Mutable heap PyStructSequence types can segfault after modifying n_fields

Crash report

What happened?

Several heap PyStructSequence types can be crashed from pure Python by modifying the writable n_fields attribute on the type and then constructing a new instance.

The issue reproduces with at least:

• os.terminal_size
• os.stat_result
• time.struct_time
• resource.struct_rusage

Minimal reproducer:

import os

os.terminal_size.n_fields = 100000
os.terminal_size((1, 2))

This consistently terminates the interpreter with:

Segmentation fault (core dumped)

The crash also reproduces for other mutable heap PyStructSequence types by assigning a large value to n_fields before construction.

The backtrace shows the crash occurring in structseq_new_impl():

#0  __strlen_avx2()
#1  PyUnicode_FromString()
#2  PyDict_GetItemStringRef()
#3  structseq_new_impl() at Objects/structseq.c:243

At the point of failure:

max_len = 100000
i = 2
n_unnamed_fields = 0

From inspecting Objects/structseq.c, structseq_new_impl() uses the type's n_fields value to determine how many member names to process. After modifying n_fields from Python, the constructor eventually reaches a NULL member name, leading to a crash through PyDict_GetItemStringRef() and PyUnicode_FromString().

During investigation I also verified that immutable builtin structseq types such as sys.version_info are not affected because their type attributes cannot be modified and new instances cannot be created.

I searched the existing issue tracker using keywords including:

• structseq_new_impl
• PyStructSequence_NewType
• n_fields
• Objects/structseq.c
• PyDict_GetItemStringRef
• tp_members

but could not find an existing report describing this behavior.

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux

Output from running 'python -VV' on the command line:

Python 3.16.0a0 (heads/main:e469fa9c807, Aug 7 2026, 09:37:12) [GCC 13.3.0]

Linked PRs

• gh-155361

從 Objects/structseq.c 和 structseq_new_impl() 開始，然後執行 issue 中描述的 os.terminal_size 最小重現程式。比較可寫入的 n_fields 值與可用成員名稱的處理方式；完成標準是建構受影響的型別不再當機，且回歸涵蓋範圍證明了此行為。