Signal handlers don't skip incomplete frames

[kumaraditya303]

Currently signal handlers don't skip incomplete frames and this leads to crash when creating frame object. Incomplete frames should be skipped in _PyErr_CheckSignalsTstate.

[mdboom]

To make sure we aren't playing "whac-a-mole", is there a way to find all of these cases and/or improve the internal API?

[kumaraditya303]

It depends upon if the function calls back into python with a frame reference. I don't think these can be found all at once, it requires following the entire call chain and if the callback can be overridden.

[mdboom]

Ok, thanks for the clarification, @kumaraditya303.

[markshannon]

This looks possible, but extremely unlikely, as AFAICT it can only happen in a C extension called during GC or from a finalizer, which then calls Py_MakePendingCalls

There may be other cases, but I suspect they are just as obscure.
We should fix this, and backport it. But I don't think we need worry about it too much.

@mdboom Looking through the code, there are two other places we don't check whether the frame is valid.
There aren't that many calls to _PyFrame_GetFrameObject so it easy to check them all.

@pablogsal Want the fix in 3.11.0 or is 3.11.1 OK?

[pablogsal]

3.11.1 unless you think it absolutely needs to go on 3.11.0

[kumaraditya303]

To be clear I was only able to reproduce this once, so yeah extremely unlikely but can be hit in pure python but I have no small reproducer.