Eryk Sun added the comment:
Unix waitpid() packs the process exit status and terminating signal number into a single status value. As specified by POSIX , the WEXITSTATUS function returns only the lower 8 bits of the process exit status. In theory, waitid() and wait6() can return the full 32-bit status value in the siginfo_t si_status. Apparently NetBSD does this , but evidently Linux does not:
>>> p = subprocess.Popen(['python3', '-c','import os; os._exit(257)']) >>> os.waitid(os.P_PID, p.pid, os.WEXITED).si_status 1
For Windows, programs sometimes return 16-bit WinAPI error codes or 32-bit HRESULT error codes. A critical error or unhandled exception may even result in returning a 32-bit NTSTATUS code (e.g. if you cancel the Windows error reporting dialog).
Currently the Python 3 implementation of sys.exit and os._exit converts to a signed long, but the Windows exit code is an unsigned long. To use integer return codes in the range 0x80000000-0xFFFFFFFF requires manually converting to the corresponding negative signed value.
Technically the two error types that require 32-bit values are signed (i.e. HRESULT and NTSTATUS), so this shouldn't be a problem. However, in practice sometimes a function -- and definitely subprocess.Popen -- returns an HRESULT error code as an unsigned value.
It would be convenient, at least on Windows, if handle_system_exit in Python/pythonrun.c converted the value using PyLong_AsUnsignedLongMask instead of PyLong_AsLong. Similarly os._exit could use Argument Clinic's "unsigned int" format, which also calls PyLong_AsUnsignedLongMask.
The problem with using PyLong_AsLong on Windows (32-bit or 64-bit) is that it overflows with an error value of -1 for values greater than 0x7FFFFFFF. In this case, os._exit raises OverflowError. handle_system_exit, on the other hand, ignores the exception. It just uses the -1 error value as the exit code, i.e. 0xFFFFFFFF.
---------- nosy: +eryksun