hyperhive/nix/module-eval.nix
atlas 7eb5c92d15 fix(#3849): scrape the path openbao actually serves
The store's scrape target rendered as `host:port` alone, so the collector
requested `/metrics` — a path openbao does not serve on any listener. Every
scrape since the target landed has 404'd, which reads as a dead exporter
rather than a wrong address, and no bao sample has ever reached the store.

`scrapeTargets` values may now carry a path and query. Both are optional and
omitted when absent, so every existing target renders the config it rendered
before. The query cannot ride in `metrics_path`: prometheus percent-encodes
the `?`, so it has to become `params`.

Measured against openbao 2.6.2 and otelcol-contrib 0.151.0, the pinned
versions: `/metrics` 404s ("this listener only serves the metrics endpoint"),
`/v1/sys/metrics` answers JSON, and only `?format=prometheus` — or an Accept
header the collector happens to send today — returns exposition text. The
query param is the half that does not depend on content negotiation.
2026-09-01 11:10:30 +02:00

453 lines
20 KiB
Nix

# `checks.module-eval` — the flake check that covers **nix**.
#
# Why this exists: every other check in ./checks.nix is a Rust
# derivation, so a `.nix`-only diff moves no hash, every check is a
# cache hit, and `nix flake check` reports green **without evaluating
# what changed**. This one's derivation hash is a function of the
# evaluated *results* below, so a nix change that flips a property
# rebuilds it and the builder fails naming that property.
#
# ## What belongs here, and what does not
#
# Anything expressible as a module `assertion` **should be one instead**:
# an assertion fires at deploy time for a real operator, not only in CI.
# What cannot be an assertion is the **absence class** — "a hive that
# hasn't opted in renders exactly what it did before", "this unit does
# not exist unless X". Those are claims about the *rendered config*
# rather than about a config being invalid, so they need an evaluator.
#
# ⚠️ **Cases are named by the PROPERTY they defend, never by the ticket
# that prompted them.** A case named after a ticket has the ticket's
# lifetime; a case named after a property lives as long as the property.
#
# ⚠️ **This check evaluates. It does not execute.** Where the artifact is
# a command line, an HTTP request or a certificate, a value assertion
# cannot stand in — those need something that *runs* them. And a case
# that needs a **rendered file** must stub the packages that file drags
# in (`swarm.ui.package = pkgs.emptyDirectory`), or it costs a full
# frontend build to answer a question about a listen directive.
{
pkgs,
lib,
self,
nixosSystem,
}:
let
# Stub host, same shape ./docs/default.nix already uses: enough for a
# `nixosSystem` to evaluate, nothing that pulls a real disk or
# bootloader in.
hive =
extra:
(nixosSystem {
system = pkgs.stdenv.hostPlatform.system;
modules = [
self.nixosModules.default
{
fileSystems."/" = {
device = "/dev/null";
fsType = "tmpfs";
};
boot.loader.grub.enable = false;
system.stateVersion = "25.11";
# `recursiveUpdate`, not `//`: a plain `//` only merges the
# *top-level* keys of `extra` in, so an `extra` that touches a
# nested attr under an existing top-level key (e.g. `swarm.*`)
# silently drops every sibling under that key instead of merging
# into it — the same class of bug as the `services.hyperhive`
# double-nesting mistake this stub already had to dodge once,
# just one level down. `recursiveUpdate` merges nested attrsets
# all the way down instead.
services.hyperhive = lib.recursiveUpdate {
enable = true;
hiveName = "h1";
swarm.domain = "t.local";
swarm.hives.h1.domain = "h1.t.local";
} extra;
}
];
}).config;
allLocal = hive { deploy.singleHostSwarm = true; };
bare = hive { };
withCi = hive { deploy.forgejo.ci.enable = true; };
baoPkcs11 = hive {
deploy.bao.enable = true;
deploy.bao.seal = "pkcs11";
};
baoShamir = hive {
deploy.bao.enable = true;
deploy.bao.seal = "shamir";
};
baoExplicitCerts = hive {
deploy.bao.enable = true;
deploy.bao.serverCertFile = "/etc/pki/bao.pem";
deploy.bao.serverKeyFile = "/etc/pki/bao-key.pem";
};
