Improvements of Node Security

This document describes an enhancement of Ganeti’s security by restricting the distribution of security-sensitive data to the master and master candidates only.

Note: In this document, we will use the term ‘normal node’ for a node that is neither master nor master-candidate.

Contents

• Improvements of Node Security
  • Objective
  • Detailed design
    • Current state and shortcomings
    • Proposal concerning SSH host key distribution
      • RAPI
      • Cluster initialization
      • (Re-)Adding nodes to a cluster
      • Pro- and demoting a node to/from master candidate
      • Offlining and onlining a node
      • Cluster verify
      • Upgrades
      • Downgrades
      • Renew-Crypto
      • Compliance to –no-ssh-init and –no-node-setup
    • Proposal regarding node daemon certificates
      • Cluster initialization
      • (Re-)Adding nodes
      • Promotion and demotion of master candidates
      • Cluster verify
      • Crypto renewal
    • Further considerations
      • Watcher
      • Other Keys and Daemons
      • Related and Future Work

Objective

Up till 2.10, Ganeti distributes security-relevant keys to all nodes, including nodes that are neither master nor master-candidates. Those keys are the private and public SSH keys for node communication and the SSL certficate and private key for RPC communication. Objective of this design is to limit the set of nodes that can establish ssh and RPC connections to the master and master candidates.

As pointed out in issue 377, this is a security risk. Since all nodes have these keys, compromising any of those nodes would possibly give an attacker access to all other machines in the cluster. Reducing the set of nodes that are able to make ssh and RPC connections to the master and master candidates would significantly reduce the risk simply because fewer machines would be a valuable target for attackers.

Note: For bigger installations of Ganeti, it is advisable to run master candidate nodes as non-vm-capable nodes. This would reduce the attack surface for the hypervisor exploitation.

Detailed design

Current state and shortcomings

Currently (as of 2.10), all nodes hold the following information:

• the ssh host keys (public and private)
• the ssh root keys (public and private)
• node daemon certificate (the SSL client certificate and its corresponding private key)

Concerning ssh, this setup contains the following security issue. Since all nodes of a cluster can ssh as root into any other cluster node, one compromised node can harm all other nodes of a cluster.

Regarding the SSL encryption of the RPC communication with the node daemon, we currently have the following setup. There is only one certificate which is used as both, client and server certificate. Besides the SSL client verification, we check if the used client certificate is the same as the certificate stored on the server.

This means that any node running a node daemon can also act as an RPC client and use it to issue RPC calls to other cluster nodes. This in turn means that any compromised node could be used to make RPC calls to any node (including itself) to gain full control over VMs. This could be used by an attacker to for example bring down the VMs or exploit bugs in the virtualization stacks to gain access to the host machines as well.

Proposal concerning SSH host key distribution

We propose the following design regarding the SSH host key handling. The root keys are untouched by this design.

Each node gets its own ssh private/public key pair, but only the public keys of the master candidates get added to the authorized_keys file of all nodes. This has the advantages, that:

• Only master candidates can ssh into other nodes, thus compromised nodes cannot compromise the cluster further.
• One can remove a compromised master candidate from a cluster (including removing it’s public key from all nodes’ authorized_keys file) without having to regenerate and distribute new ssh keys for all master candidates. (Even though it is be good practice to do that anyway, since the compromising of the other master candidates might have taken place already.)
• If a (uncompromised) master candidate is offlined to be sent for repair due to a hardware failure before Ganeti can remove any keys from it (for example when the network adapter of the machine is broken), we don’t have to worry about the keys being on a machine that is physically accessible.

To ensure security while transferring public key information and updating the authorized_keys, there are several other changes necessary:

• Any distribution of keys (in this case only public keys) is done via SSH and not via RPC. An attacker who has RPC control should not be able to get SSH access where he did not have SSH access before already.
• The only RPC calls that are made in this context are from the master daemon to the node daemon on its own host and noded ensures as much as possible that the change to be made does not harm the cluster’s security boundary.
• The nodes that are potential master candidates keep a list of public keys of potential master candidates of the cluster in a separate file called ganeti_pub_keys to keep track of which keys could possibly be added authorized_keys files of the nodes. We come to what “potential” means in this case in the next section. The key list is only transferred via SSH or written directly by noded. It is not stored in the cluster config, because the config is distributed via RPC.

The following sections describe in detail which Ganeti commands are affected by the proposed changes.

