Crash when detaching dict from inline values and no memory is available

[DinoV]

Bug report

Bug description:

Add to test_class.py in TestInlineValues:

 def test_detach_materialized_dict_no_memory(self):
 a = WithAttrs()
 d = a.__dict__
 _testcapi.set_nomemory(0)
 del a
 print(d["a"])

This will crash because the dictionary will no longer points to valid memory:

#0 Py_XINCREF (op=<unknown at remote 0xdddddddddddddddd>) at ./Include/refcount.h:456
#1 _Py_XNewRef (obj=<unknown at remote 0xdddddddddddddddd>) at ./Include/refcount.h:488
#2 _Py_dict_lookup_threadsafe (mp=mp@entry=0x7ffff778ff50, key=key@entry='a', hash=<optimized out>, value_addr=value_addr@entry=0x7fffffffb078) at Objects/dictobject.c:1544
#3 0x0000000000518457 in dict_subscript (self={'b': <unknown at remote 0xdddddddddddddddd>, 'd': <unknown at remote 0xdddddddddddddddd>, 'a': <unknown at remote 0xdddddddddddddddd>, 'c': <unknown at remote 0xdddddddddddddddd>}, key='a')
 at Objects/dictobject.c:3325
#4 0x000000000049d8c6 in PyObject_GetItem (o=o@entry={'b': <unknown at remote 0xdddddddddddddddd>, 'd': <unknown at remote 0xdddddddddddddddd>, 'a': <unknown at remote 0xdddddddddddddddd>, 'c': <unknown at remote 0xdddddddddddddddd>},
 key=key@entry='a') at Objects/abstract.c:158

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux

Linked PRs

• GH-124547: Clear instance dictionary if memory error occurs during object dealloc #124627
• [3.13] GH-124547: Clear instance dictionary if memory error occurs during object dealloc (GH-124627) #124714

[DinoV]

In 3.13 this could be fixed by having the dict keep the object alive, and then decref it when the dict goes away (recovering the object from the inline values). In 3.14 it's no longer possible to recover the inline values because we support inline values on objects of various sizes.

[rruuaanng]

It seems the issue originates from _testcapi.set_nomemory(0).

When I run this code

from test.test_class import WithAttrs
import _testcapi


a = WithAttrs()
d = a.__dict__
# _testcapi.set_nomemory(0)
del a
print(d, d["a"])

It's output is as follows

Running Release|x64 interpreter...
object address : 000001C8DC77D1E0
object refcount : 3
object type : 00007FFFC0F03F80
object type name: MemoryError
object repr :
lost sys.stderr
object address : 000001C8DC77D120
object refcount : 3
object type : 00007FFFC0F03F80
object type name: MemoryError
object repr :
lost sys.stderr
object address : 000001C8DC77D1E0
object refcount : 3
object type : 00007FFFC0F03F80
object type name: MemoryError
object repr :
lost sys.stderr

It seems that nomemory(0) is used to simulate normal conditions and insert debugging hooks, but I don’t understand why commenting it out appears to resolve the issue.