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 $SINGLEMDS) = "zfs" ] &&
28 # bug number for skipped test: LU-951
29 ALWAYS_EXCEPT="$ALWAYS_EXCEPT 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 > $BEFOREUSED + $(fs_log_size) )) &&
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 # always need connection to MDS to verify layout during IO. LU-2628.
800 lctl get_param mdc.*.connect_flags | grep -q layout_lock &&
801 skip "layout_lock needs MDS connection for IO" && return 0
803 $LCTL mark multiop $MOUNT/$tfile OS_c
804 multiop $MOUNT/$tfile OS_c &
806 writeme -s $MOUNT/${tfile}-2 &
809 facet_failover $SINGLEMDS
810 #define OBD_FAIL_MDS_CONNECT_NET 0x117
811 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000117"
813 stat1=`count_ost_writes`
815 stat2=`count_ost_writes`
816 echo "$stat1, $stat2"
817 if [ $stat1 -lt $stat2 ]; then
818 echo "writes continuing during recovery"
821 echo "writes not continuing during recovery, bug 2477"
824 echo "waiting for writeme $WRITE_PID"
828 echo "waiting for multiop $PID"
829 wait $PID || return 2
830 do_facet client munlink $MOUNT/$tfile || return 3
831 do_facet client munlink $MOUNT/${tfile}-2 || return 3
834 run_test 40 "cause recovery in ptlrpc, ensure IO continues"
838 # make sure that a read to one osc doesn't try to double-unlock its page just
839 # because another osc is invalid. trigger_group_io used to mistakenly return
840 # an error if any oscs were invalid even after having successfully put rpcs
841 # on valid oscs. This was fatal if the caller was ll_readpage who unlocked
842 # the page, guarnateeing that the unlock from the RPC completion would
843 # assert on trying to unlock the unlocked page.
845 [ $OSTCOUNT -lt 2 ] &&
846 skip_env "skipping test 41: we don't have a second OST to test with" &&
849 local f=$MOUNT/$tfile
850 # make sure the start of the file is ost1
851 $SETSTRIPE -S $((128 * 1024)) -i 0 $f
852 do_facet client dd if=/dev/zero of=$f bs=4k count=1 || return 3
854 # fail ost2 and read from ost1
855 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $ost2_svc)
856 local osc2dev=$(do_facet $SINGLEMDS "lctl get_param -n devices" | \
857 grep $mdtosc | awk '{print $1}')
858 [ -z "$osc2dev" ] && echo "OST: $ost2_svc" && lctl get_param -n devices &&
860 do_facet $SINGLEMDS $LCTL --device $osc2dev deactivate || return 1
861 do_facet client dd if=$f of=/dev/null bs=4k count=1 || return 3
862 do_facet $SINGLEMDS $LCTL --device $osc2dev activate || return 2
865 run_test 41 "read from a valid osc while other oscs are invalid"
867 # test MDS recovery after ost failure
869 blocks=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
870 createmany -o $DIR/$tfile-%d 800
872 unlinkmany $DIR/$tfile-%d 0 400
874 lctl set_param debug=-1
877 # osc is evicted, fs is smaller (but only with failout OSTs (bug 7287)
878 #blocks_after=`df -P $MOUNT | tail -n 1 | awk '{ print $2 }'`
879 #[ $blocks_after -lt $blocks ] || return 1
880 echo wait for MDS to timeout and recover
881 sleep $((TIMEOUT * 2))
883 unlinkmany $DIR/$tfile-%d 400 400
884 $CHECKSTAT -t file $DIR/$tfile-* && return 2 || true
886 run_test 42 "recovery after ost failure"
888 # timeout in MDS/OST recovery RPC will LBUG MDS
889 test_43() { # bug 2530
890 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
892 replay_barrier $SINGLEMDS
894 # OBD_FAIL_OST_CREATE_NET 0x204
895 do_facet ost1 "lctl set_param fail_loc=0x80000204"
898 do_facet ost1 "lctl set_param fail_loc=0"
902 run_test 43 "mds osc import failure during recovery; don't LBUG"
904 test_44a() { # was test_44
907 local mdcdev=$($LCTL get_param -n devices |
908 awk "/ ${FSNAME}-MDT0000-mdc-/ {print \$1}")
909 [ "$mdcdev" ] || return 2
910 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
911 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
913 # adaptive timeouts slow this way down
914 if at_is_enabled; then
915 at_max_saved=$(at_max_get mds)
919 for i in `seq 1 10`; do
920 echo "$i of 10 ($(date +%s))"
921 do_facet $SINGLEMDS "lctl get_param -n mdt.*.mdt.timeouts | grep service"
922 #define OBD_FAIL_TGT_CONN_RACE 0x701
923 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000701"
924 # lctl below may fail, it is valid case
925 $LCTL --device $mdcdev recover
928 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
929 [ $at_max_saved -ne 0 ] && at_max_set $at_max_saved mds
932 run_test 44a "race in target handle connect"
935 local mdcdev=$($LCTL get_param -n devices |
936 awk "/ ${FSNAME}-MDT0000-mdc-/ {print \$1}")
937 [ "$mdcdev" ] || return 2
938 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
939 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
941 for i in `seq 1 10`; do
942 echo "$i of 10 ($(date +%s))"
943 do_facet $SINGLEMDS "lctl get_param -n mdt.*.mdt.timeouts | grep service"
944 #define OBD_FAIL_TGT_DELAY_RECONNECT 0x704
945 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000704"
946 # lctl below may fail, it is valid case
947 $LCTL --device $mdcdev recover
950 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
953 run_test 44b "race in target handle connect"
956 replay_barrier $SINGLEMDS
957 createmany -m $DIR/$tfile-%d 100 || error "failed to create directories"
958 #define OBD_FAIL_TGT_RCVG_FLAG 0x712
959 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000712"
960 fail_abort $SINGLEMDS
961 unlinkmany $DIR/$tfile-%d 100 && error "unliked after fail abort"
963 unlinkmany $DIR/$tfile-%d 100 && error "unliked after fail"
966 run_test 44c "race in target handle connect"
968 # Handle failed close
970 local mdcdev=$($LCTL get_param -n devices |
971 awk "/ ${FSNAME}-MDT0000-mdc-/ {print \$1}")
972 [ "$mdcdev" ] || return 2
973 [ $(echo $mdcdev | wc -w) -eq 1 ] ||
974 { echo mdcdev=$mdcdev; $LCTL dl; return 3; }
976 $LCTL --device $mdcdev recover || return 6
978 multiop_bg_pause $DIR/$tfile O_c || return 1
981 # This will cause the CLOSE to fail before even
982 # allocating a reply buffer
983 $LCTL --device $mdcdev deactivate || return 4
987 wait $pid || return 1
989 $LCTL --device $mdcdev activate || return 5
992 $CHECKSTAT -t file $DIR/$tfile || return 2
995 run_test 45 "Handle failed close"
999 drop_reply "touch $DIR/$tfile"
1001 # ironically, the previous test, 45, will cause a real forced close,
1002 # so just look for one for this test
1003 dmesg | grep -i "force closing client file handle for $tfile" && return 1
1006 run_test 46 "Don't leak file handle after open resend (3325)"
1008 test_47() { # bug 2824
1009 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1011 # create some files to make sure precreate has been done on all
1012 # OSTs. (just in case this test is run independently)
1013 createmany -o $DIR/$tfile 20 || return 1
1015 # OBD_FAIL_OST_CREATE_NET 0x204
1017 do_facet ost1 "lctl set_param fail_loc=0x80000204"
1018 client_up || return 2
1020 # let the MDS discover the OST failure, attempt to recover, fail
1021 # and recover again.
