7 # This test needs to be run on the client
10 export MULTIOP=${MULTIOP:-multiop}
11 LUSTRE=${LUSTRE:-$(cd $(dirname $0)/..; echo $PWD)}
14 . $LUSTRE/tests/test-framework.sh
16 . ${CONFIG:=$LUSTRE/tests/cfg/$NAME.sh}
18 CHECK_GRANT=${CHECK_GRANT:-"yes"}
19 GRANT_CHECK_LIST=${GRANT_CHECK_LIST:-""}
21 require_dsh_mds || exit 0
24 # bug number: 17466 18857 LU1867
25 ALWAYS_EXCEPT="61d 33a 33b 89 $REPLAY_SINGLE_EXCEPT"
27 [ $(facet_fstype $SINGLEMDS) = "zfs" ] &&
28 # bug number for skipped test: LU-2342 LU-951
29 ALWAYS_EXCEPT="$ALWAYS_EXCEPT 20b 70a 73a"
31 # 63 min 7 min AT AT AT AT"
32 [ "$SLOW" = "no" ] && EXCEPT_SLOW="1 2 3 4 6 12 16 44a 44b 65 66 67 68"
36 check_and_setup_lustre
41 rm -rf $DIR/[df][0-9]* $DIR/f.$TESTSUITE.*
43 # LU-482 Avert LVM and VM inability to flush caches in pre .33 kernels
44 if [ $LINUX_VERSION_CODE -lt $(version_code 2.6.33) ]; then
46 do_facet $SINGLEMDS sync
49 test_0a() { # was test_0
51 replay_barrier $SINGLEMDS
55 run_test 0a "empty replay"
58 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
60 # this test attempts to trigger a race in the precreation code,
61 # and must run before any other objects are created on the filesystem
63 createmany -o $DIR/$tfile 20 || return 1
64 unlinkmany $DIR/$tfile 20 || return 2
66 run_test 0b "ensure object created after recover exists. (3284)"
69 replay_barrier $SINGLEMDS
72 facet_failover $SINGLEMDS
73 zconf_mount `hostname` $MOUNT || error "mount fails"
74 client_up || error "post-failover df failed"
75 # file shouldn't exist if replay-barrier works as expected
76 rm $DIR/$tfile && error "File exists and it shouldn't"
79 run_test 0c "check replay-barrier"
82 replay_barrier $SINGLEMDS
84 facet_failover $SINGLEMDS
85 zconf_mount `hostname` $MOUNT || error "mount fails"
86 client_up || error "post-failover df failed"
88 run_test 0d "expired recovery with no clients"
91 replay_barrier $SINGLEMDS
94 $CHECKSTAT -t file $DIR/$tfile || return 1
97 run_test 1 "simple create"
100 replay_barrier $SINGLEMDS
103 $CHECKSTAT -t file $DIR/$tfile || return 1
110 replay_barrier $SINGLEMDS
113 $CHECKSTAT -t file $DIR/$tfile || return 1
119 local file=$DIR/$tfile
120 replay_barrier $SINGLEMDS
122 openfile -f O_DIRECTORY $file
124 $CHECKSTAT -t file $file || return 2
127 run_test 3a "replay failed open(O_DIRECTORY)"
130 replay_barrier $SINGLEMDS
131 #define OBD_FAIL_MDS_OPEN_PACK | OBD_FAIL_ONCE
132 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000114"
134 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
136 $CHECKSTAT -t file $DIR/$tfile && return 2
139 run_test 3b "replay failed open -ENOMEM"
142 replay_barrier $SINGLEMDS
143 #define OBD_FAIL_MDS_ALLOC_OBDO | OBD_FAIL_ONCE
144 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000128"
146 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
149 $CHECKSTAT -t file $DIR/$tfile && return 2
152 run_test 3c "replay failed open -ENOMEM"
154 test_4a() { # was test_4
155 replay_barrier $SINGLEMDS
156 for i in `seq 10`; do
157 echo "tag-$i" > $DIR/$tfile-$i
160 for i in `seq 10`; do
161 grep -q "tag-$i" $DIR/$tfile-$i || error "$tfile-$i"
164 run_test 4a "|x| 10 open(O_CREAT)s"
167 replay_barrier $SINGLEMDS
170 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
172 run_test 4b "|x| rm 10 files"
174 # The idea is to get past the first block of precreated files on both
175 # osts, and then replay.
177 replay_barrier $SINGLEMDS
178 for i in `seq 220`; do
179 echo "tag-$i" > $DIR/$tfile-$i
182 for i in `seq 220`; do
183 grep -q "tag-$i" $DIR/$tfile-$i || error "$tfile-$i"
187 # waiting for commitment of removal
189 run_test 5 "|x| 220 open(O_CREAT)"
192 test_6a() { # was test_6
194 replay_barrier $SINGLEMDS
195 mcreate $DIR/$tdir/$tfile
197 $CHECKSTAT -t dir $DIR/$tdir || return 1
198 $CHECKSTAT -t file $DIR/$tdir/$tfile || return 2
200 # waiting for log process thread
202 run_test 6a "mkdir + contained create"
206 replay_barrier $SINGLEMDS
209 $CHECKSTAT -t dir $DIR/$tdir && return 1 || true
211 run_test 6b "|X| rmdir"
215 replay_barrier $SINGLEMDS
216 mcreate $DIR/$tdir/$tfile
218 $CHECKSTAT -t dir $DIR/$tdir || return 1
219 $CHECKSTAT -t file $DIR/$tdir/$tfile || return 2
222 run_test 7 "mkdir |X| contained create"
225 # make sure no side-effect from previous test.
227 replay_barrier $SINGLEMDS
228 multiop_bg_pause $DIR/$tfile mo_c || return 4
232 $CHECKSTAT -t file $DIR/$tfile || return 1
233 kill -USR1 $MULTIPID || return 2
234 wait $MULTIPID || return 3
237 run_test 8 "creat open |X| close"
240 replay_barrier $SINGLEMDS
242 local old_inum=`ls -i $DIR/$tfile | awk '{print $1}'`
244 local new_inum=`ls -i $DIR/$tfile | awk '{print $1}'`
246 echo " old_inum == $old_inum, new_inum == $new_inum"
247 if [ $old_inum -eq $new_inum ] ;
249 echo " old_inum and new_inum match"
251 echo "!!!! old_inum and new_inum NOT match"
256 run_test 9 "|X| create (same inum/gen)"
260 replay_barrier $SINGLEMDS
261 mv $DIR/$tfile $DIR/$tfile-2
264 $CHECKSTAT $DIR/$tfile && return 1
265 $CHECKSTAT $DIR/$tfile-2 ||return 2
269 run_test 10 "create |X| rename unlink"
273 echo "old" > $DIR/$tfile
274 mv $DIR/$tfile $DIR/$tfile-2
275 replay_barrier $SINGLEMDS
276 echo "new" > $DIR/$tfile
278 grep old $DIR/$tfile-2
280 grep new $DIR/$tfile || return 1
281 grep old $DIR/$tfile-2 || return 2
283 run_test 11 "create open write rename |X| create-old-name read"
287 multiop_bg_pause $DIR/$tfile o_tSc || return 3
290 replay_barrier $SINGLEMDS
292 wait $pid || return 1
295 [ -e $DIR/$tfile ] && return 2
298 run_test 12 "open, unlink |X| close"
301 # 1777 - replay open after committed chmod that would make
302 # a regular open a failure
305 multiop_bg_pause $DIR/$tfile O_wc || return 3
308 $CHECKSTAT -p 0 $DIR/$tfile
309 replay_barrier $SINGLEMDS
312 wait $pid || return 1
314 $CHECKSTAT -s 1 -p 0 $DIR/$tfile || return 2
315 rm $DIR/$tfile || return 4
318 run_test 13 "open chmod 0 |x| write close"
321 multiop_bg_pause $DIR/$tfile O_tSc || return 4
324 replay_barrier $SINGLEMDS
325 kill -USR1 $pid || return 1
326 wait $pid || return 2
329 [ -e $DIR/$tfile ] && return 3
332 run_test 14 "open(O_CREAT), unlink |X| close"
335 multiop_bg_pause $DIR/$tfile O_tSc || return 5
338 replay_barrier $SINGLEMDS
339 touch $DIR/g11 || return 1
341 wait $pid || return 2
344 [ -e $DIR/$tfile ] && return 3
345 touch $DIR/h11 || return 4
348 run_test 15 "open(O_CREAT), unlink |X| touch new, close"
352 replay_barrier $SINGLEMDS
355 mcreate $DIR/$tfile-2
357 [ -e $DIR/$tfile ] && return 1
358 [ -e $DIR/$tfile-2 ] || return 2
359 munlink $DIR/$tfile-2 || return 3
361 run_test 16 "|X| open(O_CREAT), unlink, touch new, unlink new"
364 replay_barrier $SINGLEMDS
365 multiop_bg_pause $DIR/$tfile O_c || return 4
368 kill -USR1 $pid || return 1
369 wait $pid || return 2
370 $CHECKSTAT -t file $DIR/$tfile || return 3
373 run_test 17 "|X| open(O_CREAT), |replay| close"
376 replay_barrier $SINGLEMDS
377 multiop_bg_pause $DIR/$tfile O_tSc || return 8
380 touch $DIR/$tfile-2 || return 1
381 echo "pid: $pid will close"
383 wait $pid || return 2
386 [ -e $DIR/$tfile ] && return 3
387 [ -e $DIR/$tfile-2 ] || return 4
388 # this touch frequently fails
389 touch $DIR/$tfile-3 || return 5
390 munlink $DIR/$tfile-2 || return 6
391 munlink $DIR/$tfile-3 || return 7
394 run_test 18 "|X| open(O_CREAT), unlink, touch new, close, touch, unlink"
396 # bug 1855 (a simpler form of test_11 above)
398 replay_barrier $SINGLEMDS
400 echo "old" > $DIR/$tfile
401 mv $DIR/$tfile $DIR/$tfile-2
402 grep old $DIR/$tfile-2
404 grep old $DIR/$tfile-2 || return 2
406 run_test 19 "|X| mcreate, open, write, rename "
408 test_20a() { # was test_20
409 replay_barrier $SINGLEMDS
410 multiop_bg_pause $DIR/$tfile O_tSc || return 3
416 wait $pid || return 1
417 [ -e $DIR/$tfile ] && return 2
420 run_test 20a "|X| open(O_CREAT), unlink, replay, close (test mds_cleanup_orphans)"
422 test_20b() { # bug 10480
423 BEFOREUSED=`df -P $DIR | tail -1 | awk '{ print $3 }'`
425 dd if=/dev/zero of=$DIR/$tfile bs=4k count=10000 &
427 while [ ! -e $DIR/$tfile ] ; do
428 usleep 60 # give dd a chance to start
431 $GETSTRIPE $DIR/$tfile || return 1
432 rm -f $DIR/$tfile || return 2 # make it an orphan
434 client_up || client_up || true # reconnect
436 fail $SINGLEMDS # start orphan recovery
437 wait_recovery_complete $SINGLEMDS || error "MDS recovery not done"
438 wait_mds_ost_sync || return 3
439 AFTERUSED=`df -P $DIR | tail -1 | awk '{ print $3 }'`
