tbv 0.2.14: rail addr helper resolves live ifname via XDomain parent (udev rename race)

udev renames u4r<N> -> tbr-<peer> BEFORE running RUN+=, so the passed kernel name
is already stale and the helper bailed (rail got no per-link /64, only fe80::).
Resolve the XDomain parent device from DEVPATH (stable across rename) and read the
current ifname back out of <xdomain>/net. Validated on appmana-018: both rails now
carry a routable fd../64 RoCE GID (usb4_rdma5 fd67:..::1, usb4_rdma15 fd72:..::1).

Author: doctorpangloss

drivers/thunderbolt_ibverbs/dkms.conf

@@ -1,5 +1,5 @@
 PACKAGE_NAME="thunderbolt-ibverbs"
-PACKAGE_VERSION="0.2.13"
+PACKAGE_VERSION="0.2.14"
 
 BUILT_MODULE_NAME[0]="thunderbolt_ibverbs"
 BUILT_MODULE_LOCATION[0]="kernel"

drivers/thunderbolt_ibverbs/packaging/udev/tbv-rdma-addr

@@ -2,42 +2,47 @@
 # tbv-rdma-addr <ifname> -- assign the deterministic per-link routable /64 to a
 # usb4_rdma GID-only rail netdev, so both ends of a cabled link share a subnet.
 # Invoked from udev (RUN+=) after the rail netdev appears, and safe to re-run by
-# hand. Pure derivation lives in tbv-rdma-addr-lib.sh; this only does the sysfs
+# hand. Pure derivation lives in tbv-rdma-addr-lib.sh; this does the sysfs
 # lookups and the `ip` assignment.
 #
 # The rail netdev is parented (SET_NETDEV_DEV) to its XDomain *device* (e.g.
 # "0-1"), whose unique_id is the remote peer's host-router UUID. The local
 # host-router UUID is "<domain>-0" in the same thunderbolt domain. Both ends
 # therefore hold the same unordered UUID pair -> the same /64 (see the lib).
+#
+# udev renames the netdev (u4r<N> -> tbr-<peer>) BEFORE running RUN+=, so the
+# passed kernel name is stale by the time we run. We therefore resolve the live
+# interface from the stable XDomain parent device (via DEVPATH), not the name.
 set -eu
+export PATH=/usr/sbin:/usr/bin:/sbin:/bin:${PATH:-}
 
 LIB="$(dirname "$0")/tbv-rdma-addr-lib.sh"
 [ -r "$LIB" ] || LIB=/usr/lib/thunderbolt-ibverbs/tbv-rdma-addr-lib.sh
 . "$LIB"
 
 ifc="${1:-}"
-[ -n "$ifc" ] || exit 0
-
-# Resolve the netdev's parent device, tolerating a concurrent udev rename: prefer
-# the passed name, fall back to udev's DEVPATH for this uevent.
-dev="$(readlink -f "/sys/class/net/${ifc}/device" 2>/dev/null || true)"
-if [ -z "${dev}" ] && [ -n "${DEVPATH:-}" ]; then
-	dev="$(readlink -f "/sys${DEVPATH}/device" 2>/dev/null || true)"
+xdev=""
+# Prefer the passed name if it still resolves; else recover the XDomain parent
+# from DEVPATH (stable across the rename) and read the live name back out of it.
+if [ -n "$ifc" ] && [ -e "/sys/class/net/${ifc}/device" ]; then
+	xdev="$(readlink -f "/sys/class/net/${ifc}/device" 2>/dev/null || true)"
+elif [ -n "${DEVPATH:-}" ]; then
+	xdev="$(readlink -f "/sys${DEVPATH%/net/*}" 2>/dev/null || true)"
+	[ -n "$xdev" ] && ifc="$(ls "${xdev}/net" 2>/dev/null | head -n1)"
 fi
-[ -n "${dev}" ] || exit 0
+[ -n "$xdev" ] && [ -n "$ifc" ] && [ -e "/sys/class/net/${ifc}" ] || exit 0
 
-remote="$(cat "${dev}/unique_id" 2>/dev/null || true)"
-[ -n "${remote}" ] || exit 0                       # not an XDomain-parented rail
+remote="$(cat "${xdev}/unique_id" 2>/dev/null || true)"
+[ -n "$remote" ] || exit 0                       # not an XDomain-parented rail
 
-xname="$(basename "${dev}")"                        # e.g. 0-1
-domain="${xname%%-*}"                               # e.g. 0
+domain="$(basename "$xdev")"; domain="${domain%%-*}"
 local="$(cat "/sys/bus/thunderbolt/devices/${domain}-0/unique_id" 2>/dev/null || true)"
-[ -n "${local}" ] || exit 0
+[ -n "$local" ] || exit 0
 
-addr="$(tbv_link_addr "${local}" "${remote}")"
+addr="$(tbv_link_addr "$local" "$remote")"
 
 # nodad: link-local DAD is pointless on a point-to-point GID-only netdev and only
 # adds latency before the GID appears. `replace` keeps the helper idempotent.
-ip -6 addr replace "${addr}" dev "${ifc}" nodad 2>/dev/null || \
-	ip -6 addr replace "${addr}" dev "${ifc}"
-ip link set "${ifc}" up
+ip -6 addr replace "$addr" dev "$ifc" nodad 2>/dev/null || \
+	ip -6 addr replace "$addr" dev "$ifc"
+ip link set "$ifc" up