1022 sleep $((3 * TIMEOUT))
1024 # Without 2824, this createmany would hang
1025 createmany -o $DIR/$tfile 20 || return 3
1026 unlinkmany $DIR/$tfile 20 || return 4
1028 do_facet ost1 "lctl set_param fail_loc=0"
1031 run_test 47 "MDS->OSC failure during precreate cleanup (2824)"
1034 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1035 [ "$OSTCOUNT" -lt "2" ] && skip_env "$OSTCOUNT < 2 OSTs -- skipping" && return
1037 replay_barrier $SINGLEMDS
1038 createmany -o $DIR/$tfile 20 || return 1
1039 # OBD_FAIL_OST_EROFS 0x216
1040 facet_failover $SINGLEMDS
1041 do_facet ost1 "lctl set_param fail_loc=0x80000216"
1042 client_up || return 2
1044 createmany -o $DIR/$tfile 20 20 || return 2
1045 unlinkmany $DIR/$tfile 40 || return 3
1048 run_test 48 "MDS->OSC failure during precreate cleanup (2824)"
1051 local mdtosc=$(get_mdtosc_proc_path $SINGLEMDS $ost1_svc)
1052 local oscdev=$(do_facet $SINGLEMDS "lctl get_param -n devices" | \
1053 grep $mdtosc | awk '{print $1}')
1054 [ "$oscdev" ] || return 1
1055 do_facet $SINGLEMDS $LCTL --device $oscdev recover || return 2
1056 do_facet $SINGLEMDS $LCTL --device $oscdev recover || return 3
1057 # give the mds_lov_sync threads a chance to run
1060 run_test 50 "Double OSC recovery, don't LASSERT (3812)"
1062 # b3764 timed out lock replay
1065 cancel_lru_locks mdc
1067 multiop $DIR/$tfile s || return 1
1068 replay_barrier $SINGLEMDS
1069 #define OBD_FAIL_LDLM_REPLY 0x30c
1070 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000030c"
1071 fail $SINGLEMDS || return 2
1072 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1074 $CHECKSTAT -t file $DIR/$tfile-* && return 3 || true
1076 run_test 52 "time out lock replay (3764)"
1078 # bug 3462 - simultaneous MDC requests
1080 cancel_lru_locks mdc # cleanup locks from former test cases
1081 mkdir -p $DIR/${tdir}-1
1082 mkdir -p $DIR/${tdir}-2
1083 multiop $DIR/${tdir}-1/f O_c &
1085 # give multiop a change to open
1088 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1089 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1090 kill -USR1 $close_pid
1091 cancel_lru_locks mdc # force the close
1092 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1094 mcreate $DIR/${tdir}-2/f || return 1
1096 # close should still be here
1097 [ -d /proc/$close_pid ] || return 2
1099 replay_barrier_nodf $SINGLEMDS
1101 wait $close_pid || return 3
1103 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1104 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1105 rm -rf $DIR/${tdir}-*
1107 run_test 53a "|X| close request while two MDC requests in flight"
1110 cancel_lru_locks mdc # cleanup locks from former test cases
1111 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1113 mkdir -p $DIR/${tdir}-1
1114 mkdir -p $DIR/${tdir}-2
1115 multiop_bg_pause $DIR/${tdir}-1/f O_c || return 6
1118 #define OBD_FAIL_MDS_REINT_NET 0x107
1119 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1120 mcreate $DIR/${tdir}-2/f &
1124 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1125 kill -USR1 $close_pid
1126 cancel_lru_locks mdc # force the close
1127 wait $close_pid || return 1
1128 # open should still be here
1129 [ -d /proc/$open_pid ] || return 2
1131 replay_barrier_nodf $SINGLEMDS
1133 wait $open_pid || return 3
1135 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1136 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1137 rm -rf $DIR/${tdir}-*
1139 run_test 53b "|X| open request while two MDC requests in flight"
1142 cancel_lru_locks mdc # cleanup locks from former test cases
1143 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1145 mkdir -p $DIR/${tdir}-1
1146 mkdir -p $DIR/${tdir}-2
1147 multiop $DIR/${tdir}-1/f O_c &
1150 #define OBD_FAIL_MDS_REINT_NET 0x107
1151 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1152 mcreate $DIR/${tdir}-2/f &
1156 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1157 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1158 kill -USR1 $close_pid
1159 cancel_lru_locks mdc # force the close
1161 #bz20647: make sure all pids are exists before failover
1162 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1163 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1164 replay_barrier_nodf $SINGLEMDS
1165 fail_nodf $SINGLEMDS
1166 wait $open_pid || return 1
1168 # close should be gone
1169 [ -d /proc/$close_pid ] && return 2
1170 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1172 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1173 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1174 rm -rf $DIR/${tdir}-*
1176 run_test 53c "|X| open request and close request while two MDC requests in flight"
1179 cancel_lru_locks mdc # cleanup locks from former test cases
1180 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1182 mkdir -p $DIR/${tdir}-1
1183 mkdir -p $DIR/${tdir}-2
1184 multiop $DIR/${tdir}-1/f O_c &
1186 # give multiop a chance to open
1189 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1190 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1191 kill -USR1 $close_pid
1192 cancel_lru_locks mdc # force the close
1193 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1194 mcreate $DIR/${tdir}-2/f || return 1
