Galaxy IPC unblock — live dev-box E2E path

Three root-cause fixes to get an elevated dev-box shell past session open
through to real MXAccess reads:

1. PipeAcl — drop BUILTIN\Administrators deny ACE. UAC's filtered token
   carries the Admins SID as deny-only, so the deny fired even from
   non-elevated admin-account shells. The per-connection SID check in
   PipeServer.VerifyCaller remains the real authorization boundary.

2. PipeServer — swap the Hello-read / VerifyCaller order. ImpersonateNamedPipeClient
   returns ERROR_CANNOT_IMPERSONATE until at least one frame has been read
   from the pipe; reading Hello first satisfies that rule. Previously the
   ACL deny-first path masked this race — removing the deny ACE exposed it.

3. GalaxyIpcClient — add a background reader + single pending-response
   slot. A RuntimeStatusChange event between OpenSessionRequest and
   OpenSessionResponse used to satisfy the caller's single ReadFrameAsync
   and fail CallAsync with "Expected OpenSessionResponse, got
   RuntimeStatusChange". The reader now routes response kinds (and
   ErrorResponse) to the pending TCS and everything else to a handler the
   driver registers in InitializeAsync. The Proxy was already set up to
   raise managed events from RaiseDataChange / RaiseAlarmEvent /
   OnHostConnectivityUpdate — those helpers had no caller until now.

4. RedundancyPublisherHostedService — swallow BadServerHalted while
   polling host.Server.CurrentInstance. StandardServer throws that code
   during startup rather than returning null, so the first poll attempt
   crashed the BackgroundService (and the host) before OnServerStarted
   ran. This race was latent behind the Galaxy init failure above.

Updates docs that described the Admins deny ACE + mandatory non-elevated
shells, and drops the admin-skip guards from every Galaxy integration +
E2E fixture that had them (IpcHandshakeIntegrationTests, EndToEndIpcTests,
ParityFixture, LiveStackFixture, HostSubprocessParityTests).

Adds GalaxyIpcClientRoutingTests covering the router's
request/response match, ErrorResponse, event-between-call, idle event,
and peer-close paths.

Verified live on the dev box against the p7-smoke cluster (gen 6):
driver registered=1 failedInit=0, Phase 7 bridge subscribed, OPC UA
server up on 4840, MXAccess read round-trip returns real data with
Status=0x00000000.

Task #112 — partial: Galaxy live stack is functional end-to-end. The
supplied test-galaxy.ps1 script still fails because the UNS walker
encodes TagConfig JSON as the tag's NodeId instead of the seeded TagId
(pre-existing; separate issue from this commit).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Joseph Doherty
2026-04-24 16:30:16 -04:00
parent fb6dd3478d
commit d11dd0520b
17 changed files with 443 additions and 130 deletions

View File

@@ -35,13 +35,14 @@ Multi-project test topology:
## How tests skip
- **E2E parity**: `ParityFixture.SkipIfUnavailable()` runs at class init and
checks Windows-only, non-admin user, ZB SQL reachable on
`localhost:1433`, Host EXE built in the expected `bin/` folder. Any miss
→ tests skip.
checks Windows-only, ZB SQL reachable on `localhost:1433`, Host EXE built
in the expected `bin/` folder. Any miss → tests skip.
- **Live-smoke** (`GalaxyRepositoryLiveSmokeTests`): `Assert.Skip` when ZB
unreachable. A `per project_galaxy_host_installed` memory on this repo's
dev box notes the MXAccess runtime is installed + pipe ACL denies Admins,
so live tests must run from a non-elevated shell.
dev box notes the MXAccess runtime is installed. The pipe ACL allows the
configured SID outright; elevation of the caller doesn't matter because
the per-connection SID check in `PipeServer.VerifyCaller` only compares
user SIDs (not group membership or integrity level).
- **Unit** tests (Shared, Proxy contract, most Host.Tests) have no skip —
they run anywhere.

View File

@@ -31,7 +31,7 @@ The same Tier-C isolation story applies to FOCAS (decision record in `docs/v2/pl
- Pipe name: `otopcua-galaxy-{DriverInstanceId}` (localhost-only, no TCP surface)
- Wire format: MessagePack-CSharp, length-prefixed frames
- ACL: pipe is created with a DACL that grants only the Server's service identity; the Admins group is explicitly denied so a live-smoke test running from an elevated shell fails fast rather than silently bypassing the handshake
- ACL: pipe is created with a DACL that grants `ReadWrite | Synchronize` only to the configured Server service-principal SID + denies `LocalSystem`. The per-connection SID check in `PipeServer.VerifyCaller` is the real authorization boundary — any caller whose impersonated token SID doesn't match the allowed SID is dropped before the first frame is read.
- Handshake: Proxy presents a shared secret at `OpenSessionRequest`; Host rejects anything else with `MessageKind.OpenSessionResponse{Success=false}`
- Heartbeat: Proxy sends a periodic ping; missed heartbeats trigger the Proxy-side crash-loop supervisor to restart the Host