RFH: SIGBUS in buildd but not in porterbox

[Felipe Sateler]

On Tue, Feb 18, 2014 at 2:50 PM, David Daney <ddaney at caviumnetworks.com> wrote:
> On 02/17/2014 01:28 PM, Andreas Barth wrote:
>>
>> * Felipe Sateler (fsateler at gmail.com) [140217 20:33]:
>>>
>>> The package liblo runs a test suite at build time. This test suite
>>> failed in mipsel (but not mips)[1]. I tried to reproduce the issue in
>>> eder.d.o but it passed.
>>>
>>> Liblo does use unaligned access (liblo had to be removed from sparc
>>> for this reason).
>>> But I'm confused as to why does the buildd error out and the porterbox
>>> doesn't. Is it possible to (programatically) instruct the kernel to
>>> not raise a SIGBUS?
>>>
>>>
>>> [1]
>>> https://buildd.debian.org/status/fetch.php?pkg=liblo&arch=mipsel&ver=0.28-3&stamp=1392064303
>>
>>
>> rem and eder are different hardware.
>
>
> Yes, Some MIPS CPUs (like OCTEON) can be, and are, configured to handle
> unaligned accesses.  Others cannot.
>
> The kernel may be able to emulate unaligned accesses for integer loads and
> stores, but I think the problem may be that the FPU emulator cannot do
> unaligned accesses.

I'm not sure this is the cause. Liblo doesn't do much math on its own.
What it does do is read 64bit ints or doubles from a 4-byte aligned
data stream.

>
> Really the application that is doing this should be fixed.

Yes, but this is not so easy because on my amd64 I can't enable sigbus
traps as I can't even do a printf without triggering it:

#include <stdio.h>

main() {
#if defined(__GNUC__)
# if defined(__i386__)
         /* Enable Alignment Checking on x86 */
         __asm__("pushf\norl $0x40000,(%esp)\npopf");
# elif defined(__x86_64__)
        /* Enable Alignment Checking on x86_64 */
         __asm__("pushf\norl $0x40000,(%rsp)\npopf");
# endif
#endif

        printf("hello\n");
        printf("goodbye\n");
}


This program never prints goodbye.

-- 

Saludos,
Felipe Sateler