René's URL Explorer Experiment

Title: gh-149146: Add dealloc-depth counter fallback to trashcan trigger by bhuvi27 · Pull Request #151595 · python/cpython · GitHub

Open Graph Title: gh-149146: Add dealloc-depth counter fallback to trashcan trigger by bhuvi27 · Pull Request #151595 · python/cpython

X Title: gh-149146: Add dealloc-depth counter fallback to trashcan trigger by bhuvi27 · Pull Request #151595 · python/cpython

Description: Fixes #149146. while True: b = (b, None) under RLIMIT_AS segfaults during cleanup: b = 9 while True: b = (b, None) Docker with ulimit -v 524288: MemoryError → lost sys.stderr → segfault, exit 139. The trashcan deposit in _Py_Dealloc only fires when _Py_RecursionLimit_GetMargin(tstate) < 2, which compares the stack pointer against c_stack_soft_limit. Under RLIMIT_AS the kernel SIGSEGVs trying to grow the stack while SP is still nowhere near the soft limit — issue #150722 has an LLDB trace from the same reporter showing SP ~7 MB above c_stack_hard_limit at the fault. So the trashcan never deposits and the recursive tuple_dealloc → Py_XDECREF → _Py_Dealloc → tuple_dealloc chain just runs into the wall. This adds a per-thread c_dealloc_depth counter and uses it as a secondary trigger inside _Py_Dealloc, depositing at depth 50 (matches the old _PyTrash_UNWIND_LEVEL). The stack-pointer margin stays the primary signal; the counter only kicks in when SP can't. One thing I missed on the first pass: the drain side has the same problem. margin >= 4 is also true in this scenario, so every _Py_Dealloc exit would re-enter _PyTrash_thread_destroy_chain from inside the deallocator it just called. Fix: destroy_chain bumps the counter for the duration of the drain (same idea as the old delete_nesting), and the drain check in _Py_Dealloc becomes just c_dealloc_depth == 0. Counter is only touched for GC types so non-GC types pay nothing. Tests: new test_trashcan_nested_tuple_deep in test_gc.py (100k nested tuples, must clean up without crash); test_gc test_exceptions test_capi all green locally. Docker repro now exits 1 with MemoryError instead of 139 with Segmentation fault. #150722 is the same SP-signal blindness in the comparison path (Py_EnterRecursiveCall) — separate code path, separate PR.

Open Graph Description: Fixes #149146. while True: b = (b, None) under RLIMIT_AS segfaults during cleanup: b = 9 while True: b = (b, None) Docker with ulimit -v 524288: MemoryError → lost sys.stderr → segfault, exit 1...

X Description: Fixes #149146. while True: b = (b, None) under RLIMIT_AS segfaults during cleanup: b = 9 while True: b = (b, None) Docker with ulimit -v 524288: MemoryError → lost sys.stderr → segfault, exit 1...

Opengraph URL: https://github.com/python/cpython/pull/151595

X: @github

direct link

Domain: github.com

Links:

Viewport: width=device-width