440 log "before $BEFOREUSED, after $AFTERUSED"
441 [ $AFTERUSED -gt $((BEFOREUSED + 20)) ] && \
442 error "after $AFTERUSED > before $BEFOREUSED"
445 run_test 20b "write, unlink, eviction, replay, (test mds_cleanup_orphans)"
447 test_20c() { # bug 10480
448 multiop_bg_pause $DIR/$tfile Ow_c || return 1
454 client_up || client_up || true # reconnect
457 wait $pid || return 1
458 [ -s $DIR/$tfile ] || error "File was truncated"
462 run_test 20c "check that client eviction does not affect file content"
465 replay_barrier $SINGLEMDS
466 multiop_bg_pause $DIR/$tfile O_tSc || return 5
469 touch $DIR/g11 || return 1
473 wait $pid || return 2
474 [ -e $DIR/$tfile ] && return 3
475 touch $DIR/h11 || return 4
478 run_test 21 "|X| open(O_CREAT), unlink touch new, replay, close (test mds_cleanup_orphans)"
481 multiop_bg_pause $DIR/$tfile O_tSc || return 3
484 replay_barrier $SINGLEMDS
489 wait $pid || return 1
490 [ -e $DIR/$tfile ] && return 2
493 run_test 22 "open(O_CREAT), |X| unlink, replay, close (test mds_cleanup_orphans)"
496 multiop_bg_pause $DIR/$tfile O_tSc || return 5
499 replay_barrier $SINGLEMDS
501 touch $DIR/g11 || return 1
505 wait $pid || return 2
506 [ -e $DIR/$tfile ] && return 3
507 touch $DIR/h11 || return 4
510 run_test 23 "open(O_CREAT), |X| unlink touch new, replay, close (test mds_cleanup_orphans)"
513 multiop_bg_pause $DIR/$tfile O_tSc || return 3
516 replay_barrier $SINGLEMDS
520 wait $pid || return 1
521 [ -e $DIR/$tfile ] && return 2
524 run_test 24 "open(O_CREAT), replay, unlink, close (test mds_cleanup_orphans)"
527 multiop_bg_pause $DIR/$tfile O_tSc || return 3
531 replay_barrier $SINGLEMDS
534 wait $pid || return 1
535 [ -e $DIR/$tfile ] && return 2
538 run_test 25 "open(O_CREAT), unlink, replay, close (test mds_cleanup_orphans)"
541 replay_barrier $SINGLEMDS
542 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
544 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
549 wait $pid2 || return 1
553 wait $pid1 || return 2
554 [ -e $DIR/$tfile-1 ] && return 3
555 [ -e $DIR/$tfile-2 ] && return 4
558 run_test 26 "|X| open(O_CREAT), unlink two, close one, replay, close one (test mds_cleanup_orphans)"
561 replay_barrier $SINGLEMDS
562 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
564 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
571 wait $pid1 || return 1
573 wait $pid2 || return 2
574 [ -e $DIR/$tfile-1 ] && return 3
575 [ -e $DIR/$tfile-2 ] && return 4
578 run_test 27 "|X| open(O_CREAT), unlink two, replay, close two (test mds_cleanup_orphans)"
581 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
583 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
585 replay_barrier $SINGLEMDS
589 wait $pid2 || return 1
593 wait $pid1 || return 2
594 [ -e $DIR/$tfile-1 ] && return 3
595 [ -e $DIR/$tfile-2 ] && return 4
598 run_test 28 "open(O_CREAT), |X| unlink two, close one, replay, close one (test mds_cleanup_orphans)"
601 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
603 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
605 replay_barrier $SINGLEMDS
611 wait $pid1 || return 1
613 wait $pid2 || return 2
614 [ -e $DIR/$tfile-1 ] && return 3
615 [ -e $DIR/$tfile-2 ] && return 4
618 run_test 29 "open(O_CREAT), |X| unlink two, replay, close two (test mds_cleanup_orphans)"
621 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
623 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
628 replay_barrier $SINGLEMDS
631 wait $pid1 || return 1
633 wait $pid2 || return 2
634 [ -e $DIR/$tfile-1 ] && return 3
635 [ -e $DIR/$tfile-2 ] && return 4
638 run_test 30 "open(O_CREAT) two, unlink two, replay, close two (test mds_cleanup_orphans)"
641 multiop_bg_pause $DIR/$tfile-1 O_tSc || return 5
643 multiop_bg_pause $DIR/$tfile-2 O_tSc || return 6
647 replay_barrier $SINGLEMDS
651 wait $pid1 || return 1
653 wait $pid2 || return 2
654 [ -e $DIR/$tfile-1 ] && return 3
655 [ -e $DIR/$tfile-2 ] && return 4
658 run_test 31 "open(O_CREAT) two, unlink one, |X| unlink one, close two (test mds_cleanup_orphans)"
660 # tests for bug 2104; completion without crashing is success. The close is
661 # stale, but we always return 0 for close, so the app never sees it.
663 multiop_bg_pause $DIR/$tfile O_c || return 2
665 multiop_bg_pause $DIR/$tfile O_c || return 3
668 client_up || client_up || return 1
671 wait $pid1 || return 4
672 wait $pid2 || return 5
675 run_test 32 "close() notices client eviction; close() after client eviction"
678 createmany -o $DIR/$tfile-%d 10
679 replay_barrier_nosync $SINGLEMDS
680 fail_abort $SINGLEMDS
681 # recreate shouldn't fail
682 createmany -o $DIR/$tfile--%d 10 || return 1
686 run_test 33a "fid seq shouldn't be reused after abort recovery"
689 #define OBD_FAIL_SEQ_ALLOC 0x1311
690 do_facet $SINGLEMDS "lctl set_param fail_loc=0x1311"
692 createmany -o $DIR/$tfile-%d 10
693 replay_barrier_nosync $SINGLEMDS
694 fail_abort $SINGLEMDS
695 # recreate shouldn't fail
696 createmany -o $DIR/$tfile--%d 10 || return 1
700 run_test 33b "test fid seq allocation"
703 multiop_bg_pause $DIR/$tfile O_c || return 2
707 replay_barrier $SINGLEMDS
708 fail_abort $SINGLEMDS
710 wait $pid || return 3
711 [ -e $DIR/$tfile ] && return 1
715 run_test 34 "abort recovery before client does replay (test mds_cleanup_orphans)"
717 # bug 2278 - generate one orphan on OST, then destroy it during recovery from llog
721 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
722 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000119"
727 # give a chance to remove from MDS
728 fail_abort $SINGLEMDS
729 $CHECKSTAT -t file $DIR/$tfile && return 1 || true
731 run_test 35 "test recovery from llog for unlink op"
733 # b=2432 resent cancel after replay uses wrong cookie,
734 # so don't resend cancels
736 replay_barrier $SINGLEMDS
738 checkstat $DIR/$tfile
739 facet_failover $SINGLEMDS
741 if dmesg | grep "unknown lock cookie"; then
742 echo "cancel after replay failed"
746 run_test 36 "don't resend cancel"
749 # directory orphans can't be unlinked from PENDING directory
751 rmdir $DIR/$tfile 2>/dev/null
752 multiop_bg_pause $DIR/$tfile dD_c || return 2
756 replay_barrier $SINGLEMDS
757 # clear the dmesg buffer so we only see errors from this recovery
758 do_facet $SINGLEMDS dmesg -c >/dev/null
759 fail_abort $SINGLEMDS
761 do_facet $SINGLEMDS dmesg | grep "error .* unlinking .* from PENDING" &&
763 wait $pid || return 3
767 start_full_debug_logging
768 run_test 37 "abort recovery before client does replay (test mds_cleanup_orphans for directories)"
769 stop_full_debug_logging
772 createmany -o $DIR/$tfile-%d 800
773 unlinkmany $DIR/$tfile-%d 0 400
774 replay_barrier $SINGLEMDS
776 unlinkmany $DIR/$tfile-%d 400 400
778 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
780 run_test 38 "test recovery from unlink llog (test llog_gen_rec) "
782 test_39() { # bug 4176
783 createmany -o $DIR/$tfile-%d 800
784 replay_barrier $SINGLEMDS
785 unlinkmany $DIR/$tfile-%d 0 400
787 unlinkmany $DIR/$tfile-%d 400 400
789 $CHECKSTAT -t file $DIR/$tfile-* && return 1 || true
791 run_test 39 "test recovery from unlink llog (test llog_gen_rec) "
794 lctl get_param -n osc.*.stats | awk -vwrites=0 '/ost_write/ { writes += $2 } END { print writes; }'
799 $LCTL mark multiop $MOUNT/$tfile OS_c
800 multiop $MOUNT/$tfile OS_c &
802 writeme -s $MOUNT/${tfile}-2 &
805 facet_failover $SINGLEMDS
806 #define OBD_FAIL_MDS_CONNECT_NET 0x117
807 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000117"
809 stat1=`count_ost_writes`
811 stat2=`count_ost_writes`
812 echo "$stat1, $stat2"
813 if [ $stat1 -lt $stat2 ]; then
814 echo "writes continuing during recovery"
817 echo "writes not continuing during recovery, bug 2477"
820 echo "waiting for writeme $WRITE_PID"
824 echo "waiting for multiop $PID"
825 wait $PID || return 2
826 do_facet client munlink $MOUNT/$tfile || return 3
827 do_facet client munlink $MOUNT/${tfile}-2 || return 3
830 run_test 40 "cause recovery in ptlrpc, ensure IO continues"
834 # make sure that a read to one osc doesn't try to double-unlock its page just
835 # because another osc is invalid. trigger_group_io used to mistakenly return
836 # an error if any oscs were invalid even after having successfully put rpcs
837 # on valid oscs. This was fatal if the caller was ll_readpage who unlocked
838 # the page, guarnateeing that the unlock from the RPC completion would
839 # assert on trying to unlock the unlocked page.