RAPI

The design goal to limit SSH powers to master candidates conflicts with the current powers a user of the RAPI interface would have. The master_capable flag of nodes can be modified via RAPI. That means, an attacker that has access to the RAPI interface, can make all non-master-capable nodes master-capable, and then increase the master candidate pool size till all machines are master candidates (or at least a particular machine that he is aming for). This means that with RAPI access and a compromised normal node, one can make this node a master candidate and then still have the power to compromise the whole cluster.

Various options have been explored to mitigate this, with no feasible solution so far. We generally advise to not expose RAPI to the Internet. For more details on making Ganeti secure, see Security in Ganeti.

Alternatively, there was the idea of adding a flag to the cluster config that would ‘freeze’ the master_capable state of nodes. This turned out to be infeasible, as promoting a node from not master_capable to master_capable would mean to add the nodes’s key to the ganeti_pub_keys file. Due to security reasons, this needed to be done in the client (similar to when adding a node). That would have meant that it would no longer be possible to set this flag via RAPI. As setting this flag via RAPI is a feature our users depend on and that has been available in the past, we refrain from breaking this feature.

Cluster initialization

On cluster initialization, the following steps are taken in bootstrap.py:

• A public/private key pair is generated (as before), but only used by the first (and thus master) node. In particular, the private key never leaves the node.
• A mapping of node UUIDs to public SSH keys is created and stored as text file in /var/lib/ganeti/ganeti_pub_keys only accessible by root (permissions 0600). The master node’s uuid and its public key is added as first entry. The format of the file is one line per node, each line composed as node_uuid ssh_key.
• The node’s public key is added to it’s own authorized_keys file.

(Re-)Adding nodes to a cluster

According to Design for adding a node to a cluster, Ganeti transfers the ssh keys to every node that gets added to the cluster.

Adding a new node will require the following steps.

In gnt_node.py:

• On the new node, a new public/private SSH key pair is generated.
• The public key of the new node is fetched (via SSH) to the master node and if it is a potential master candidate (see definition above), it is added to the ganeti_pub_keys list on the master node.
• The public keys of all current master candidates are added to the new node’s authorized_keys file (also via SSH).

In LUNodeAdd in cmdlib/node.py:

• The LUNodeAdd determines whether or not the new node is a master candidate and in any case updates the cluster’s configuration with the new nodes information. (This is not changed by the proposed design.)
• If the new node is a master candidate, we make an RPC call to the node daemon of the master node to add the new node’s public key to all nodes’ authorized_keys files. The implementation of this RPC call has to be extra careful as described in the next steps, because compromised RPC security should not compromise SSH security.

RPC call execution in noded (on master node):

• Check that the public key of the new node is in the ganeti_pub_keys file of the master node to make sure that no keys of nodes outside the Ganeti cluster and no keys that are not potential master candidates gain SSH access in the cluster.
• Via SSH, transfer the new node’s public key to all nodes (including the new node) and add it to their authorized_keys file.
• The ganeti_pub_keys file is transferred via SSH to all potential master candidates nodes except the master node (including the new one).

In case of readding a node that used to be in the cluster before, handling of the SSH keys would basically be the same, in particular also a new SSH key pair is generated for the node, because we cannot be sure that the old key pair has not been compromised while the node was offlined. Note that for reasons of data hygiene, a node’s ganeti_pub_keys file is cleared before the node is readded. Also, Ganeti attempts to remove any Ganeti keys from the authorized_keys file before the node is readded. However, since Ganeti does not keep a list of all keys ever used in the cluster, this applies only to keys which are currently used in the cluster. Note that Ganeti won’t touch any keys that were added to the authorized_keys by other systems than Ganeti.

Pro- and demoting a node to/from master candidate

If the role of a node is changed from ‘normal’ to ‘master_candidate’, the procedure is the same as for adding nodes from the step “In LUNodeAdd ...” on.

If a node gets demoted to ‘normal’, the master daemon makes a similar RPC call to the master node’s node daemon as for adding a node.

In the RPC call, noded will perform the following steps:

• Check that the public key of the node to be demoted is indeed in the ganeti_pub_keys file to avoid deleting ssh keys of machines that don’t belong to the cluster (and thus potentially lock out the administrator).
• Via SSH, remove the key from all node’s authorized_keys files.

This affected the behavior of the following commands:

::

gnt-node modify –master-candidate=yes gnt-node modify –master-candidate=no [–auto-promote]

If the node has been master candidate already before the command to promote it was issued, Ganeti does not do anything.

Note that when you demote a node from master candidate to normal node, another master-capable and normal node will be promoted to master candidate. For this newly promoted node, the same changes apply as if it was explicitely promoted.

