Merge pull request #7460 from AkihiroSuda/allow-rootless-cni

rootless: support `podman network create` (CNI-in-slirp4netns)

3 files changed, 204 insertions, 0 deletions

diff --git a/contrib/rootless-cni-infra/Containerfile b/contrib/rootless-cni-infra/Containerfile
new file mode 100644
index 000000000..c5d812a6e
--- /dev/null
+++ b/contrib/rootless-cni-infra/Containerfile
@@ -0,0 +1,35 @@
+ARG GOLANG_VERSION=1.15
+ARG ALPINE_VERSION=3.12
+ARG CNI_VERSION=v0.8.0
+ARG CNI_PLUGINS_VERSION=v0.8.7
+# Aug 20, 2020
+ARG DNSNAME_VESION=78b4da7bbfc51c27366da630e1df1c4f2e8b1b5b
+
+FROM golang:${GOLANG_VERSION}-alpine${ALPINE_VERSION} AS golang-base
+RUN apk add --no-cache git
+
+FROM golang-base AS cnitool
+RUN git clone https://github.com/containernetworking/cni /go/src/github.com/containernetworking/cni
+WORKDIR /go/src/github.com/containernetworking/cni
+ARG CNI_VERSION
+RUN git checkout ${CNI_VERSION}
+RUN go build -o /cnitool ./cnitool
+
+FROM golang-base AS dnsname
+RUN git clone https://github.com/containers/dnsname /go/src/github.com/containers/dnsname
+WORKDIR /go/src/github.com/containers/dnsname
+ARG DNSNAME_VERSION
+RUN git checkout ${DNSNAME_VERSION}
+RUN go build -o /dnsname ./plugins/meta/dnsname
+
+FROM alpine:${ALPINE_VERSION}
+RUN apk add --no-cache curl dnsmasq iptables ip6tables iproute2
+ARG TARGETARCH
+ARG CNI_PLUGINS_VERSION
+RUN mkdir -p /opt/cni/bin && \
+    curl -fsSL https://github.com/containernetworking/plugins/releases/download/${CNI_PLUGINS_VERSION}/cni-plugins-linux-${TARGETARCH}-${CNI_PLUGINS_VERSION}.tgz | tar xz -C /opt/cni/bin
+COPY --from=cnitool /cnitool /usr/local/bin
+COPY --from=dnsname /dnsname /opt/cni/bin
+COPY rootless-cni-infra /usr/local/bin
+ENV CNI_PATH=/opt/cni/bin
+CMD ["sleep", "infinity"]
diff --git a/contrib/rootless-cni-infra/README.md b/contrib/rootless-cni-infra/README.md
new file mode 100644
index 000000000..937e057fb
--- /dev/null
+++ b/contrib/rootless-cni-infra/README.md
@@ -0,0 +1,22 @@
+# rootless-cni-infra
+
+Infra container for CNI-in-slirp4netns.
+
+## How it works
+
+When a CNI network is specified for `podman run` in rootless mode, Podman launches the `rootless-cni-infra` container to execute CNI plugins inside slirp4netns.
+
+The infra container is created per user, by executing an equivalent of:
+`podman run -d --name rootless-cni-infra --pid=host --privileged -v $HOME/.config/cni/net.d:/etc/cni/net.d rootless-cni-infra`.
+The infra container is automatically deleted when no CNI network is in use.
+
+Podman then allocates a CNI netns in the infra container, by executing an equivalent of:
+`podman exec rootless-cni-infra rootless-cni-infra alloc $CONTAINER_ID $NETWORK_NAME $POD_NAME`.
+
+The allocated netns is deallocated when the container is being removed, by executing an equivalent of:
+`podman exec rootless-cni-infra rootless-cni-infra dealloc $CONTAINER_ID $NETWORK_NAME`.
+
+## Directory layout
+
+* `/run/rootless-cni-infra/${CONTAINER_ID}/pid`: PID of the `sleep infinity` process that corresponds to the allocated netns
+* `/run/rootless-cni-infra/${CONTAINER_ID}/attached/${NETWORK_NAME}`: CNI result
diff --git a/contrib/rootless-cni-infra/rootless-cni-infra b/contrib/rootless-cni-infra/rootless-cni-infra
new file mode 100755
index 000000000..5a574d2eb
--- /dev/null
+++ b/contrib/rootless-cni-infra/rootless-cni-infra