841 [ $OSTCOUNT -lt 2 ] &&
842 skip_env "skipping test 41: we don't have a second OST to test with" &&
845 local f=$MOUNT/$tfile
846 # make sure the start of the file is ost1
847 $SETSTRIPE -S $((128 * 1024)) -i 0 $f
848 do_facet client dd if=/dev/zero of=$f bs=4k count=1 || return 3
850 # fail ost2 and read from ost1
851 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $ost2_svc)
852 local osc2dev=$(do_facet $SINGLEMDS "lctl get_param -n devices" | \
853 grep $mdtosc | awk '{print $1}')
854 [ -z "$osc2dev" ] && echo "OST: $ost2_svc" && lctl get_param -n devices &&
856 do_facet $SINGLEMDS $LCTL --device $osc2dev deactivate || return 1
857 do_facet client dd if=$f of=/dev/null bs=4k count=1 || return 3
858 do_facet $SINGLEMDS $LCTL --device $osc2dev activate || return 2
861 run_test 41 "read from a valid osc while other oscs are invalid"
863 # test MDS recovery after ost failure
865 blocks=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
866 createmany -o $DIR/$tfile-%d 800
868 unlinkmany $DIR/$tfile-%d 0 400
870 lctl set_param debug=-1
873 # osc is evicted, fs is smaller (but only with failout OSTs (bug 7287)
874 #blocks_after=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
875 #[ $blocks_after -lt $blocks ] || return 1
876 echo wait for MDS to timeout and recover
877 sleep $((TIMEOUT * 2))
879 unlinkmany $DIR/$tfile-%d 400 400
880 $CHECKSTAT -t file $DIR/$tfile-* && return 2 || true
882 run_test 42 "recovery after ost failure"
884 # timeout in MDS/OST recovery RPC will LBUG MDS
885 test_43() { # bug 2530
886 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
888 replay_barrier $SINGLEMDS
890 # OBD_FAIL_OST_CREATE_NET 0x204
891 do_facet ost1 "lctl set_param fail_loc=0x80000204"
894 do_facet ost1 "lctl set_param fail_loc=0"
898 run_test 43 "mds osc import failure during recovery; don't LBUG"
900 test_44a() { # was test_44
903 local mdcdev=$($LCTL get_param -n devices |
904 awk "/ ${FSNAME}-MDT0000-mdc-/ {print \$1}")
905 [ "$mdcdev" ] || return 2
906 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
907 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
909 # adaptive timeouts slow this way down
910 if at_is_enabled; then
911 at_max_saved=$(at_max_get mds)
915 for i in `seq 1 10`; do
916 echo "$i of 10 ($(date +%s))"
917 do_facet $SINGLEMDS "lctl get_param -n mdt.*.mdt.timeouts | grep service"
918 #define OBD_FAIL_TGT_CONN_RACE 0x701
919 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000701"
920 # lctl below may fail, it is valid case
921 $LCTL --device $mdcdev recover
924 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
925 [ $at_max_saved -ne 0 ] && at_max_set $at_max_saved mds
928 run_test 44a "race in target handle connect"
931 local mdcdev=$($LCTL get_param -n devices |
932 awk "/ ${FSNAME}-MDT0000-mdc-/ {print \$1}")
933 [ "$mdcdev" ] || return 2
934 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
935 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
937 for i in `seq 1 10`; do
938 echo "$i of 10 ($(date +%s))"
939 do_facet $SINGLEMDS "lctl get_param -n mdt.*.mdt.timeouts | grep service"
940 #define OBD_FAIL_TGT_DELAY_RECONNECT 0x704
941 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000704"
942 # lctl below may fail, it is valid case
943 $LCTL --device $mdcdev recover
946 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
949 run_test 44b "race in target handle connect"
952 replay_barrier $SINGLEMDS
953 createmany -m $DIR/$tfile-%d 100 || error "failed to create directories"
954 #define OBD_FAIL_TGT_RCVG_FLAG 0x712
955 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000712"
956 fail_abort $SINGLEMDS
957 unlinkmany $DIR/$tfile-%d 100 && error "unliked after fail abort"
959 unlinkmany $DIR/$tfile-%d 100 && error "unliked after fail"
962 run_test 44c "race in target handle connect"
964 # Handle failed close
966 local mdcdev=$($LCTL get_param -n devices |
967 awk "/ ${FSNAME}-MDT0000-mdc-/ {print \$1}")
968 [ "$mdcdev" ] || return 2
969 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
970 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
972 $LCTL --device $mdcdev recover || return 6
974 multiop_bg_pause $DIR/$tfile O_c || return 1
977 # This will cause the CLOSE to fail before even
978 # allocating a reply buffer
979 $LCTL --device $mdcdev deactivate || return 4
983 wait $pid || return 1
985 $LCTL --device $mdcdev activate || return 5
988 $CHECKSTAT -t file $DIR/$tfile || return 2
991 run_test 45 "Handle failed close"
995 drop_reply "touch $DIR/$tfile"
997 # ironically, the previous test, 45, will cause a real forced close,
998 # so just look for one for this test
999 dmesg | grep -i "force closing client file handle for $tfile" && return 1
1002 run_test 46 "Don't leak file handle after open resend (3325)"
1004 test_47() { # bug 2824
1005 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1007 # create some files to make sure precreate has been done on all
1008 # OSTs. (just in case this test is run independently)
1009 createmany -o $DIR/$tfile 20 || return 1
1011 # OBD_FAIL_OST_CREATE_NET 0x204
1013 do_facet ost1 "lctl set_param fail_loc=0x80000204"
1014 client_up || return 2
1016 # let the MDS discover the OST failure, attempt to recover, fail
1017 # and recover again.
1018 sleep $((3 * TIMEOUT))
1020 # Without 2824, this createmany would hang
1021 createmany -o $DIR/$tfile 20 || return 3
1022 unlinkmany $DIR/$tfile 20 || return 4
1024 do_facet ost1 "lctl set_param fail_loc=0"
1027 run_test 47 "MDS->OSC failure during precreate cleanup (2824)"
1030 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1031 [ "$OSTCOUNT" -lt "2" ] && skip_env "$OSTCOUNT < 2 OSTs -- skipping" && return
1033 replay_barrier $SINGLEMDS
1034 createmany -o $DIR/$tfile 20 || return 1
1035 # OBD_FAIL_OST_EROFS 0x216
1036 facet_failover $SINGLEMDS
1037 do_facet ost1 "lctl set_param fail_loc=0x80000216"
1038 client_up || return 2
1040 createmany -o $DIR/$tfile 20 20 || return 2
1041 unlinkmany $DIR/$tfile 40 || return 3
1044 run_test 48 "MDS->OSC failure during precreate cleanup (2824)"
1047 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $ost1_svc)
1048 local oscdev=$(do_facet $SINGLEMDS "lctl get_param -n devices" | \
1049 grep $mdtosc | awk '{print $1}')
1050 [ "$oscdev" ] || return 1
1051 do_facet $SINGLEMDS $LCTL --device $oscdev recover || return 2
1052 do_facet $SINGLEMDS $LCTL --device $oscdev recover || return 3
1053 # give the mds_lov_sync threads a chance to run
1056 run_test 50 "Double OSC recovery, don't LASSERT (3812)"
1058 # b3764 timed out lock replay
1061 cancel_lru_locks mdc
1063 multiop $DIR/$tfile s || return 1
1064 replay_barrier $SINGLEMDS
1065 #define OBD_FAIL_LDLM_REPLY 0x30c
1066 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000030c"
1067 fail $SINGLEMDS || return 2
1068 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1070 $CHECKSTAT -t file $DIR/$tfile-* && return 3 || true
1072 run_test 52 "time out lock replay (3764)"
1074 # bug 3462 - simultaneous MDC requests
1076 cancel_lru_locks mdc # cleanup locks from former test cases
1077 mkdir -p $DIR/${tdir}-1
1078 mkdir -p $DIR/${tdir}-2
1079 multiop $DIR/${tdir}-1/f O_c &
1081 # give multiop a change to open
1084 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1085 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1086 kill -USR1 $close_pid
1087 cancel_lru_locks mdc # force the close
1088 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1090 mcreate $DIR/${tdir}-2/f || return 1
1092 # close should still be here
1093 [ -d /proc/$close_pid ] || return 2
1095 replay_barrier_nodf $SINGLEMDS
1097 wait $close_pid || return 3
1099 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1100 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1101 rm -rf $DIR/${tdir}-*
1103 run_test 53a "|X| close request while two MDC requests in flight"
1106 cancel_lru_locks mdc # cleanup locks from former test cases
1107 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1109 mkdir -p $DIR/${tdir}-1
1110 mkdir -p $DIR/${tdir}-2
1111 multiop_bg_pause $DIR/${tdir}-1/f O_c || return 6
1114 #define OBD_FAIL_MDS_REINT_NET 0x107
1115 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1116 mcreate $DIR/${tdir}-2/f &
1120 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1121 kill -USR1 $close_pid
1122 cancel_lru_locks mdc # force the close
1123 wait $close_pid || return 1
1124 # open should still be here
1125 [ -d /proc/$open_pid ] || return 2
1127 replay_barrier_nodf $SINGLEMDS
1129 wait $open_pid || return 3
1131 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1132 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1133 rm -rf $DIR/${tdir}-*
1135 run_test 53b "|X| open request while two MDC requests in flight"
1138 cancel_lru_locks mdc # cleanup locks from former test cases
1139 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1141 mkdir -p $DIR/${tdir}-1
1142 mkdir -p $DIR/${tdir}-2
1143 multiop $DIR/${tdir}-1/f O_c &
1146 #define OBD_FAIL_MDS_REINT_NET 0x107
1147 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1148 mcreate $DIR/${tdir}-2/f &
1152 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1153 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1154 kill -USR1 $close_pid
1155 cancel_lru_locks mdc # force the close
1157 #bz20647: make sure all pids are exists before failover
1158 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1159 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1160 replay_barrier_nodf $SINGLEMDS
1161 fail_nodf $SINGLEMDS
1162 wait $open_pid || return 1
1164 # close should be gone
1165 [ -d /proc/$close_pid ] && return 2
1166 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1168 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1169 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1170 rm -rf $DIR/${tdir}-*
1172 run_test 53c "|X| open request and close request while two MDC requests in flight"
1175 cancel_lru_locks mdc # cleanup locks from former test cases
1176 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1178 mkdir -p $DIR/${tdir}-1
1179 mkdir -p $DIR/${tdir}-2
1180 multiop $DIR/${tdir}-1/f O_c &
1182 # give multiop a chance to open
1185 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1186 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1187 kill -USR1 $close_pid
1188 cancel_lru_locks mdc # force the close
