fix(updater): hand the host docker socket in under a name of its own

An update runs the updater image the target configuration names, so a
downgrade runs the updater of the older release - and that one starts a
dockerd of its own, unconditionally, on /var/run/docker.sock. The socket
do-update.sh bind-mounts there makes that impossible: the path cannot be
unlinked, so the daemon never binds it.

The host's socket now goes in as /run/cuos-docker.sock, which an updater
that does not know about it simply does not see - the situation it was
written for. docker_environment.sh prefers it, still accepts the default
path, because image-factory and every system older than this updater
mount the socket there, and otherwise starts a daemon of its own. A
socket that is present but does not answer has to unset DOCKER_HOST, or
the client keeps addressing it while the retry loop brings up a dockerd
on the default path.

tool.sh's image factory keeps the default path deliberately: it builds on
the host's daemon because that is faster and platform-independent, and it
always runs the factory image it pins itself.

Signed-off-by: Simon Walz <contact@simonwalz.de>

Author: simonwalz

system/cuos/do-update-shell.sh

@@ -25,13 +25,14 @@ ROOT_DEV="$(findmnt -n -o SOURCE -T "/" | sed 's/\[.*\]//')"
 ROOT_DISK="/dev/$(lsblk -no PKNAME "$ROOT_DEV" | head -n1)"
 
 touch "/root/.docker/config.json"
+# Same socket path as do-update.sh, and for the same reason.
 docker run --rm -it \
 	--pull=never \
 	--log-driver=journald \
 	--privileged \
 	--device "${ROOT_DISK}" \
 	-v "/root/.docker/config.json:/root/.docker/config.json:ro" \
-	-v "/var/run/docker.sock:/var/run/docker.sock:ro" \
+	-v "/var/run/docker.sock:/run/cuos-docker.sock:ro" \
 	-v "/usr/local/share/ca-certificates/custom:/usr/local/share/ca-certificates/custom:ro" \
 	-v "/etc/image:/etc/image:ro" \
 	-e "TARGET_DEVICE=${ROOT_DISK}" \

system/cuos/do-update.sh

@@ -50,13 +50,18 @@ ROOT_DEV="$(findmnt -n -o SOURCE -T "/" | sed 's/\[.*\]//')"
 ROOT_DISK="/dev/$(lsblk -no PKNAME "$ROOT_DEV" | head -n1)"
 
 touch "/root/.docker/config.json"
+# The host's docker socket goes in under a name of its own, because
+# /var/run/docker.sock inside the updater has to stay free: an updater that
+# brings no daemon along starts a dockerd there, and a bind mount over that
+# path cannot be unlinked, so the daemon never comes up. The updater of an
+# older release - the one a downgrade runs - is such an updater.
 docker run --rm \
   --pull=never \
   --log-driver=journald \
   --privileged \
   --device "${ROOT_DISK}" \
   -v "/root/.docker/config.json:/root/.docker/config.json:ro" \
-  -v "/var/run/docker.sock:/var/run/docker.sock:ro" \
+  -v "/var/run/docker.sock:/run/cuos-docker.sock:ro" \
   -v "/usr/local/share/ca-certificates/custom:/usr/local/share/ca-certificates/custom:ro" \
   -v "/etc/image:/etc/image:ro" \
   -v "${CONFIG_PATH}:/system.json" \

updater/docker_environment.sh

@@ -92,7 +92,19 @@ start_dockerd() {
 	return 0
 }
 
+# The host's daemon, if the caller handed one in. /var/run/docker.sock is left
+# alone so that start_dockerd below can bind it; the plain path is still
+# honoured, because image-factory and any system older than this updater mount
+# the socket there.
+CUOS_DOCKER_SOCKET="${CUOS_DOCKER_SOCKET:-"/run/cuos-docker.sock"}"
+if [[ -S "${CUOS_DOCKER_SOCKET}" ]]; then
+	export DOCKER_HOST="unix://${CUOS_DOCKER_SOCKET}"
+fi
+
 if ! docker info >/dev/null 2>&1; then
+	# A socket that does not answer must not keep the client from reaching the
+	# dockerd started below, which listens on the default path.
+	unset DOCKER_HOST
 	if ! start_dockerd; then
 		if ! start_dockerd; then
 			if ! start_dockerd; then