# The store and a service that reads from it, versus the store alone. The
# pair is what makes the reader's absence arm mean anything.
baoWithMatrix = hive {
deploy.bao.enable = true;
deploy.matrix.enable = true;
};
# A hive that reads from a store it does not run: no `deploy.bao.enable`, so
# nothing here mints a leaf and the operator names one placed by hand. The
# deployment this pairing exists to serve, and the one that was previously
# inexpressible — the gate asked whether the store was a neighbour.
baoRemoteReader = hive {
deploy.matrix.enable = true;
deploy.bao.clientCertFile = "/etc/pki/bao-client.pem";
deploy.bao.clientKeyFile = "/etc/pki/bao-client-key.pem";
};
# The same hive with the identity taken away, which separates "a homeserver
# is deployed" from "this host can authenticate to the store".
matrixNoBaoIdentity = hive { deploy.matrix.enable = true; };
baoNames = machine: machine.services.hyperhive.gateway.localNames;
baoTwoAddresses = hive {
deploy.bao.enable = true;
deploy.bao.extraListenAddresses = [ "10.0.0.1" ];
# Pinned, not incidental: the case counting these listeners is about the
# declared addresses, and a collector on this host would add one of its own.
deploy.swarm-otel.enable = false;
};
# The store with and without a collector on the same host. `scrapeTargets`
# is only ever read by a local collector, so the metrics endpoint is a
# function of the pairing rather than of the store.
baoWithCollector = hive {
deploy.bao.enable = true;
deploy.swarm-otel.enable = true;
# A second job declared as bare `host:port`, so the pair of cases below
# reads one rendered scrape list: the store's entry carries a path, this
# one carries none.
swarm.otel.scrapeTargets.plain = "127.0.0.1:9999";
};
baoNoCollector = hive {
deploy.bao.enable = true;
deploy.swarm-otel.enable = false;
};
# The config file openbao parses, not the nix that produces it: a setting it
# requires is absent here without anything in the module system minding, so
# the daemon's own startup is otherwise the first reader.
baoSettings = machine: machine.containers.swarm-bao.config.services.openbao.settings;
# The store's units live inside its container, so the gates below have to
# look there rather than at the host's service set.
baoUnits = machine: machine.containers.swarm-bao.config.systemd.services;
# The scrape list prometheus is handed, not the option a service declared:
# the address, the path and the query are one string on the way in and three
# fields on the way out, and only the second shape is what gets requested.
scrapeJob =
machine: job:
lib.findFirst (c: c.job_name == job) null
machine.containers.swarm-otel.config.services.opentelemetry-collector.settings.receivers.prometheus.config.scrape_configs;
# 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
# both necessary and sufficient. `deepSeq` over-specifies this: it keeps
# walking *into* the resulting value after the merge already succeeded,
# and a package/derivation-shaped value's `override`/`overrideAttrs`
# self-reference sends it into nixpkgs' fixpoint machinery and blows the
# stack (measured — this is not a hypothetical).
forceCiServiceConfigs =
let
svcs = withCi.containers.hive-ci.config.systemd.services;
vals = lib.concatMap (s: builtins.attrValues (s.serviceConfig or { })) (builtins.attrValues svcs);
in
builtins.foldl' (acc: v: builtins.seq v acc) true vals;
# Each case: a name stating the property, and `ok`.
cases = [
{
name = "a hive that has not opted into all-local runs no swarm controller";
ok = !bare.services.hyperhive.deploy.swarm-controller.enable;
}
{
name = "the all-local mode turns the swarm controller on";
ok = allLocal.services.hyperhive.deploy.swarm-controller.enable;
}
{
# The gateway's per-name issuer choice. If this ever collapses to a
# constant, every swarm-service vhost serves a certificate its CA
# is name-constrained out of — which evaluates cleanly and fails in
# a browser.
name = "a swarm service name gets the swarm-services leaf and the default server does not";
ok =
let
l = allLocal.services.hyperhive.gateway.lib;
in
(l.tlsFor "t.local").sslCertificate != (l.tlsFor "_").sslCertificate;
}
{
# nixos asserts when a vhost declares both, so this is also a
# statement that the `removeAttrs` upstream of it still happens.
name = "the swarm UI vhost forces TLS instead of merely adding it";
ok =
let
v = allLocal.services.nginx.virtualHosts."t.local";
in
v.forceSSL && !(v.addSSL or false);
}
{
name = "a hive with matrix off serves no matrix discovery endpoint";
ok =
!(builtins.hasAttr "= /.well-known/matrix/client" bare.services.nginx.virtualHosts."_".locations);
}
{
# main got eval-borked twice by this exact class of bug (once on the
# unit's `Restart` key, once on `RestartSec`) — a nixpkgs bump to
# `gitea-actions-runner.nix` adds a plain `serviceConfig.*`
# definition that collides with one of ours, and nix refuses to
# merge two plain definitions at *host* eval. No other check
# instantiates a host with `containers.hive-ci` actually enabled, so
# the collision only surfaces on operator deploy, not in CI.
name = "the CI container's unit definitions merge without a priority collision";
ok = forceCiServiceConfigs;
}
{
# The store's seal is spread over six gates — the stanza, the
# provisioning unit, two bind mounts, a device and an EnvironmentFile.
# Rendering only some of them is the dangerous state: a store that
# says hardware-backed and seals with a software key, which no
# assertion can catch because every value is individually valid.
name = "a shamir store renders no TPM provisioning unit";
ok = !(baoUnits baoShamir ? swarm-bao-token);
}
{
# Presence control for the case above. Without it, a typo in the
# option name would satisfy the absence arm forever. The second half is
# the fix itself: the unit has to run where openbao's `DynamicUser` is
# allocated, and a host unit writing the same bytes has no name to hand
# them to.
name = "a pkcs11 store provisions the token in the container, not on the host";
ok = (baoUnits baoPkcs11 ? swarm-bao-token) && !(baoPkcs11.systemd.services ? swarm-bao-token);
}
{
# `allowedDevices` renders `DeviceAllow=` and nothing else — nspawn
# mounts its own /dev and cannot create device nodes, so permission to
# use a device that was never bound in opens nothing. Neither half
# fails on its own, which is why they are asserted as a pair.
name = "a pkcs11 store gets the TPM device bound in, not merely allowed";
ok =
let
c = baoPkcs11.containers.swarm-bao;
in
(c.bindMounts ? "/dev/tpmrm0") && builtins.any (d: d.node == "/dev/tpmrm0") c.allowedDevices;
}
{
# The stanza's label and the label the unit creates are two literals that
# have to name one object, and the mechanism is asserted with them
# because it is valid only for an RSA key: openbao takes AES-GCM or
# RSA-OAEP and a TPM 2.0 has neither GCM nor an opinion about which the
# seal asked for. Every wrong combination renders and deploys, and
# surfaces as a pkcs11 error at `operator init`.
name = "the seal asks for the RSA key the provisioning unit creates";
ok =
let
p = (baoSettings baoPkcs11).seal.pkcs11;
in
(p.mechanism or "") == "CKM_RSA_PKCS_OAEP"
&& lib.hasInfix "--algorithm=rsa2048 --key-label=${p.key_label or ""}" (
(baoUnits baoPkcs11).swarm-bao-token.script or ""
);
