Skip to content

Under-estimated stack size for recursion limit (macOS + python >= 3.14)#139231

@rokm

Description

@rokm

Bug report

Bug description:

GH-130398 introduced estimation of stack size (for the purpose of recursion-limit management) via pthread_getattr_np() on platforms that support it. This non-portable function is not available on macOS, and so the following fallback codepath is used:

cpython/Python/ceval.c

Lines 485 to 487 in 4fb338d

_tstate->c_stack_top=_Py_SIZE_ROUND_UP(here_addr, 4096);
_tstate->c_stack_soft_limit=_tstate->c_stack_top-Py_C_STACK_SIZE;
_tstate->c_stack_hard_limit=_tstate->c_stack_top- (Py_C_STACK_SIZE+_PyOS_STACK_MARGIN_BYTES);

with Py_C_STACK_SIZE being set to 4 MB on macOS:

cpython/Python/ceval.c

Lines 369 to 381 in 4fb338d

#if defined(__s390x__)
# definePy_C_STACK_SIZE 320000
#elif defined(_WIN32)
// Don't define Py_C_STACK_SIZE, ask the O/S
#elif defined(__ANDROID__)
# definePy_C_STACK_SIZE 1200000
#elif defined(__sparc__)
# definePy_C_STACK_SIZE 1600000
#elif defined(__hppa__) || defined(__powerpc64__)
# definePy_C_STACK_SIZE 2000000
#else
# definePy_C_STACK_SIZE 4000000
#endif

However, this is too conservative, as the python's build script explicitly increases the stack to 16 MB:

cpython/configure.ac

Lines 3601 to 3628 in 4fb338d

Darwin/*|iOS/*)
LINKFORSHARED="$extra_undefs -framework CoreFoundation"
# Issue #18075: the default maximum stack size (8MBytes) is too
# small for the default recursion limit. Increase the stack size
# to ensure that tests don't crash
stack_size="1000000" # 16 MB
if test "$with_ubsan" = "yes"
then
# Undefined behavior sanitizer requires an even deeper stack
stack_size="4000000" # 64 MB
fi
AC_DEFINE_UNQUOTED([THREAD_STACK_SIZE],
[0x$stack_size],
[Custom thread stack size depending on chosen sanitizer runtimes.])
if test $ac_sys_system = "Darwin"; then
LINKFORSHARED="-Wl,-stack_size,$stack_size $LINKFORSHARED"
if test "$enable_framework"; then
LINKFORSHARED="$LINKFORSHARED "'$(PYTHONFRAMEWORKDIR)/Versions/$(VERSION)/$(PYTHONFRAMEWORK)'
fi
LINKFORSHARED="$LINKFORSHARED"
elif test $ac_sys_system = "iOS"; then
LINKFORSHARED="-Wl,-stack_size,$stack_size $LINKFORSHARED "'$(PYTHONFRAMEWORKDIR)/$(PYTHONFRAMEWORK)'
fi
;

This was observed in GH-131543, but was not investigated further, because the primary issue there turned out to be excessive stack consumption due to inlining (GH-137573).

Nevertheless, the under-estimated stack size still negatively impacts the recursion limit.

Using slightly modified reproducer from #131543 (comment) that keeps trying to dynamically increase recursion limit as it keeps recursing:

// stackpointer.c//// gcc -shared -fpic -o stackpointer.dylib stackpointer.c//// import ctypes// sp = ctypes.CDLL('./stackpointer.dylib')// address = sp.get_machine_stack_pointer()#include<stdint.h>uintptr_tget_machine_stack_pointer(void){return (uintptr_t)__builtin_frame_address(0)}
# recursion_limit_test.pyimportsysimportctypesimportctypes.util# Obtain stack address and stack size as reported by non-portable pthread functions:libc=ctypes.CDLL(ctypes.util.find_library('c')) libc.pthread_self.restype=ctypes.c_void_plibc.pthread_get_stackaddr_np.argtypes= [ctypes.c_void_p] libc.pthread_get_stackaddr_np.restype=ctypes.c_void_plibc.pthread_get_stacksize_np.argtypes= [ctypes.c_void_p] libc.pthread_get_stacksize_np.restype=ctypes.c_ulonglongthis_thread=libc.pthread_self() stack_address=libc.pthread_get_stackaddr_np(this_thread) stack_size=libc.pthread_get_stacksize_np(this_thread) print(f"Stack address: {stack_address} = 0x{stack_address:X}") print(f"Stack size: {stack_size} = {stack_size/1024.0} kB = {stack_size/1024**2} MB") # Helper for tracking stack pointer locationsplib=ctypes.CDLL('./stackpointer.dylib') splib.get_machine_stack_pointer.restype=ctypes.c_void_pstack_pointer=splib.get_machine_stack_pointer() print(f"Stack pointer: 0x{stack_pointer:X}, depth: {(stack_address-stack_pointer)/1024:.2f} kB") # Recursion limit testlimit=sys.getrecursionlimit() counter=0classA: def__getattribute__(self, name): globalcountercounter+=1stack_pointer=splib.get_machine_stack_pointer() print(f"Recursion level: {counter}, stack pointer: 0x{stack_pointer:X}, depth: {(stack_address-stack_pointer)/1024:.2f} kB") # Increase recursion limit, if necessarygloballimitifcounter+1>=limit: limit*=2print(f"Increasing recursion limit: {limit}") sys.setrecursionlimit(limit) # Recursereturngetattr(self, name) a=A() print("Testing Recursion Limit") print(f"Initial limit: {limit}") try: a.testexceptRecursionError: print(f"Recursion Limit ok (reached level {counter})")

Running with python 3.13:

% python3.13 recursion_limit_test.py Stack address: 6101024768 = 0x16BA64000 Stack size: 16777216 = 16384.0 kB = 16.0 MB Stack pointer: 0x16BA622A0, depth: 7.34 kB Testing Recursion Limit Initial limit: 1000 Recursion level: 1, stack pointer: 0x16BA61CD0, depth: 8.80 kB Recursion level: 2, stack pointer: 0x16BA616A0, depth: 10.34 kB Recursion level: 3, stack pointer: 0x16BA610B0, depth: 11.83 kB ... Recursion level: 4997, stack pointer: 0x16B323CD0, depth: 7424.80 kB Recursion level: 4998, stack pointer: 0x16B3236E0, depth: 7426.28 kB Recursion Limit ok (reached level 4999) 

Note that even with the old approach, only about half of the actual 16 MB stack was used before the recursion limit kicked in (as evident from estimated stack depth).


With 3.14(.0rc2):

% python3.14 recursion_limit_test.py Stack address: 6091063296 = 0x16B0E4000 Stack size: 16777216 = 16384.0 kB = 16.0 MB Stack pointer: 0x16B0E1CC0, depth: 8.81 kB Testing Recursion Limit Initial limit: 1000 Recursion level: 1, stack pointer: 0x16B0E1120, depth: 11.72 kB Recursion level: 2, stack pointer: 0x16B0E0550, depth: 14.67 kB ... Recursion level: 1320, stack pointer: 0x16AD13470, depth: 3906.89 kB Recursion level: 1321, stack pointer: 0x16AD128A0, depth: 3909.84 kB Recursion Limit ok (reached level 1322) 

With 3.14, we don't get far above the original limit before the recursion limit kicks in, under assumption of 4 MB stack.

CPython versions tested on:

3.14

Operating systems tested on:

macOS

Linked PRs

Metadata

Metadata

Assignees

No one assigned

    Labels

    3.14bugs and security fixes3.15new features, bugs and security fixesOS-macinterpreter-core(Objects, Python, Grammar, and Parser dirs)type-bugAn unexpected behavior, bug, or error

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions