From: Klemens Nanni <kn@openbsd.org>
Subject: Re: devel/ruby-ffi failure on arm64 BTI
To: ports <ports@openbsd.org>, Jeremy Evans <jeremy@openbsd.org>
Date: Mon, 23 Mar 2026 19:39:12 +0000

22.03.2026 22:23, Stuart Henderson пишет:
> On 2026/03/22 15:51, Klemens Nanni wrote:
>> 	# pkg_add ruby34-ffi
>> 	$ ktrace -di ruby34 /usr/local/lib/ruby/gems/3.4/gems/ffi-1.17.2/samples/getpid.rb
>> 	[ big dump of loaded files ]
>> 	Illegal instruction (core dumped)
>>
>> 	$ kdump | grep BTCFI
>> 	 88448 ruby34   PSIG  SIGILL caught handler=0x221c893fcc mask=0<> code=ILL_BTCFI addr=0x22b6114000 trapno=905969665
> ...
>> Even with USE_NOBTCFI-aarch64=Yes in lang/ruby/3.4 (as well as devel/ruby-ffi
>> and devel/libffi for good measure) it still fails.
> 
> does it make it into the binary? objdump -p | grep OPENBSD
> (compare with e.g. python3).

No, but now it does and Ruby no longer dies, getpid.rb and openfact work!

OPENBSD_NOBTCFI off    0x0000000000000000 vaddr 0x0000000000000000 paddr 0x0000000000000000 align 2**0

Remains to see why Ruby is special while ffi-getpid.c works fine with BTI/PAC.



So I only checked like this:

	$ make show=USE_NOBTCFI       
	Yes     

But this is where it fails already:

	$ make show=_LINKER_FLAGS 
	 -z wxneeded

Going over bsd.port.mk code, it really baffled me, until double checked
my Makefile changes:

	$ head -n1 Makefile
	USE_NOBTCFI-aarch64 =   Yes     # devel/libffi
	$ make show=USE_NOBTCFI-aarch64
	Yes     
	$ make show=USE_NOBTCFI-aarch64 | hexdump -C
	00000000  59 65 73 09 0a                                    |Yes..|
	00000005

That delimiting tab is consumed into the variable's value, hence why
make's ${USE_NOBTCFI:L} == "yes" never matched...

We do this all over at least for SHARED_LIBS, but apparently these
variables/values are handled in a way that doesn't have this problem.