Re: [RFC] [PATCH] Fix b43 driver build for arm

["Gordon Farquharson" <gordonfarquharson@gmail.com>]

On Feb 18, 2008 5:01 PM, Michael Buesch <mb@bu3sch.de> wrote:
>
> On Tuesday 19 February 2008 00:42:12 Sam Ravnborg wrote:
> > On Tue, Feb 19, 2008 at 12:17:04AM +0100, Michael Buesch wrote:
> > > On Tuesday 19 February 2008 00:00:58 Russell King wrote:
> > > > > > Why can't we have an array of this structure on ARM?
> > > > > >
> > > > > > struct ssb_device_id {
> > > > > >        __u16   vendor;
> > > > >
> > > > > 2 bytes
> > > > >
> > > > > >        __u16   coreid;
> > > > >
> > > > > 2 bytes
> > > > >
> > > > > >        __u8    revision;
> > > > >
> > > > > 1 byte
> > > > >
> > > > > > };
> > > > >
> > > > > and therefore sizeof this structure will be 5 bytes, but because of the
> > > > > ABI rules (which are _explicitly_ allowed by the C standard), it'll
> > > > > become 8 bytes due to padding afterwards.
> > > >
> > > > Another guess might be that, if using AEABI, this structure might
> > > > be 6 bytes in size, but the linker will align structures to 4 bytes.
> > >
> > > If the struct is padded to 6 bytes and the linker aligns it to 4 byte
> > > everything will be naturally aligned, as far as I can see.
> > >
> > > > FATAL: drivers/net/wireless/b43/b43: sizeof(struct ssb_device_id)=6 is
> > > > not a modulo of the size of section __mod_ssb_device_table=64.
> > > > Fix definition of struct ssb_device_id in mod_devicetable.h
> > >
> > > So this message tells me the table size is 64 bytes. There are 8 entries,
> > > so it seems the structure is padded to 8 bytes.
> > > But above that it says that sizeof(struct ssb_device_id)=6
> > >
> > > IMO this sanity check is broken and not the code.
> > >
> > > Where does this sanity check message come from? The linker?
> > $ git grep 'not a modulo'
> > scripts/mod/file2alias.c:               fatal("%s: sizeof(struct %s_device_id)=%lu is not a modulo "
> >
> > It is a consistencycheck between host and target
> > layout of data.
> > You need to pad the structure so it becomes 8 byte in size.
>
> Ok, I looked at the code and it is hightly questionable to me that this
> check does work in a crosscompile environment (which the ARM build
> most likely is).
>
> It seems to check the size of the structure in the .o file against
> the size of the structure on the _host_ where it is compiled.
> I can't see why we would want to check _anything_ of the target stuff
> to the host this stuff is compiled on.
> I can compile an ARM kernel on any machine I want.
>
> There actually is a comment:
>  * Check that sizeof(device_id type) are consistent with size of section
>  * in .o file. If in-consistent then userspace and kernel does not agree
>  * on actual size which is a bug.
>
> So it seems what this check _wants_ to compare the sizeof the structure
> in the kernel to the size of the stucture in the userland of the target system.
> But it does _not_ do that.
> It does compare the size of the structure in the kernel against the size of
> the stucture in userland on the machine it is _compiled_ on.
> That is wrong.

Does this thread [1] provide any clues as to the Right Thing (TM) to do?

It should be noted that Linus and Andrew signed off on the m68k fix
[2]. I'm CC'ing them and Al Viro on this email to solicit their input.

Gordon

[1] http://www.gossamer-threads.com/lists/linux/kernel/801528
[2] http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=7492d4a416d68ab4bd254b36ffcc4e0138daa8ff

-- 
Gordon Farquharson
GnuPG Key ID: 32D6D676