nvme/050: add support for NVMe multipath devices

common/multipath-over-rdma

@@ -156,7 +156,7 @@ dev_to_mpath() {
 	[ -e "$1" ] || return $?
 
 	if [ -L "$1" ]; then
-		e=$(readlink -f "$1")
+		e=$(readlink --canonicalize "$1")
 	else
 		e="$1"
 	fi
@@ -167,7 +167,7 @@ dev_to_mpath() {
 
 	for d in /dev/mapper/mpath*; do
 		if [ -L "$d" ]; then
-			e=$(readlink -f "$d")
+			e=$(readlink --canonicalize "$d")
 		elif [ -e "$d" ]; then
 			e="$d"
 		else
@@ -240,7 +240,7 @@ remove_mpath_dev() {
 				break
 			fi
 			ls -l "$1"
-			stop_bdev_users "$(readlink -f "$1")"
+			stop_bdev_users "$(readlink --canonicalize "$1")"
 			sleep .5
 		done
 		if [[ $i = 0 ]]; then

common/nvme

@@ -1122,7 +1122,7 @@ _nvmet_target_cleanup() {
 }
 
 _require_test_dev_is_nvme() {
-	if ! readlink -f "$TEST_DEV_SYSFS/device" | grep -q nvme; then
+	if ! readlink --canonicalize "$TEST_DEV_SYSFS/device" | grep -q nvme; then
 		SKIP_REASONS+=("$TEST_DEV is not a NVMe device")
 		return 1
 	fi

common/rc

@@ -392,16 +392,16 @@ _test_dev_set_scheduler() {
 }
 
 _require_test_dev_is_pci() {
-	if ! readlink -f "$TEST_DEV_SYSFS/device" | grep -q pci; then
+	if ! readlink --canonicalize "$TEST_DEV_SYSFS/device" | grep -q pci; then
 		# nvme needs some special casing
-		if readlink -f "$TEST_DEV_SYSFS/device" | grep -q nvme; then
+		if readlink --canonicalize "$TEST_DEV_SYSFS/device" | grep -q nvme; then
 			local blkdev
 			local ctrldev
 			# First get the controller device from the namespace blockdev
 			blkdev="$(echo "$TEST_DEV_SYSFS" | cut -d '/' -f 7)"
 			ctrldev="$(echo "$blkdev" | grep -Eo 'nvme[0-9]+')"
 			# Then get the pci device from the controller device
-			if readlink -f "$ctrldev/device" | grep -1 pci; then
+			if readlink --canonicalize "$ctrldev/device" | grep -1 pci; then
 				return 0
 			fi
 		fi
@@ -413,7 +413,7 @@ _require_test_dev_is_pci() {
 }
 
 _get_pci_from_dev_sysfs() {
-	readlink -f "$1/device" | \
+	readlink --canonicalize "$1/device" | \
 		grep -Eo '[0-9a-f]{4,5}:[0-9a-f]{2}:[0-9a-f]{2}\.[0-9a-f]' | \
 		tail -1
 }
@@ -423,7 +423,7 @@ _get_pci_dev_from_blkdev() {
 }
 
 _get_pci_parent_from_blkdev() {
-	readlink -f "$TEST_DEV_SYSFS/device" | \
+	readlink --canonicalize "$TEST_DEV_SYSFS/device" | \
 		grep -Eo '[0-9a-f]{4,5}:[0-9a-f]{2}:[0-9a-f]{2}\.[0-9a-f]' | \
 		tail -2 | head -1
 }
@@ -734,7 +734,7 @@ _real_dev()
 {
 	local dev=$1
 	if [ -b "$dev" ] && [ -L "$dev" ]; then
-		dev=$(readlink -f "$dev")
+		dev=$(readlink --canonicalize "$dev")
 	fi
 	echo "$dev"
 }

tests/nvme/050

@@ -20,13 +20,11 @@ requires() {
 }
 
 test_device() {
-	local nvme_ns
-	local pdev
-	local i
+	local nvme_ns pdev io_error i
 
 	echo "Running ${TEST_NAME}"
 
-	pdev=$(_get_pci_dev_from_blkdev)
+	pdev=$(_nvme_get_pci_from_dev_sysfs)
 	nvme_ns="$(basename "${TEST_DEV}")"
 	echo 1 > /sys/block/"${nvme_ns}"/io-timeout-fail
 
@@ -40,10 +38,19 @@ test_device() {
 	    --name=reads --direct=1 --filename="${TEST_DEV}" --group_reporting \
 	    --time_based --runtime=1m >& "$FULL"
 