The same behavior should be ensured for the corresponding rapi command.

Offlining and onlining a node

When offlining a node, it immediately loses its role as master or master candidate as well. When it is onlined again, it will become master candidate again if it was so before. The handling of the keys should be done in the same way as when the node is explicitely promoted or demoted to or from master candidate. See the previous section for details.

This affects the command:

::

gnt-node modify –offline=yes gnt-node modify –offline=no [–auto-promote]

For offlining, the removal of the keys is particularly important, as the detection of a compromised node might be the very reason for the offlining. Of course we cannot guarantee that removal of the key is always successful, because the node might not be reachable anymore. Even though it is a best-effort operation, it is still an improvement over the status quo, because currently Ganeti does not even try to remove any keys.

The same behavior should be ensured for the corresponding rapi command.

Cluster verify

So far, gnt-cluster verify checks the SSH connectivity of all nodes to all other nodes. We propose to replace this by the following checks:

• For all master candidates, we check if they can connect any other node in the cluster (other master candidates and normal nodes).
• We check if the ganeti_pub_keys file contains keys of nodes that are no longer in the cluster or that are not potential master candidates.
• For all normal nodes, we check if their key does not appear in other node’s authorized_keys. For now, we will only emit a warning rather than an error if this check fails, because Ganeti might be run in a setup where Ganeti is not the only system manipulating the SSH keys.

Upgrades

When upgrading from a version that has the previous SSH setup to the one proposed in this design, the upgrade procedure has to involve the following steps in the post-upgrade hook:

• For all nodes, new SSH key pairs are generated.
• All nodes and their public keys are added to the ganeti_pub_keys file and the file is copied to all nodes.
• All keys of master candidate nodes are added to the authorized_keys files of all other nodes.

Since this upgrade significantly changes the configuration of the clusters’ nodes, we will add a note to the UPGRADE notes to make the administrator aware of this fact (in case he intends to enable access from normal nodes to master candidates for other reasons than Ganeti uses the machines).

Also, in any operation where Ganeti creates new SSH keys, the old keys will be backed up and not simply overridden.

Downgrades

These downgrading steps will be implemtented from 2.13 to 2.12:

• The master node’s private/public key pair will be distributed to all nodes (via SSH) and the individual SSH keys will be backed up.
• The obsolete individual ssh keys will be removed from all nodes’ authorized_keys file.

Renew-Crypto

The gnt-cluster renew-crypto command will be extended by a new option --new-ssh-keys, which will renew all SSH keys on all nodes and rebuild the authorized_keys files and the ganeti_pub_keys files according to the previous sections. This operation will be performed considering the already stated security considerations, for example minimizing RPC calls, distribution of keys via SSH only etc.

Compliance to –no-ssh-init and –no-node-setup

With this design, Ganeti will do more manipulations of SSH keys and authorized_keys files than before. If this is not feasible in a Ganeti environment, the administrator has the option to prevent Ganeti from performing any manipulations on the SSH setup of the nodes. The options for doing so, are --no-ssh-init for gnt-cluster init, and --no-node-setup for gnt-node add. Note that these options already existed before the implementation of this design, we just confirm that they will be complied to with the new design as well.

Proposal regarding node daemon certificates

Regarding the node daemon certificates, we propose the following changes in the design.

• Instead of using the same certificate for all nodes as both, server and client certificate, we generate a common server certificate (and the corresponding private key) for all nodes and a different client certificate (and the corresponding private key) for each node. The server certificate will be self-signed. The client certficate will be signed by the server certificate. The client certificates will use the node UUID as serial number to ensure uniqueness within the cluster. They will use the host’s hostname as the certificate common name (CN).
• In addition, we store a mapping of (node UUID, client certificate digest) in the cluster’s configuration and ssconf for hosts that are master or master candidate. The client certificate digest is a hash of the client certificate. We suggest a ‘sha1’ hash here. We will call this mapping ‘candidate map’ from here on.
• The node daemon will be modified in a way that on an incoming RPC request, it first performs a client verification (same as before) to ensure that the requesting host is indeed the holder of the corresponding private key. Additionally, it compares the digest of the certificate of the incoming request to the respective entry of the candidate map. If the digest does not match the entry of the host in the mapping or is not included in the mapping at all, the SSL connection is refused.

