1 -v bar1 $DIR/$tdir/$tfile" || \
1435 VAL=`getfattr --absolute-names --only-value -n trusted.foo1 $MOUNT2/$tdir/$tfile`
1436 [ x$VAL = x"bar1" ] || return 4
1437 rm -f $DIR/$tdir/$tfile
1439 zconf_umount `hostname` $MOUNT2
1441 run_test 58c "resend/reconstruct setxattr op"
1443 # log_commit_thread vs filter_destroy race used to lead to import use after free
1446 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1449 createmany -o $DIR/$tdir/$tfile-%d 200
1451 unlinkmany $DIR/$tdir/$tfile-%d 200
1452 #define OBD_FAIL_PTLRPC_DELAY_RECOV 0x507
1453 do_facet ost1 "lctl set_param fail_loc=0x507"
1456 do_facet ost1 "lctl set_param fail_loc=0x0"
1460 run_test 59 "test log_commit_thread vs filter_destroy race"
1462 # race between add unlink llog vs cat log init in post_recovery (only for b1_6)
1463 # bug 12086: should no oops and No ctxt error for this test
1466 createmany -o $DIR/$tdir/$tfile-%d 200
1467 replay_barrier $SINGLEMDS
1468 unlinkmany $DIR/$tdir/$tfile-%d 0 100
1470 unlinkmany $DIR/$tdir/$tfile-%d 100 100
1471 local no_ctxt=`dmesg | grep "No ctxt"`
1472 [ -z "$no_ctxt" ] || error "ctxt is not initialized in recovery"
1474 run_test 60 "test llog post recovery init vs llog unlink"
1476 #test race llog recovery thread vs llog cleanup
1477 test_61a() { # was test_61
1478 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1481 createmany -o $DIR/$tdir/$tfile-%d 800
1483 # OBD_FAIL_OST_LLOG_RECOVERY_TIMEOUT 0x221
1484 unlinkmany $DIR/$tdir/$tfile-%d 800
1485 set_nodes_failloc "$(osts_nodes)" 0x80000221
1490 set_nodes_failloc "$(osts_nodes)" 0x0
1492 $CHECKSTAT -t file $DIR/$tdir/$tfile-* && return 1
1495 run_test 61a "test race llog recovery vs llog cleanup"
1497 #test race mds llog sync vs llog cleanup
1499 # OBD_FAIL_MDS_LLOG_SYNC_TIMEOUT 0x13a
1500 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013a"
1501 facet_failover $SINGLEMDS
1504 do_facet client dd if=/dev/zero of=$DIR/$tfile bs=4k count=1 || return 1
1506 run_test 61b "test race mds llog sync vs llog cleanup"
1508 #test race cancel cookie cb vs llog cleanup
1510 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1512 # OBD_FAIL_OST_CANCEL_COOKIE_TIMEOUT 0x222
1514 set_nodes_failloc "$(osts_nodes)" 0x80000222
1518 set_nodes_failloc "$(osts_nodes)" 0x0
1520 run_test 61c "test race mds llog sync vs llog cleanup"
1522 test_61d() { # bug 16002 # bug 17466
1523 shutdown_facet $SINGLEMDS
1524 #define OBD_FAIL_OBD_LLOG_SETUP 0x605
1525 do_facet $SINGLEMDS "lctl set_param fail_loc=0x605"
1526 start $SINGLEMDS `mdsdevname 1` $MDS_MOUNT_OPTS && error "mds start should have failed"
1527 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1528 start $SINGLEMDS `mdsdevname 1` $MDS_MOUNT_OPTS || error "cannot restart mds"
1530 run_test 61d "error in llog_setup should cleanup the llog context correctly"
1532 test_62() { # Bug 15756 - don't mis-drop resent replay
1534 replay_barrier $SINGLEMDS
1535 createmany -o $DIR/$tdir/$tfile- 25
1536 #define OBD_FAIL_TGT_REPLAY_DROP 0x707
1537 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000707"
1538 facet_failover $SINGLEMDS
1539 df $MOUNT || return 1
1540 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1541 unlinkmany $DIR/$tdir/$tfile- 25 || return 2
1544 run_test 62 "don't mis-drop resent replay"
1546 #Adaptive Timeouts (bug 3055)
1554 echo "Cleaning up AT ..."