1196 # close should still be here
1197 [ -d /proc/$close_pid ] || return 2
1199 wait $close_pid || return 3
1201 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1202 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1203 rm -rf $DIR/${tdir}-*
1205 run_test 53d "|X| close reply while two MDC requests in flight"
1208 cancel_lru_locks mdc # cleanup locks from former test cases
1209 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1211 mkdir -p $DIR/${tdir}-1
1212 mkdir -p $DIR/${tdir}-2
1213 multiop $DIR/${tdir}-1/f O_c &
1216 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1217 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1218 mcreate $DIR/${tdir}-2/f &
1222 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1223 kill -USR1 $close_pid
1224 cancel_lru_locks mdc # force the close
1225 wait $close_pid || return 1
1226 # open should still be here
1227 [ -d /proc/$open_pid ] || return 2
1229 replay_barrier_nodf $SINGLEMDS
1231 wait $open_pid || return 3
1233 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 4
1234 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 5
1235 rm -rf $DIR/${tdir}-*
1237 run_test 53e "|X| open reply while two MDC requests in flight"
1240 cancel_lru_locks mdc # cleanup locks from former test cases
1241 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1243 mkdir -p $DIR/${tdir}-1
1244 mkdir -p $DIR/${tdir}-2
1245 multiop $DIR/${tdir}-1/f O_c &
1248 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1249 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1250 mcreate $DIR/${tdir}-2/f &
1254 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1255 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1256 kill -USR1 $close_pid
1257 cancel_lru_locks mdc # force the close
1259 #bz20647: make sure all pids are exists before failover
1260 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1261 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1262 replay_barrier_nodf $SINGLEMDS
1263 fail_nodf $SINGLEMDS
1264 wait $open_pid || return 1
1266 # close should be gone
1267 [ -d /proc/$close_pid ] && return 2
1268 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1270 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1271 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1272 rm -rf $DIR/${tdir}-*
1274 run_test 53f "|X| open reply and close reply while two MDC requests in flight"
1277 cancel_lru_locks mdc # cleanup locks from former test cases
1278 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1280 mkdir -p $DIR/${tdir}-1
1281 mkdir -p $DIR/${tdir}-2
1282 multiop $DIR/${tdir}-1/f O_c &
1285 #define OBD_FAIL_MDS_REINT_NET_REP 0x119
1286 do_facet $SINGLEMDS "lctl set_param fail_loc=0x119"
1287 mcreate $DIR/${tdir}-2/f &
1291 #define OBD_FAIL_MDS_CLOSE_NET 0x115
1292 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000115"
1293 kill -USR1 $close_pid
1294 cancel_lru_locks mdc # force the close
1295 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1297 #bz20647: make sure all pids are exists before failover
1298 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1299 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1300 replay_barrier_nodf $SINGLEMDS
1301 fail_nodf $SINGLEMDS
1302 wait $open_pid || return 1
1304 # close should be gone
1305 [ -d /proc/$close_pid ] && return 2
1307 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1308 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1309 rm -rf $DIR/${tdir}-*
1311 run_test 53g "|X| drop open reply and close request while close and open are both in flight"
1314 cancel_lru_locks mdc # cleanup locks from former test cases
1315 rm -rf $DIR/${tdir}-1 $DIR/${tdir}-2
1317 mkdir -p $DIR/${tdir}-1
1318 mkdir -p $DIR/${tdir}-2
1319 multiop $DIR/${tdir}-1/f O_c &
1322 #define OBD_FAIL_MDS_REINT_NET 0x107
1323 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000107"
1324 mcreate $DIR/${tdir}-2/f &
1328 #define OBD_FAIL_MDS_CLOSE_NET_REP 0x13b
1329 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013b"
1330 kill -USR1 $close_pid
1331 cancel_lru_locks mdc # force the close
1334 #bz20647: make sure all pids are exists before failover
1335 [ -d /proc/$close_pid ] || error "close_pid doesn't exist"
1336 [ -d /proc/$open_pid ] || error "open_pid doesn't exists"
1337 replay_barrier_nodf $SINGLEMDS
1338 fail_nodf $SINGLEMDS
1339 wait $open_pid || return 1
1341 # close should be gone
1342 [ -d /proc/$close_pid ] && return 2
1343 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1345 $CHECKSTAT -t file $DIR/${tdir}-1/f || return 3
1346 $CHECKSTAT -t file $DIR/${tdir}-2/f || return 4
1347 rm -rf $DIR/${tdir}-*
1349 run_test 53h "|X| open request and close reply while two MDC requests in flight"
1351 #b_cray 54 "|X| open request and close reply while two MDC requests in flight"
1353 #b3761 ASSERTION(hash != 0) failed
1355 # OBD_FAIL_MDS_OPEN_CREATE | OBD_FAIL_ONCE
1356 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012b"
1358 # give touch a chance to run
1360 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1364 run_test 55 "let MDS_CHECK_RESENT return the original return code instead of 0"