1189 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1190 mcreate $DIR/${tdir}-2/f || return 1
1192 # close should still be here
1193 [ -d /proc/$close_pid ] || return 2
1195 wait $close_pid || return 3
1197 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1198 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1199 rm -rf $DIR/${tdir}-*
1201 run_test 53d "|X| close reply while two MDC requests in flight"
1204 cancel_lru_locks mdc # cleanup locks from former test cases
1205 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1207 mkdir -p $DIR/${tdir}-1
1208 mkdir -p $DIR/${tdir}-2
1209 multiop $DIR/${tdir}-1/f O_c &
1212 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1213 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1214 mcreate $DIR/${tdir}-2/f &
1218 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1219 kill -USR1 $close_pid
1220 cancel_lru_locks mdc # force the close
1221 wait $close_pid || return 1
1222 # open should still be here
1223 [ -d /proc/$open_pid ] || return 2
1225 replay_barrier_nodf $SINGLEMDS
1227 wait $open_pid || return 3
1229 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1230 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1231 rm -rf $DIR/${tdir}-*
1233 run_test 53e "|X| open reply while two MDC requests in flight"
1236 cancel_lru_locks mdc # cleanup locks from former test cases
1237 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1239 mkdir -p $DIR/${tdir}-1
1240 mkdir -p $DIR/${tdir}-2
1241 multiop $DIR/${tdir}-1/f O_c &
1244 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1245 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1246 mcreate $DIR/${tdir}-2/f &
1250 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1251 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1252 kill -USR1 $close_pid
1253 cancel_lru_locks mdc # force the close
1255 #bz20647: make sure all pids are exists before failover
1256 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1257 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1258 replay_barrier_nodf $SINGLEMDS
1259 fail_nodf $SINGLEMDS
1260 wait $open_pid || return 1
1262 # close should be gone
1263 [ -d /proc/$close_pid ] && return 2
1264 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1266 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1267 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1268 rm -rf $DIR/${tdir}-*
1270 run_test 53f "|X| open reply and close reply while two MDC requests in flight"
1273 cancel_lru_locks mdc # cleanup locks from former test cases
1274 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1276 mkdir -p $DIR/${tdir}-1
1277 mkdir -p $DIR/${tdir}-2
1278 multiop $DIR/${tdir}-1/f O_c &
1281 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1282 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1283 mcreate $DIR/${tdir}-2/f &
1287 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1288 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1289 kill -USR1 $close_pid
1290 cancel_lru_locks mdc # force the close
1291 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1293 #bz20647: make sure all pids are exists before failover
1294 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1295 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1296 replay_barrier_nodf $SINGLEMDS
1297 fail_nodf $SINGLEMDS
1298 wait $open_pid || return 1
1300 # close should be gone
1301 [ -d /proc/$close_pid ] && return 2
1303 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1304 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1305 rm -rf $DIR/${tdir}-*
1307 run_test 53g "|X| drop open reply and close request while close and open are both in flight"
1310 cancel_lru_locks mdc # cleanup locks from former test cases
1311 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1313 mkdir -p $DIR/${tdir}-1
1314 mkdir -p $DIR/${tdir}-2
1315 multiop $DIR/${tdir}-1/f O_c &
1318 #define OBD_FAIL_MDS_REINT_NET 0x107
1319 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1320 mcreate $DIR/${tdir}-2/f &
1324 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1325 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1326 kill -USR1 $close_pid
1327 cancel_lru_locks mdc # force the close
1330 #bz20647: make sure all pids are exists before failover
1331 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1332 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1333 replay_barrier_nodf $SINGLEMDS
1334 fail_nodf $SINGLEMDS
1335 wait $open_pid || return 1
1337 # close should be gone
1338 [ -d /proc/$close_pid ] && return 2
1339 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1341 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1342 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1343 rm -rf $DIR/${tdir}-*
1345 run_test 53h "|X| open request and close reply while two MDC requests in flight"
1347 #b_cray 54 "|X| open request and close reply while two MDC requests in flight"
1349 #b3761 ASSERTION(hash != 0) failed
1351 # OBD_FAIL_MDS_OPEN_CREATE | OBD_FAIL_ONCE
1352 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012b"
1354 # give touch a chance to run
1356 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1360 run_test 55 "let MDS_CHECK_RESENT return the original return code instead of 0"
1362 #b3440 ASSERTION(rec->ur_fid2->id) failed
1364 ln -s foo $DIR/$tfile
1365 replay_barrier $SINGLEMDS
1366 #drop_reply "cat $DIR/$tfile"
1370 run_test 56 "don't replay a symlink open request (3440)"
1372 #recovery one mds-ost setattr from llog
1374 #define OBD_FAIL_MDS_OST_SETATTR 0x12c
1375 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012c"
1377 replay_barrier $SINGLEMDS
1380 $CHECKSTAT -t file $DIR/$tfile || return 1
1381 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1384 run_test 57 "test recovery from llog for setattr op"
1386 #recovery many mds-ost setattr from llog
1389 #define OBD_FAIL_MDS_OST_SETATTR 0x12c
1390 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012c"
1391 createmany -o $DIR/$tdir/$tfile-%d 2500
1392 replay_barrier $SINGLEMDS
1395 $CHECKSTAT -t file $DIR/$tdir/$tfile-* >/dev/null || return 1
1396 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1397 unlinkmany $DIR/$tdir/$tfile-%d 2500
1400 run_test 58a "test recovery from llog for setattr op (test llog_gen_rec)"
1406 large_xattr_enabled &&
1407 orig="$(generate_string $(max_xattr_size))" || orig="bar"
1409 mount_client $MOUNT2
1411 touch $DIR/$tdir/$tfile
1412 replay_barrier $SINGLEMDS
1413 setfattr -n trusted.foo -v $orig $DIR/$tdir/$tfile
1415 new=$(get_xattr_value trusted.foo $MOUNT2/$tdir/$tfile)
1416 [[ "$new" = "$orig" ]] || return 1
1417 rm -f $DIR/$tdir/$tfile
1419 zconf_umount `hostname` $MOUNT2
1421 run_test 58b "test replay of setxattr op"
1423 test_58c() { # bug 16570
1428 if large_xattr_enabled; then
1429 local xattr_size=$(max_xattr_size)
1430 orig="$(generate_string $((xattr_size / 2)))"
1431 orig1="$(generate_string $xattr_size)"
1437 mount_client $MOUNT2
1439 touch $DIR/$tdir/$tfile
1440 drop_request "setfattr -n trusted.foo -v $orig $DIR/$tdir/$tfile" ||
1442 new=$(get_xattr_value trusted.foo $MOUNT2/$tdir/$tfile)
1443 [[ "$new" = "$orig" ]] || return 2
1444 drop_reint_reply "setfattr -n trusted.foo1 -v $orig1 $DIR/$tdir/$tfile" ||
1446 new=$(get_xattr_value trusted.foo1 $MOUNT2/$tdir/$tfile)
1447 [[ "$new" = "$orig1" ]] || return 4
1448 rm -f $DIR/$tdir/$tfile
1450 zconf_umount $HOSTNAME $MOUNT2
1452 run_test 58c "resend/reconstruct setxattr op"
1454 # log_commit_thread vs filter_destroy race used to lead to import use after free
1457 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1460 createmany -o $DIR/$tdir/$tfile-%d 200
1462 unlinkmany $DIR/$tdir/$tfile-%d 200
1463 #define OBD_FAIL_PTLRPC_DELAY_RECOV 0x507
1464 do_facet ost1 "lctl set_param fail_loc=0x507"
1467 do_facet ost1 "lctl set_param fail_loc=0x0"
1471 run_test 59 "test log_commit_thread vs filter_destroy race"
1473 # race between add unlink llog vs cat log init in post_recovery (only for b1_6)
1474 # bug 12086: should no oops and No ctxt error for this test
1477 createmany -o $DIR/$tdir/$tfile-%d 200
1478 replay_barrier $SINGLEMDS
1479 unlinkmany $DIR/$tdir/$tfile-%d 0 100
1481 unlinkmany $DIR/$tdir/$tfile-%d 100 100
1482 local no_ctxt=`dmesg | grep "No ctxt"`
1483 [ -z "$no_ctxt" ] || error "ctxt is not initialized in recovery"
1485 run_test 60 "test llog post recovery init vs llog unlink"
1487 #test race llog recovery thread vs llog cleanup
1488 test_61a() { # was test_61
1489 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1492 createmany -o $DIR/$tdir/$tfile-%d 800
1494 # OBD_FAIL_OST_LLOG_RECOVERY_TIMEOUT 0x221
1495 unlinkmany $DIR/$tdir/$tfile-%d 800
1496 set_nodes_failloc "$(osts_nodes)" 0x80000221
1501 set_nodes_failloc "$(osts_nodes)" 0x0
1503 $CHECKSTAT -t file $DIR/$tdir/$tfile-* && return 1
1506 run_test 61a "test race llog recovery vs llog cleanup"
1508 #test race mds llog sync vs llog cleanup
1510 # OBD_FAIL_MDS_LLOG_SYNC_TIMEOUT 0x13a
1511 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013a"
1512 facet_failover $SINGLEMDS
1515 do_facet client dd if=/dev/zero of=$DIR/$tfile bs=4k count=1 || return 1
1517 run_test 61b "test race mds llog sync vs llog cleanup"
1519 #test race cancel cookie cb vs llog cleanup
1521 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1523 # OBD_FAIL_OST_CANCEL_COOKIE_TIMEOUT 0x222
1525 set_nodes_failloc "$(osts_nodes)" 0x80000222
1529 set_nodes_failloc "$(osts_nodes)" 0x0
1531 run_test 61c "test race mds llog sync vs llog cleanup"
1533 test_61d() { # bug 16002 # bug 17466 # bug 22137
1534 # OBD_FAIL_OBD_LLOG_SETUP 0x605
1536 do_facet mgs "lctl set_param fail_loc=0x80000605"
1537 start mgs $MGSDEV $MGS_MOUNT_OPTS && error "mgs start should have failed"
1538 do_facet mgs "lctl set_param fail_loc=0"
1539 start mgs $MGSDEV $MGS_MOUNT_OPTS || error "cannot restart mgs"
1541 run_test 61d "error in llog_setup should cleanup the llog context correctly"
1543 test_62() { # Bug 15756 - don't mis-drop resent replay
1545 replay_barrier $SINGLEMDS
1546 createmany -o $DIR/$tdir/$tfile- 25
1547 #define OBD_FAIL_TGT_REPLAY_DROP 0x707
1548 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000707"
1550 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1551 unlinkmany $DIR/$tdir/$tfile- 25 || return 2
1554 run_test 62 "don't mis-drop resent replay"
1556 #Adaptive Timeouts (bug 3055)
1564 echo "Cleaning up AT ..."