1555 if [ -n "$ATOLDBASE" ]; then
1556 local at_history=$($LCTL get_param -n at_history)
1557 do_facet mds "lctl set_param at_history=$at_history" || true
1558 do_facet ost1 "lctl set_param at_history=$at_history" || true
1561 if [ $AT_MAX_SET -ne 0 ]; then
1562 for facet in mds client ost; do
1563 var=AT_MAX_SAVE_${facet}
1564 echo restore AT on $facet to saved value ${!var}
1565 at_max_set ${!var} $facet
1566 at_new=$(at_max_get $facet)
1567 echo Restored AT value on $facet $at_new
1568 [ $at_new -eq ${!var} ] || \
1569 error "$facet : AT value was not restored SAVED ${!var} NEW $at_new"
1576 local at_max_new=600
1578 # Save at_max original values
1580 if [ $AT_MAX_SET -eq 0 ]; then
1581 # Suppose that all osts have the same at_max
1582 for facet in mds client ost; do
1583 eval AT_MAX_SAVE_${facet}=$(at_max_get $facet)
1587 for facet in mds client ost; do
1588 at_max=$(at_max_get $facet)
1589 if [ $at_max -ne $at_max_new ]; then
1590 echo "AT value on $facet is $at_max, set it by force temporarily to $at_max_new"
1591 at_max_set $at_max_new $facet
1596 if [ -z "$ATOLDBASE" ]; then
1597 ATOLDBASE=$(do_facet mds "lctl get_param -n at_history")
1598 # speed up the timebase so we can check decreasing AT
1599 do_facet mds "lctl set_param at_history=8" || true
1600 do_facet ost1 "lctl set_param at_history=8" || true
1602 # sleep for a while to cool down, should be > 8s and also allow
1603 # at least one ping to be sent. simply use TIMEOUT to be safe.
1608 test_65a() #bug 3055
1610 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1612 at_start || return 0
1613 $LCTL dk > /dev/null
1615 sysctl -w lnet.debug="+other"
1616 # Slow down a request to the current service time, this is critical
1617 # because previous tests may have caused this value to increase.
1618 REQ_DELAY=`lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts |
1619 awk '/portal 12/ {print $5}'`
1620 REQ_DELAY=$((${REQ_DELAY} + ${REQ_DELAY} / 4 + 5))
1622 do_facet mds lctl set_param fail_val=$((${REQ_DELAY} * 1000))
1623 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1624 do_facet mds sysctl -w lustre.fail_loc=0x8000050a
1625 createmany -o $DIR/$tfile 10 > /dev/null
1626 unlinkmany $DIR/$tfile 10 > /dev/null
1627 # check for log message
1628 $LCTL dk | grep "Early reply #" || error "No early reply"
1630 # client should show REQ_DELAY estimates
1631 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep portal
1633 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep portal
1635 run_test 65a "AT: verify early replies"
1637 test_65b() #bug 3055
1639 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1641 at_start || return 0
1644 sysctl -w lnet.debug="other trace"
1645 $LCTL dk > /dev/null
1646 # Slow down a request to the current service time, this is critical
1647 # because previous tests may have caused this value to increase.
1648 REQ_DELAY=`lctl get_param -n osc.${FSNAME}-OST0000-osc-*.timeouts |
1649 awk '/portal 6/ {print $5}'`
1650 REQ_DELAY=$((${REQ_DELAY} + ${REQ_DELAY} / 4 + 5))
1652 do_facet ost1 lctl set_param fail_val=${REQ_DELAY}
1653 #define OBD_FAIL_OST_BRW_PAUSE_PACK 0x224
1654 do_facet ost1 sysctl -w lustre.fail_loc=0x224
1657 lfs setstripe $DIR/$tfile --index=0 --count=1
1658 # force some real bulk transfer
1659 multiop $DIR/$tfile oO_CREAT:O_RDWR:O_SYNC:w4096c
1661 do_facet ost1 sysctl -w lustre.fail_loc=0
1662 # check for log message
1663 $LCTL dk | grep "Early reply #" || error "No early reply"
1665 # client should show REQ_DELAY estimates
1666 lctl get_param -n osc.${FSNAME}-OST0000-osc-*.timeouts | grep portal
1668 run_test 65b "AT: verify early replies on packed reply / bulk"
1670 test_66a() #bug 3055
1672 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1674 at_start || return 0
1675 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1676 # adjust 5s at a time so no early reply is sent (within deadline)
1677 do_facet mds "sysctl -w lustre.fail_val=5000"
1678 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1679 do_facet mds "sysctl -w lustre.fail_loc=0x8000050a"
1680 createmany -o $DIR/$tfile 20 > /dev/null
1681 unlinkmany $DIR/$tfile 20 > /dev/null
1682 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1683 do_facet mds "sysctl -w lustre.fail_val=10000"
1684 do_facet mds "sysctl -w lustre.fail_loc=0x8000050a"
1685 createmany -o $DIR/$tfile 20 > /dev/null
1686 unlinkmany $DIR/$tfile 20 > /dev/null