1366 #b3440 ASSERTION(rec->ur_fid2->id) failed
1368 ln -s foo $DIR/$tfile
1369 replay_barrier $SINGLEMDS
1370 #drop_reply "cat $DIR/$tfile"
1374 run_test 56 "don't replay a symlink open request (3440)"
1376 #recovery one mds-ost setattr from llog
1378 #define OBD_FAIL_MDS_OST_SETATTR 0x12c
1379 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012c"
1381 replay_barrier $SINGLEMDS
1384 $CHECKSTAT -t file $DIR/$tfile || return 1
1385 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1388 run_test 57 "test recovery from llog for setattr op"
1390 #recovery many mds-ost setattr from llog
1393 #define OBD_FAIL_MDS_OST_SETATTR 0x12c
1394 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000012c"
1395 createmany -o $DIR/$tdir/$tfile-%d 2500
1396 replay_barrier $SINGLEMDS
1399 $CHECKSTAT -t file $DIR/$tdir/$tfile-* >/dev/null || return 1
1400 do_facet $SINGLEMDS "lctl set_param fail_loc=0x0"
1401 unlinkmany $DIR/$tdir/$tfile-%d 2500
1404 run_test 58a "test recovery from llog for setattr op (test llog_gen_rec)"
1410 large_xattr_enabled &&
1411 orig="$(generate_string $(max_xattr_size))" || orig="bar"
1413 mount_client $MOUNT2
1415 touch $DIR/$tdir/$tfile
1416 replay_barrier $SINGLEMDS
1417 setfattr -n trusted.foo -v $orig $DIR/$tdir/$tfile
1419 new=$(get_xattr_value trusted.foo $MOUNT2/$tdir/$tfile)
1420 [[ "$new" = "$orig" ]] || return 1
1421 rm -f $DIR/$tdir/$tfile
1423 zconf_umount `hostname` $MOUNT2
1425 run_test 58b "test replay of setxattr op"
1427 test_58c() { # bug 16570
1432 if large_xattr_enabled; then
1433 local xattr_size=$(max_xattr_size)
1434 orig="$(generate_string $((xattr_size / 2)))"
1435 orig1="$(generate_string $xattr_size)"
1441 mount_client $MOUNT2
1443 touch $DIR/$tdir/$tfile
1444 drop_request "setfattr -n trusted.foo -v $orig $DIR/$tdir/$tfile" ||
1446 new=$(get_xattr_value trusted.foo $MOUNT2/$tdir/$tfile)
1447 [[ "$new" = "$orig" ]] || return 2
1448 drop_reint_reply "setfattr -n trusted.foo1 -v $orig1 $DIR/$tdir/$tfile" ||
1450 new=$(get_xattr_value trusted.foo1 $MOUNT2/$tdir/$tfile)
1451 [[ "$new" = "$orig1" ]] || return 4
1452 rm -f $DIR/$tdir/$tfile
1454 zconf_umount $HOSTNAME $MOUNT2
1456 run_test 58c "resend/reconstruct setxattr op"
1458 # log_commit_thread vs filter_destroy race used to lead to import use after free
1461 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1464 createmany -o $DIR/$tdir/$tfile-%d 200
1466 unlinkmany $DIR/$tdir/$tfile-%d 200
1467 #define OBD_FAIL_PTLRPC_DELAY_RECOV 0x507
1468 do_facet ost1 "lctl set_param fail_loc=0x507"
1471 do_facet ost1 "lctl set_param fail_loc=0x0"
1475 run_test 59 "test log_commit_thread vs filter_destroy race"
1477 # race between add unlink llog vs cat log init in post_recovery (only for b1_6)
1478 # bug 12086: should no oops and No ctxt error for this test
1481 createmany -o $DIR/$tdir/$tfile-%d 200
1482 replay_barrier $SINGLEMDS
1483 unlinkmany $DIR/$tdir/$tfile-%d 0 100
1485 unlinkmany $DIR/$tdir/$tfile-%d 100 100
1486 local no_ctxt=`dmesg | grep "No ctxt"`
1487 [ -z "$no_ctxt" ] || error "ctxt is not initialized in recovery"
1489 run_test 60 "test llog post recovery init vs llog unlink"
1491 #test race llog recovery thread vs llog cleanup
1492 test_61a() { # was test_61
1493 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1496 createmany -o $DIR/$tdir/$tfile-%d 800
1498 # OBD_FAIL_OST_LLOG_RECOVERY_TIMEOUT 0x221
1499 unlinkmany $DIR/$tdir/$tfile-%d 800
1500 set_nodes_failloc "$(osts_nodes)" 0x80000221
1505 set_nodes_failloc "$(osts_nodes)" 0x0
1507 $CHECKSTAT -t file $DIR/$tdir/$tfile-* && return 1
1510 run_test 61a "test race llog recovery vs llog cleanup"
1512 #test race mds llog sync vs llog cleanup
1514 # OBD_FAIL_MDS_LLOG_SYNC_TIMEOUT 0x13a
1515 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000013a"
1516 facet_failover $SINGLEMDS
1519 do_facet client dd if=/dev/zero of=$DIR/$tfile bs=4k count=1 || return 1
1521 run_test 61b "test race mds llog sync vs llog cleanup"
1523 #test race cancel cookie cb vs llog cleanup
1525 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1527 # OBD_FAIL_OST_CANCEL_COOKIE_TIMEOUT 0x222
1529 set_nodes_failloc "$(osts_nodes)" 0x80000222
1533 set_nodes_failloc "$(osts_nodes)" 0x0
1535 run_test 61c "test race mds llog sync vs llog cleanup"
1537 test_61d() { # bug 16002 # bug 17466 # bug 22137
1538 # OBD_FAIL_OBD_LLOG_SETUP 0x605
1540 do_facet mgs "lctl set_param fail_loc=0x80000605"
1541 start mgs $MGSDEV $MGS_MOUNT_OPTS && error "mgs start should have failed"
1542 do_facet mgs "lctl set_param fail_loc=0"
1543 start mgs $MGSDEV $MGS_MOUNT_OPTS || error "cannot restart mgs"
1545 run_test 61d "error in llog_setup should cleanup the llog context correctly"
1547 test_62() { # Bug 15756 - don't mis-drop resent replay
1549 replay_barrier $SINGLEMDS
1550 createmany -o $DIR/$tdir/$tfile- 25
1551 #define OBD_FAIL_TGT_REPLAY_DROP 0x707
1552 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000707"
1554 do_facet $SINGLEMDS "lctl set_param fail_loc=0"
1555 unlinkmany $DIR/$tdir/$tfile- 25 || return 2
1558 run_test 62 "don't mis-drop resent replay"
1560 #Adaptive Timeouts (bug 3055)
1568 echo "Cleaning up AT ..."
