[issue36116] test_multiprocessing_spawn fails on AMD64 Windows8 3.x

[Eryk Sun]

Eryk Sun <eryksun at gmail.com> added the comment:

> And depending on the OS, abort() calls (via Py_FatalError) sometimes 
> appear to be segfaults, so it could be any number of issues. 
> (Aside - I'd love to replace the abort() calls with specific exit
> codes for configuration errors - they really mess up the crash data 
> we see on Windows.)

In particular, with the Universal CRT, an unhandled abort() is implemented by a __fastfail intrinsic [1] (int 0x29 instruction in x86) with the argument FAST_FAIL_FATAL_APP_EXIT (7). 

Prior to Windows 8 this appears as an access violation. In Windows 8+ it's implemented as a second-chance STATUS_STACK_BUFFER_OVERRUN (0xC0000409) exception, which is overloaded from its previous use to support failure codes. (The old usage appears as the failure code FAST_FAIL_LEGACY_GS_VIOLATION, defined to be 0.) It starts as a second-chance exception in order to bypass normal exception handling (i.e. SEH, VEH, UnhandledExceptionFilter). The second-chance exception event is sent to an attached debugger and/or the session server (csrss.exe).

Python's normal signal handling for SIGABRT can't prevent this, since the C handler just sets a flag and returns. But enabling faulthandler sets a C signal handler that restores the previous handler and calls raise(SIGABRT). The default SIGABRT handler for the explicit raise() code path simply calls _exit(3). 

Alternatively, we could prevent the __fastfail call via _set_abort_behavior [2], if implemented in msvcrt. For example: msvcrt.set_abort_behavior(0, msvcrt.CALL_REPORTFAULT).

[1]: https://docs.microsoft.com/en-us/cpp/intrinsics/fastfail
[2]: https://docs.microsoft.com/en-us/cpp/c-runtime-library/reference/set-abort-behavior

----------
nosy: +eryksun

_______________________________________
Python tracker <report at bugs.python.org>
<https://bugs.python.org/issue36116>
_______________________________________