Re: segfault in ffi_data_to_code_pointer

[DJ Delorie <dj@redhat.com>]

> DJ Delorie <dj@redhat.com> writes:
>> In src/closures.c, ffi_data_to_code_pointer() calls segment_holding()
>> to get a pointer to the code segment for a data segment.  It doesn't
>> check for a NULL return, and I've got a test case where I run Ruby's
>> test suite (on a non-selinux aarch64 machine, if that matters) and
>> segment_holding() returns NULL and much hilarity ensues.

Anthony Green <green@redhat.com> writes:
> This suggests that the ffi_closure object used to invoke
> ffi_data_to_code_pointer wasn't allocated by ffi_closure_alloc().  Is
> that something you can check?

It seems to be so:

 [root ruby-2.6.3]# ./miniruby ./tool/runruby.rb -Itest/lib -r 'test/unit' "./test/fiddle/test_import.rb"
 < 0000ffffaa9b0000
 < 0000ffffaa900000
 Run options: 

The '<' lines indicate calls to ffi_data_to_code_pointer that had a NULL
segment, and the non-existing '>' lines indicated non-existing calls to
ffi_closure_allocate.

However, a simple NULL check may be the right fix anyway:

* it fixes the ruby build/test problem
* it does not re-introduce the illegal instruction bug

Which means that the non-code-seg'd closure isn't the one that triggers
the illegal instruction... but I don't know why still ;-)