1569 if [ -n "$ATOLDBASE" ]; then
1570 local at_history=$($LCTL get_param -n at_history)
1571 do_facet $SINGLEMDS "lctl set_param at_history=$at_history" || true
1572 do_facet ost1 "lctl set_param at_history=$at_history" || true
1575 if [ $AT_MAX_SET -ne 0 ]; then
1576 for facet in mds client ost; do
1577 var=AT_MAX_SAVE_${facet}
1578 echo restore AT on $facet to saved value ${!var}
1579 at_max_set ${!var} $facet
1580 at_new=$(at_max_get $facet)
1581 echo Restored AT value on $facet $at_new
1582 [ $at_new -eq ${!var} ] || \
1583 error "$facet : AT value was not restored SAVED ${!var} NEW $at_new"
1590 local at_max_new=600
1592 # Save at_max original values
1594 if [ $AT_MAX_SET -eq 0 ]; then
1595 # Suppose that all osts have the same at_max
1596 for facet in mds client ost; do
1597 eval AT_MAX_SAVE_${facet}=$(at_max_get $facet)
1601 for facet in mds client ost; do
1602 at_max=$(at_max_get $facet)
1603 if [ $at_max -ne $at_max_new ]; then
1604 echo "AT value on $facet is $at_max, set it by force temporarily to $at_max_new"
1605 at_max_set $at_max_new $facet
1610 if [ -z "$ATOLDBASE" ]; then
1611 ATOLDBASE=$(do_facet $SINGLEMDS "lctl get_param -n at_history")
1612 # speed up the timebase so we can check decreasing AT
1613 do_facet $SINGLEMDS "lctl set_param at_history=8" || true
1614 do_facet ost1 "lctl set_param at_history=8" || true
1616 # sleep for a while to cool down, should be > 8s and also allow
1617 # at least one ping to be sent. simply use TIMEOUT to be safe.
1622 test_65a() #bug 3055
1624 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1626 at_start || return 0
1627 $LCTL dk > /dev/null
1629 $LCTL set_param debug="other"
1630 # Slow down a request to the current service time, this is critical
1631 # because previous tests may have caused this value to increase.
1632 REQ_DELAY=`lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts |
1633 awk '/portal 12/ {print $5}'`
1634 REQ_DELAY=$((${REQ_DELAY} + ${REQ_DELAY} / 4 + 5))
1636 do_facet $SINGLEMDS lctl set_param fail_val=$((${REQ_DELAY} * 1000))
1637 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1638 do_facet $SINGLEMDS $LCTL set_param fail_loc=0x8000050a
1639 createmany -o $DIR/$tfile 10 > /dev/null
1640 unlinkmany $DIR/$tfile 10 > /dev/null
1641 # check for log message
1642 $LCTL dk | grep "Early reply #" || error "No early reply"
1644 # client should show REQ_DELAY estimates
1645 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep portal
1647 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep portal
1649 run_test 65a "AT: verify early replies"
1651 test_65b() #bug 3055
1653 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1655 at_start || return 0
1658 $LCTL set_param debug="other trace"
1659 $LCTL dk > /dev/null
1660 # Slow down a request to the current service time, this is critical
1661 # because previous tests may have caused this value to increase.
1662 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tfile
1663 multiop $DIR/$tfile Ow1yc
1664 REQ_DELAY=`lctl get_param -n osc.${FSNAME}-OST0000-osc-*.timeouts |
1665 awk '/portal 6/ {print $5}'`
1666 REQ_DELAY=$((${REQ_DELAY} + ${REQ_DELAY} / 4 + 5))
1668 do_facet ost1 lctl set_param fail_val=${REQ_DELAY}
1669 #define OBD_FAIL_OST_BRW_PAUSE_PACK 0x224
1670 do_facet ost1 $LCTL set_param fail_loc=0x224
1673 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tfile
1674 # force some real bulk transfer
1675 multiop $DIR/$tfile oO_CREAT:O_RDWR:O_SYNC:w4096c
1677 do_facet ost1 $LCTL set_param fail_loc=0
1678 # check for log message
1679 $LCTL dk | grep "Early reply #" || error "No early reply"
1681 # client should show REQ_DELAY estimates
1682 lctl get_param -n osc.${FSNAME}-OST0000-osc-*.timeouts | grep portal
1684 run_test 65b "AT: verify early replies on packed reply / bulk"
1686 test_66a() #bug 3055
1688 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1690 at_start || return 0
1691 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1692 # adjust 5s at a time so no early reply is sent (within deadline)
1693 do_facet $SINGLEMDS "$LCTL set_param fail_val=5000"
1694 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1695 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0x8000050a"
1696 createmany -o $DIR/$tfile 20 > /dev/null
1697 unlinkmany $DIR/$tfile 20 > /dev/null
1698 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1699 do_facet $SINGLEMDS "$LCTL set_param fail_val=10000"
1700 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0x8000050a"
1701 createmany -o $DIR/$tfile 20 > /dev/null
1702 unlinkmany $DIR/$tfile 20 > /dev/null
1703 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1704 do_facet $SINGLEMDS "$LCTL set_param fail_loc=0"
1706 createmany -o $DIR/$tfile 20 > /dev/null
1707 unlinkmany $DIR/$tfile 20 > /dev/null
1708 lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | grep "portal 12"
1709 CUR=$(lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | awk '/portal 12/ {print $5}')
1710 WORST=$(lctl get_param -n mdc.${FSNAME}-MDT0000-mdc-*.timeouts | awk '/portal 12/ {print $7}')
1711 echo "Current MDT timeout $CUR, worst $WORST"
1712 [ $CUR -lt $WORST ] || error "Current $CUR should be less than worst $WORST"
1714 run_test 66a "AT: verify MDT service time adjusts with no early replies"
1716 test_66b() #bug 3055
1718 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1720 at_start || return 0
1721 ORIG=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1722 awk '/network/ {print $4}')
1723 $LCTL set_param fail_val=$(($ORIG + 5))
1724 #define OBD_FAIL_PTLRPC_PAUSE_REP 0x50c
1725 $LCTL set_param fail_loc=0x50c
1726 ls $DIR/$tfile > /dev/null 2>&1
1727 $LCTL set_param fail_loc=0
1728 CUR=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1729 awk '/network/ {print $4}')
1730 WORST=$(lctl get_param -n mdc.${FSNAME}-MDT0000*.timeouts |
1731 awk '/network/ {print $6}')
1732 echo "network timeout orig $ORIG, cur $CUR, worst $WORST"
1733 [ $WORST -gt $ORIG ] ||
1734 error "Worst $WORST should be worse than orig $ORIG"
1736 run_test 66b "AT: verify net latency adjusts"
1738 test_67a() #bug 3055
1740 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1742 at_start || return 0
1743 CONN1=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1744 # sleeping threads may drive values above this
1745 do_facet ost1 "$LCTL set_param fail_val=400"
1746 #define OBD_FAIL_PTLRPC_PAUSE_REQ 0x50a
1747 do_facet ost1 "$LCTL set_param fail_loc=0x50a"
1748 createmany -o $DIR/$tfile 20 > /dev/null
1749 unlinkmany $DIR/$tfile 20 > /dev/null
1750 do_facet ost1 "$LCTL set_param fail_loc=0"
1751 CONN2=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1752 ATTEMPTS=$(($CONN2 - $CONN1))
1753 echo "$ATTEMPTS osc reconnect attempts on gradual slow"
1754 [ $ATTEMPTS -gt 0 ] && error_ignore 13721 "AT should have prevented reconnect"
1757 run_test 67a "AT: verify slow request processing doesn't induce reconnects"
1759 test_67b() #bug 3055
1761 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1763 at_start || return 0
1764 CONN1=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1766 # exhaust precreations on ost1
1767 local OST=$(ostname_from_index 0)
1768 local mdtosc=$(get_mdtosc_proc_path mds $OST)
1769 local last_id=$(do_facet $SINGLEMDS lctl get_param -n \
1770 osc.$mdtosc.prealloc_last_id)
1771 local next_id=$(do_facet $SINGLEMDS lctl get_param -n \
1772 osc.$mdtosc.prealloc_next_id)
1774 mkdir -p $DIR/$tdir/${OST}
1775 $SETSTRIPE -i 0 -c 1 $DIR/$tdir/${OST} || error "$SETSTRIPE"
1776 echo "Creating to objid $last_id on ost $OST..."