1687 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1688 do_facet mds "sysctl -w lustre.fail_loc=0"
1690 createmany -o $DIR/$tfile 20 > /dev/null
1691 unlinkmany $DIR/$tfile 20 > /dev/null
1692 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1693 CUR=$(lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | awk '/portal 12/ {print $5}')
1694 WORST=$(lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | awk '/portal 12/ {print $7}')
1695 echo "Current MDT timeout $CUR, worst $WORST"
1696 [ $CUR -lt $WORST ] || error "Current $CUR should be less than worst $WORST"
1698 run_test 66a "AT: verify MDT service time adjusts with no early replies"
1700 test_66b() #bug 3055
1702 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1704 at_start || return 0
1705 ORIG=$(lctl get_param -n mdc.${FSNAME}-*.timeouts | awk '/network/ {print $4}')
1706 sysctl -w lustre.fail_val=$(($ORIG + 5))
1707 #define OBD_FAIL_PTLRPC_PAUSE_REP 0x50c
1708 sysctl -w lustre.fail_loc=0x50c
1709 ls $DIR/$tfile > /dev/null 2>&1
1710 sysctl -w lustre.fail_loc=0
1711 CUR=$(lctl get_param -n mdc.${FSNAME}-*.timeouts | awk '/network/ {print $4}')
1712 WORST=$(lctl get_param -n mdc.${FSNAME}-*.timeouts | awk '/network/ {print $6}')
1713 echo "network timeout orig $ORIG, cur $CUR, worst $WORST"
1714 [ $WORST -gt $ORIG ] || error "Worst $WORST should be worse than orig $ORIG"
1716 run_test 66b "AT: verify net latency adjusts"
1718 test_67a() #bug 3055
1720 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1722 at_start || return 0
1723 CONN1=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1724 # sleeping threads may drive values above this
1725 do_facet ost1 "sysctl -w lustre.fail_val=400"
1726 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1727 do_facet ost1 "sysctl -w lustre.fail_loc=0x50a"
1728 createmany -o $DIR/$tfile 20 > /dev/null
1729 unlinkmany $DIR/$tfile 20 > /dev/null
1730 do_facet ost1 "sysctl -w lustre.fail_loc=0"
1731 CONN2=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1732 ATTEMPTS=$(($CONN2 - $CONN1))
1733 echo "$ATTEMPTS osc reconnect attempts on gradual slow"
1734 [ $ATTEMPTS -gt 0 ] && error_ignore 13721 "AT should have prevented reconnect"
1737 run_test 67a "AT: verify slow request processing doesn't induce reconnects"
1739 test_67b() #bug 3055
1741 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1743 at_start || return 0
1744 CONN1=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1745 #define OBD_FAIL_OST_PAUSE_CREATE 0x223
1746 do_facet ost1 "sysctl -w lustre.fail_val=20000"
1747 do_facet ost1 "sysctl -w lustre.fail_loc=0x80000223"
1748 cp /etc/profile $DIR/$tfile || error "cp failed"
1750 do_facet ost1 "lctl get_param -n ost.OSS.ost_create.timeouts"
1752 CONN2=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1753 ATTEMPTS=$(($CONN2 - $CONN1))
1754 echo "$ATTEMPTS osc reconnect attempts on instant slow"
1755 # do it again; should not timeout
1756 do_facet ost1 "sysctl -w lustre.fail_loc=0x80000223"
1757 cp /etc/profile $DIR/$tfile || error "cp failed"
1758 do_facet ost1 "sysctl -w lustre.fail_loc=0"
1760 do_facet ost1 "lctl get_param -n ost.OSS.ost_create.timeouts"
1761 CONN3=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1762 ATTEMPTS=$(($CONN3 - $CONN2))
1763 echo "$ATTEMPTS osc reconnect attempts on 2nd slow"
1764 [ $ATTEMPTS -gt 0 ] && error "AT should have prevented reconnect"
1767 run_test 67b "AT: verify instant slowdown doesn't induce reconnects"
1769 test_68 () #bug 13813
1771 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1773 at_start || return 0
1774 local ldlm_enqueue_min=$(find /sys -name ldlm_enqueue_min)
1775 [ -z "$ldlm_enqueue_min" ] && skip "missing /sys/.../ldlm_enqueue_min" && return 0
1776 local ldlm_enqueue_min_r=$(do_facet ost1 "find /sys -name ldlm_enqueue_min")
1777 [ -z "$ldlm_enqueue_min_r" ] && skip "missing /sys/.../ldlm_enqueue_min in the ost1" && return 0
1778 local ENQ_MIN=$(cat $ldlm_enqueue_min)
1779 local ENQ_MIN_R=$(do_facet ost1 "cat $ldlm_enqueue_min_r")
1780 echo $TIMEOUT >> $ldlm_enqueue_min
1781 do_facet ost1 "echo $TIMEOUT >> $ldlm_enqueue_min_r"
1785 lfs setstripe $DIR/$tdir --index=0 --count=1
1786 #define OBD_FAIL_LDLM_PAUSE_CANCEL 0x312
1787 sysctl -w lustre.fail_val=$(($TIMEOUT - 1))
1788 sysctl -w lustre.fail_loc=0x80000312
1789 cp /etc/profile $DIR/$tdir/${tfile}_1 || error "1st cp failed $?"
1790 sysctl -w lustre.fail_val=$((TIMEOUT * 5 / 4))
1791 sysctl -w lustre.fail_loc=0x80000312
1792 cp /etc/profile $DIR/$tdir/${tfile}_2 || error "2nd cp failed $?"
1793 sysctl -w lustre.fail_loc=0
1795 echo $ENQ_MIN >> $ldlm_enqueue_min
1796 do_facet ost1 "echo $ENQ_MIN_R >> $ldlm_enqueue_min_r"
1800 run_test 68 "AT: verify slowing locks"
1803 # end of AT tests includes above lines
1806 # start multi-client tests
1808 [ -z "$CLIENTS" ] && \
1809 { skip "Need two or more clients." && return; }
1810 [ $CLIENTCOUNT -lt 2 ] && \
1811 { skip "Need two or more clients, have $CLIENTCOUNT" && return; }
1813 echo "mount clients $CLIENTS ..."
1814 zconf_mount_clients $CLIENTS $DIR
1816 local clients=${CLIENTS//,/ }
1817 echo "Write/read files on $DIR ; clients $CLIENTS ... "
1818 for CLIENT in $clients; do
1819 do_node $CLIENT dd bs=1M count=10 if=/dev/zero \
1820 of=$DIR/${tfile}_${CLIENT} 2>/dev/null || \
1821 error "dd failed on $CLIENT"
1824 local prev_client=$(echo $clients | sed 's/^.* \(.\+\)$/\1/')
1825 for C in ${CLIENTS//,/ }; do
1826 do_node $prev_client dd if=$DIR/${tfile}_${C} of=/dev/null 2>/dev/null || \
1827 error "dd if=$DIR/${tfile}_${C} failed on $prev_client"
1833 run_test 70a "check multi client t-f"
1836 local clients=${CLIENTS:-$HOSTNAME}
1838 zconf_mount_clients $clients $DIR
1841 [ "$SLOW" = "no" ] && duration=60
1842 local cmd="rundbench 1 -t $duration"
1844 do_nodes --verbose $clients "set -x; MISSING_DBENCH_OK=$MISSING_DBENCH_OK \
1845 PATH=:$PATH:$LUSTRE/utils:$LUSTRE/tests/:$DBENCH_LIB \
1846 DBENCH_LIB=$DBENCH_LIB TESTSUITE=$TESTSUITE TESTNAME=$TESTNAME \
1849 log "Started rundbench load PID=$PID ..."
1852 START_TS=$(date +%s)
1853 CURRENT_TS=$START_TS
1854 while [ $ELAPSED -lt $duration ]; do
1856 replay_barrier $SINGLEMDS
1857 sleep 1 # give clients a time to do operations
1858 # Increment the number of failovers
1859 NUM_FAILOVERS=$((NUM_FAILOVERS+1))
1860 log "$TESTNAME fail mds1 $NUM_FAILOVERS times"
1862 CURRENT_TS=$(date +%s)
1863 ELAPSED=$((CURRENT_TS - START_TS))
1865 wait $PID || error "rundbench load on $CLIENTS failed!"