1565 if [ -n "$ATOLDBASE" ]; then
1566 local at_history=$($LCTL get_param -n at_history)
1567 do_facet $SINGLEMDS "lctl set_param at_history=$at_history" || true
1568 do_facet ost1 "lctl set_param at_history=$at_history" || true
1571 if [ $AT_MAX_SET -ne 0 ]; then
1572 for facet in mds client ost; do
1573 var=AT_MAX_SAVE_${facet}
1574 echo restore AT on $facet to saved value ${!var}
1575 at_max_set ${!var} $facet
1576 at_new=$(at_max_get $facet)
1577 echo Restored AT value on $facet $at_new
1578 [ $at_new -eq ${!var} ] || \
1579 error "$facet : AT value was not restored SAVED ${!var} NEW $at_new"
1586 local at_max_new=600
1588 # Save at_max original values
1590 if [ $AT_MAX_SET -eq 0 ]; then
1591 # Suppose that all osts have the same at_max
1592 for facet in mds client ost; do
1593 eval AT_MAX_SAVE_${facet}=$(at_max_get $facet)
1597 for facet in mds client ost; do
1598 at_max=$(at_max_get $facet)
1599 if [ $at_max -ne $at_max_new ]; then
1600 echo "AT value on $facet is $at_max, set it by force temporarily to $at_max_new"
1601 at_max_set $at_max_new $facet
1606 if [ -z "$ATOLDBASE" ]; then
1607 ATOLDBASE=$(do_facet $SINGLEMDS "lctl get_param -n at_history")
1608 # speed up the timebase so we can check decreasing AT
1609 do_facet $SINGLEMDS "lctl set_param at_history=8" || true
1610 do_facet ost1 "lctl set_param at_history=8" || true
1612 # sleep for a while to cool down, should be > 8s and also allow
1613 # at least one ping to be sent. simply use TIMEOUT to be safe.
1618 test_65a() #bug 3055
1620 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1622 at_start || return 0
1623 $LCTL dk > /dev/null
1625 $LCTL set_param debug="other"
1626 # Slow down a request to the current service time, this is critical
1627 # because previous tests may have caused this value to increase.
1628 REQ_DELAY=`lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts |
1629 awk '/portal 12/ {print $5}'`
1630 REQ_DELAY=$((${REQ_DELAY} + ${REQ_DELAY} / 4 + 5))
1632 do_facet $SINGLEMDS lctl set_param fail_val=$((${REQ_DELAY} * 1000))
1633 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1634 do_facet $SINGLEMDS $LCTL set_param fail_loc=0x8000050a
1635 createmany -o $DIR/$tfile 10 > /dev/null
1636 unlinkmany $DIR/$tfile 10 > /dev/null
1637 # check for log message
1638 $LCTL dk | grep "Early reply #" || error "No early reply"
1640 # client should show REQ_DELAY estimates
1641 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep portal
1643 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep portal
1645 run_test 65a "AT: verify early replies"
1647 test_65b() #bug 3055
1649 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1651 at_start || return 0
1654 $LCTL set_param debug="other trace"
1655 $LCTL dk > /dev/null
1656 # Slow down a request to the current service time, this is critical
1657 # because previous tests may have caused this value to increase.
1658 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tfile
1659 multiop $DIR/$tfile Ow1yc
1660 REQ_DELAY=`lctl get_param -n osc.${FSNAME}-OST0000-osc-*.timeouts |
1661 awk '/portal 6/ {print $5}'`
1662 REQ_DELAY=$((${REQ_DELAY} + ${REQ_DELAY} / 4 + 5))
1664 do_facet ost1 lctl set_param fail_val=${REQ_DELAY}
1665 #define OBD_FAIL_OST_BRW_PAUSE_PACK 0x224
1666 do_facet ost1 $LCTL set_param fail_loc=0x224
1669 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tfile
1670 # force some real bulk transfer
1671 multiop $DIR/$tfile oO_CREAT:O_RDWR:O_SYNC:w4096c
1673 do_facet ost1 $LCTL set_param fail_loc=0
1674 # check for log message
1675 $LCTL dk | grep "Early reply #" || error "No early reply"
1677 # client should show REQ_DELAY estimates
1678 lctl get_param -n osc.${FSNAME}-OST0000-osc-*.timeouts | grep portal
1680 run_test 65b "AT: verify early replies on packed reply / bulk"
1682 test_66a() #bug 3055
1684 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1686 at_start || return 0
1687 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1688 # adjust 5s at a time so no early reply is sent (within deadline)
1689 do_facet $SINGLEMDS "$LCTL set_param fail_val=5000"
1690 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1691 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0x8000050a"
1692 createmany -o $DIR/$tfile 20 > /dev/null
1693 unlinkmany $DIR/$tfile 20 > /dev/null
1694 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1695 do_facet $SINGLEMDS "$LCTL set_param fail_val=10000"
1696 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0x8000050a"
1697 createmany -o $DIR/$tfile 20 > /dev/null
1698 unlinkmany $DIR/$tfile 20 > /dev/null
1699 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1700 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0"
1702 createmany -o $DIR/$tfile 20 > /dev/null
1703 unlinkmany $DIR/$tfile 20 > /dev/null
1704 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1705 CUR=$(lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | awk '/portal 12/ {print $5}')
1706 WORST=$(lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | awk '/portal 12/ {print $7}')
1707 echo "Current MDT timeout $CUR, worst $WORST"
1708 [ $CUR -lt $WORST ] || error "Current $CUR should be less than worst $WORST"
1710 run_test 66a "AT: verify MDT service time adjusts with no early replies"
1712 test_66b() #bug 3055
1714 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1716 at_start || return 0
1717 ORIG=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1718 awk '/network/ {print $4}')
1719 $LCTL set_param fail_val=$(($ORIG + 5))
1720 #define OBD_FAIL_PTLRPC_PAUSE_REP 0x50c
1721 $LCTL set_param fail_loc=0x50c
1722 ls $DIR/$tfile > /dev/null 2>&1
1723 $LCTL set_param fail_loc=0
1724 CUR=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1725 awk '/network/ {print $4}')
1726 WORST=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1727 awk '/network/ {print $6}')
1728 echo "network timeout orig $ORIG, cur $CUR, worst $WORST"
1729 [ $WORST -gt $ORIG ] ||
1730 error "Worst $WORST should be worse than orig $ORIG"
1732 run_test 66b "AT: verify net latency adjusts"
1734 test_67a() #bug 3055
1736 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1738 at_start || return 0
1739 CONN1=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1740 # sleeping threads may drive values above this
1741 do_facet ost1 "$LCTL set_param fail_val=400"
1742 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1743 do_facet ost1 "$LCTL set_param fail_loc=0x50a"
1744 createmany -o $DIR/$tfile 20 > /dev/null
1745 unlinkmany $DIR/$tfile 20 > /dev/null
1746 do_facet ost1 "$LCTL set_param fail_loc=0"
1747 CONN2=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1748 ATTEMPTS=$(($CONN2 - $CONN1))
1749 echo "$ATTEMPTS osc reconnect attempts on gradual slow"
1750 [ $ATTEMPTS -gt 0 ] && error_ignore 13721 "AT should have prevented reconnect"
1753 run_test 67a "AT: verify slow request processing doesn't induce reconnects"
1755 test_67b() #bug 3055
1757 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1759 at_start || return 0
1760 CONN1=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1762 # exhaust precreations on ost1
1763 local OST=$(ostname_from_index 0)
1764 local mdtosc=$(get_mdtosc_proc_path mds $OST)
1765 local last_id=$(do_facet $SINGLEMDS lctl get_param -n \
1766 osc.$mdtosc.prealloc_last_id)
1767 local next_id=$(do_facet $SINGLEMDS lctl get_param -n \
1768 osc.$mdtosc.prealloc_next_id)
1770 mkdir -p $DIR/$tdir/${OST}
1771 $SETSTRIPE -i 0 -c 1 $DIR/$tdir/${OST} || error "$SETSTRIPE"
1772 echo "Creating to objid $last_id on ost $OST..."
