[default] Fix possible crash at RunLoop::wake()

[pozdnyakov]

The runOnce() call during run loop destruction might cause invoke() calls.

Fixes #16250

[alexshalamov]

lgtm

[springmeyer]

Thanks for this fix! I'm curious:

• What is the root problem that this is fixing?
• What it introduced by 9cb19b9 or was 9cb19b9 simply what uncovered a lurking problem that predated it?
• Is there a way to write a test that would have failed before this fix (so that the problem would be caught in the gl-native tests if it regresses again, rather than only downstream)?

[springmeyer]

@pozdnyakov @tmpsantos bump on my questions above.

[pozdnyakov]

@springmeyer sorry for missing your questions!

What is the root problem that this is fixing?

The runOnce() call during run loop destruction caused the invoke() call on the same partially destructed run loop (i.e. impl->async was already deleted).

What it introduced by 9cb19b9 or was 9cb19b9 simply what uncovered a lurking problem that predated it?

the latter, it was a long-standing problem

Is there a way to write a test that would have failed before this fix (so that the problem would be caught in the gl-native tests if it regresses again, rather than only downstream)?

Yep, I'm planning to add a unit test probably next week

[springmeyer]

Thanks @pozdnyakov - I will watch for your unit test.