mmp_on_uberblocks: Use kstat for uberblock counts

Use kstat to get a more accurate count of uberblock updates.
Using a loop with zdb can potentially miss some uberblocks.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Giuseppe Di Natale <dinatale2@llnl.gov>
Closes #6407 
Closes #6419
This commit is contained in:
Giuseppe Di Natale 2017-07-31 16:54:34 -07:00 committed by Brian Behlendorf
parent c7a7601c08
commit af0f842883
4 changed files with 35 additions and 30 deletions

View File

@ -182,3 +182,19 @@ function import_activity_check # pool opts
return $rc
}
function clear_mmp_history
{
log_must set_tunable64 zfs_multihost_history $MMP_HISTORY_OFF
log_must set_tunable64 zfs_multihost_history $MMP_HISTORY
}
function count_uberblocks # pool duration
{
typeset pool=$1
typeset -i duration=$2
typeset hist_path="/proc/spl/kstat/zfs/$pool/multihost"
log_must sleep $duration
echo $(cat "$hist_path" | sed '1,2d' | wc -l)
}

View File

@ -24,8 +24,8 @@
# STRATEGY:
# 1. Set zfs_txg_timeout to large value
# 2. Create a zpool
# 3. Find the current "best" uberblock
# 4. Loop for 10 seconds, increment counter for each change in UB
# 3. Clear multihost history
# 4. Sleep, then collect count of uberblocks written
# 5. If number of changes seen is less than min threshold, then fail
# 6. If number of changes seen is more than max threshold, then fail
#
@ -37,12 +37,15 @@
verify_runnable "both"
UBER_CHANGES=0
EXPECTED=$(($(echo $DISKS | wc -w) * 10))
FUDGE=$((EXPECTED * 20 / 100))
MIN=$((EXPECTED - FUDGE))
MAX=$((EXPECTED + FUDGE))
function cleanup
{
default_cleanup_noexit
set_tunable64 zfs_txg_timeout $TXG_TIMEOUT_DEFAULT
log_must rm -f $PREV_UBER $CURR_UBER
log_must mmp_clear_hostid
}
@ -52,28 +55,19 @@ log_onexit cleanup
log_must set_tunable64 zfs_txg_timeout $TXG_TIMEOUT_LONG
log_must mmp_set_hostid $HOSTID1
default_setup_noexit $DISK
default_setup_noexit "$DISKS"
log_must zpool set multihost=on $TESTPOOL
log_must zdb -u $TESTPOOL > $PREV_UBER
SECONDS=0
while [[ $SECONDS -le 10 ]]; do
log_must zdb -u $TESTPOOL > $CURR_UBER
if ! diff -u "$CURR_UBER" "$PREV_UBER"; then
(( UBER_CHANGES = UBER_CHANGES + 1 ))
log_must mv "$CURR_UBER" "$PREV_UBER"
fi
done
clear_mmp_history
UBER_CHANGES=$(count_uberblocks $TESTPOOL 10)
log_note "Uberblock changed $UBER_CHANGES times"
if [[ $UBER_CHANGES -lt 8 ]]; then
log_fail "Fewer uberblock writes occured than expected (10)"
if [ $UBER_CHANGES -lt $MIN ]; then
log_fail "Fewer uberblock writes occured than expected ($EXPECTED)"
fi
if [[ $UBER_CHANGES -gt 12 ]]; then
log_fail "More uberblock writes occured than expected (10)"
if [ $UBER_CHANGES -gt $MAX ]; then
log_fail "More uberblock writes occured than expected ($EXPECTED)"
fi
log_pass "Ensure MMP uberblocks update at the correct interval passed"

View File

@ -53,14 +53,11 @@ log_must mmp_set_hostid $HOSTID1
default_setup_noexit $DISK
log_must zpool set multihost=on $TESTPOOL
prev_count=$(wc -l /proc/spl/kstat/zfs/$TESTPOOL/multihost | cut -f1 -d' ')
clear_mmp_history
log_must set_tunable64 zfs_multihost_interval $MMP_INTERVAL_DEFAULT
uber_count=$(count_uberblocks $TESTPOOL 1)
# slight delay to allow time for the mmp write to complete
sleep 1
curr_count=$(wc -l /proc/spl/kstat/zfs/$TESTPOOL/multihost | cut -f1 -d' ')
if [ $curr_count -eq $prev_count ]; then
if [ $uber_count -eq 0 ]; then
log_fail "mmp writes did not start when zfs_multihost_interval reduced"
fi

View File

@ -48,12 +48,10 @@ log_must mmp_set_hostid $HOSTID1
default_mirror_setup_noexit $DISKS
log_must zpool set multihost=on $TESTPOOL
log_must zinject -d ${DISK[0]} -e io -T write -f 50 $TESTPOOL -L uber
clear_mmp_history
uber_count=$(count_uberblocks $TESTPOOL 3)
prev_count=$(wc -l /proc/spl/kstat/zfs/$TESTPOOL/multihost | cut -f1 -d' ')
log_must sleep 3
curr_count=$(wc -l /proc/spl/kstat/zfs/$TESTPOOL/multihost | cut -f1 -d' ')
if [ $curr_count -eq $prev_count ]; then
if [ $uber_count -eq 0 ]; then
log_fail "mmp writes did not occur when uberblock IO errors injected"
fi