1777 #define OBD_FAIL_OST_PAUSE_CREATE 0x223
1778 do_facet ost1 "$LCTL set_param fail_val=20000"
1779 do_facet ost1 "$LCTL set_param fail_loc=0x80000223"
1780 createmany -o $DIR/$tdir/${OST}/f $next_id $((last_id - next_id + 2))
1783 do_facet ost1 "lctl get_param -n ost.OSS.ost_create.timeouts"
1785 CONN2=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1786 ATTEMPTS=$(($CONN2 - $CONN1))
1787 echo "$ATTEMPTS osc reconnect attempts on instant slow"
1788 # do it again; should not timeout
1789 do_facet ost1 "$LCTL set_param fail_loc=0x80000223"
1790 cp /etc/profile $DIR/$tfile || error "cp failed"
1791 do_facet ost1 "$LCTL set_param fail_loc=0"
1793 do_facet ost1 "lctl get_param -n ost.OSS.ost_create.timeouts"
1794 CONN3=$(lctl get_param -n osc.*.stats | awk '/_connect/ {total+=$2} END {print total}')
1795 ATTEMPTS=$(($CONN3 - $CONN2))
1796 echo "$ATTEMPTS osc reconnect attempts on 2nd slow"
1797 [ $ATTEMPTS -gt 0 ] && error "AT should have prevented reconnect"
1800 run_test 67b "AT: verify instant slowdown doesn't induce reconnects"
1802 test_68 () #bug 13813
1804 remote_ost_nodsh && skip "remote OST with nodsh" && return 0
1806 at_start || return 0
1807 local ldlm_enqueue_min=$(find /sys -name ldlm_enqueue_min)
1808 [ -z "$ldlm_enqueue_min" ] && skip "missing /sys/.../ldlm_enqueue_min" && return 0
1809 local ldlm_enqueue_min_r=$(do_facet ost1 "find /sys -name ldlm_enqueue_min")
1810 [ -z "$ldlm_enqueue_min_r" ] && skip "missing /sys/.../ldlm_enqueue_min in the ost1" && return 0
1811 local ENQ_MIN=$(cat $ldlm_enqueue_min)
1812 local ENQ_MIN_R=$(do_facet ost1 "cat $ldlm_enqueue_min_r")
1813 echo $TIMEOUT >> $ldlm_enqueue_min
1814 do_facet ost1 "echo $TIMEOUT >> $ldlm_enqueue_min_r"
1818 $SETSTRIPE --stripe-index=0 --count=1 $DIR/$tdir
1819 #define OBD_FAIL_LDLM_PAUSE_CANCEL 0x312
1820 $LCTL set_param fail_val=$(($TIMEOUT - 1))
1821 $LCTL set_param fail_loc=0x80000312
1822 cp /etc/profile $DIR/$tdir/${tfile}_1 || error "1st cp failed $?"
1823 $LCTL set_param fail_val=$((TIMEOUT * 5 / 4))
1824 $LCTL set_param fail_loc=0x80000312
1825 cp /etc/profile $DIR/$tdir/${tfile}_2 || error "2nd cp failed $?"
1826 $LCTL set_param fail_loc=0
1828 echo $ENQ_MIN >> $ldlm_enqueue_min
1829 do_facet ost1 "echo $ENQ_MIN_R >> $ldlm_enqueue_min_r"
1833 run_test 68 "AT: verify slowing locks"
1836 # end of AT tests includes above lines
1839 # start multi-client tests
1841 [ -z "$CLIENTS" ] && \
1842 { skip "Need two or more clients." && return; }
1843 [ $CLIENTCOUNT -lt 2 ] && \
1844 { skip "Need two or more clients, have $CLIENTCOUNT" && return; }
1846 echo "mount clients $CLIENTS ..."
1847 zconf_mount_clients $CLIENTS $MOUNT
1849 local clients=${CLIENTS//,/ }
1850 echo "Write/read files on $DIR ; clients $CLIENTS ... "
1851 for CLIENT in $clients; do
1852 do_node $CLIENT dd bs=1M count=10 if=/dev/zero \
1853 of=$DIR/${tfile}_${CLIENT} 2>/dev/null || \
1854 error "dd failed on $CLIENT"
1857 local prev_client=$(echo $clients | sed 's/^.* \(.\+\)$/\1/')
1858 for C in ${CLIENTS//,/ }; do
1859 do_node $prev_client dd if=$DIR/${tfile}_${C} of=/dev/null 2>/dev/null || \
1860 error "dd if=$DIR/${tfile}_${C} failed on $prev_client"
1866 run_test 70a "check multi client t-f"
1868 check_for_process () {
1873 killall_process $clients "$prog" -0
1876 killall_process () {
1877 local clients=${1:-$(hostname)}
1882 do_nodes $clients "killall $signal $name"
1886 local clients=${CLIENTS:-$HOSTNAME}
1888 zconf_mount_clients $clients $MOUNT
1891 [ "$SLOW" = "no" ] && duration=120
1892 # set duration to 900 because it takes some time to boot node
1893 [ "$FAILURE_MODE" = HARD ] && duration=900
1896 local start_ts=$(date +%s)
1897 local cmd="rundbench 1 -t $duration"
1899 do_nodesv $clients "set -x; MISSING_DBENCH_OK=$MISSING_DBENCH_OK \
1900 PATH=\$PATH:$LUSTRE/utils:$LUSTRE/tests/:$DBENCH_LIB \
1901 DBENCH_LIB=$DBENCH_LIB TESTSUITE=$TESTSUITE TESTNAME=$TESTNAME \
1902 MOUNT=$MOUNT DIR=$DIR/$tdir/\\\$(hostname) LCTL=$LCTL $cmd" &
1905 #LU-1897 wait for all dbench copies to start
1906 while ! check_for_process $clients dbench; do
1907 elapsed=$(($(date +%s) - start_ts))
1908 if [ $elapsed -gt $duration]; then
1909 killall_process $clients dbench
1910 error "dbench failed to start on $clients!"
