Bug#609371: linux-image-2.6.37-trunk-sparc64: module scsi_mod: Unknown relocation: 36
* David Miller (davem@davemloft.net) wrote:
> From: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
> Date: Wed, 19 Jan 2011 00:08:45 -0500
>
> > - No aligned() type attribute nor variable attribute. I get a crash on x86_64
> > (NULL pointer exception when executing __trace_add_event_call, the 5th call).
> > __alignof__(struct ftrace_event_call) is worth 8.
>
> I think I figured it out.
>
> It's the static vs. non-static thing, or some other crazyness wrt.
> how x86-64 implements it's alignment rules.
>
> GCC on x86-64 uses a completely different policy for aligning local
> (ie. "static") data objects vs. globally visible ones, for one thing.
> It also has different alignment policies for objects that are part
> of an array vs. those which are not.
>
> On both counts, we're lying to the compiler, so maybe it's sort of our
> fault.
Yep, we're in strong agreement here.
>
> As far as GCC can see, the object is static and also not part of an
> array or any other C construct for which things like this could matter
> as long as the alignment it chooses meets the minimum alignment
> requirements of the ABI.
>
> So it doesn't let us do this trick where we put the individual event
> markers into a special section, yet mark it __used and static, then
> later access them as if they were part of a globally visible array.
>
> If you look these static objects, they are emitted with a leading
> ".align 32" directive. This is what screws everything up.
Ah, yep, good way to identify the culprit.
>
> When the linker sees this, it aligns the start of every individual
> "_ftrace_events" section, and that's where the "gaps" come from and
> the crashes.
OK (more to come in the following email response).
Mathieu
--
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com
Reply to: