Re: [linux-next][PATCH] revert headers_check fix: ia64, fpu.h

From: Russell King - ARM Linux
Date: Fri Feb 06 2009 - 11:25:24 EST


On Fri, Feb 06, 2009 at 05:12:29PM +0100, Ingo Molnar wrote:
>
> * Russell King - ARM Linux <linux@xxxxxxxxxxxxxxxx> wrote:
>
> > > We cannot see any downside of this patch.
> > >
> > > But we can see upside of this patch is:
> > > 1. No need to protect linux/types.h with #ifndef __ASSEMBLY__ in many
> > > files
> > > 2. So we trying to replace multiple #ifndef __ASSEMBLY__ with one.
> >
> > The point is:
> >
> > 1. If the parent include needs to include linux/types.h to get at C
> > types _and_ the include file needs to also be included by assembly
> > code, it itself needs to have #ifndef __ASSEMBLY__ to protect those
> > uses from the assembly code.
> >
> > In that case, the linux/types.h include should be contained within
> > the #ifndef __ASSEMBLY__ .. #endif block along with all C only
> > parts of the header file.
>
> That makes the code much less clean: putting #include's in the middle of a
> header is poor style and leads to people failing to consider dependencies.
> We generally put them to the header portion.
>
> Putting an #include line in the middle of a header file is a receipe for a
> dependency hell (it can easily fall inside #ifdefs, can be overlooked,
> etc.), so it's _strongly_ discouraged (at least on arch/x86).

Put them at the top then with an additional ifndef.

> > 2. if it doesn't need C types from linux/types.h, then that header has
> > no business including linux/types.h, and the include should be
> > eliminated to save the already dirbolically slow compiler from
> > having to read and parse that file, and more importantly allowing
> > it to eliminate linux/types.h from the build dependencies.
> >
> > Yes, you can wrap linux/types.h with that ifndef, and yes it will fix any
> > problems, but I view it as a hack rather than fixing the real problem
> > which is lazyness by code writers to get their include dependencies right.
>
> It is not about include dependencies at all - it is about the existing and
> accepted practice which you did not consider in your argument: the use of
> mixed-mode headers. A linux/types.h include there is perfectly clean and
> should not break the build.

It _is_ about include dependencies.

> Again, i repeat: there is nothing wrong about making a small number of very
> commonly used C header assembly-invariant. It results in better structured
> header files and cleaner code.
>
> The argument is as simple as that, and up until this email you wrote roughly
> 10 replies and it's not even that you disagreed with our point on some
> honest basis that we could argue with - the ting is that you failed to even
> _realize_ this argument of us and you tried to force your partial (and
> trivially flawed) world view on us impatiently. In view of that you need to
> be more careful before calling people 'stupidly obtuse' ;-)

If you think that I'm ignoring your argument, then screw you.

I'm making a counter argument which I believe is _equally_ valid and which
I believe is actually far more important - at least it is _to_ _me_.

I don't wish to see my build times extended any more than they already
have been by sloppy and lazy programming in the kernel. And if I didn't
make that point earlier, what a shame. I'm making it NOW.

I've been trying over the last six months to (a) reduce the number of
files included in the ARM sub-tree and (b) reduce the namespace pollution
from ARM headers into the generic kernel both with the aim of trying to
improve build time. That's something you can verify by looking through
the commits in arch/arm/include/asm if you think I'm making this up,
which clearly you will do.

Oh why bring it up after 11 replies. I guess I'm just useless at
expressing myself clearly.
--
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/