Skip to content

Commit ef52ba1

Browse files
devdanzinclaude
andcommitted
OOM-0006: link to upstream python/cpython#152107 (repro_direct.py contributed there)
The bug is a pre-existing independent upstream report (#152107); the simplified repro_direct.py was posted to it and acked by the author. Set upstream_reported=true, upstream_issue=152107, status=reported (in addition to the older gist). Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
1 parent bbfd964 commit ef52ba1

1 file changed

Lines changed: 4 additions & 4 deletions

File tree

  • reports/OOM-0006-dictiter-gc-untrack

‎reports/OOM-0006-dictiter-gc-untrack/meta.json‎

Lines changed: 4 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -16,11 +16,11 @@
1616
],
1717
"matrix": {"ft_debug_asan": "abort", "ft_release": "n/a", "jit": "abort", "upstream": "segv"},
1818
"repro_start": 2,
19-
"upstream_reported": false,
20-
"upstream_issue": null,
19+
"upstream_reported": true,
20+
"upstream_issue": "152107",
2121
"gist_url": "https://gist.github.com/devdanzin/c809eb4072c0c787c0c890f54ba1c843",
22-
"status": "gisted",
22+
"status": "reported",
2323
"found_date": "2026-06-18",
2424
"confirmed_commit": "15d7406",
25-
"notes": "Assertion bug: fires only on assertion-enabled builds (ft_debug_asan/gil_debug_asan/jit abort at dictobject.c:5532). ft_release=n/a (NDEBUG strips assert; no segv on that run). Release builds segv at the untrack: upstream (release-gil-nojit) segv rc139 in _Py_Dealloc via PyObject_GetIter; release-gil-nojit-asan ASan SEGV at _PyObject_GC_UNTRACK (identical frames) -- NDEBUG corrupts the GC list instead of asserting; this is why the zipapp vehicle was labeled 'segfault' yet trips the identical dictiter assert on the debug build (verified). Only item iterators affected (dict_itemiterator) -- only kind whose di_result needs an alloc between PyObject_GC_New and _PyObject_GC_TRACK. TWO repros: (1) repro.py = realistic stdlib path `_strptime._strptime('', '')` under the set_nomemory sweep (this is the form published in the gist); (2) repro_direct.py = DIRECT minimal, stdlib-only, no _strptime (deterministic, re-verified 8/8 on ft_debug_asan @1b9fe5c, crashes ~start=2): iter(d.items()) under the sweep after draining the size-2 tuple freelist (hold a list of live 2-tuples) so di_result's _PyTuple_FromPairSteal actually hits the failing allocator; uses the iter() builtin (not GET_ITER) to avoid the adjacent PyStackRef_XCLOSE negrefcount bug; range(16) covers the per-build offset. The earlier 'minimization: partial' assumed a warm pymalloc/freelist steady state -- the freelist drain (plus the ASan triage builds being --without-pymalloc, so PyObject_GC_New is already counted) is the missing trick; minimization is now COMPLETE. signature/backtrace.txt unchanged: iter() reaches the SAME dictiter_new->Py_DECREF->dictiter_dealloc defect frames as the vehicle's GET_ITER path. repro_direct.py from another instance's handoff, independently re-verified (8/8 abort on ft_debug_asan; no-drain control = no crash; release-gil-nojit-asan SEGV confirmed)."
25+
"notes": "Assertion bug: fires only on assertion-enabled builds (ft_debug_asan/gil_debug_asan/jit abort at dictobject.c:5532). ft_release=n/a (NDEBUG strips assert; no segv on that run). Release builds segv at the untrack: upstream (release-gil-nojit) segv rc139 in _Py_Dealloc via PyObject_GetIter; release-gil-nojit-asan ASan SEGV at _PyObject_GC_UNTRACK (identical frames) -- NDEBUG corrupts the GC list instead of asserting; this is why the zipapp vehicle was labeled 'segfault' yet trips the identical dictiter assert on the debug build (verified). Only item iterators affected (dict_itemiterator) -- only kind whose di_result needs an alloc between PyObject_GC_New and _PyObject_GC_TRACK. TWO repros: (1) repro.py = realistic stdlib path `_strptime._strptime('', '')` under the set_nomemory sweep (this is the form published in the gist); (2) repro_direct.py = DIRECT minimal, stdlib-only, no _strptime (deterministic, re-verified 8/8 on ft_debug_asan @1b9fe5c, crashes ~start=2): iter(d.items()) under the sweep after draining the size-2 tuple freelist (hold a list of live 2-tuples) so di_result's _PyTuple_FromPairSteal actually hits the failing allocator; uses the iter() builtin (not GET_ITER) to avoid the adjacent PyStackRef_XCLOSE negrefcount bug; range(16) covers the per-build offset. The earlier 'minimization: partial' assumed a warm pymalloc/freelist steady state -- the freelist drain (plus the ASan triage builds being --without-pymalloc, so PyObject_GC_New is already counted) is the missing trick; minimization is now COMPLETE. signature/backtrace.txt unchanged: iter() reaches the SAME dictiter_new->Py_DECREF->dictiter_dealloc defect frames as the vehicle's GET_ITER path. repro_direct.py from another instance's handoff, independently re-verified (8/8 abort on ft_debug_asan; no-drain control = no crash; release-gil-nojit-asan SEGV confirmed). UPSTREAM: this bug is python/cpython#152107 (a pre-existing INDEPENDENT report, not ours); the simplified repro_direct.py was posted there 2026-06-24 and acked by the issue author. status=reported reflects that upstream issue (we also have the older gist)."
2626
}

0 commit comments

Comments
 (0)