ceph-csi/scripts/rook.sh
ShyamsundarR 1a8f8e3c24 Add support for erasure coded pools
This commit adds support to mention dataPool parameter for the
topology constrained pools in the StorageClass, that can be
leveraged to mention erasure coded pool names to use for RBD
data instead of the replica pools.

Signed-off-by: ShyamsundarR <srangana@redhat.com>
2020-04-14 14:14:29 +00:00

155 lines
4.6 KiB
Bash
Executable File

#!/bin/bash -e
ROOK_VERSION=${ROOK_VERSION:-"v1.1.7"}
ROOK_DEPLOY_TIMEOUT=${ROOK_DEPLOY_TIMEOUT:-300}
ROOK_URL="https://raw.githubusercontent.com/rook/rook/${ROOK_VERSION}/cluster/examples/kubernetes/ceph"
ROOK_BLOCK_POOL_NAME=${ROOK_BLOCK_POOL_NAME:-"newrbdpool"}
function deploy_rook() {
kubectl create -f "${ROOK_URL}/common.yaml"
kubectl create -f "${ROOK_URL}/operator.yaml"
kubectl create -f "${ROOK_URL}/cluster-test.yaml"
kubectl create -f "${ROOK_URL}/toolbox.yaml"
kubectl create -f "${ROOK_URL}/filesystem-test.yaml"
kubectl create -f "${ROOK_URL}/pool-test.yaml"
# Check if CephCluster is empty
if ! kubectl -n rook-ceph get cephclusters -oyaml | grep 'items: \[\]' &>/dev/null; then
check_ceph_cluster_health
fi
# Check if CephFileSystem is empty
if ! kubectl -n rook-ceph get cephfilesystems -oyaml | grep 'items: \[\]' &>/dev/null; then
check_mds_stat
fi
# Check if CephBlockPool is empty
if ! kubectl -n rook-ceph get cephblockpools -oyaml | grep 'items: \[\]' &>/dev/null; then
check_rbd_stat ""
fi
}
function teardown_rook() {
kubectl delete -f "${ROOK_URL}/pool-test.yaml"
kubectl delete -f "${ROOK_URL}/filesystem-test.yaml"
kubectl delete -f "${ROOK_URL}/toolbox.yaml"
kubectl delete -f "${ROOK_URL}/cluster-test.yaml"
kubectl delete -f "${ROOK_URL}/operator.yaml"
kubectl delete -f "${ROOK_URL}/common.yaml"
}
function create_block_pool() {
curl -o newpool.yaml "${ROOK_URL}/pool-test.yaml"
sed -i "s/replicapool/$ROOK_BLOCK_POOL_NAME/g" newpool.yaml
kubectl create -f "./newpool.yaml"
rm -f "./newpool.yaml"
check_rbd_stat "$ROOK_BLOCK_POOL_NAME"
}
function delete_block_pool() {
curl -o newpool.yaml "${ROOK_URL}/pool-test.yaml"
sed -i "s/replicapool/$ROOK_BLOCK_POOL_NAME/g" newpool.yaml
kubectl delete -f "./newpool.yaml"
rm -f "./newpool.yaml"
}
function check_ceph_cluster_health() {
for ((retry = 0; retry <= ROOK_DEPLOY_TIMEOUT; retry = retry + 5)); do
echo "Wait for rook deploy... ${retry}s" && sleep 5
CEPH_STATE=$(kubectl -n rook-ceph get cephclusters -o jsonpath='{.items[0].status.state}')
CEPH_HEALTH=$(kubectl -n rook-ceph get cephclusters -o jsonpath='{.items[0].status.ceph.health}')
echo "Checking CEPH cluster state: [$CEPH_STATE]"
if [ "$CEPH_STATE" = "Created" ]; then
if [ "$CEPH_HEALTH" = "HEALTH_OK" ]; then
echo "Creating CEPH cluster is done. [$CEPH_HEALTH]"
break
fi
fi
done
if [ "$retry" -gt "$ROOK_DEPLOY_TIMEOUT" ]; then
echo "[Timeout] CEPH cluster not in a healthy state (timeout)"
exit 1
fi
echo ""
}
function check_mds_stat() {
for ((retry = 0; retry <= ROOK_DEPLOY_TIMEOUT; retry = retry + 5)); do
FS_NAME=$(kubectl -n rook-ceph get cephfilesystems.ceph.rook.io -ojsonpath='{.items[0].metadata.name}')
echo "Checking MDS ($FS_NAME) stats... ${retry}s" && sleep 5
ACTIVE_COUNT=$(kubectl -n rook-ceph get cephfilesystems myfs -ojsonpath='{.spec.metadataServer.activeCount}')
ACTIVE_COUNT_NUM=$((ACTIVE_COUNT + 0))
echo "MDS ($FS_NAME) active_count: [$ACTIVE_COUNT_NUM]"
if ((ACTIVE_COUNT_NUM < 1)); then
continue
else
if kubectl -n rook-ceph get pod -l rook_file_system=myfs | grep Running &>/dev/null; then
echo "Filesystem ($FS_NAME) is successfully created..."
break
fi
fi
done
if [ "$retry" -gt "$ROOK_DEPLOY_TIMEOUT" ]; then
echo "[Timeout] Failed to get ceph filesystem pods"
exit 1
fi
echo ""
}
function check_rbd_stat() {
for ((retry = 0; retry <= ROOK_DEPLOY_TIMEOUT; retry = retry + 5)); do
if [ -z "$1" ]; then
RBD_POOL_NAME=$(kubectl -n rook-ceph get cephblockpools -ojsonpath='{.items[0].metadata.name}')
else
RBD_POOL_NAME=$1
fi
echo "Checking RBD ($RBD_POOL_NAME) stats... ${retry}s" && sleep 5
TOOLBOX_POD=$(kubectl -n rook-ceph get pods -l app=rook-ceph-tools -o jsonpath='{.items[0].metadata.name}')
TOOLBOX_POD_STATUS=$(kubectl -n rook-ceph get pod "$TOOLBOX_POD" -ojsonpath='{.status.phase}')
[[ "$TOOLBOX_POD_STATUS" != "Running" ]] && \
{ echo "Toolbox POD ($TOOLBOX_POD) status: [$TOOLBOX_POD_STATUS]"; continue; }
if kubectl exec -n rook-ceph "$TOOLBOX_POD" -it -- rbd pool stats "$RBD_POOL_NAME" &>/dev/null; then
echo "RBD ($RBD_POOL_NAME) is successfully created..."
break
fi
done
if [ "$retry" -gt "$ROOK_DEPLOY_TIMEOUT" ]; then
echo "[Timeout] Failed to get RBD pool stats"
exit 1
fi
echo ""
}
case "${1:-}" in
deploy)
deploy_rook
;;
teardown)
teardown_rook
;;
create-block-pool)
create_block_pool
;;
delete-block-pool)
delete_block_pool
;;
*)
echo " $0 [command]
Available Commands:
deploy Deploy a rook
teardown Teardown a rook
create-block-pool Create a rook block pool
delete-block-pool Delete a rook block pool
" >&2
;;
esac