While the app was running, working set grew ~20 MB/min (reaching tens of GB
over a day). dotnet-gcdump traced it to COM-callable wrappers accumulating from
the per-tick binding interop: SensorTreeItemViewModel.RefreshValues raised
PropertyChanged for Value/Min/Max/ToolTip on every sensor every update tick, and
ViewModelBase allocated a new PropertyChangedEventArgs per raise. WinUI's binding
engine is native, so each raised event marshals its args across the boundary and
retains a wrapper.
- ViewModelBase: reuse one cached PropertyChangedEventArgs per property name.
- SensorTreeItemViewModel.RefreshValues: raise PropertyChanged only when the
formatted text actually changed.
Also close a leaked D3DKMT adapter handle: D3DDisplayDevice.GetDeviceInfoByIdentifier
opened the adapter but skipped CloseAdapter on every early-return failure path,
and it runs on each GPU's Update() tick. Close it in a finally block.
Verified with dotnet-gcdump over an 18-minute soak: managed heap and live object
count stay flat (~50 MB) instead of climbing 47 -> 470 MB.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Concurrency crashes
- H1 HardwareMonitorService.RebuildTree now builds the tree under _updateLock, so it can't enumerate a hardware's _active HashSet while the update loop mutates it.
- A1 AppSettings now guards every dictionary read/write (and snapshots in Save) with a lock — safe under concurrent access from the parallel discovery threads.
- R2 Plumbed that same lock (HardwareMonitorService.SensorReadLock) into RemoteWebServer and wrapped the Prometheus sensor.Values enumeration with it.
- L1 Computer — refactored Add into AddCore, which performs the cancellation/enabled re-check and the _groups insertion atomically under _lock. A deferred task can
no longer add (and leak) a group after Close() drained the list; if it loses the race it closes the group instead.
- M1 UpdateTimer_Tick now bails before/after the await when _isShuttingDown is set in MainWindow_Closed, so an in-flight tick won't touch the disposed
view-model/Computer.
Broken behavior
- T1 Tray callback now decodes NOTIFYICON_VERSION_4 correctly (message = LOWORD(lParam), icon id = HIWORD(lParam)) — right-click menu and double-click work again.
- M2 A transient update exception no longer calls _timer.Stop(); the loop keeps running.
- H2 Newly discovered (deferred) storage devices get the current ForceDriveWakeup setting applied in HardwareChanged.
- V2 Sensor items carry a parent reference; toggling IsVisible recomputes the parent group's visibility, so no empty group headers. (Strengthened the existing test
that had skipped this assertion.)
- H3 Tree-rebuild coalescing now uses a dirty flag with a re-check, so a change arriving during a rebuild isn't lost.
- R4 Web routing matches endpoints exactly on the query-stripped path (Url.AbsolutePath), so static assets like metrics.html aren't hijacked.
- L8 IntelCpu.Update skips the bus/core-clock math while TimeStampCounterFrequency is still 0 (deferred-TSC window), so clocks keep their prior value instead of
reporting 0 MHz.
- V1 Existing plot series keep their assigned color; only an explicit user pen color updates them (no per-tick color shifting).
- M3 Runtime errors write to a dedicated runtime.log (once), instead of overwriting the shared startup.log.
- T2 CreateSensorIcon returns IntPtr.Zero on DIB failure instead of the shared main-icon handle (which callers DestroyIcon).
I also set _isOpen = false in HardwareMonitorService.Dispose so the rebuild guard actually holds during shutdown (the latent after-close-rebuild issue adjacent to
H3
JMR-dev