This design has the following advantages:

• A compromised normal node cannot issue RPC calls, because it will not be in the candidate map. (See the Drawbacks section regarding an indirect way of achieving this though.)
• A compromised master candidate would be able to issue RPC requests, but on detection of its compromised state, it can be removed from the cluster (and thus from the candidate map) without the need for redistribution of any certificates, because the other master candidates can continue using their own certificates. However, it is best practice to issue a complete key renewal even in this case, unless one can ensure no actions compromising other nodes have not already been carried out.
• A compromised node would not be able to use the other (possibly master candidate) nodes’ information from the candidate map to issue RPCs, because the config just stores the digests and not the certificate itself.
• A compromised node would be able to obtain another node’s certificate by waiting for incoming RPCs from this other node. However, the node cannot use the certificate to issue RPC calls, because the SSL client verification would require the node to hold the corresponding private key as well.

Drawbacks of this design:

• Complexity of node and certificate management will be increased (see following sections for details).
• If the candidate map is not distributed fast enough to all nodes after an update of the configuration, it might be possible to issue RPC calls from a compromised master candidate node that has been removed from the Ganeti cluster already. However, this is still a better situation than before and an inherent problem when one wants to distinguish between master candidates and normal nodes.
• A compromised master candidate would still be able to issue RPC calls, if it uses ssh to retrieve another master candidate’s client certificate and the corresponding private SSL key. This is an issue even with the first part of the improved handling of ssh keys in this design (limiting ssh keys to master candidates), but it will be eliminated with the second part of the design (separate ssh keys for each master candidate).
• Even though this proposal is an improvement towards the previous situation in Ganeti, it still does not use the full power of SSL. For further improvements, see Section “Related and future work”.
• Signing the client certificates with the server certificate will increase the complexity of the renew-crypto, as a renewal of the server certificates requires the renewal (and signing) of all client certificates as well.

Alternative proposals:

• The initial version of this document described a setup where the client certificates were also self-signed. This led to a serious problem (Issue 1094), which would only have been solvable by distributing all client certificates to all nodes and load them as trusted CAs. As this would have resulted in having to restart noded on all nodes every time a node is added, removed, demoted or promoted, this was not feasible and we switched to client certficates which are signed by the server certificate.
• Instead of generating a client certificate per node, one could think of just generating two different client certificates, one for normal nodes and one for master candidates. Noded could then just check if the requesting node has the master candidate certificate. Drawback of this proposal is that once one master candidate gets compromised, all master candidates would need to get a new certificate even if the compromised master candidate had not yet fetched the certificates from the other master candidates via ssh.
• In addition to our main proposal, one could think of including a piece of data (for example the node’s host name or UUID) in the RPC call which is encrypted with the requesting node’s private key. The node daemon could check if the datum can be decrypted using the node’s certificate. However, this would ensure similar functionality as SSL’s built-in client verification and add significant complexity to Ganeti’s RPC protocol.

In the following sections, we describe how our design affects various Ganeti operations.

Cluster initialization

On cluster initialization, so far only the node daemon certificate was created. With our design, two certificates (and corresponding keys) need to be created, a server certificate to be distributed to all nodes and a client certificate only to be used by this particular node. In the following, we use the term node daemon certificate for the server certficate only.

In the cluster configuration, the candidate map is created. It is populated with the respective entry for the master node. It is also written to ssconf.

(Re-)Adding nodes

When a node is added, the server certificate is copied to the node (as before). Additionally, a new client certificate (and the corresponding private key) is created on the new node to be used only by the new node as client certifcate.

If the new node is a master candidate, the candidate map is extended by the new node’s data. As before, the updated configuration is distributed to all nodes (as complete configuration on the master candidates and ssconf on all nodes). Note that distribution of the configuration after adding a node is already implemented, since all nodes hold the list of nodes in the cluster in ssconf anyway.

If the configuration for whatever reason already holds an entry for this node, it will be overriden.

When readding a node, the procedure is the same as for adding a node.

Promotion and demotion of master candidates

When a normal node gets promoted to be master candidate, an entry to the candidate map has to be added and the updated configuration has to be distributed to all nodes. If there was already an entry for the node, we override it.

On demotion of a master candidate, the node’s entry in the candidate map gets removed and the updated configuration gets redistibuted.

The same procedure applied to onlining and offlining master candidates.

Cluster verify

Cluster verify will be extended by the following checks:

• Whether each entry in the candidate map indeed corresponds to a master candidate.
• Whether the master candidate’s certificate digest match their entry in the candidate map.
• Whether no node tries to use the certificate of another node. In particular, it is important to check that no normal node tries to use the certificate of a master candidate.
• Whether there are still self-signed client certificates in use (from a pre 2.12.4 Ganeti version).

