SystemEvents deadlock if UI thread shuts down early

[weltkante]

Description

The first thread that accesses SystemEvents is used for initialization. If its a MTA thread then a private STA thread with message loop is created for listening to system events, but if the first thread is an STA thread the initialization happens on this thread.

The code is not prepared for an STA thread that stops pumping messages or shuts down before process exit.

• WinForms is running UI tests on multiple threads. Initialization of SystemEvents can happen directly in a test or indirectly by usage of WinForms API. If the first usage happens on a UI test (STA) then SystemEvents will bind to that thread. If this is a test thread which gets terminated earily this can lead to deadlocks in the CI framework (Flaky test: ProfessionalColorTable_ChangeUserPreferences_GetColor_ReturnsExpected deadlock winforms#3254)
• Some applications are using a separate UI thread for the initial splash screen during application startup. This UI thread is likely to terminate early, if SystemEvents were initialized on that thread they would stop working.
• Some applications use multiple UI threads to implement responsive progress/notification dialogs when the main application thread is busy and work cannot be easily moved off-thread. If such a dialog shows up early (e.g. progress dialog for some update during application startup) SystemEvents may be initialized on a secondary thread that could be terminated early.
• There apparently have been Watson reports indicating deadlocking STA threads and/or threads that stop pumping messages.

Regression?

No

[Dotnet-GitSync-Bot]

I couldn't figure out the best area label to add to this issue. Please help me learn by adding exactly one area label.

[weltkante]

Some options to consider are:

• always create a private event thread (could do that always, or could make it opt-in)
• expose a public API that can be used to notify SystemEvents if a thread goes down (basically an opt-in API)
• make use of CLR internal API to get notified when a thread exits (assuming such an API exists or can be created; wouldn't require to opt-in)

If the solution to the issue is opt-in it would probably not solve third party deadlocks, since it will be hard to identify the opt-in would fix the deadlocks.

[weltkante]

SystemEvents is not part of WinForms, the WinForms issue is dotnet/winforms#3254, I was asked to open one in the runtime repo who owns SystemEvents.

[Dotnet-GitSync-Bot]

I couldn't figure out the best area label to add to this issue. Please help me learn by adding exactly one area label.

[ericstj]

always create a private event thread (could do that always, or could make it opt-in)

This seems reasonable, any drawbacks? cc @JeremyKuhne

Given this isn't a regression from and desktop behaved this way for a decade I don't think we need to fix this for 5.0.