1915 log "Started rundbench load pid=$pid ..."
1917 elapsed=$(($(date +%s) - start_ts))
1918 local num_failovers=0
1919 while [ $elapsed -lt $duration ]; do
1920 if ! check_for_process $clients dbench; then
1921 error_noexit "dbench stopped on some of $clients!"
1922 killall_process $clients dbench
1926 replay_barrier $SINGLEMDS
1927 sleep 1 # give clients a time to do operations
1928 # Increment the number of failovers
1929 num_failovers=$((num_failovers+1))
1930 log "$TESTNAME fail $SINGLEMDS $num_failovers times"
1932 elapsed=$(($(date +%s) - start_ts))
1935 wait $pid || error "rundbench load on $clients failed!"
1937 run_test 70b "mds recovery; $CLIENTCOUNT clients"
1938 # end multi-client tests
1941 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1945 replay_barrier $SINGLEMDS
1946 #define OBD_FAIL_LDLM_ENQUEUE_NET 0x302
1947 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000302"
1950 wait $pid || return 1
1951 [ -e $DIR/$tfile ] && return 2
1954 run_test 73a "open(O_CREAT), unlink, replay, reconnect before open replay , close"
1957 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1961 replay_barrier $SINGLEMDS
1962 #define OBD_FAIL_LDLM_REPLY 0x30c
1963 do_facet $SINGLEMDS "lctl set_param fail_loc=0x8000030c"
1966 wait $pid || return 1
1967 [ -e $DIR/$tfile ] && return 2
1970 run_test 73b "open(O_CREAT), unlink, replay, reconnect at open_replay reply, close"
1973 multiop_bg_pause $DIR/$tfile O_tSc || return 3
1977 replay_barrier $SINGLEMDS
1978 #define OBD_FAIL_TGT_LAST_REPLAY 0x710
1979 do_facet $SINGLEMDS "lctl set_param fail_loc=0x80000710"
1982 wait $pid || return 1
1983 [ -e $DIR/$tfile ] && return 2
1986 run_test 73c "open(O_CREAT), unlink, replay, reconnect at last_replay, close"
1990 local clients=${CLIENTS:-$HOSTNAME}
1992 zconf_umount_clients $clients $MOUNT
1994 facet_failover $SINGLEMDS
1995 zconf_mount_clients $clients $MOUNT
1997 touch $DIR/$tfile || return 1
1998 rm $DIR/$tfile || return 2
1999 clients_up || error "client evicted: $?"
2002 run_test 74 "Ensure applications don't fail waiting for OST recovery"
2004 remote_dir_check_80() {
2006 local diridx=$($GETSTRIPE -M $remote_dir)
2007 [ $diridx -eq $MDTIDX ] || error "$diridx != $MDTIDX"
2009 createmany -o $remote_dir/f-%d 20 || error "creation failed"
2010 local fileidx=$($GETSTRIPE -M $remote_dir/f-1)
2011 [ $fileidx -eq $MDTIDX ] || error "$fileidx != $MDTIDX"
2017 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2018 ([ $FAILURE_MODE == "HARD" ] &&
2019 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2020 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2024 local remote_dir=$DIR/$tdir/remote_dir
2027 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2028 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1701
2029 $LFS mkdir -i $MDTIDX $remote_dir &
2032 fail mds$((MDTIDX + 1))
2034 wait $CLIENT_PID || error "remote creation failed"
2036 remote_dir_check_80 || error "remote dir check failed"
2037 rm -rf $DIR/$tdir || error "rmdir failed"
2041 run_test 80a "DNE: create remote dir, drop update rep from MDT1, fail MDT1"
2044 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2045 ([ $FAILURE_MODE == "HARD" ] &&
2046 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2047 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2051 local remote_dir=$DIR/$tdir/remote_dir
2054 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2055 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1701
2056 $LFS mkdir -i $MDTIDX $remote_dir &
2061 wait $CLIENT_PID || error "remote creation failed"
2063 remote_dir_check_80 || error "remote dir check failed"
2064 rm -rf $DIR/$tdir || error "rmdir failed"
2068 run_test 80b "DNE: create remote dir, drop update rep from MDT1, fail MDT0"
2071 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2072 ([ $FAILURE_MODE == "HARD" ] &&
2073 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2074 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2078 local remote_dir=$DIR/$tdir/remote_dir
2081 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2082 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1701
2083 $LFS mkdir -i $MDTIDX $remote_dir &
2087 fail mds$((MDTIDX + 1))
2089 wait $CLIENT_PID || error "remote creation failed"
2091 remote_dir_check_80 || error "remote dir check failed"
2092 rm -rf $DIR/$tdir || error "rmdir failed"
2096 run_test 80c "DNE: create remote dir, drop update rep from MDT1, fail MDT[0,1]"
2099 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2101 local remote_dir=$DIR/$tdir/remote_dir
2104 #define OBD_FAIL_UPDATE_OBJ_NET_REP 0x1701
2105 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x1701
2106 $LFS mkdir -i $MDTIDX $remote_dir &
2109 fail mds${MDTIDX},mds$((MDTIDX + 1))
2111 wait $CLIENT_PID || error "remote creation failed"
2113 remote_dir_check_80 || error "remote dir check failed"
2114 rm -rf $DIR/$tdir || error "rmdir failed"
2118 run_test 80d "DNE: create remote dir, drop update rep from MDT1, fail 2 MDTs"
2121 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2122 ([ $FAILURE_MODE == "HARD" ] &&
2123 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2124 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2128 local remote_dir=$DIR/$tdir/remote_dir
2131 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2132 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2133 $LFS mkdir -i $MDTIDX $remote_dir &
2138 wait $CLIENT_PID || error "remote creation failed"
2140 remote_dir_check_80 || error "remote dir check failed"
2141 rm -rf $DIR/$tdir || error "rmdir failed"
2145 run_test 80e "DNE: create remote dir, drop MDT0 rep, fail MDT0"
2148 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2149 ([ $FAILURE_MODE == "HARD" ] &&
2150 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2151 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2154 local remote_dir=$DIR/$tdir/remote_dir
2157 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2158 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2159 $LFS mkdir -i $MDTIDX $remote_dir &
2162 fail mds$((MDTIDX + 1))
2164 wait $CLIENT_PID || error "remote creation failed"
2166 remote_dir_check_80 || error "remote dir check failed"
2167 rm -rf $DIR/$tdir || error "rmdir failed"
2171 run_test 80f "DNE: create remote dir, drop MDT0 rep, fail MDT1"
