Watch
0
0
Fork
You've already forked hyperhive
0

sign ssh certificates on demand from ssh itself, warn before setup, say what to do when a host key changes

This commit is contained in:
atlas 2026-11-16 19:42:44 +01:00 • committed by mara
commit 3d7541c0af

View file

@ -3,18 +3,26 @@
#
# On every agent that has a store address (the same switch as ./bao.nix: with no
# store there is nothing to sign with). The agent has no standing SSH credential:
# `${unitName}.service` generates an ed25519 key in its own `/run` directory
# (tmpfs, never on disk), logs in to the swarm secret store with the identity
# ./bao.nix keeps current, and has the public half signed by the store's SSH
# engine (`swarm_secret_client::policy::SSH_MOUNT`). The role it signs through
# issues a user certificate with this agent's name as its only principal, valid
# for an hour; a timer signs again well before that runs out.
# the helper below generates an ed25519 key in `/run/${unitName}` (tmpfs, never
# on disk), logs in to the swarm secret store with the identity ./bao.nix keeps
# current, and has the public half signed by the store's SSH engine
# (`swarm_secret_client::policy::SSH_MOUNT`). The role it signs through issues a
# user certificate with this agent's name as its only principal, valid for an
# hour.
#
# Whether the store signs is the operator's per-agent switch in the swarm UI
# (the `ssh` permission, `swarm-controller`'s `agent_ssh`). With it off the store
# refuses, the unit logs one line and ends successfully, and the next tick asks
# again; with it on the agent has a certificate within one tick. The unit
# needs no change to be switched.
# (the `ssh` permission, `swarm-controller`'s `agent_ssh`). Every signing makes a
# fresh login, so a token never outlives the policies the role had when it was
# issued by more than the helper's own run. Two things run the helper:
#
# - `${unitName}.service`, from a timer, renews well before the certificate
# runs out. With the switch off the store refuses: the unit logs one line, ends
# successfully and asks again at the next tick.
# - `ssh` itself, through a `Match exec` line in the system config: when the
# certificate is missing or close to its end, the helper signs before the
# connection starts. That is what makes the first `ssh` after the operator
# switches SSH on work at once. It is silent and its exit status is whether a
# usable certificate exists; nothing in the config depends on the answer.
#
# A target machine trusts the CA, not this key: ../target-modules/agent-ssh.nix.
{
@ -38,28 +46,57 @@ let
unitName = "hive-agent-ssh-cert";
runtimeDir = unitName;
keyFile = "/run/${runtimeDir}/key";
dir = "/run/${runtimeDir}";
keyFile = "${dir}/key";
pubFile = "${keyFile}.pub";
# `ssh` reads `<IdentityFile>-cert.pub` by itself, but the config names it so
# that the file a Host block uses is visible in the config.
certFile = "${keyFile}-cert.pub";
stagingFile = "/run/${runtimeDir}/cert.new";
errFile = "/run/${runtimeDir}/bao.err";
stagingFile = "${dir}/cert.new";
errFile = "${dir}/bao.err";
# The store's CA, copied out of the unit's credentials so that the helper can
# reach the store from an `ssh` run, which has no credentials directory.
serverCaCopy = "${dir}/server-ca.pem";
# Written when the store refuses, so that a loop of `ssh` calls does not log in
# to the store each time the switch is off.
refusedFile = "${dir}/refused";
# In the state dir, which is bound from the host, so a pin survives a restart.
knownHostsFile = "/agents/${agentName}/state/ssh_known_hosts";
signPath = "${ca.mountPath}/sign/${ca.roleName}";
# A certificate with less than this left is signed again by `ssh`. The timer
# fires every `ca.renewEveryMinutes`, so a healthy agent never gets here.
ensureMarginSec = 600;
# How long a refusal is remembered. The delay the operator sees after
# switching on is at most this, then one login and one signing.
refusedForSec = 10;
storeRetry = import ../host-modules/lib/store-retry.nix { };
binPath = lib.makeBinPath [
pkgs.openbao
pkgs.openssh
pkgs.coreutils
pkgs.gnused
pkgs.util-linux
];
# The helper, and the unit's script: one text. `HIVE_SSH_CERT_ENSURE=1` is the
# `ssh` entry; without it this is the timer's. The tools go after the caller's
# `PATH`, so a caller that has its own wins.
helper = pkgs.writeShellScript unitName config.systemd.services.${unitName}.script;
in
{
config = lib.mkIf (baoCfg.addr != null) {
# Every host: the certificate, and the pins in the state dir, which outlives
# the container root. `accept-new` records a host's key at first contact and
# refuses a changed one after that.
# refuses a changed one after that. The `Match` line has no settings: ssh
# runs its command while it reads the config, before it loads any identity,
# and the answer is not used.
programs.ssh.extraConfig = ''
Match exec "HIVE_SSH_CERT_ENSURE=1 ${helper} </dev/null"
Host *
IdentityFile ${keyFile}
CertificateFile ${certFile}
@ -77,11 +114,6 @@ in
"hive-agent-bao-identity.service"
];
wantedBy = [ "multi-user.target" ];
path = [
pkgs.openbao
pkgs.openssh
pkgs.coreutils
];
# ../host-modules/lib/store-retry.nix.
inherit (storeRetry) startLimitBurst startLimitIntervalSec;
serviceConfig = storeRetry.serviceConfig // {
@ -98,19 +130,77 @@ in
UMask = "0377";
LoadCredential = [ serverCaCredential ];
};
environment = {
BAO_ADDR = toString baoCfg.addr;
BAO_CLIENT_CERT = baoCfg.clientCertFile;
BAO_CLIENT_KEY = baoCfg.clientKeyFile;
};
script = ''
set -euo pipefail
# `UMask=0377` in the unit makes every file this script creates `0400`,
# so each one is removed before the command that creates it. `ssh` runs
# the same text, which therefore sets it too.
umask 0377
export PATH="$PATH:${binPath}"
export BAO_ADDR=${lib.escapeShellArg (toString baoCfg.addr)}
export BAO_CLIENT_CERT=${lib.escapeShellArg baoCfg.clientCertFile}
export BAO_CLIENT_KEY=${lib.escapeShellArg baoCfg.clientKeyFile}
dir=${lib.escapeShellArg dir}
key=${lib.escapeShellArg keyFile}
pub=${lib.escapeShellArg pubFile}
cert=${lib.escapeShellArg certFile}
staged=${lib.escapeShellArg stagingFile}
err=${lib.escapeShellArg errFile}
server_ca=${lib.escapeShellArg serverCaCopy}
refused=${lib.escapeShellArg refusedFile}
ensure=
if [ -n "''${HIVE_SSH_CERT_ENSURE-}" ]; then ensure=1; fi
# The `ssh` entry is silent and says yes or no by its exit status; the
# unit's output is its log.
if [ -n "$ensure" ]; then exec >/dev/null 2>&1; fi
# `ssh` can run before the unit's first start has made the directory.
if [ ! -d "$dir" ]; then mkdir -m 0700 "$dir"; fi
# One run at a time, the unit's or an `ssh`'s: two would race on the key
# and the staged file.
exec {lock}<"$dir"
if ! flock -w 20 "$lock"; then
echo "another run holds $dir and did not finish in 20 seconds." >&2
exit 1
fi
# Epoch second the certificate ends, 0 when there is none.
cert_end() {
local end
end=$(ssh-keygen -L -f "$cert" 2>/dev/null | sed -n 's/^ *Valid: from .* to //p') || end=
if [ -n "$end" ]; then date -d "$end" +%s 2>/dev/null || echo 0; else echo 0; fi
}
if [ -n "$ensure" ]; then
now=$(date +%s)
# Enough left: nothing to do, and nothing asked of the store.
if [ "$(cert_end)" -gt $((now + ${toString ensureMarginSec})) ]; then exit 0; fi
# The store refused a moment ago: the switch is off.
if [ -e "$refused" ] && [ $((now - $(stat -c %Y "$refused"))) -lt ${toString refusedForSec} ]; then
exit 1
fi
fi
# The unit has the credentials directory and keeps the store's CA where
# an `ssh` run can find it.
if [ -n "''${CREDENTIALS_DIRECTORY-}" ] && [ -s "$CREDENTIALS_DIRECTORY/${serverCaCredential}" ]; then
rm -f "$server_ca"
cp "$CREDENTIALS_DIRECTORY/${serverCaCredential}" "$server_ca"
fi
if [ -s "$server_ca" ]; then export BAO_CACERT="$server_ca"; fi
trap 'rm -f "$err" "$staged"' EXIT
# No identity delivered: ./bao.nix's check reports that; saying it
# twice adds nothing.
for f in "$BAO_CLIENT_CERT" "$BAO_CLIENT_KEY"; do
if [ ! -s "$f" ]; then
echo "this agent has no store identity, so it cannot get an SSH certificate." >&2
if [ -n "$ensure" ]; then exit 1; fi
exit 0
fi
if [ ! -r "$f" ]; then
@ -119,19 +209,6 @@ in
fi
done
if [ -s "$CREDENTIALS_DIRECTORY/${serverCaCredential}" ]; then
export BAO_CACERT="$CREDENTIALS_DIRECTORY/${serverCaCredential}"
fi
# `UMask=0377` makes every file this script creates `0400`, so each one
# is removed before the command that creates it.
key=${lib.escapeShellArg keyFile}
pub=${lib.escapeShellArg pubFile}
cert=${lib.escapeShellArg certFile}
staged=${lib.escapeShellArg stagingFile}
err=${lib.escapeShellArg errFile}
trap 'rm -f "$err" "$staged"' EXIT
# The key lives as long as /run does, which is as long as this
# container: a restart is a new key. Only a pair with both halves is
# kept.
@ -141,6 +218,8 @@ in
chmod 0400 "$key"
fi
# A fresh login every time: a token carries the policies the role had
# when it was issued, so only a new one sees the operator's switch.
rm -f "$err"
if ! BAO_TOKEN="$(bao login -method=cert -token-only 2>"$err")"; then
echo "the swarm secret store at $BAO_ADDR refused or could not answer this agent's certificate login:" >&2
@ -157,7 +236,10 @@ in
# so no failed unit. Anything else (a sealed or unreachable store) is
# one, and retried.
if [[ "$(<"$err")" == *"Code: 403"* ]]; then
rm -f "$refused"
: >"$refused"
echo "SSH certificates are not enabled for this agent (the swarm store refused to sign); asking again next tick."
if [ -n "$ensure" ]; then exit 1; fi
exit 0
fi
echo "the swarm secret store would not sign this agent's SSH key at ${signPath}:" >&2
@ -173,6 +255,7 @@ in
# One rename: ssh never reads a half-written certificate.
mv -f "$staged" "$cert"
rm -f "$refused"
echo "signed this agent's SSH key as ${agentName}, valid for ${toString ca.certTtlMinutes} minutes."
'';
};