swarm-tls: narrow each gateway's services leaf to the names it fronts

Every gateway asked the store's `pki/issue/swarm-services` for the whole
swarm's service set, so a private key on any gateway host could serve a
valid certificate for services that host does not front and never has.

`swarm.localServiceDomains` derives the per-host subset by filtering
`swarm.serviceDomains` against the vhosts this host actually renders —
the deploy flags those vhosts are already guarded on, read once rather
than copied into a second filter. The leaf request and the coverage
guard that decides whether to re-issue both read it, so they cannot
disagree about which names the leaf owes.

The sub-CA's name constraint and the role's `allowed_domains` stay the
swarm-wide set: every host's subset is inside it, and narrowing the
constraint per host would turn one signing into N.

docs/swarm/ca.md

@@ -54,10 +54,12 @@ copies from. Every hive does this with its own identity, so holding the
 swarm root's private key stopped being what decides whether a hive can
 serve its swarm's names.
 
-The same `services.hyperhive.swarm.serviceDomains` that builds the SANs
+`services.hyperhive.swarm.serviceDomains` populates the role's
-also populates the role's `allowed_domains`, so asking for a name nobody
+`allowed_domains`, so asking for a name nobody configured is a refusal
-configured is a refusal from the store naming that name — not a
+from the store naming that name — not a certificate quietly issued for
-certificate quietly issued for it.
+it. A host's own SANs are narrower still: it asks only for the names it
+fronts a vhost for (`swarm.localServiceDomains`), so the key on one
+gateway can't serve a swarm service that runs behind another.
 
 **The root's public certificate is a file, on every hive:**
 `/var/lib/hive-tls/swarm-services-root.pem` (0644), written beside the

docs/swarm/ui.md

@@ -92,7 +92,8 @@ own options, the way `swarm-ui.nix` and `swarm-authelia.nix` do.
 
 ⚠️ The certificate one is the hardest to predict and the most visible when
 missed. `serviceDomains` is _both_ the `allowed_domains` the secret
-store's `pki/roles/swarm-services` narrows to and the leaf's SAN list,
+store's `pki/roles/swarm-services` narrows to and the list each gateway's
+leaf draws its SANs from (it carries the ones that host fronts),
 and the apex is a **sibling** of `forge.<swarm>` / `chat.<swarm>` /
 `auth.<swarm>`, not a parent — no CA in the hierarchy issues for it
 implicitly. Left out, the vhost falls back to the hive leaf and the

nix/host-modules/hive-tls.nix

@@ -21,10 +21,19 @@ let
   baoServicesPkiMount = baoDeploy.servicesPkiMountPath;
   baoServicesPkiRole = baoDeploy.servicesPkiRoleName;
 
-  # Derived once in ./swarm.nix and read here + in ./swarm-bao.nix, so
+  # Derived once in ./swarm.nix: the swarm service names THIS host fronts
-  # the names this leaf carries as SANs and the names the store's
+  # a vhost for. Every one of them is inside the swarm-wide set that
-  # `pki/roles/swarm-services` is narrowed to cannot disagree.
+  # ./swarm-bao.nix narrows `pki/roles/swarm-services` to, so the leaf
-  swarmServiceDomains = hyperhiveCfg.swarm.serviceDomains;
+  # stays issuable while carrying no name this gateway does not serve.
+  localServiceDomains = hyperhiveCfg.swarm.localServiceDomains;
+
+  # The one name the request's `common_name` carries; the rest ride as
+  # SANs. Not `builtins.head`: a host fronting no swarm service renders
+  # this script too — its unit exits on `want_svc` long before running
+  # the request, but a bare `head []` would have failed the EVALUATION.
+  leafCommonName = lib.optionalString (localServiceDomains != [ ]) (
+    builtins.head localServiceDomains
+  );
 
   # The host-managed hive CA is the trust anchor for self-signed mode.
   # It is only stood up when the gateway actually serves a self-signed
@@ -195,15 +204,16 @@ let
     svcleaf="$d/swarm-services.pem"
     svcroot="$d/swarm-services-root.pem"
     hiveNames=${lib.escapeShellArg "${domain} *.${domain}"}
-    svcNames=${lib.escapeShellArg (lib.concatStringsSep " " swarmServiceDomains)}
+    svcNames=${lib.escapeShellArg (lib.concatStringsSep " " localServiceDomains)}
 
-    # A hive with no configured service names has no services leaf to
+    # A host fronting none of the swarm's service names has no services
-    # hold, and its absence there is the correct state rather than a
+    # leaf to hold, and its absence there is the correct state rather
-    # stale one. Anywhere else it is expected: the leaf comes from the
+    # than a stale one — its own vhosts are covered by the hive leaf
+    # above. Anywhere else it is expected: the leaf comes from the
     # secret store's `pki` mount now, not from a sub-CA that exists on
     # one host in the swarm, so "this host cannot issue it" is no longer
     # one of the answers.
-    want_svc=${if swarmServiceDomains == [ ] then "0" else "1"}
+    want_svc=${if localServiceDomains == [ ] then "0" else "1"}
 
