Repository navigation
Isolate the Default Object Allocator between Interpreters #101659
Description
Activity
FYI, one consequence of debugging CI failures in gh-101660 is that I was reminded that we may be able to take advantage of an object's corresponding obmalloc pool/arena to determine to which interpreter that object belongs. I'll open an issue about that once this one is settled.
- added a commit that references this issue
on Mar 13, 2023 - added a commit that references this issue
on Mar 14, 2023 - added 4 commits that reference this issue
on Apr 11, 2023
Metadata
Metadata
Assignees
Labels
Projects
- Status
(see gh-100227)
By default, the allocator for the "object" and "mem" domains is historically known as "obmalloc". The implementation of the allocator is
_PyObject_Malloc(), etc., for which the runtime state is found in_PyRuntimeState.obmalloc. Thus all interpreters currently share the runtime state for the default allocator.Isolating that state to each interpreter is important for a per-interpreter GIL (my short-term motivation here), but there are other benefits. For example, it helps us manage resources (e.g. objects) relative to the lifecycle of each interpreter, and to do so more efficiently. Furthermore, any situation where we share memory between interpreters is an invitation for memory leaks and other bugs.
The solution here should be as simple as moving the "obmalloc" state to
PyInterpreterState.Linked PRs