mirror of
https://github.com/OPCFoundation/UA-.NETStandard.git
synced 2026-10-19 15:22:15 +08:00
romanett
efad0ddeff
# Description Fixes defects found in a spec-grounded audit of the server **Subscription / MonitoredItem** implementation (`src/Opc.Ua.Server/Subscription/**`, `NodeManager/MonitoredItem/**`, `EventManager`, monitored-item service dispatch and filter validation). Each behaviour change was checked against OPC 10000-4 (§5.13, §5.14, §7.21, §7.22), Part 5 (subscription diagnostics) and Part 9 (ConditionRefresh). Every fix has a regression test. The audit found 56 defects (1 Critical, 3 High, 20 Medium, 32 Low). A follow-up max-effort code review of the fix branch found 27 more issues, mostly introduced by the first round of fixes. All but one are fixed here. ### Most important fixes - **Critical — remote hang:** an `AggregateFilter` with `processingInterval = NaN` was not revised, and `AggregateCalculator` looped forever while holding the node-manager lock, so one CreateMonitoredItems request blocked the node manager. Non-finite, ≤0 and sub-tick intervals are now revised (Part 4 §5.13.2.1), and the calculator rejects them as a backstop. - **Monitored-item limits:** new `ServerConfiguration.MaxMonitoredItemCount` / `MaxMonitoredItemsPerSubscription`; Bad_TooManyMonitoredItems is returned when a limit is hit, and ServerCapabilities now report the limits. Both default to `0` (no limit), so existing behaviour is unchanged. - **Sampling groups:** a client could get one dedicated thread per distinct sampling interval above 1 h; intervals above the last rate group are now rounded, and the loop is async. An item could also end up in two sampling groups and keep being sampled after it was deleted. - **Subscription state machine** (Part 4 §5.14.1.2): - the lifetime counter resets while a Publish is queued and counts otherwise; - the lifetime and keep-alive stay within `MaxSubscriptionLifetime`; - LatePublishRequestCount is counted once per entry into LATE; - messages held while publishing is disabled are no longer sent, and are not offered by Transfer/Republish either; - ConditionRefresh returns Bad_RefreshInProgress for any refresh already queued or running on the subscription. - **Triggering** (Part 4 §5.13.1.6): - linked items are triggered by an item still draining a backlog, by the first value after re-enable, and by an item created Disabled; - links are persisted for durable subscriptions; - Disabled/Sampling items follow the publish rules in the shared pipeline, so custom `IMonitoredItem` implementations follow them too. - **Queues** (Part 4 §5.13.1.5, §7.21): - no duplicate value after a queue-size increase; - EventQueueOverflow events are placed where the events were lost, including on queue shrink; - the duplicate-event check scans the newest entries; - EventQueueOverflowCount and MonitoringQueueOverflowCount are maintained correctly; - changing select clauses rebuilds queued events in the new layout, and is safe on durable queues and on failure. - **Filters:** - per-clause EventFilterResult instead of rejecting the whole item (Part 4 §7.22.3); - corrected status codes: Bad_DeadbandFilterInvalid, Bad_MonitoredItemFilterInvalid, Bad_EventFilterInvalid, Bad_MonitoringModeInvalid, Bad_TimestampsToReturnInvalid; - aggregate filter in subscriptions: processing interval at least 2× the revised sampling interval, revised StartTime on the requested grid, ServerTimestamp = end of interval (Part 4 §7.22.4). - **Lifecycle and races:** - expired subscriptions are always cleaned up, even when the session closes during cleanup; - Create/Transfer racing CloseSession returns Bad_SessionClosed; - session diagnostics counts no longer drift; - a large but valid Publish TimeoutHint no longer fails the request; - a ready subscription is served by queued Publish requests instead of going Late; - restored persisted state is validated. - **Access:** a value change notification is checked against each subscriber's UserAccessLevel, not the writer's. - **Redundancy / durable store:** - the first-unsent sequence number is mirrored; - sequence numbers are no longer reused after a second failover; - triggering links are stored (definition format 6, written only when links exist, so older replicas can still read link-less subscriptions). ### Behaviour changes reviewers should know - With the default config, publishing intervals above `MaxSubscriptionLifetime / 3` (1,200,000 ms) are revised down, so lifetime ≥ 3×keep-alive fits the configured maximum. - The revised keep-alive is capped at a third of the max lifetime. - Idle subscriptions expire after exactly MaxLifetimeCount cycles without a queued Publish. - A partly invalid EventFilter select list is accepted with per-clause errors. The public `EventFilter.Validate` keeps Status Good if at least one clause is valid; callers that cannot return per-clause results must also check the new `EventFilter.Result.HasSelectClauseErrors`. There is a migration note in `docs/migrate/2.0.x/sessions-subscriptions.md`. - Aggregate processing intervals may be revised higher (≥ 2× sampling interval); aggregate ServerTimestamp is now the interval end. - Sampling intervals above 3,600,000 ms are rounded to multiples of 900,000 ms (default rate groups). ### Public API Additive only: - `ServerConfiguration.MaxMonitoredItemCount` and `MaxMonitoredItemsPerSubscription`; - `EventFilter.Result.HasSelectClauseErrors`; - `IStoredSubscriptionTriggering`, `ISubscriptionRetransmissionSendStateStore`, `SubscriptionRetransmissionState.FirstUnsentSequenceNumber` (optional store capabilities, detected with `as`); - new `EventQueueHandler` constructor overloads. ### Known open items (not in this PR) - **Part 4 §5.13.2.1:** CreateMonitoredItems should create an item even when the user lacks Read permission and report the error in Publish. This is not done: the create-time RolePermissions check is the only role check on the initial value, so following the spec would leak values until every node manager re-checks per sample. - **Sampling loop:** it runs on thread-pool threads; node managers with blocking synchronous reads should not use sampling groups heavily. The alternative, a dedicated thread per group, was the thread-exhaustion issue fixed here. - **Restarts:** a pending QueueSize-1 value and a pending overflow flag are not kept across a restart; that needs persistence interface extensions. - **Limit defaults:** decide whether the new limits should get finite defaults (currently 0 = unlimited). ## Related Issues - No tracking issue; the findings come from an internal audit of the server subscription code. ## Checklist - [ ] I have signed the [CLA](https://opcfoundation.org/license/cla/ContributorLicenseAgreementv1.0.pdf) and read the [CONTRIBUTING](https://github.com/OPCFoundation/UA-.NETStandard/blob/master/CONTRIBUTING.md) doc. - [x] I have added tests that prove my fix is effective or that my feature works and increased code coverage. - [x] I have added all necessary documentation. - [ ] I have verified that my changes do not introduce (new) build or analyzer warnings. - [ ] I ran **all** tests locally using the **UA.slnx** solution against at least .net **framework** and .net **10**, and all passed. - [ ] I fixed **all** failing and flaky tests in the CI pipelines and **all** CodeQL warnings. - [ ] I have addressed **all** PR feedback received. ### Local verification (net10.0, all green) | Suite | Passed | Failed | |---|---|---| | Opc.Ua.Server.Tests | 7963 | 0 | | Opc.Ua.Subscriptions.Tests | 643 | 0 | | Opc.Ua.Subscriptions.Classic.Tests | 57 | 0 | | Opc.Ua.Subscriptions.Durable.Tests | 36 | 0 | | Opc.Ua.Redundancy.Server.Tests | 847 | 0 | | Opc.Ua.Core.Tests (filter tests) | 283 | 0 | net48 build of `Opc.Ua.Server.Tests`: 0 errors. net48 tests and the full UA.slnx were not run locally. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Marc Schier <marcschier@users.noreply.github.com>