Re: Fwd: evbarm ldscript vs. KERNEL_BASE_PHYS != KERNEL_BASE_VIRT

On Wed, Mar 12, 2008 at 06:11:19PM +0000, Richard Earnshaw wrote:
> 
> On Wed, 2008-03-12 at 07:48 -0400, Rafal Boni wrote:
> > My question is whether it was only done for text & data (not bss, sbss, 
> > etc.) due to an oversight, or was this intentional?  It certainly seems 
> > wrong to me, but I don't own any of the evbarm-supported eval boards, so 
> > can't say it has no side effects there.
> 
> The LMA view is only used until the MMU is turned on.  Since nobody
> would want to try and zero the BSS with the MMU/caches off (far too
> slow) I don't think anybody has thought about it. 

Heh, on CE using hpcboot, the load is *much much* slower than any zeroing
we might do with the MMU off.  Part of that might be that it's getting non-
cached pages from CE (it passes the no-cache flag explicitly), but I don't
think that's close to all of it, since the second copy it does with the
MMU off (to get the kernel into the actual correct physically contiguous
space) goes a *lot* faster.

And if nobody bothers zero'ing the BSS with the caches/MMU off, they also
shouldn't care if/what the phys. addr. of the section is set to... so my
tweak should be safe :)

Thanks,
--rafal