On Mon, Feb 25, 2008 at 11:14:33AM +0100, Ingo Molnar wrote:
> 
> * Florian Fainelli <florian.fainelli@telecomint.eu> wrote:
> 
> > This patch adds the default kernel configuration for the RDC R-321x 
> > SoC.
> 
> hm, i'm not sure. Right now we just have a 32-bit defconfig and a 64-bit 
> defconfig - but there are about 8 subarchitectures in arch/x86. Given 
> the amount of variety in PC hardware, i doubt it makes sense to start 
> collecting defconfigs for hardware variants - we'd end up having 
> hundreds or thousands of them. Even ARM has only 75 defconfigs.

What I want is at least one defconfig per subarchitecture for compile 
tests.

And especially considering the original purpose "configuration users can 
use as a starting point for configuring their kernel" I even wouldn't 
mind if we had a few dozen x86 defconfigs.

> what i do is i regularly test whether "make allyesconfig" boots all the 
> way up to general user-space in regular whitebox PC hardware. For 
> example the attached config is such a config, i successfully booted it 
> on 2.6.25-rc3 on a stock PC.

You are testing something completely different here.

What I want is that e.g. after fiddling with kernel headers I want an 
easy way of having much compile coverage. And my script that builds all 
defconfig's is trivial (although it takes a day to finish).

> This way we can ensure that the (near-) totality of the config space is 
> bootable on regular PCs, and the subarch support is basically just 
> bootstrap and quirks differences.

You miss our headers mess.

You remember how your big x86 merge this merge window broke 8 or 9 other 
architectures? Change one file under include/ and watch how many 
configurations no longer build.

Or other subtle differences between the subarchs that have in the past 
led to compile errors.

I do consider them useful for the way I'm doing kernel tests, and even 
if you don't consider them that useful can we agree that adding a 
defconfig is neither a big deal for the subarchitecture maintainer nor 
imposes any maintainance work on you as maintainer (except for sometimes 
applying patches adding/updating them)?

> Longer term we should get rid of the 
> subarchitecture distinction altogether and turn them into regular 
> quirks/callbacks/drivers.
>...

Generally agreed (with my biggest worry being whether changing 
CLOCK_TICK_RATE from a compile time constant to a runtime
variable has any performance effects).

> 	Ingo

cu
Adrian

-- 

       "Is there not promise of rain?" Ling Tan asked suddenly out
        of the darkness. There had been need of rain for many days.
       "Only a promise," Lao Er said.
                                       Pearl S. Buck - Dragon Seed