@@ -0,0 +1,147 @@
+#!/bin/sh
+set -eu
+
+ARG0="$0"
+VERSION="0.1.0"
+BASE="/run/rootless-cni-infra"
+
+# CLI subcommand: "alloc $CONTAINER_ID $NETWORK_NAME $POD_NAME"
+cmd_entrypoint_alloc() {
+	if [ "$#" -ne 3 ]; then
+		echo >&2 "Usage: $ARG0 alloc CONTAINER_ID NETWORK_NAME POD_NAME"
+		exit 1
+	fi
+
+	ID="$1"
+	NET="$2"
+	K8S_POD_NAME="$3"
+
+	dir="${BASE}/${ID}"
+	mkdir -p "${dir}/attached"
+
+	pid=""
+	if [ -f "${dir}/pid" ]; then
+		pid=$(cat "${dir}/pid")
+	else
+		unshare -n sleep infinity &
+		pid="$!"
+		echo "${pid}" >"${dir}/pid"
+		nsenter -t "${pid}" -n ip link set lo up
+	fi
+	CNI_ARGS="IgnoreUnknown=1;K8S_POD_NAME=${K8S_POD_NAME}"
+	nwcount=$(find "${dir}/attached" -type f | wc -l)
+	CNI_IFNAME="eth${nwcount}"
+	export CNI_ARGS CNI_IFNAME
+	cnitool add "${NET}" "/proc/${pid}/ns/net" >"${dir}/attached/${NET}"
+
+	# return the result
+	ns="/proc/${pid}/ns/net"
+	echo "{\"ns\":\"${ns}\"}"
+}
+
+# CLI subcommand: "dealloc $CONTAINER_ID $NETWORK_NAME"
+cmd_entrypoint_dealloc() {
+	if [ "$#" -ne 2 ]; then
+		echo >&2 "Usage: $ARG0 dealloc CONTAINER_ID NETWORK_NAME"
+		exit 1
+	fi
+
+	ID=$1
+	NET=$2
+
+	dir="${BASE}/${ID}"
+	if [ ! -f "${dir}/pid" ]; then
+		exit 0
+	fi
+	pid=$(cat "${dir}/pid")
+	cnitool del "${NET}" "/proc/${pid}/ns/net"
+	rm -f "${dir}/attached/${NET}"
+
+	nwcount=$(find "${dir}/attached" -type f | wc -l)
+	if [ "${nwcount}" = 0 ]; then
+		kill -9 "${pid}"
+		rm -rf "${dir}"
+	fi
+
+	# return empty json
+	echo "{}"
+}
+
+# CLI subcommand: "is-idle"
+cmd_entrypoint_is_idle() {
+	if [ ! -d ${BASE} ]; then
+		echo '{"idle": true}'
+	elif [ -z "$(ls -1 ${BASE})" ]; then
+		echo '{"idle": true}'
+	else
+		echo '{"idle": false}'
+	fi
+}
+
+# CLI subcommand: "print-cni-result $CONTAINER_ID $NETWORK_NAME"
+cmd_entrypoint_print_cni_result() {
+	if [ "$#" -ne 2 ]; then
+		echo >&2 "Usage: $ARG0 print-cni-result CONTAINER_ID NETWORK_NAME"
+		exit 1
+	fi
+
+	ID=$1
+	NET=$2
+
+	# the result shall be CNI JSON
+	cat "${BASE}/${ID}/attached/${NET}"
+}
+
+# CLI subcommand: "print-netns-path $CONTAINER_ID"
+cmd_entrypoint_print_netns_path() {
+	if [ "$#" -ne 1 ]; then
+		echo >&2 "Usage: $ARG0 print-netns-path CONTAINER_ID"
+		exit 1
+	fi
+
+	ID=$1
+
+	pid=$(cat "${BASE}/${ID}/pid")
+	path="/proc/${pid}/ns/net"
+
+	# return the result
+	echo "{\"path\":\"${path}\"}"
+}
+
+# CLI subcommand: "help"
+cmd_entrypoint_help() {
+	echo "Usage: ${ARG0} COMMAND"
+	echo
+	echo "Rootless CNI Infra container"
+	echo
+	echo "Commands:"
+	echo "  alloc             Allocate a netns"
+	echo "  dealloc           Deallocate a netns"
+	echo "  is-idle           Print whether the infra container is idle"
+	echo "  print-cni-result  Print CNI result"
+	echo "  print-netns-path  Print netns path"
+	echo "  help              Print help"
+	echo "  version           Print version"
+}
+
+# CLI subcommand: "version"
+cmd_entrypoint_version() {
+	echo "{\"version\": \"${VERSION}\"}"
+}
+
+# parse args
+command="${1:-}"
+if [ -z "$command" ]; then
+	echo >&2 "No command was specified. Run \`${ARG0} help\` to see the usage."
+	exit 1
+fi
+
+command_func=$(echo "cmd_entrypoint_${command}" | sed -e "s/-/_/g")
+if ! command -v "${command_func}" >/dev/null 2>&1; then
+	echo >&2 "Unknown command: ${command}. Run \`${ARG0} help\` to see the usage."
+	exit 1
+fi
+
+# start the command func
+shift
+"${command_func}" "$@"