-	if grep -q "Input/output error" "$FULL"; then
-		echo "Test complete"
+	io_error=false
+	grep -q "Input/output error" "$FULL" && io_error=true
+
+	# The timeout failure injection causes an I/O to fail immediately. For
+	# a single-path device, the failed I/O is propagated up the stack and
+	# eventually reported to user space as an error. For multipath devices,
+	# the block layer evaluates whether the I/O is eligible for retry via
+	# failover to an alternate path. Because the I/O fails before the
+	# per-I/O timeout expires, it remains eligible for retry.
+	if _nvme_dev_support_native_multipath; then
+		$io_error && echo "Test failed" || echo "Test complete"
 	else
-		echo "Test failed"
+		$io_error && echo "Test complete" || echo "Test failed"
 	fi
 
 	# Remove and rescan the NVME device to ensure that it has come back

tests/nvme/rc

@@ -72,21 +72,38 @@ group_device_requires() {
 }
 
 _require_test_dev_is_nvme_pci() {
-	if [[ ! "$(readlink -f "$TEST_DEV_SYSFS/device")" =~ devices/pci ]]; then
+	if [[ ! "$(readlink --canonicalize "$TEST_DEV_SYSFS/device")" =~ devices/pci ]]; then
 		SKIP_REASONS+=("$TEST_DEV is not a PCI NVMe device")
 		return 1
 	fi
 	return 0
 }
 
 _require_test_dev_is_not_nvme_multipath() {
-	if [[ "$(readlink -f "$TEST_DEV_SYSFS/device")" =~ /nvme-subsystem/ ]]; then
+	if [[ "$(readlink --canonicalize "$TEST_DEV_SYSFS/device")" =~ /nvme-subsystem/ ]]; then
 		SKIP_REASONS+=("$TEST_DEV is a NVMe multipath device")
 		return 1
 	fi
 	return 0
 }
 
+_nvme_dev_support_native_multipath() {
+	if [[ "$(readlink --canonicalize "$TEST_DEV_SYSFS/device")" =~ /nvme-subsystem/ ]]; then
+		return 0
+	fi
+	return 1
+}
+
+_nvme_get_pci_from_dev_sysfs() {
+	if _nvme_dev_support_native_multipath; then
+		readlink --canonicalize "$TEST_DEV_SYSFS"/multipath/nvme*c*n*/device | \
+			grep -Eo '[0-9a-f]{4,5}:[0-9a-f]{2}:[0-9a-f]{2}\.[0-9a-f]' | \
+			tail -1
+	else
+		_get_pci_dev_from_blkdev
+	fi
+}
+
 _require_test_dev_support_sed() {
 	if ! nvme sed discover "$TEST_DEV" &> /dev/null; then
 		SKIP_REASONS+=("$TEST_DEV doesn't support SED operations")
@@ -170,12 +187,12 @@ _nvme_get_ctrl_list() {
 	subsys=$(readlink  "${TEST_DEV_SYSFS}/device/subsystem")
 	case $subsys in
 		*/nvme)
-			readlink -f "${TEST_DEV_SYSFS}/device"
+			readlink --canonicalize "${TEST_DEV_SYSFS}/device"
 			;;
 		*/nvme-subsystem)
 			for c in "${TEST_DEV_SYSFS}"/device/nvme*; do
 				[[ -L "$c" ]] || continue
-				[[ -f "$c/dev" ]] && readlink -f "$c"
+				[[ -f "$c/dev" ]] && readlink --canonicalize "$c"
 			done
 			;;
 	esac

tests/srp/012

@@ -21,7 +21,7 @@ test_io_schedulers() {
 		fi
 		use_blk_mq ${mq} ${mq} || return $?
 		dev=$(get_bdev 0) || return $?
-		dm=$(basename "$(readlink -f "${dev}")") || return $?
+		dm=$(basename "$(readlink --canonicalize "${dev}")") || return $?
 		scheds="$(_io_schedulers "$dm")" || return $?
 		for sched in $scheds; do
 			set_scheduler "$dm" "$sched" \

tests/srp/rc

@@ -356,7 +356,7 @@ configure_lio_vdev() {
 			cd "$dirname" &&
 			mkdir "$vdev" &&
 			cd "$vdev" &&
-			if [ -b "$(readlink -f "$path")" ]; then
+			if [ -b "$(readlink --canonicalize "$path")" ]; then
 				echo "udev_path=$path," >control
 			elif [ -e "$path" ]; then
 				size=$(stat -c %s "${path}") &&
@@ -365,7 +365,7 @@ configure_lio_vdev() {
 			else
 				{
 					ls -l "$path"
-					readlink -f "$path"
+					readlink --canonicalize "$path"
 				} &>>"$FULL"
 				false
 			fi &&
@@ -506,7 +506,7 @@ start_lio_srpt() {
 	modprobe ib_srpt "${opts[@]}" || return $?
 	i=0
 	for r in "${vdev_path[@]}"; do
-		if [ -b "$(readlink -f "$r")" ]; then
+		if [ -b "$(readlink --canonicalize "$r")" ]; then
 			oflag=("oflag=direct")
 		else
 			oflag=()