1773 #define OBD_FAIL_OST_PAUSE_CREATE 0x223
1774 do_facet ost1 "$LCTL set_param fail_val=20000"
1775 do_facet ost1 "$LCTL set_param fail_loc=0x80000223"
1776 createmany -o $DIR/$tdir/${OST}/f $next_id $((last_id - next_id + 2))
1779 do_facet ost1 "lctl get_param -n ost.OSS.ost_create.timeouts"
1781 CONN2=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1782 ATTEMPTS=$(($CONN2 - $CONN1))
1783 echo "$ATTEMPTS osc reconnect attempts on instant slow"
1784 # do it again; should not timeout
1785 do_facet ost1 "$LCTL set_param fail_loc=0x80000223"
1786 cp /etc/profile $DIR/$tfile || error "cp failed"
1787 do_facet ost1 "$LCTL set_param fail_loc=0"
1789 do_facet ost1 "lctl get_param -n ost.OSS.ost_create.timeouts"
1790 CONN3=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1791 ATTEMPTS=$(($CONN3 - $CONN2))
1792 echo "$ATTEMPTS osc reconnect attempts on 2nd slow"
1793 [ $ATTEMPTS -gt 0 ] && error "AT should have prevented reconnect"
1796 run_test 67b "AT: verify instant slowdown doesn't induce reconnects"
1798 test_68 () #bug 13813
1800 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1802 at_start || return 0
1803 local ldlm_enqueue_min=$(find /sys -name ldlm_enqueue_min)
1804 [ -z "$ldlm_enqueue_min" ] && skip "missing /sys/.../ldlm_enqueue_min" && return 0
1805 local ldlm_enqueue_min_r=$(do_facet ost1 "find /sys -name ldlm_enqueue_min")
1806 [ -z "$ldlm_enqueue_min_r" ] && skip "missing /sys/.../ldlm_enqueue_min in the ost1" && return 0
1807 local ENQ_MIN=$(cat $ldlm_enqueue_min)
1808 local ENQ_MIN_R=$(do_facet ost1 "cat $ldlm_enqueue_min_r")
1809 echo $TIMEOUT >> $ldlm_enqueue_min
1810 do_facet ost1 "echo $TIMEOUT >> $ldlm_enqueue_min_r"
1814 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tdir
1815 #define OBD_FAIL_LDLM_PAUSE_CANCEL 0x312
1816 $LCTL set_param fail_val=$(($TIMEOUT - 1))
1817 $LCTL set_param fail_loc=0x80000312
1818 cp /etc/profile $DIR/$tdir/${tfile}_1 || error "1st cp failed $?"
1819 $LCTL set_param fail_val=$((TIMEOUT * 5 / 4))
1820 $LCTL set_param fail_loc=0x80000312
1821 cp /etc/profile $DIR/$tdir/${tfile}_2 || error "2nd cp failed $?"
1822 $LCTL set_param fail_loc=0
1824 echo $ENQ_MIN >> $ldlm_enqueue_min
1825 do_facet ost1 "echo $ENQ_MIN_R >> $ldlm_enqueue_min_r"
1829 run_test 68 "AT: verify slowing locks"
1832 # end of AT tests includes above lines
1835 # start multi-client tests
1837 [ -z "$CLIENTS" ] && \
1838 { skip "Need two or more clients." && return; }
1839 [ $CLIENTCOUNT -lt 2 ] && \
1840 { skip "Need two or more clients, have $CLIENTCOUNT" && return; }
1842 echo "mount clients $CLIENTS ..."
1843 zconf_mount_clients $CLIENTS $MOUNT
1845 local clients=${CLIENTS//,/ }
1846 echo "Write/read files on $DIR ; clients $CLIENTS ... "
1847 for CLIENT in $clients; do
1848 do_node $CLIENT dd bs=1M count=10 if=/dev/zero \
1849 of=$DIR/${tfile}_${CLIENT} 2>/dev/null || \
1850 error "dd failed on $CLIENT"
1853 local prev_client=$(echo $clients | sed 's/^.* \(.\+\)$/\1/')
1854 for C in ${CLIENTS//,/ }; do
1855 do_node $prev_client dd if=$DIR/${tfile}_${C} of=/dev/null 2>/dev/null || \
1856 error "dd if=$DIR/${tfile}_${C} failed on $prev_client"
1862 run_test 70a "check multi client t-f"
1864 check_for_process () {
1869 killall_process $clients "$prog" -0
1872 killall_process () {
1873 local clients=${1:-$(hostname)}
1878 do_nodes $clients "killall $signal $name"
1882 local clients=${CLIENTS:-$HOSTNAME}
1884 zconf_mount_clients $clients $MOUNT
1887 [ "$SLOW" = "no" ] && duration=120
1888 # set duration to 900 because it takes some time to boot node
1889 [ "$FAILURE_MODE" = HARD ] && duration=900
1891 local start_ts=$(date +%s)
1892 local cmd="rundbench 1 -t $duration"
1894 do_nodesv $clients "set -x; MISSING_DBENCH_OK=$MISSING_DBENCH_OK \
1895 PATH=\$PATH:$LUSTRE/utils:$LUSTRE/tests/:$DBENCH_LIB \
1896 DBENCH_LIB=$DBENCH_LIB TESTSUITE=$TESTSUITE TESTNAME=$TESTNAME \
1897 MOUNT=$MOUNT DIR=$DIR/$tdir/\\\$(hostname) LCTL=$LCTL $cmd" &
1899 log "Started rundbench load pid=$pid ..."
1901 # give rundbench a chance to start, bug 24118
1903 local elapsed=$(($(date +%s) - start_ts))
1904 local num_failovers=0
1905 while [ $elapsed -lt $duration ]; do
1906 if ! check_for_process $clients dbench; then
1907 error_noexit "dbench not running on some of $clients!"
1908 killall_process $clients dbench
1912 replay_barrier $SINGLEMDS
1913 sleep 1 # give clients a time to do operations
1914 # Increment the number of failovers
1915 num_failovers=$((num_failovers+1))
1916 log "$TESTNAME fail $SINGLEMDS $num_failovers times"
1918 elapsed=$(($(date +%s) - start_ts))
1921 wait $pid || error "rundbench load on $clients failed!"
1923 run_test 70b "mds recovery; $CLIENTCOUNT clients"
1924 # end multi-client tests
1927 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1931 replay_barrier $SINGLEMDS
1932 #define OBD_FAIL_LDLM_ENQUEUE_NET 0x302
1933 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000302"
1936 wait $pid || return 1
1937 [ -e $DIR/$tfile ] && return 2
1940 run_test 73a "open(O_CREAT), unlink, replay, reconnect before open replay , close"
1943 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1947 replay_barrier $SINGLEMDS
1948 #define OBD_FAIL_LDLM_REPLY 0x30c
1949 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000030c"
1952 wait $pid || return 1
1953 [ -e $DIR/$tfile ] && return 2
1956 run_test 73b "open(O_CREAT), unlink, replay, reconnect at open_replay reply, close"
1959 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1963 replay_barrier $SINGLEMDS
1964 #define OBD_FAIL_TGT_LAST_REPLAY 0x710
1965 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000710"
1968 wait $pid || return 1
1969 [ -e $DIR/$tfile ] && return 2
1972 run_test 73c "open(O_CREAT), unlink, replay, reconnect at last_replay, close"
1976 local clients=${CLIENTS:-$HOSTNAME}
1978 zconf_umount_clients $clients $MOUNT
1980 facet_failover $SINGLEMDS
1981 zconf_mount_clients $clients $MOUNT
1983 touch $DIR/$tfile || return 1
1984 rm $DIR/$tfile || return 2
1985 clients_up || error "client evicted: $?"