1867 run_test 70b "mds recovery; $CLIENTCOUNT clients"
1868 # end multi-client tests
1871 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1875 replay_barrier $SINGLEMDS
1876 #define OBD_FAIL_LDLM_ENQUEUE 0x302
1877 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000302"
1880 wait $pid || return 1
1881 [ -e $DIR/$tfile ] && return 2
1884 run_test 73a "open(O_CREAT), unlink, replay, reconnect before open replay , close"
1887 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1891 replay_barrier $SINGLEMDS
1892 #define OBD_FAIL_LDLM_REPLY 0x30c
1893 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000030c"
1896 wait $pid || return 1
1897 [ -e $DIR/$tfile ] && return 2
1900 run_test 73b "open(O_CREAT), unlink, replay, reconnect at open_replay reply, close"
1903 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1907 replay_barrier $SINGLEMDS
1908 #define OBD_FAIL_TGT_LAST_REPLAY 0x710
1909 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000710"
1912 wait $pid || return 1
1913 [ -e $DIR/$tfile ] && return 2
1916 run_test 73c "open(O_CREAT), unlink, replay, reconnect at last_replay, close"
1920 local clients=${CLIENTS:-$HOSTNAME}
1923 zconf_umount_clients $clients $MOUNT
1924 facet_failover $SINGLEMDS
1925 zconf_mount_clients $clients $MOUNT
1927 touch $DIR/$tfile || return 1
1928 rm $DIR/$tfile || return 2
1929 client_df || error "df failed: $?"
1932 run_test 74 "Ensure applications don't fail waiting for OST recovery"
1935 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
1939 $CHECKSTAT -t dir $DIR/$tdir || error "$CHECKSTAT -t dir $DIR/$tdir failed"
1940 rmdir $DIR/$tdir || error "rmdir $DIR/$tdir failed"
1942 stat $DIR/$tdir 2&>/dev/null && error "$DIR/$tdir still exist after recovery!"
1945 run_test 80a "CMD: unlink cross-node dir (fail mds with inode)"
1948 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
1952 $CHECKSTAT -t dir $DIR/$tdir || error "$CHECKSTAT -t dir $DIR/$tdir failed"
1953 rmdir $DIR/$tdir || error "rmdir $DIR/$tdir failed"
1955 stat $DIR/$tdir 2&>/dev/null && error "$DIR/$tdir still exist after recovery!"
1958 run_test 80b "CMD: unlink cross-node dir (fail mds with name)"
1961 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
1964 createmany -o $DIR/$tdir/f 3000 || error "createmany failed"
1966 $CHECKSTAT -t dir $DIR/$tdir || error "$CHECKSTAT -t dir failed"
1967 $CHECKSTAT -t file $DIR/$tdir/f1002 || error "$CHECKSTAT -t file failed"
1969 rm $DIR/$tdir/f1002 || error "rm $DIR/$tdir/f1002 failed"
1971 stat $DIR/$tdir/f1002
1973 run_test 81a "CMD: unlink cross-node file (fail mds with name)"
1976 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
1980 mkdir $dir || error "mkdir $dir failed"
1984 $CHECKSTAT -t dir $dir || error "$CHECKSTAT -t dir $dir failed"
1986 run_test 82a "CMD: mkdir cross-node dir (fail mds with inode)"
1989 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
1993 mkdir $dir || error "mkdir $dir failed"
1997 $CHECKSTAT -t dir $dir || error "$CHECKSTAT -t dir $dir failed"
1999 run_test 82b "CMD: mkdir cross-node dir (fail mds with name)"
2003 createmany -o $DIR/$tdir/$tfile- 10 || return 1
2004 #define OBD_FAIL_MDS_FAIL_LOV_LOG_ADD 0x140
2005 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000140"
2006 unlinkmany $DIR/$tdir/$tfile- 10 || return 2
2008 run_test 83a "fail log_add during unlink recovery"
2012 createmany -o $DIR/$tdir/$tfile- 10 || return 1
2013 replay_barrier $SINGLEMDS
2014 unlinkmany $DIR/$tdir/$tfile- 10 || return 2
2015 #define OBD_FAIL_MDS_FAIL_LOV_LOG_ADD 0x140
2016 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000140"
2019 run_test 83b "fail log_add during unlink recovery"
2022 #define OBD_FAIL_MDS_OPEN_WAIT_CREATE 0x143
2023 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000143"
2024 createmany -o $DIR/$tfile- 1 &
2028 df -P $DIR || df -P $DIR || true # reconnect
2030 run_test 84a "stale open during export disconnect"
2032 equals_msg `basename $0`: test complete, cleaning up
2033 check_and_cleanup_lustre
2034 [ -f "$TESTSUITELOG" ] && cat $TESTSUITELOG && grep -q FAIL $TESTSUITELOG && exit 1 || true