Stack overflow detection failure

[straight-shoota]

#6928 added a cool feature which detects whether an invalid memory access is actually a stack overflow error.

However, this detection does not yet seem to work correctly in all cases.

On alpine edge the spec for SO detection on main stack fails:

  1) seg fault detects stack overflow on the main stack
     Failure/Error: error.should contain("Stack overflow")

       Expected:   "Invalid memory access (signal 11) at address 0xffffc59b7e30\n[0xaab6d0310cb8] *CallStack::print_backtrace:Int32 +100\n[0xaab6d03034d0] __crystal_sigfault_handler +208\n[0xfff785619964] __setjmp +56\n[0xaab6d03444c0] *Pointer(Hash::Entry(String, NamedTuple(time: Time, location: Time::Location)) | Nil) +136\n[0xaab6d0303530] *foo:NoReturn +48\n[0xaab6d0308fdc] *foo:NoReturn +23260 (1522 times)\n[0xaab6d02f9130] __crystal_main +2072\n[0xaab6d034690c] *Crystal::main_user_code<Int32, Pointer(Pointer(UInt8))>:Nil +8\n[0xaab6d0346864] *Crystal::main<Int32, Pointer(Pointer(UInt8))>:Int32 +40\n[0xaab6d02ffedc] main +8\n[0xfff7855ee734] ???\n"
       to include: "Stack overflow"

     # spec/std/kernel_spec.cr:249

The spec for SO detection on fiber stack passes, though.

There was a similar failure reported for Ubuntu 14.04 which doesn't seem to have been fixed (#6928 (comment)).

If this can't be fixed right away, a short term play would be to simply disable the spec on affected platforms. They're not essential.

/cc @damaxwell

[HertzDevil]

Stack overflow detection relies on the current thread's stack bounds. On Linux this is obtained using LibC.pthread_getattr_np and LibC.pthread_attr_getstack:

crystal/src/crystal/system/unix/pthread.cr

Lines 118 to 123 in d58ede5

	{% elsif flag?(:linux) %}
	if LibC.pthread_getattr_np(@system_handle, out attr) == 0
	LibC.pthread_attr_getstack(pointerof(attr), pointerof(address), out _)
	end
	ret = LibC.pthread_attr_destroy(pointerof(attr))
	raise RuntimeError.from_os_error("pthread_attr_destroy", Errno.new(ret)) unless ret == 0

However, on musl-libc the internal stack size is always hardcoded to a very small default of 128 KiB, and pthread_getattr_np returns this value unmodified, presumably for the main thread as well. The result is that the main thread has an entirely incorrect @stack, as it respects ulimit -s, and that non-main threads are really that small, as we don't pass any attr to GC.pthread_create on thread creation. This is also mentioned on Ruby's bug tracker.

The stack trace itself is wrong as well, even though Exception::CallStack.print_backtrace works correctly as long as it is outside a signal handler. This appears to be an intentional won't-fix. If we need a meaningful stack trace inside a signal handler then we have to bypass _Unwind_Backtrace and mimic whatever a debugger like LLDB does.