Re: [microblaze-uclinux] [PATCH 03/11] microblaze: fall back on generic header files for the ABI

[Arnd Bergmann]

On Thursday 02 July 2009, John Williams wrote:
> However, we've just "broken" the ABI in 2.6.31, if we merge further
> ABI breakage in 2.6.32 it's more pain and confusion.
> 
> So, unless we can merge and validate this ABI breakage during the
> 2.6.31-rc cycle, I think we need to hold on changes that would break
> the ABI again, so soon.
> 
> Longer term there will be a complete redo of glibc up to the latest
> version, which will obviously require a new toolchain for users - I
> think that is the right place to do the next round of ABI breakage.
> 
> Any thoughts?

Normally the rule is to never break the ABI at all, even in a major
update of glibc.

Initially I had hoped that we could do the change before 2.6.30 and
I worked hard on getting the ABI patches to you early enough for
that. After that failed, I made sure that you had everything in
place for 2.6.31 and I believed that Michal said he would integrate
that in the microblaze-mmu merge, before you actually start seeing
users on the mainline kernel.

If we don't get that into 2.6.31 any more (and the chances
have slimmed down a lot after the end of the merge window), I
think we should just declare complete failure and not touch
the ABI any more, however broken it may be. 

In over 14 months, not even the most basic fixes to the ABI
that I mentioned in the first review have been applied:

On Tuesday 15 April 2008, Michal Simek wrote:
> Arnd Bergmann wrote:
> > * Your syscall ABI is largely obsolete. A third of the syscalls you
> > define should not even be there in the first place as they have been
> > replaced by newer versions. E.g. you have select, _newselect and pselect6,
> > where just pselect6 would be sufficient -- you don't need to worry about
> > backwards compatibility if you don't have existing code. A good start
> > would be to take the arch/blackfin syscall list and further reduce it
> > from there. Further examples are:
> > - replace socketcall with separate syscall entry points
> > - replace ipc with a separate entry points
> > - remove all the legacy signal handling from arch/microblaze/kernel/signal.c
> > - remove sys_mmap, sys_olduname and sys_vfork
> > - finally define a generic sys_mmap2 and sys_uname in kernel/ so you don't
> >   need another copy in arch/microblaze/kernel
> > - Use 64 bit off_t, and 32 bit uid_t, gid_t etc.
> 
> This kernel don't need to keep backward compatibility. No one will port to
> previous version. I'll look at your points and I'll send you what I do.

	Arnd <><
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/