Crypto renewal

Currently, when the cluster’s cryptographic tokens are renewed using the gnt-cluster renew-crypto command the node daemon certificate is renewed (among others). Option --new-cluster-certificate renews the node daemon certificate only.

By adding an option --new-node-certificates we offer to renew the client certificate. Whenever the client certificates are renewed, the candidate map has to be updated and redistributed.

If for whatever reason, the candidate map becomes inconsistent, for example due inconsistent updating after a demotion or offlining), the user can use this option to renew the client certificates and update the candidate certificate map.

Note that renewing the server certificate requires all client certificates being renewed and signed by the new server certificate, because otherwise their signature can not be verified by the server who only has the new server certificate then.

As there was a different design in place in Ganeti 2.12.4 and previous versions, we have to ensure that renew-crypto works on pre 2.12 versions and 2.12.1-4. Users that got hit by Issue 1094 will be encouraged to run renew-crypto at least once after switching to 2.12.5. Those who did not encounter this bug yet, will still get nagged friendly by gnt-cluster verify.

Further considerations

Watcher

The watcher is a script that is run on all nodes in regular intervals. The changes proposed in this design will not affect the watcher’s implementation, because it behaves differently on the master than on non-master nodes.

Only on the master, it issues query calls which would require a client certificate of a node in the candidate mapping. This is the case for the master node. On non-master node, it’s only external communication is done via the ConfD protocol, which uses the hmac key, which is present on all nodes. Besides that, the watcher does not make any ssh connections, and thus is not affected by the changes in ssh key handling either.

Other Keys and Daemons

Ganeti handles a couple of other keys/certificates that have not been mentioned in this design so far. Also, other daemons than the ones mentioned so far perform intra-cluster communication. Neither the keys nor the daemons will be affected by this design for several reasons:

• The hmac key used by ConfD (see Ganeti 2.1 design): the hmac key is still distributed to all nodes, because it was designed to be used for communicating with ConfD, which should be possible from all nodes. For example, the monitoring daemon which runs on all nodes uses it to retrieve information from ConfD. However, since communication with ConfD is read-only, a compromised node holding the hmac key does not enable an attacker to change the cluster’s state.
• The WConfD daemon writes the configuration to all master candidates via RPC. Since it only runs on the master node, it’s ability to run RPC requests is maintained with this design.
• The rapi SSL key certificate and rapi user/password file ‘rapi_users’ is already only copied to the master candidates (see Ganeti 2.1 design, Section Redistribute Config).
• The spice certificates are still distributed to all nodes, since it should be possible to use spice to access VMs on any cluster node.
• The cluster domain secret is used for inter-cluster instance moves. Since instances can be moved from any normal node of the source cluster to any normal node of the destination cluster, the presence of this secret on all nodes is necessary.

Related and Future Work

There a couple of suggestions on how to improve the SSL setup even more. As a trade-off wrt to complexity and implementation effort, we did not implement them yet (as of version 2.11) but describe them here for future reference.

• The server certificate is currently self-signed and the client certificates are signed by the server certificate. It would increase the security if they were signed by a common CA. There is already a design doc for a Ganeti CA which was suggested in a different context (related to import/export). This would also be a benefit for the RPC calls. See design doc Design for import/export version 2 for more information. Implementing a CA is rather complex, because it would mean also to support renewing the CA certificate and providing and supporting infrastructure to revoke compromised certificates.
• An extension of the previous suggestion would be to even enable the system administrator to use an external CA. Especially in bigger setups, where already an SSL infrastructure exists, it would be useful if Ganeti can simply be integrated with it, rather than forcing the user to use the Ganeti CA.
• Ganeti RPC calls are currently done without checking if the hostname of the node complies with the common name of the certificate. This might be a desirable feature, but would increase the effort when a node is renamed.
• The typical use case for SSL is to have one certificate per node rather than one shared certificate (Ganeti’s noded server certificate) and a client certificate. One could change the design in a way that only one certificate per node is used, but this would require a common CA so that the validity of the certificate can be established by every node in the cluster.
• With the proposed design, the serial numbers of the client certificates are set to the node UUIDs. This is technically also not complying to how SSL is supposed to be used, as the serial numbers should reflect the enumeration of certificates created by the CA. Once a CA is implemented, it might be reasonable to change this accordingly. The implementation of the proposed design also has the drawback of the serial number not changing even if the certificate is replaced by a new one (for example when calling gnt-cluster renew- crypt), which also does not comply to way SSL was designed to be used.