}
{
# `DynamicUser` implies `ProtectSystem=strict`, so the token directory is
# read-only to the seal however it is owned, and the group is the only
# handle on a uid allocated at start. Dropping either surfaces as a
# pkcs11 error deep in a library, naming neither the mount nor the user.
name = "the store's seal may write the token directory, and is in both its groups";
ok =
let
sc = (baoUnits baoPkcs11).openbao.serviceConfig;
in
# `or [ ]` rather than a bare select: the interesting mutation is the
# key being gone, and a select would abort the whole run with a nix
# trace instead of failing this case by name.
builtins.elem "/var/lib/swarm-bao-token" (sc.ReadWritePaths or [ ])
&& builtins.elem "swarm-bao-token" (sc.SupplementaryGroups or [ ])
&& builtins.elem "swarm-bao-tpm" (sc.SupplementaryGroups or [ ]);
}
{
# The device node belongs to the HOST and is matched by NUMBER, while the
# unit that opens it lives in the container — so the two sides holding
# the same gid is the entire mechanism. Letting either side auto-allocate
# renders cleanly, deploys cleanly, and leaves a 0660 node the seal
# cannot open. Compared rather than each checked against a literal: the
# property is that they AGREE, not what they agree on.
name = "the TPM group has the same gid on the host and inside the container";
ok =
let
host = baoPkcs11.users.groups.swarm-bao-tpm.gid or null;
inner = baoPkcs11.containers.swarm-bao.config.users.groups.swarm-bao-tpm.gid or null;
in
host != null && host == inner;
}
{
# Absence arm for the case above — a shamir store never opens a TPM, so
# it must not claim a device node's group. Without this, pinning the gid
# unconditionally would look identical.
name = "a shamir store claims no TPM device group";
ok = !(baoShamir.users.groups ? swarm-bao-tpm);
}
{
# The store's mTLS identity is a separate trust domain from both CAs in
# this tree, because it must not come from an authority the store will
# itself distribute. What supplies it is the glue, which mints a CA of
# the store's own — so an enabled store has all three paths, and if this
# ever reads null again the store stops coming up on its own.
name = "a deployed store is given its own certificate, key and client CA";
ok =
let
b = baoPkcs11.services.hyperhive.deploy.bao;
in
b.serverCertFile != null && b.serverKeyFile != null && b.clientCaFile != null;
}
{
# Everything the glue sets is `mkDefault`, and this is the case that
# says so: a deployment whose certificates come from somewhere the glue
# has never heard of must win. Also the presence control for the case
# above — a renamed option would read `null` on both and satisfy
# neither, but only this one names a value.
name = "an operator's own certificate path beats the glue's default";
ok = baoExplicitCerts.services.hyperhive.deploy.bao.serverCertFile == "/etc/pki/bao.pem";
}
{
# The store's first reader. Its unit belongs to the pairing, not to
# either service: matrix must not learn the store exists, and the store
# must not know who reads it.
name = "a store deployed beside the homeserver fetches its registration token";
ok = baoWithMatrix.systemd.services ? swarm-bao-matrix-token;
}
{
# Absence arm. A store with nothing to serve renders no reader, so the
# unit is a function of the PAIRING rather than of the store — which is
# the property that makes it glue instead of a feature of either side.
name = "a store with no homeserver beside it renders no token reader";
ok = !(baoPkcs11.systemd.services ? swarm-bao-matrix-token);
}
{
name = "a hive that names a client identity reads from a store it does not run";
ok = baoRemoteReader.systemd.services ? swarm-bao-matrix-token;
}
{
# Absence arm for the one above, and the reason the gate is the identity
# rather than the homeserver: without it, deploying matrix anywhere would
# render a reader that cannot authenticate.
name = "a homeserver with no way to authenticate to the store renders no token reader";
ok = !(matrixNoBaoIdentity.systemd.services ? swarm-bao-matrix-token);
