mirror of
https://github.com/OPCFoundation/UA-.NETStandard.git
synced 2026-10-16 21:02:51 +08:00
romanett
ececea15c8
# Description Builds on #4428, which is now merged; this branch is rebased onto master. The GDS `ApplicationsNodeManager` becomes a source-generated `[NodeManager]`: the companion model, its loader and the fluent plumbing are emitted, and what stays hand-written is the behaviour. Getting there needed three things from the shared surface, which are the first two commits. ## Commits **1. Add the fluent hooks a hand-written manager needs at startup** - `OnReadRolePermissions` / `OnReadUserRolePermissions` on the node builder. Every other `NodeState` `On*` hook was already fluent; these were not, and a server that grants access on something the static model cannot express has to compute the permission set per request. - `Node<TState>(TState)` returns a builder over a node the caller already holds, for nodes materialised after `Configure` resolved its graph. - `OnAddressSpaceReadyAsync` on `FluentNodeManagerBase` — the one seam between the address space existing and the wiring being applied, for managers whose wiring depends on asynchronous setup. `Configure` is a `partial void` and cannot await. `DiNodeManager` had already invented the hook privately, so its declaration becomes an `override`. **2. Let a `[NodeManager]` bind to a model a referenced assembly supplies** A model belongs to the assembly that emits its types, which leaves a node manager for it homeless whenever that assembly cannot reference `Opc.Ua.Server` — a model-only package shared with clients, for instance. Binding was rejected twice over: the design was skipped before the binding was matched (`MODELGEN010`), and fluent-accessors-only mode refused to run alongside manager generation (`MODELGEN014`). The binding is now matched before the skip decision, and a matched binding puts that design into accessors-only mode. A design nothing binds to is still skipped, so this cannot start duplicating types by accident. The emitted manager also gains what a real manager needs: `DefaultNamespaceUris()` plus a protected constructor taking a replacement set; `GenerateDefaultConstructor=false` to suppress the two-argument form; its own logger category; and the call to `OnAddressSpaceReadyAsync`. Generated fluent accessors also stop re-enabling nullable warnings — a bare `#nullable enable` undid the shared header's `disable warnings`, so any model with a structure-typed method argument reported CS8600 on every generated `out T` in a project that builds warnings as errors. **3. Reimplement the GDS `ApplicationsNodeManager` as a generated node manager** The design stays owned by `Opc.Ua.Gds.Common`; the server assembly only binds a manager to it. **NodeIds do not move.** `NamespaceIndexes[0]` must stay the application record namespace — it is the one the application database, the certificate request store and the base allocator mint ids in — so the constructor passes the namespace order explicitly rather than adopting the generated model-first default. A test pins both ends against a live server. Four overrides go, leaving `OnAddressSpaceReadyAsync` and `Dispose(bool)` — no address-space overrides at all: | Removed | Why it was unnecessary | | --- | --- | | `New` | The base allocator already mints into `NamespaceIndexes[0]`, with a thread-safe counter. The override existed only because `Create(assignNodeIds: true)` asks the factory with the declaration id still set; `AssignInstanceNodeId` nulls it and re-asks, documented as being there for exactly this allocator. | | `AddBehaviourToPredefinedNodeAsync` | Its `AuthorizationServiceType` branch re-wired a node that was already wired, and its `KeyCredentialServiceType` branch had nothing to wire — the model ships that folder empty. | | `GetManagerHandleAsync`, `ValidateNodeAsync` | Verbatim reimplementations of the base that also blocked the fluent base's virtual-node support. | | `DeleteAddressSpaceAsync` | A `// TBD` passthrough. | Everything else — all sixteen directory methods with their self-administration permissions, the certificate groups' concrete types and trust-list flags, and the authorization service (created through this PR's `Add` surface and wired in the same place) — is one `Configure` pass. **4. Bind the certificate groups from the `Configure` pass** Acquiring a certificate group and binding it to its nodes are different kinds of work: the first is I/O (`InitAsync` opens certificate stores and creates CA certificates), the second only touches the address space. Each now runs in the phase it belongs to, which puts the whole of the wiring in one place. Ownership no longer rides on the binding. A group joins `m_ownedCertificateGroups` the moment it is constructed, before `InitAsync` can throw, so a group that fails half way through is still disposed — which the previous code could not manage, keying ownership off a dictionary it filled only after initialization succeeded. > Commit 3's message says this move is impossible and blames `PushTest`. **That was a wrong diagnosis, and this commit supersedes it.** Two real defects were hiding behind those failures: - **`Create(..., NodeId.Null, ..., assignNodeIds: false)` does not null the NodeId.** `NodeState.Create` overrides it only when handed a non-null one, so a custom group node kept the namespace-0 id of its type declaration. `AssignInstanceNodeId` hid this by re-asking the factory with the id nulled; the builder validates instead, and rejected it. The node is now created with an id in the manager's own namespace, as the `Default` authorization service already was. - **The custom certificate group node was not browseable.** It answered `GetCertificateGroups` and `ReadNode`, but a type-filtered recursive browse from the Objects folder never returned it — which OPC 10000-12 §7.8.2 requires of a certificate group. Staging the node through the builder gives it the reference it was missing. ### Two `PushTest` methods change — please review these closely The second defect had been masking what those tests actually did. Selecting the group positionally, `groups[3]` resolved to the server's **own `ServerConfiguration` group** while the custom one stayed invisible, so `UpdateTrustListOfCustomGroupAsync` and `AddRemoveCertOfCustomGroupAsync` were exercising a *transactional* trust list under a custom-group name. Browse results, before and after: | # | Before (baseline) | After | | --- | --- | --- | | 0-2 | GDS Directory groups (`ns=3`) | GDS Directory groups (`ns=3`) | | 3 | **`DefaultApplicationGroup` `i=14156`** — the server's own | **`MyCustomGroup` `ns=2`** | | 4 | — | `DefaultApplicationGroup` `i=14156` | They now resolve the group by browse name — which asserts it is browseable at all — and expect the immediate-apply semantics the GDS actually gives the groups of its own `Directory`: `CloseAndUpdate` writes the stores at once and reports `applyChangesRequired == false`, so the `ApplyChanges` calls go away (they would return `Bad_NothingToDo`). The read-back assertions are what prove the writes landed, and they are unchanged. This is a genuine loss of the *transactional* trust-list coverage those two methods were accidentally providing — the server's own group is still covered by the other `PushTest` methods, which target it deliberately. **5. Select `GetCertificates` results by type instead of by position** `GetCertificatesAsync` checked entry zero of a parallel-array result. Which entry that is depends on the order the server reports its certificate types in, an empty result throws `IndexOutOfRange` rather than failing an assertion, and a second type going missing or unparseable would not be noticed. Every returned pair is now checked, and a failure names the certificate type. **6. Drop the `.Common` suffix from the three GDS projects** Every other companion specification ships as `<Family>` / `<Family>.Client` / `<Family>.Server` — Di, ISA95, Positioning, Robotics, Vision, WotCon, XRegistry. GDS was the exception, with a `.Common` suffix that named nothing. | Before | After | | --- | --- | | `src/Opc.Ua.Gds.Common` | `src/Opc.Ua.Gds` | | `src/Opc.Ua.Gds.Client.Common` | `src/Opc.Ua.Gds.Client` | | `src/Opc.Ua.Gds.Server.Common` | `src/Opc.Ua.Gds.Server` | **No source changes were needed.** `RootNamespace` was already the final name in all three projects — the namespaces have always been `Opc.Ua.Gds`, `Opc.Ua.Gds.Client` and `Opc.Ua.Gds.Server`, and only the assembly and package ids disagreed. Nothing compiling against these assemblies changes a `using`. The package ids do change, which is consumer-visible, so `docs/migrate/2.0.x/packages.md` gains a renamed-packages section — it is the document that undertakes to cover NuGet renames — and the migration skill's package table gains the two rows a 1.5.378 consumer needs. Build infrastructure that names the projects or their output moves with them: both solution files, `expected-packages.txt` (re-sorted — it is maintained alphabetically), both signing manifests, the reference-server Dockerfile, the metapackage nuspecs, and the analyzer's package dependencies. `preview-pack.slnx` is derived from `UA.slnx` and needs no edit. The migration analyzer's shim tree mirrors the source layout by its own documented convention, so `Gds.Client.Common/` becomes `Gds.Client/`. Prose naming past work is left alone — plan 23 still records that its deferral happened during the "GDS Client.Common modernization", which is the name that effort had. ## Validation `net10.0`, on master + these four commits, from a clean GDS PKI: - `UA.slnx`: 0 errors, 0 warnings in changed files (the solution's 30 pre-existing warning lines are CA1861/CA1850/CA1307 in `XRegistry`, `Types.Tests` and `WotCon.Tests` — same rules and projects as the base). - `Opc.Ua.Gds.Tests`: 1,117 passed, 0 failed, 54 skipped — identical to the base branch. - `Opc.Ua.Server.Tests`: 4,961 passed, 0 failed, 5 skipped (measured before commit 4, which touches only the GDS assembly and its tests). - `Opc.Ua.SourceGeneration.Tests`: 169 passed. - `Opc.Ua.SourceGeneration.Core.Tests`, NodeManager/Generators/Fluent: 84 passed; `GenerateCode`: 21 passed, 4 skipped. - `Opc.Ua.MigrationAnalyzer.Core.Tests`: 10 passed, 3 skipped. - After the rename: assemblies build as `Opc.Ua.Gds.dll` / `.Client.dll` / `.Server.dll`, with no `*.Common.dll` produced anywhere. - Commits 1-3 build standalone. - `Opc.Ua.Gds.Server.Common` also builds for `net48` and `netstandard2.1` (checked before the rebase onto #4428). Not yet run: the full `Opc.Ua.SourceGeneration.Core.Tests` suite (~3,800), and the older target frameworks since the rebase. ## Checklist - [x] Existing tests pass - [x] New tests cover the new surface (role permissions, resolved-node builder, referenced-model binding, constructor/namespace control, GDS namespace invariant, custom-group node namespace) - [ ] Full Core.Tests sweep and net48/netstandard2.1 re-run after the rebase 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>