     # Expiry is not the only way a leaf goes wrong. One signed when the
     # configured name set was smaller stays valid for its whole lifetime
@@ -681,10 +691,11 @@ in
     # for it to be an intermediate OF.
     #
     # The names it may carry are not this unit's to assert: the role
-    # refuses anything outside `allowed_domains`, which is read from the
+    # refuses anything outside `allowed_domains`, read from the
-    # same `swarm.serviceDomains` the SANs below are built from. A
+    # swarm-wide `swarm.serviceDomains` that the SANs below are a
-    # mismatch is a refusal from the store naming the offending name,
+    # per-host SUBSET of. A mismatch is a refusal from the store naming
-    # not a certificate quietly issued for something nobody configured.
+    # the offending name, not a certificate quietly issued for something
+    # nobody configured.
     systemd.services.swarm-services-cert = {
       description = "Issue the swarm-services TLS leaf from the secret store's PKI";
       wantedBy = [ "multi-user.target" ];
@@ -847,8 +858,8 @@ in
         # the store, not this host, decides what the certificate says.
         if ! bao write -format=json \
             ${lib.escapeShellArg "${baoServicesPkiMount}/issue/${baoServicesPkiRole}"} \
-            common_name=${lib.escapeShellArg (builtins.head swarmServiceDomains)} \
+            common_name=${lib.escapeShellArg leafCommonName} \
-            alt_names=${lib.escapeShellArg (lib.concatStringsSep "," swarmServiceDomains)} \
+            alt_names=${lib.escapeShellArg (lib.concatStringsSep "," localServiceDomains)} \
             > "$resp" 2>"$err"; then
           echo "the store refused to issue the swarm-services certificate." >&2
           cat "$err" >&2

nix/host-modules/swarm.nix

@@ -299,6 +299,27 @@ in
     '';
   };
 
+  options.services.hyperhive.swarm.localServiceDomains = lib.mkOption {
+    type = lib.types.listOf lib.types.str;
+    readOnly = true;
+    internal = true;
+    description = ''
+      Read-only: the names in `serviceDomains` **this host actually
+      fronts**, which is what its services leaf may carry. The swarm-wide
+      set stays what the sub-CA is name-constrained to and what the
+      store's issuing role permits; narrowing the leaf is what stops a
+      private key on one gateway presenting a valid certificate for
+      services that host does not serve and never has.
+
+      Read off the rendered `services.nginx.virtualHosts` rather than off
+      the per-service deploy flags: those flags are what the vhosts are
+      already guarded on, so a second reading of them is a copy that
+      drifts from the thing it claims to describe. Filtering
+      `serviceDomains` rather than assembling a list beside it is that
+      option's own constraint, and it applies here for the same reason.
+    '';
+  };
+
   config = {
     services.hyperhive.swarm.peerHives = lib.filterAttrs (name: _: name != cfg.hiveName) swarmCfg.hives;
 
@@ -306,6 +327,10 @@ in
       lib.unique (lib.filter (d: d != null && d != "") serviceDomains')
     );
 
+    services.hyperhive.swarm.localServiceDomains = lib.filter (
+      d: config.services.nginx.virtualHosts ? ${d}
+    ) swarmCfg.serviceDomains;
+
     assertions = [
       {
         # An EMPTY `hives` fires this too, deliberately: since

nix/module-eval/core-toggle.nix

@@ -52,6 +52,10 @@ let
     deploy.forgejo.ci.enable = true;
   };
 
+  # A hive whose one gateway-published swarm service sits on another host:
+  # every swarm name still configured, not one of them served here.
+  forgeElsewhere = hive { deploy.forgejo.behindGateway = false; };
+
   # A priority collision is a property of the *option*, not
   # of the merged value's interior — nix throws the moment the value is
   # demanded at all, so `seq`-ing each `serviceConfig` value to WHNF is
@@ -199,6 +203,32 @@ let
         in
         (l.tlsFor "t.local").sslCertificate != (l.tlsFor "_").sslCertificate;
     }
+    {
+      # A leaf is a key that can SERVE every name in it, so the names it
+      # asks for are that key's blast radius. `bare` fronts forge and
+      # nothing else; the last conjunct is the control, since `auth.t.local`
+      # is in the swarm's set and an unnarrowed request would carry it here
+      # too.
+      name = "a gateway's services leaf asks for only the swarm names that host fronts";
+      ok =
+        let
+          s = bare.services.hyperhive.swarm.serviceDomains;
+          u = bare.systemd.services.swarm-services-cert.script;
+        in
+        lib.hasInfix "alt_names=forge.t.local" u
+        && !(lib.hasInfix "auth.t.local" u)
+        && lib.elem "auth.t.local" s;
+    }
+    {
+      # The narrowing's floor: no names left means no request, rather than a
+      # request for an empty SAN set that the store would refuse. Its own
+      # vhosts are the hive leaf's, which this host still signs.
+      name = "a host fronting none of the swarm's service names requests no services leaf";
+      ok =
+        forgeElsewhere.services.hyperhive.swarm.localServiceDomains == [ ]
+        && forgeElsewhere.services.hyperhive.swarm.serviceDomains != [ ]
+        && lib.hasInfix "want_svc=0" forgeElsewhere.systemd.services.swarm-services-cert.script;
+    }
     {
       # nixos asserts when a vhost declares both, so this is also a
       # statement that the `removeAttrs` upstream of it still happens.