7.9 KiB
The SSH CA for agents
Agents SSH out with certificates signed by an OpenBao SSH secrets engine mount. The operator view, how to enable it and how to trust it on a machine, is SSH access to a machine.
The mount and its role
swarm-bao-agent-ssh-policy.service runs on the store's host under the
granter's identity, like the agent PKI unit in
the store page. Each run:
- enables the
ssh-agentsmount (typessh) when it isn't there; - generates the CA key inside the mount, once, with
config/ca generate_signing_key=true key_type=ssh-ed25519. It does that only whenssh-agents/issuerslists none. Nothing replaces or deletes the CA, and the granter holds no read on it; - writes the role
ssh-agents/roles/swarm-agent.
The mount and role names come from nix/host-modules/lib/ssh-agent-ca.nix and
are SSH_MOUNT and SSH_ROLE in swarm-secret-client/src/policy.rs; the
bao-grants module-eval suite fails when they differ.
The role signs for every agent:
| field | value |
|---|---|
key_type |
ca |
allow_user_certificates / allow_host_certificates |
true / false |
allowed_users, default_user |
{{identity.entity.name}}, with allowed_users_template and default_user_template on |
allowed_user_key_lengths |
{"ssh-ed25519": 0}: ed25519 keys only |
allowed_extensions, default_extensions |
permit-pty |
allow_user_key_ids |
false |
ttl, max_ttl |
60 minutes, 180 minutes |
The entity name is the agent's name: swarm-controller names the entity of
each agent's cert login after the agent
(swarm-controller/src/agent_entity.rs). A request for any other principal,
a glob, host certificates, other key types or another extension is a refusal
from the store.
The switch and the signing policy
The operator's switch is the ssh field of the agent's record in agent-permissions
(swarm_wire::agent_permissions), set with PUT /api/agents/{name}/permissions
from the swarm UI's ssh row. Absent is off. A PUT or DELETE that leaves
the field out keeps it.
Signing is a separate policy, hive-agent-ssh (policy::SSH_POLICY,
render_ssh_signer): update on ssh-agents/sign/swarm-agent and nothing
else. The shared agent policy hive-agent holds no SSH stanza. swarm-controller
writes hive-agent-ssh at every identity mint and every pass of agent_ssh.rs,
and attaches it to the agent's cert-auth role hive-agent-<name> as a second
token_policies entry only while the switch is on.
- Rewrite. A change of the field wakes
agent_ssh's pass at once, and the pass also runs every five minutes. It reads each role's policies and rewrites onlytoken_policies(a partial write: the role's certificate, names and lifetimes stay) when they differ from the shared policy plus, if on,hive-agent-ssh. The identity mint (agent_identity.rs) keeps whatever the role has, so a certificate renewal doesn't switch it off. The pass doesn't create a missing role. - When it takes effect. A token carries the policies of the role it logged in
under and keeps them until it expires, at most the role's
token_ttlof one hour. Switching on reaches the agent at its next login, which its certificate unit makes on every run (every 20 minutes), so within about 25 minutes of the pass. Switching off stops every new login at once; an agent already holding a token that carries the policy can still sign until that token expires, at most an hour. Neither touches a certificate already signed: it stays valid until it expires, at most three hours. - The CA key.
GET /api/ssh/careturns{ "public_key": "ssh-ed25519 …" }, read fromssh-agents/config/cawith the controller's own identity. The controller holdsreadon that path and nothing else on the mount: it can't sign. It answers 404 while the store holds no CA.
The agent's unit
nix/agent-modules/ssh.nix is on in every agent that has a store address
(services.hyperhive.agent.bao.addr), with no option. It adds
hive-agent-ssh-cert.service and its timer, as the agent user:
- the key is
/run/hive-agent-ssh-cert/key(directory0700, files0400,UMask=0377), generated when absent; the certificate iskey-cert.pubbeside it; - the unit's script is also the helper
sshruns: one text, two entries. Each run logs in afresh to the store with the cert identityhive-agent-bao-identitykeeps, signskey.pubthrough the role, checks the answer parses as a certificate and renames it into place. A403from the signing call is the switch being off: the unit prints one line and exits 0, and the next tick asks again. Any other failure fails the unit, which retries. One run at a time (flockon the directory), the unit's or anssh's; - the unit copies the store's CA out of its credentials into
server-ca.pembeside the key, so that ansshrun, which has no credentials directory, can reach the store; - the timer fires 20 minutes after each run, a third of the 60-minute lifetime, from the same file as the role's TTL;
- the system
/etc/ssh/ssh_confighas oneHost *block:IdentityFileandCertificateFilefrom/run,IdentitiesOnly yes,StrictHostKeyChecking accept-new, andUserKnownHostsFile /agents/<name>/state/ssh_known_hosts. The host binds the state directory in, so the pins survive a restart; the container root doesn't. The agent trusts the first key a host shows and refuses a changed one; - the same config starts with
Match exec "HIVE_SSH_CERT_ENSURE=1 <helper> </dev/null", with no settings of its own.sshruns the command while it reads the config, before it loads an identity. WithHIVE_SSH_CERT_ENSUREthe helper is silent and its exit status says whether a usable certificate exists: it exits 0 at once when the certificate has more than ten minutes left, asks the store nothing for ten seconds after a refusal, and otherwise logs in and signs. Nothing in the config depends on the answer; the point is the certificate it leaves behind. Everysshon the agent runs this, so a loop of them while the switch is off costs one login per ten seconds.
The target module
nixosModules.agent-ssh-target (nix/target-modules/agent-ssh.nix) sets
services.hyperhive.agentSshTarget: caPublicKey, written to
/etc/ssh/hyperhive-agent-ssh-ca.pub and named by TrustedUserCAKeys, and
principals.<user>, which becomes users.users.<user>.openssh.authorizedPrincipals.
NixOS then writes /etc/ssh/authorized_principals.d/<user> and sets
AuthorizedPrincipalsFile to it.
Not covered
Egress. The default egress.forward = "log" lets a direct connection from an
agent to a machine's SSH port through, and reject blocks it; there is no
per-agent allowance for it yet.
Revocation
Unticking the switch stops signing as described above. Destroying an agent
disables its identity entity and deletes its cert-auth role
(agent_cutoff.rs), after which its tokens stop working at once. A target's
principals list is the immediate control per machine. A certificate already
signed stays valid for at most max_ttl (180 minutes).