hyperhive/nix/host-modules/glue-bao-readers-policy-order.nix
atlas 92e1909caf swarm-bao: give the store forwarder's OIDC reader its own bao identity
`swarm-bao-forwarder-oidc` fetches the store container's collector secret,
one path, and was the last reader still logging in with
`deploy.bao.clientCertFile`: the hive's own leaf, whose policy reads every
agent's credentials, the hive's tree and every service's OIDC secret. The
four-way split gave grafana's and the swarm collector's readers leaves of
their own and left this one behind.

It now holds `forwarder-oidc.pem`, minted by `swarm-bao-pki`, and logs in
under the `swarm-forwarder-oidc` cert-auth role, whose policy reads
`secret/data/swarm/services/<store forwarder client id>/oidc/client` and
nothing else. The role is written by `swarm-bao-forwarder-oidc-policy`
from the bootstrap token, which gains the two grants that unit calls, and
the reader is ordered after it. The subject is reserved as a hive name. A
store host whose pair is null is refused at eval rather than falling back
to the hive's leaf.

The hive's own role and `client.pem` are untouched; nothing is revoked.
2026-09-25 00:37:31 +02:00

65 lines
2.5 KiB
Nix

# Glue: where the store and one of its readers share a host, the reader waits
# for the unit that writes the cert-auth role it logs in with.
#
# ONE PAIRING PER FILE — the store's reader policy units ← the readers
# they grant, and nothing else. Deleting this leaves every reader as it is on a
# host whose store is remote: it may log in before its role exists, and its own
# retries are what carry it past that.
#
# ⚠️ Gated on BOTH the store and that reader being here. Off the store's host
# there is no local policy unit to order against. On the store's host without
# the reader, setting `systemd.services.<reader>.after` would define a unit
# with no ExecStart, so each gate below restates the one the reader's own
# module puts on it. A reader whose gate changes must change here too.
#
# Ordering, never a requirement: a policy unit skips once the bootstrap token
# is gone, and a skipped unit counts as done. `wants` as well as `after`, so a
# reader started on its own pulls its policy unit into the same transaction.
{
lib,
config,
...
}:
let
hyperhiveCfg = config.services.hyperhive;
deployCfg = hyperhiveCfg.deploy;
baoDeploy = deployCfg.bao;
havePair = cert: key: cert != null && key != null;
# Reader unit → the gate its own module defines it under.
readersHere = {
# ./glue-matrix-bao-token.nix
swarm-bao-matrix-token =
hyperhiveCfg.enable
&& havePair baoDeploy.matrixTokenClientCertFile baoDeploy.matrixTokenClientKeyFile
&& deployCfg.matrix.enable;
# ./glue-queue-agent-credential.nix
swarm-bao-queue-agent =
hyperhiveCfg.enable
&& havePair baoDeploy.queueAgentClientCertFile baoDeploy.queueAgentClientKeyFile;
# ./swarm-grafana.nix
swarm-bao-grafana-oidc = hyperhiveCfg.enable && deployCfg.grafana.enable;
# ./swarm-otel.nix
swarm-bao-otel-oidc =
deployCfg.swarm-otel.enable
&& havePair baoDeploy.otelOidcClientCertFile baoDeploy.otelOidcClientKeyFile;
# ./swarm-bao.nix: its block is gated on the store, which `orderAfterPolicy`
# already checks.
swarm-bao-forwarder-oidc = hyperhiveCfg.enable;
# ./swarm-nats.nix: the queue's TLS leaf, not a secret, but the same wait.
swarm-bao-nats-tls = deployCfg.nats.enable;
};
orderAfterPolicy =
reader: here:
lib.mkIf (baoDeploy.enable && here) {
systemd.services.${reader} = {
after = [ "${reader}-policy.service" ];
wants = [ "${reader}-policy.service" ];
};
};
in
{
config = lib.mkMerge (lib.mapAttrsToList orderAfterPolicy readersHere);
}