Bug#609371: linux-image-2.6.37-trunk-sparc64: module scsi_mod: Unknown relocation: 36
* Steven Rostedt (rostedt@goodmis.org) wrote:
> On Wed, 2011-01-19 at 11:15 -0500, Mathieu Desnoyers wrote:
> > * Steven Rostedt (rostedt@goodmis.org) wrote:
> > > After applying David's "remove align" patch, I got it to boot on x86_64
> > > with the following two patches. I thought just adding the "align" to the
> > > structure declaration would work, but it still failed on the syscall for
> > > init_module. By removing the double "declaration" of event_exit_##sname,
> > > removed this problem.
> > >
> > > I'll test this on x86 32bit and PPC 64. If it works there, I'll push all
> > > of them out for 38. Should these go to 37 stable too?
> >
> > Please hold before adding these patches into git. They don't seem to address the
> > underlying problem correctly. See the latest exchanges between David Miller and
> > myself for more info.
> >
> > We need to come up with something better than "it boots" as an explanation for
> > the fix.
>
> Yes, I agree that we should solve this issue correctly. But if there is
> a work around to the problem, we could implement that if the real
> solution is not in our grasp yet.
A known working workaround (used in tracepoints for a few years) is to align the
type declaration on 32 bytes. It wastes space, but works. With this solution,
you should remove all the per-variable alignment attributes.
Now what I'm discussing with David Miller is if creating a
__long_packed_aligned
and using it for *both* type and variable alignment would be more palatable (it
also works, and is more compact).
David proposed a solution with an array of pointers (extra indirection) which I
don't really like for 3 reasons I exposed in my reply to him.
So it's not that the solution is not in our grasp yet, it's more that we have to
choose the right one.
Thanks,
Mathieu
--
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com
Reply to: