Bug#609371: linux-image-2.6.37-trunk-sparc64: module scsi_mod: Unknown relocation: 36
* David Miller (email@example.com) wrote:
> From: Mathieu Desnoyers <firstname.lastname@example.org>
> Date: Wed, 19 Jan 2011 00:08:45 -0500
> > The following works fine for me now. Comments are welcome.
> Thanks for doing this work Mathieu.
> > - 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.
> This is really bizarre. Does it only happen on x86_64?
Sadly, my ppc32 test machine is currently broken, so I could not check on other
than x86 archs.
> I'm wondering if GCC does something bizarre like work with different
> default alignments based upon the section or something like that.
> If so, maybe adding the section attribute to the array definition will
> "fix" things?
Well, I thought about it in my sleep, and it looks like gcc is within its rights
to align these statically declared structures on a larger alignment: gcc has no
clue that we're going to do tricks with the linker to access the structures as
an array, so aligning on a larger alignment *should* be fine for the compiler,
but we suffer because we're doing something non-standard.
> > On 32-bit architectures, we really want a aligned(4), and on 64-bit
> > architectures, aligned(8). Represent this by creating:
> > #define __long_aligned __attribute__((__aligned__(__alignof__(long))))
> Do any of these datastructures have, or will have, "u64" or "long
> long" types in them? If so, then we will need to use "8"
> unconditionally or "__alignof__(long long)".
If my memory serves me correctly, I think "long long" is aligned on 4 bytes on
ppc32, but on 8 bytes on x86_32 (yeah, that's weird). How about we create a
#define __long_long_aligned __attribute__((__aligned__(__alignof__(long long))))
> I'll see if I can work out why using no align directive explodes
> on x86-64.
Operating System Efficiency R&D Consultant