1988 run_test 74 "Ensure applications don't fail waiting for OST recovery"
1990 remote_dir_check_80() {
1992 local diridx=$($GETSTRIPE -M $remote_dir)
1993 [ $diridx -eq $MDTIDX ] || error "$diridx != $MDTIDX"
1995 createmany -o $remote_dir/f-%d 20 || error "creation failed"
1996 local fileidx=$($GETSTRIPE -M $remote_dir/f-1)
1997 [ $fileidx -eq $MDTIDX ] || error "$fileidx != $MDTIDX"
2003 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2004 ([ $FAILURE_MODE == "HARD" ] &&
2005 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2006 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2010 local remote_dir=$DIR/$tdir/remote_dir
2013 # OBD_FAIL_UPDATE_OBJ_NET 0x1500
2014 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1500
2015 $LFS mkdir -i $MDTIDX $remote_dir &
2017 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0
2019 fail mds$((MDTIDX + 1))
2021 wait $CLIENT_PID || error "remote creation failed"
2023 remote_dir_check_80 || error "remote dir check failed"
2024 rm -rf $DIR/$tdir || error "rmdir failed"
2028 run_test 80a "DNE: create remote dir, drop update rep from MDT1, fail MDT1"
2031 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2032 ([ $FAILURE_MODE == "HARD" ] &&
2033 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2034 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2038 local remote_dir=$DIR/$tdir/remote_dir
2041 # OBD_FAIL_UPDATE_OBJ_NET 0x1500
2042 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1500
2043 $LFS mkdir -i $MDTIDX $remote_dir &
2045 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0
2049 wait $CLIENT_PID || error "remote creation failed"
2051 remote_dir_check_80 || error "remote dir check failed"
2052 rm -rf $DIR/$tdir || error "rmdir failed"
2056 run_test 80b "DNE: create remote dir, drop update rep from MDT1, fail MDT0"
2059 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2060 ([ $FAILURE_MODE == "HARD" ] &&
2061 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2062 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2066 local remote_dir=$DIR/$tdir/remote_dir
2069 # OBD_FAIL_UPDATE_OBJ_NET 0x1500
2070 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1500
2071 $LFS mkdir -i $MDTIDX $remote_dir &
2073 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0
2076 fail mds$((MDTIDX + 1))
2078 wait $CLIENT_PID || error "remote creation failed"
2080 remote_dir_check_80 || error "remote dir check failed"
2081 rm -rf $DIR/$tdir || error "rmdir failed"
2085 run_test 80c "DNE: create remote dir, drop update rep from MDT1, fail MDT[0,1]"
2088 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2090 local remote_dir=$DIR/$tdir/remote_dir
2093 # OBD_FAIL_UPDATE_OBJ_NET 0x1500
2094 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1500
2095 $LFS mkdir -i $MDTIDX $remote_dir &
2097 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0
2099 fail mds${MDTIDX},mds$((MDTIDX + 1))
2101 wait $CLIENT_PID || error "remote creation failed"
2103 remote_dir_check_80 || error "remote dir check failed"
2104 rm -rf $DIR/$tdir || error "rmdir failed"
2108 run_test 80d "DNE: create remote dir, drop update rep from MDT1, fail 2 MDTs"
2111 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2112 ([ $FAILURE_MODE == "HARD" ] &&
2113 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2114 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2118 local remote_dir=$DIR/$tdir/remote_dir
2121 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2122 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2123 $LFS mkdir -i $MDTIDX $remote_dir &
2125 do_facet mds${MDTIDX} lctl set_param fail_loc=0
2129 wait $CLIENT_PID || error "remote creation failed"
2131 remote_dir_check_80 || error "remote dir check failed"
2132 rm -rf $DIR/$tdir || error "rmdir failed"
2136 run_test 80e "DNE: create remote dir, drop MDT0 rep, fail MDT0"
2139 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2140 ([ $FAILURE_MODE == "HARD" ] &&
2141 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2142 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2145 local remote_dir=$DIR/$tdir/remote_dir
2148 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2149 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2150 $LFS mkdir -i $MDTIDX $remote_dir &
2152 do_facet mds${MDTIDX} lctl set_param fail_loc=0
2154 fail mds$((MDTIDX + 1))
2156 wait $CLIENT_PID || error "remote creation failed"
2158 remote_dir_check_80 || error "remote dir check failed"
2159 rm -rf $DIR/$tdir || error "rmdir failed"
2163 run_test 80f "DNE: create remote dir, drop MDT0 rep, fail MDT1"
2166 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2167 ([ $FAILURE_MODE == "HARD" ] &&
2168 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2169 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2173 local remote_dir=$DIR/$tdir/remote_dir
2176 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2177 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2178 $LFS mkdir -i $MDTIDX $remote_dir &
2180 do_facet mds${MDTIDX} lctl set_param fail_loc=0
2183 fail mds$((MDTIDX + 1))
2185 wait $CLIENT_PID || error "remote creation failed"
2187 remote_dir_check_80 || error "remote dir check failed"
2188 rm -rf $DIR/$tdir || error "rmdir failed"
2192 run_test 80g "DNE: create remote dir, drop MDT0 rep, fail MDT0, then MDT1"
2195 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2197 local remote_dir=$DIR/$tdir/remote_dir
2200 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2201 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2202 $LFS mkdir -i $MDTIDX $remote_dir &
2204 do_facet mds${MDTIDX} lctl set_param fail_loc=0
2206 fail mds${MDTIDX},mds$((MDTIDX + 1))
2208 wait $CLIENT_PID || return 1
2210 remote_dir_check_80 || error "remote dir check failed"
2211 rm -rf $DIR/$tdir || error "rmdir failed"
2215 run_test 80h "DNE: create remote dir, drop MDT0 rep, fail 2 MDTs"
2218 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2219 ([ $FAILURE_MODE == "HARD" ] &&
2220 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2221 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2225 local remote_dir=$DIR/$tdir/remote_dir
2228 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2230 # OBD_FAIL_UPDATE_OBJ_NET 0x1500
2231 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1500
2234 do_facet mds${MDTIDX} lctl set_param fail_loc=0
2236 fail mds$((MDTIDX + 1))
2238 wait $CLIENT_PID || error "rm remote dir failed"
2240 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2242 rm -rf $DIR/$tdir || error "rmdir failed"
2246 run_test 81a "DNE: unlink remote dir, drop MDT0 update rep, fail MDT1"
2249 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2250 ([ $FAILURE_MODE == "HARD" ] &&
2251 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2252 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2255 local remote_dir=$DIR/$tdir/remote_dir
2258 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2260 # OBD_FAIL_UPDATE_OBJ_NET 0x1500
2261 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1500
2264 do_facet mds${MDTIDX} lctl set_param fail_loc=0
2268 wait $CLIENT_PID || error "rm remote dir failed"
2270 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2272 rm -rf $DIR/$tdir || error "rmdir failed"
2276 run_test 81b "DNE: unlink remote dir, drop MDT0 update reply, fail MDT0"
2279 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2280 ([ $FAILURE_MODE == "HARD" ] &&
2281 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2282 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2286 local remote_dir=$DIR/$tdir/remote_dir
2289 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2291 # OBD_FAIL_UPDATE_OBJ_NET 0x1500
2292 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1500
2295 do_facet mds${MDTIDX} lctl set_param fail_loc=0
2298 fail mds$((MDTIDX + 1))
2300 wait $CLIENT_PID || error "rm remote dir failed"
2302 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2304 rm -rf $DIR/$tdir || error "rmdir failed"
2308 run_test 81c "DNE: unlink remote dir, drop MDT0 update reply, fail MDT0,MDT1"
2311 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2313 local remote_dir=$DIR/$tdir/remote_dir
2316 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2318 # OBD_FAIL_UPDATE_OBJ_NET 0x1500
2319 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1500
2322 do_facet mds${MDTIDX} lctl set_param fail_loc=0
2324 fail mds${MDTIDX},mds$((MDTIDX + 1))
2326 wait $CLIENT_PID || error "rm remote dir failed"
2328 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2330 rm -rf $DIR/$tdir || error "rmdir failed"
2334 run_test 81d "DNE: unlink remote dir, drop MDT0 update reply, fail 2 MDTs"
2337 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2338 ([ $FAILURE_MODE == "HARD" ] &&
2339 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2340 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2344 local remote_dir=$DIR/$tdir/remote_dir
2347 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2349 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2350 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2353 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0
2357 wait $CLIENT_PID || error "rm remote dir failed"
2359 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2361 rm -rf $DIR/$tdir || error "rmdir failed"
2365 run_test 81e "DNE: unlink remote dir, drop MDT1 req reply, fail MDT0"
2368 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2369 ([ $FAILURE_MODE == "HARD" ] &&
2370 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2371 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2375 local remote_dir=$DIR/$tdir/remote_dir
2378 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2380 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2381 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2384 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0
2386 fail mds$((MDTIDX + 1))
2388 wait $CLIENT_PID || error "rm remote dir failed"
2390 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2392 rm -rf $DIR/$tdir || error "rmdir failed"
2396 run_test 81f "DNE: unlink remote dir, drop MDT1 req reply, fail MDT1"
2399 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2400 ([ $FAILURE_MODE == "HARD" ] &&
2401 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2402 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2406 local remote_dir=$DIR/$tdir/remote_dir
2409 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2411 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2412 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2415 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0
2418 fail mds$((MDTIDX + 1))
2420 wait $CLIENT_PID || error "rm remote dir failed"
2422 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2424 rm -rf $DIR/$tdir || error "rmdir failed"
2428 run_test 81g "DNE: unlink remote dir, drop req reply, fail M0, then M1"
2431 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2433 local remote_dir=$DIR/$tdir/remote_dir
2436 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2438 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2439 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2442 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0
2444 fail mds${MDTIDX},mds$((MDTIDX + 1))
2446 wait $CLIENT_PID || error "rm remote dir failed"
2448 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2450 rm -rf $DIR/$tdir || error "rmdir failed"
2454 run_test 81h "DNE: unlink remote dir, drop request reply, fail 2 MDTs"
2458 createmany -o $DIR/$tdir/$tfile- 10 || return 1
2459 #define OBD_FAIL_MDS_FAIL_LOV_LOG_ADD 0x140
2460 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000140"
2461 unlinkmany $DIR/$tdir/$tfile- 10 || return 2
2463 run_test 83a "fail log_add during unlink recovery"
2467 createmany -o $DIR/$tdir/$tfile- 10 || return 1
2468 replay_barrier $SINGLEMDS
2469 unlinkmany $DIR/$tdir/$tfile- 10 || return 2
2470 #define OBD_FAIL_MDS_FAIL_LOV_LOG_ADD 0x140
2471 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000140"
2474 run_test 83b "fail log_add during unlink recovery"
2477 #define OBD_FAIL_MDS_OPEN_WAIT_CREATE 0x144
2478 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000144"
2479 createmany -o $DIR/$tfile- 1 &
2483 client_up || client_up || true # reconnect
2485 run_test 84a "stale open during export disconnect"
2487 test_85a() { #bug 16774
2488 lctl set_param -n ldlm.cancel_unused_locks_before_replay "1"
2490 for i in `seq 100`; do
2491 echo "tag-$i" > $DIR/$tfile-$i
2492 grep -q "tag-$i" $DIR/$tfile-$i || error "f2-$i"
2495 lov_id=`lctl dl | grep "clilov"`
2496 addr=`echo $lov_id | awk '{print $4}' | awk -F '-' '{print $3}'`
2497 count=`lctl get_param -n ldlm.namespaces.*MDT0000*$addr.lock_unused_count`
2498 echo "before recovery: unused locks count = $count"
2502 count2=`lctl get_param -n ldlm.namespaces.*MDT0000*$addr.lock_unused_count`
2503 echo "after recovery: unused locks count = $count2"
2505 if [ $count2 -ge $count ]; then
2506 error "unused locks are not canceled"
2509 run_test 85a "check the cancellation of unused locks during recovery(IBITS)"
2511 test_85b() { #bug 16774
2512 lctl set_param -n ldlm.cancel_unused_locks_before_replay "1"
2514 do_facet mgs $LCTL pool_new $FSNAME.$TESTNAME || return 1
2515 do_facet mgs $LCTL pool_add $FSNAME.$TESTNAME $FSNAME-OST0000 || return 2
2517 $SETSTRIPE -c 1 -p $FSNAME.$TESTNAME $DIR
2519 for i in `seq 100`; do
2520 dd if=/dev/urandom of=$DIR/$tfile-$i bs=4096 count=32 >/dev/null 2>&1
2523 cancel_lru_locks osc
2525 for i in `seq 100`; do
2526 dd if=$DIR/$tfile-$i of=/dev/null bs=4096 count=32 >/dev/null 2>&1
2529 lov_id=`lctl dl | grep "clilov"`
2530 addr=`echo $lov_id | awk '{print $4}' | awk -F '-' '{print $3}'`
2531 count=`lctl get_param -n ldlm.namespaces.*OST0000*$addr.lock_unused_count`
2532 echo "before recovery: unused locks count = $count"
2533 [ $count != 0 ] || return 3
2537 count2=`lctl get_param -n ldlm.namespaces.*OST0000*$addr.lock_unused_count`
2538 echo "after recovery: unused locks count = $count2"
2540 do_facet mgs $LCTL pool_remove $FSNAME.$TESTNAME $FSNAME-OST0000 || return 4
2541 do_facet mgs $LCTL pool_destroy $FSNAME.$TESTNAME || return 5
2543 if [ $count2 -ge $count ]; then
2544 error "unused locks are not canceled"
2547 run_test 85b "check the cancellation of unused locks during recovery(EXTENT)"
2550 local clients=${CLIENTS:-$HOSTNAME}
2552 zconf_umount_clients $clients $MOUNT
2553 do_facet $SINGLEMDS lctl set_param mdt.${FSNAME}-MDT*.exports.clear=0
2554 remount_facet $SINGLEMDS
2555 zconf_mount_clients $clients $MOUNT
2557 run_test 86 "umount server after clear nid_stats should not hit LBUG"
2560 do_facet ost1 "lctl set_param -n obdfilter.${ost1_svc}.sync_journal 0"
2563 $SETSTRIPE -i 0 -c 1 $DIR/$tfile
2564 dd if=/dev/urandom of=$DIR/$tfile bs=1024k count=8 || error "Cannot write"
2565 cksum=`md5sum $DIR/$tfile | awk '{print $1}'`
2566 cancel_lru_locks osc
2568 dd if=$DIR/$tfile of=/dev/null bs=1024k count=8 || error "Cannot read"
2569 cksum2=`md5sum $DIR/$tfile | awk '{print $1}'`
2570 if [ $cksum != $cksum2 ] ; then
2571 error "New checksum $cksum2 does not match original $cksum"
2574 run_test 87 "write replay"
2577 do_facet ost1 "lctl set_param -n obdfilter.${ost1_svc}.sync_journal 0"
2580 $SETSTRIPE -i 0 -c 1 $DIR/$tfile
2581 dd if=/dev/urandom of=$DIR/$tfile bs=1024k count=8 || error "Cannot write"
2582 sleep 1 # Give it a chance to flush dirty data
2583 echo TESTTEST | dd of=$DIR/$tfile bs=1 count=8 seek=64
2584 cksum=`md5sum $DIR/$tfile | awk '{print $1}'`
2585 cancel_lru_locks osc
2587 dd if=$DIR/$tfile of=/dev/null bs=1024k count=8 || error "Cannot read"
2588 cksum2=`md5sum $DIR/$tfile | awk '{print $1}'`
2589 if [ $cksum != $cksum2 ] ; then
2590 error "New checksum $cksum2 does not match original $cksum"
2593 run_test 87b "write replay with changed data (checksum resend)"
2595 test_88() { #bug 17485
2599 $SETSTRIPE -i 0 -c 1 $DIR/$tdir || error "$SETSTRIPE"
2602 replay_barrier $SINGLEMDS
2604 # exhaust precreations on ost1
2605 local OST=$(ostname_from_index 0)
2606 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $OST)
2607 local last_id=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_last_id)
2608 local next_id=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_next_id)
2609 echo "before test: last_id = $last_id, next_id = $next_id"
2611 echo "Creating to objid $last_id on ost $OST..."
2612 createmany -o $DIR/$tdir/f-%d $next_id $((last_id - next_id + 2))
2614 #create some files to use some uncommitted objids
2615 last_id=$(($last_id + 1))
2616 createmany -o $DIR/$tdir/f-%d $last_id 8
2618 last_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_last_id)
2619 next_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_next_id)
2620 echo "before recovery: last_id = $last_id2, next_id = $next_id2"
2622 # if test uses shutdown_facet && reboot_facet instead of facet_failover ()
2623 # it has to take care about the affected facets, bug20407
2624 local affected_mds1=$(affected_facets mds1)
2625 local affected_ost1=$(affected_facets ost1)
2627 shutdown_facet $SINGLEMDS
2630 reboot_facet $SINGLEMDS
2631 change_active $affected_mds1
2632 wait_for_facet $affected_mds1
2633 mount_facets $affected_mds1 || error "Restart of mds failed"
2636 change_active $affected_ost1
2637 wait_for_facet $affected_ost1
2638 mount_facets $affected_ost1 || error "Restart of ost1 failed"
2642 last_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_last_id)
2643 next_id2=$(do_facet $SINGLEMDS lctl get_param -n osc.$mdtosc.prealloc_next_id)
2644 echo "after recovery: last_id = $last_id2, next_id = $next_id2"
2646 # create new files, which should use new objids, and ensure the orphan
2647 # cleanup phase for ost1 is completed at the same time
2648 for i in `seq 8`; do
2649 file_id=$(($last_id + 10 + $i))
2650 dd if=/dev/urandom of=$DIR/$tdir/f-$file_id bs=4096 count=128
2653 # if the objids were not recreated, then "ls" will failed for -ENOENT
2654 ls -l $DIR/$tdir/* || error "can't get the status of precreated files"
2657 # write into previously created files
2658 for i in `seq 8`; do
2659 file_id=$(($last_id + $i))
2660 dd if=/dev/urandom of=$DIR/$tdir/f-$file_id bs=4096 count=128
2661 cp -f $DIR/$tdir/f-$file_id $TMP/$tdir/
2664 # compare the content
2665 for i in `seq 8`; do
2666 file_id=$(($last_id + $i))
2667 cmp $TMP/$tdir/f-$file_id $DIR/$tdir/f-$file_id || error "the content" \
2668 "of file is modified!"
2673 run_test 88 "MDS should not assign same objid to different files "
2676 cancel_lru_locks osc
2678 rm -f $DIR/$tdir/$tfile
2680 wait_delete_completed
2681 BLOCKS1=$(df -P $MOUNT | tail -n 1 | awk '{ print $3 }')
2682 $SETSTRIPE -i 0 -c 1 $DIR/$tdir/$tfile
2683 dd if=/dev/zero bs=1M count=10 of=$DIR/$tdir/$tfile
2686 facet_failover $SINGLEMDS
2687 rm $DIR/$tdir/$tfile
2690 zconf_mount $(hostname) $MOUNT
2691 client_up || return 1
2693 wait_delete_completed
2694 BLOCKS2=$(df -P $MOUNT | tail -n 1 | awk '{ print $3 }')
2695 [ $((BLOCKS2 - BLOCKS1)) -le 4 ] || \
2696 error $((BLOCKS2 - BLOCKS1)) blocks leaked
2699 run_test 89 "no disk space leak on late ost connection"
2705 change_active $facet
2706 wait_for_facet $facet
2707 mount_facet $facet || error "Restart of $facet failed"
2711 test_90() { # bug 19494
2712 local dir=$DIR/$tdir
2713 local ostfail=$(get_random_entry $(get_facets OST))
2715 if [[ $FAILURE_MODE = HARD ]]; then
2716 local affected=$(affected_facets $ostfail);
2717 if [[ "$affected" != $ostfail ]]; then
2718 skip not functional with FAILURE_MODE=$FAILURE_MODE, affected: $affected
2725 echo "Create the files"
2727 # file "f${index}" striped over 1 OST
2728 # file "all" striped over all OSTs
2730 $SETSTRIPE -c $OSTCOUNT $dir/all ||
2731 error "setstripe failed to create $dir/all"
2733 for (( i=0; i<$OSTCOUNT; i++ )); do
2735 $SETSTRIPE -i $i -c 1 $f || error "$SETSTRIPE failed to create $f"
2737 # confirm setstripe actually created the stripe on the requested OST
2738 local uuid=$(ostuuid_from_index $i)
2739 for file in f$i all; do
2740 if [[ $dir/$file != $($LFS find --obd $uuid --name $file $dir) ]]; then
2741 $GETSTRIPE $dir/file
2742 error wrong stripe: $file, uuid: $uuid
2747 # Before failing an OST, get its obd name and index
2748 local varsvc=${ostfail}_svc
2749 local obd=$(do_facet $ostfail lctl get_param -n obdfilter.${!varsvc}.uuid)
2750 local index=${obd:(-6):1}
2752 echo "Fail $ostfail $obd, display the list of affected files"
2753 shutdown_facet $ostfail || return 2
2755 trap "cleanup_90 $ostfail" EXIT INT
2756 echo "General Query: lfs find $dir"
2757 local list=$($LFS find $dir)
2759 for (( i=0; i<$OSTCOUNT; i++ )); do
2760 list_member "$list" $dir/f$i || error_noexit "lfs find $dir: no file f$i"
2762 list_member "$list" $dir/all || error_noexit "lfs find $dir: no file all"
2764 # focus on the missing OST,
2765 # we expect to see only two files affected: "f$(index)" and "all"
2767 echo "Querying files on shutdown $ostfail: lfs find --obd $obd"
2768 list=$($LFS find --obd $obd $dir)
2770 for file in all f$index; do
2771 list_member "$list" $dir/$file ||
2772 error_noexit "lfs find does not report the affected $obd for $file"
2775 [[ $(echo $list | wc -w) -eq 2 ]] ||
2776 error_noexit "lfs find reports the wrong list of affected files ${#list[@]}"
2778 echo "Check getstripe: $GETSTRIPE -r --obd $obd"
2779 list=$($GETSTRIPE -r --obd $obd $dir)
2781 for file in all f$index; do
2782 echo "$list" | grep $dir/$file ||
2783 error_noexit "lfs getsripe does not report the affected $obd for $file"
2788 run_test 90 "lfs find identifies the missing striped file segments"
2791 check_and_cleanup_lustre