[BUG] Segfault running pytest when a test fails

[shwina]

When a test that uses a MemoryResource with an UpstreamResourceAdaptor raises or fails, it results in a segfault in the subsequent test:

# test_segfault.py

import rmm                                                                                                                                                                                                 
import pytest                                                                                                                                                                                              
import gc                                                                                                                                                                                                  
                                                                                                                                                                                                           
@pytest.fixture(scope="function", autouse=True)                                                                                                                                                            
def rmm_auto_reinitialize(request):                                                                                                                                                                        
                                                                                                                                                                                                           
                                                                                                                                                                                                           
    # Run the test                                                                                                                                                                                         
    yield                                                                                                                                                                                                  
                                                                                                                                                                                                           
    # Automatically reinitialize the current memory resource after running each                                                                                                                            
    # test                                                                                                                                                                                                 
    rmm.reinitialize()                                                                                                                                                                                     
                                                                                                                                                                                                           
def test_one():                                                                                                                                                                                            
    mr = rmm.mr.PoolMemoryResource(rmm.mr.CudaMemoryResource())                                                                                                                                            
    rmm.mr.set_current_device_resource(mr)                                                                                                                                                                 
    buf = rmm.DeviceBuffer(size=10)                                                                                                                                                                        
    bl   # raises                                                                                                                                                                                                    
                                                                                                                                                                                                           
def test_two():                                                                                                                                                                                            
    gc.collect() 

pytest test_segfault.py  # segfaults

I'm still trying to get to the bottom of this, but I think it has something to do with traceback object corresponding to the error keeping objects alive for longer than expected, and this causing things to be destructed in the wrong order. Will report back when I know more.