2174 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2175 ([ $FAILURE_MODE == "HARD" ] &&
2176 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2177 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2181 local remote_dir=$DIR/$tdir/remote_dir
2184 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2185 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2186 $LFS mkdir -i $MDTIDX $remote_dir &
2190 fail mds$((MDTIDX + 1))
2192 wait $CLIENT_PID || error "remote creation failed"
2194 remote_dir_check_80 || error "remote dir check failed"
2195 rm -rf $DIR/$tdir || error "rmdir failed"
2199 run_test 80g "DNE: create remote dir, drop MDT0 rep, fail MDT0, then MDT1"
2202 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2204 local remote_dir=$DIR/$tdir/remote_dir
2207 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2208 do_facet mds${MDTIDX} lctl set_param fail_loc=0x119
2209 $LFS mkdir -i $MDTIDX $remote_dir &
2212 fail mds${MDTIDX},mds$((MDTIDX + 1))
2214 wait $CLIENT_PID || return 1
2216 remote_dir_check_80 || error "remote dir check failed"
2217 rm -rf $DIR/$tdir || error "rmdir failed"
2221 run_test 80h "DNE: create remote dir, drop MDT0 rep, fail 2 MDTs"
2224 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2225 ([ $FAILURE_MODE == "HARD" ] &&
2226 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2227 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2231 local remote_dir=$DIR/$tdir/remote_dir
2234 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2237 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2238 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2242 fail mds$((MDTIDX + 1))
2244 wait $CLIENT_PID || error "rm remote dir failed"
2246 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2248 rm -rf $DIR/$tdir || error "rmdir failed"
2252 run_test 81a "DNE: unlink remote dir, drop MDT0 update rep, fail MDT1"
2255 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2256 ([ $FAILURE_MODE == "HARD" ] &&
2257 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2258 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2261 local remote_dir=$DIR/$tdir/remote_dir
2264 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2266 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2267 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2273 wait $CLIENT_PID || error "rm remote dir failed"
2275 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2277 rm -rf $DIR/$tdir || error "rmdir failed"
2281 run_test 81b "DNE: unlink remote dir, drop MDT0 update reply, fail MDT0"
2284 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2285 ([ $FAILURE_MODE == "HARD" ] &&
2286 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2287 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2291 local remote_dir=$DIR/$tdir/remote_dir
2294 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2296 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2297 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2302 fail mds$((MDTIDX + 1))
2304 wait $CLIENT_PID || error "rm remote dir failed"
2306 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2308 rm -rf $DIR/$tdir || error "rmdir failed"
2312 run_test 81c "DNE: unlink remote dir, drop MDT0 update reply, fail MDT0,MDT1"
2315 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2317 local remote_dir=$DIR/$tdir/remote_dir
2320 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2322 # OBD_FAIL_OBJ_UPDATE_NET_REP 0x1701
2323 do_facet mds${MDTIDX} lctl set_param fail_loc=0x1701
2327 fail mds${MDTIDX},mds$((MDTIDX + 1))
2329 wait $CLIENT_PID || error "rm remote dir failed"
2331 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2333 rm -rf $DIR/$tdir || error "rmdir failed"
2337 run_test 81d "DNE: unlink remote dir, drop MDT0 update reply, fail 2 MDTs"
2340 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2341 ([ $FAILURE_MODE == "HARD" ] &&
2342 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2343 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2347 local remote_dir=$DIR/$tdir/remote_dir
2350 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2352 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2353 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2356 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0
2360 wait $CLIENT_PID || error "rm remote dir failed"
2362 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2364 rm -rf $DIR/$tdir || error "rmdir failed"
2368 run_test 81e "DNE: unlink remote dir, drop MDT1 req reply, fail MDT0"
2371 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2372 ([ $FAILURE_MODE == "HARD" ] &&
2373 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2374 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2378 local remote_dir=$DIR/$tdir/remote_dir
2381 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2383 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2384 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2388 fail mds$((MDTIDX + 1))
2390 wait $CLIENT_PID || error "rm remote dir failed"
2392 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2394 rm -rf $DIR/$tdir || error "rmdir failed"
2398 run_test 81f "DNE: unlink remote dir, drop MDT1 req reply, fail MDT1"
2401 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2402 ([ $FAILURE_MODE == "HARD" ] &&
2403 [ "$(facet_host mds1)" == "$(facet_host mds2)" ]) &&
2404 skip "MDTs needs to be on diff hosts for HARD fail mode" &&
2408 local remote_dir=$DIR/$tdir/remote_dir
2411 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2413 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2414 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
2419 fail mds$((MDTIDX + 1))
2421 wait $CLIENT_PID || error "rm remote dir failed"
2423 stat $remote_dir 2&>/dev/null && error "$remote_dir still exist!"
2425 rm -rf $DIR/$tdir || error "rmdir failed"
2429 run_test 81g "DNE: unlink remote dir, drop req reply, fail M0, then M1"
2432 [ $MDSCOUNT -lt 2 ] && skip "needs >= 2 MDTs" && return 0
2434 local remote_dir=$DIR/$tdir/remote_dir
2437 $LFS mkdir -i $MDTIDX $remote_dir || error "lfs mkdir failed"
2439 # OBD_FAIL_MDS_REINT_NET_REP 0x119
2440 do_facet mds$((MDTIDX + 1)) lctl set_param fail_loc=0x119
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=$(($(facet_number $ostfail) - 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