}
{
# `Requires=` on a unit that does not exist fails the job, and nothing
# local mints certificates off-host — so this orders against nothing.
# Eval cannot see that failure; only the empty list here stands in for it.
name = "an off-host reader requires no unit the store's host would have provided";
# Membership first, then the value: indexing a missing unit throws, and a
# table that reports which property broke must not be the thing that dies.
ok =
let
s = baoRemoteReader.systemd.services;
in
s ? swarm-bao-matrix-token && s.swarm-bao-matrix-token.requires == [ ];
}
{
# Presence control for the case above: the list is conditional, not gone.
name = "a co-located reader still orders after the local pki unit";
ok =
let
s = baoWithMatrix.systemd.services;
in
s ? swarm-bao-matrix-token && s.swarm-bao-matrix-token.requires == [ "swarm-bao-pki.service" ];
}
{
# The name a reader dials has to resolve where the store runs; a
# multi-host swarm resolves it upstream instead.
name = "the store's host answers for the store's name";
ok = builtins.elem "bao.t.local" (baoNames baoPkcs11);
}
{
# Absence arm, and the one that matters: claiming a name this host does
# not serve points every local reader at the wrong machine.
name = "a hive that does not run the store claims no name for it";
ok = !(builtins.elem "bao.t.local" (baoNames bare));
}
{
# Raft refuses to start without it, and says so in a message that names
# neither the setting nor the stanza.
name = "the store advertises a cluster address";
ok = lib.hasPrefix "https://" ((baoSettings baoPkcs11).cluster_addr or "");
}
{
# Control for the case above: these settings are rendered per deployment,
# not constants a passing case could be indifferent to.
name = "a declared extra address renders a second listener beside loopback";
ok = builtins.length (builtins.attrNames (baoSettings baoTwoAddresses).listener) == 2;
}
{
# Retention is what serves the endpoint at all, so the listener alone
# would be a port that answers 404.
name = "a store beside a collector serves metrics on its own listener";
ok =
let
s = baoSettings baoWithCollector;
in
s.listener ? metrics && (s.telemetry.prometheus_retention_time or "0s") != "0s";
}
{
# openbao serves no `/metrics` at all, so a scrape of the default path
# 404s: the store looks like a dead exporter, and every panel built on
# it renders empty rather than erroring.
name = "the store's scrape asks for the path openbao serves";
ok =
let
j = scrapeJob baoWithCollector "bao";
in
(j.metrics_path or "") == "/v1/sys/metrics" && (j.params.format or [ ]) == [ "prometheus" ];
}
{
# Presence control for the case above: both fields are omitted rather
# than defaulted, so a target declared as bare `host:port` renders what
# it rendered before the path grammar existed.
name = "a target with no path renders neither metrics_path nor params";
ok =
let
j = scrapeJob baoWithCollector "plain";
in
j != null && !(j ? metrics_path) && !(j ? params);
}
{
# Absence arm. Unauthenticated by design, so it must not exist where
# nothing reads it.
name = "a store with no collector beside it serves no metrics";
ok =
let
s = baoSettings baoNoCollector;
in
!(s.listener ? metrics) && !(s ? telemetry);
}
];
bad = builtins.filter (c: !c.ok) cases;
# Escaped, because a name is prose and prose contains apostrophes. Hand-quoting
# broke the builder mid-report on the first such name that failed — and only
# ever on failure, so every green run agreed the reporter was fine.
report = lib.concatMapStringsSep "\n" (
c: " echo ${lib.escapeShellArg "FAILED: ${c.name}"} >&2"
) bad;
in
# The results are embedded in the builder text on purpose: that is what
# makes this derivation's hash depend on them, so a nix-only change that
# flips a case cannot be answered from cache.
pkgs.runCommand "hyperhive-module-eval" { } ''
${report}
${
if bad == [ ] then
"echo '${toString (builtins.length cases)} module properties hold' && touch $out"
else
"echo 'module-eval: ${toString (builtins.length bad)} of ${toString (builtins.length cases)} properties broke' >&2 && exit 1"
}
''