test_sys_settrace leaked [6604, 6602, 6604] on main branch

[kumaraditya303]

Commit tested: a25a803

@kumaraditya303 ➜ /workspaces/cpython (main) $ ./python -m test -R 3:3 test_sys_settrace
0:00:00 load avg: 31.03 Run tests sequentially
0:00:00 load avg: 31.03 [1/1] test_sys_settrace
beginning 6 repetitions
123456
......
test_sys_settrace leaked [6604, 6602, 6604] references, sum=19810
test_sys_settrace leaked [1835, 1860, 1856] memory blocks, sum=5551
test_sys_settrace failed (reference leak)

== Tests result: FAILURE ==

1 test failed:
    test_sys_settrace

Total duration: 1.2 sec
Tests result: FAILURE

[kumaraditya303]

cc @markshannon @pablogsal

[kumaraditya303]

Git bisect points to 2f8bff6 as first bad commit:

commit 2f8bff6879c5d76d143068e8bc867196a7d28afc (HEAD)
Author: Mark Shannon <mark@hotpy.org>
Date:   Mon Jul 18 16:06:42 2022 +0100

    GH-94739: Mark stacks of exception handling blocks for setting frame.f_lineno in the debugger. (GH-94958)

[mdboom]

frame.setlineno strikes again.

[brandtbucher]

Checking this out now.

[brandtbucher]

The following three tests are leaking:

• test_jump_from_except_to_finally
• test_jump_out_of_bare_except_block
• test_jump_out_of_qualified_except_block

The issue is when we jump out of an exception handler into non-exception-handling code. We skip over POP_EXCEPT, which is supposed to clear tstate->exc_info->exc_value, the currently active exception. This exception references thousands of objects through its traceback frames.

I'm not sure what a good solution is. Perhaps clearing the thread's exception state if we can detect that we're jumping out of an exception handler? I'm not sure if that's always correct (or even easy to detect). Maybe we can flag every stack between an exception handler entry and a POP_EXCEPT?

Or, just disallow jumping out of exception handlers like this?