Re: [PATCH 14/15] x86: convert to use __HEAD and HEAD_TEXT macros.

[Ingo Molnar <mingo@elte.hu>]

* Linus Torvalds <torvalds@linux-foundation.org> wrote:

> On Sun, 26 Apr 2009, Sam Ravnborg wrote:
> 
> > On Sun, Apr 26, 2009 at 10:12:47AM -0700, Linus Torvalds wrote:
> > > 
> > > 
> > > On Sun, 26 Apr 2009, Linus Torvalds wrote:
> > > > 
> > > > Btw, this one really needs to unify the two lds files first. Look at 
> > > > 
> > > > 	diff -u arch/x86/boot/compressed/vmlinux_*.lds
> > > > 
> > > > output and realize that they're basially exctly the same except for 
> > > > trivial naming differences, and the fact that the64-bit version hs a 
> > > > "pgtable" thing.
> > > > 
> > > > So this really needs to be done by first unifying the thing so that there 
> > > > is _one_ arch/x86/boot/compressed/vmlinux.lds.S file with a preprocessor 
> > > > that takes care of the trivial differences [..]
> > > 
> > > Something like this?
> > 
> > Looks good/correct.
> > Acked-by: Sam Ravnborg <sam@ravnborg.org>
> > 
> > You should add your s-o-b if you expect Ingo to pick it up.
> 
> Sure. I don't tend to add SOB lines for stuff that I'd not be 
> ready to commit, but with some testing and other people looking at 
> it, I think it's good to go.
> 
> 	Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Thanks, applied to tip:x86/kbuild. I'll do some more testing of it 
before pushing it out.

> As mentioned, though, the much more interesting case would be the 
> _real_ kernel vmlinux.lds.S file, which is a lot more complex and 
> where the differences between 32-bit and 64-bit cases aren't 
> totally trivial.
> 
> Looking at 
> 
> 	diff -u arch/x86/kernel/vmlinux_*.lds.S | less -S
> 
> output, many of them are just whitespace, and others are trivial 
> and meaningless (comments in one, not the other, placement of 
> alignment etc, different ordering of sections like 
> "parainstructions"). Yet others seem to be things that we _could_ 
> do in general, but that don't matter on one architecture or other 
> (x86-64 has ".eh_frame" in the DISCARD section, i386 apparently 
> doesn't ever generate them, we could just use the x86-64 version).

We generally do these by separating the unification into at least 
2-3 distinct steps - a mechanic, low-risk cleanup first, preparatory 
changes to bring the two files in sync second, and mechanic 
unification as the third and final step.

That way any bugs are easily bisectable to a reasonably sized (and 
reasonably risky) sub-patch. Review also gets much easier.

I've yet to see a non-trivial Makefile unification in arch/x86 that 
does not regress :-) They concentrate a lot of quirks and implicit 
dependencies and small but significant tricks. [usually we catch the 
bugs early on though - but even at an early stage it's good to have 
a reasonable splitup.]

	Ingo