Re: [Qemu-devel] MMIO address changes

[Edgar E. Iglesias]

On Wed, Dec 03, 2008 at 01:17:45PM +0100, Edgar E. Iglesias wrote:
> On Mon, Dec 01, 2008 at 06:59:35PM +0000, Paul Brook wrote:
> > I've just committed a patch that changes the MMIO callback interface for 
> > devices.  Instead of being passed an absolute address these are now passed 
> > an 
> > offset from the start[1] of the memory region that was registered.
> > 
> > By itself this change has fairly neutral benefit, it just moves logic 
> > about. 
> > However it makes subsequent dynamic board configuration bits nicer, and is 
> > a 
> > step towards a proper bus level API.
> > 
> > Most of the groundwork for this is already there, from my earlier changes 
> > to 
> > separate ram and MMIO addresses TLB handling.
> > 
> > The main notable change it that the PhysPageDesc structure is not bigger. 
> > This 
> > isn't ideal, however l2_phys_map needs to go away anyway, so I'm not really 
> > worried about this.  
> > 
> > Some devices register their memory regions in multiple segments. To 
> > facilitate 
> > this I have added cpu_register_physical_memory_offset.
> > 
> > Most of the remaining changes are fairly mechanical tweaks to fix devices 
> > that 
> > explicitly compensated for the absolute address. Many devices are untouched 
> > because they ignore the high bits of the address.
> > 
> > I've tried to be fairly thorough with the changes, and tested what I can. 
> > However it's possible I missed or broke something, so please test your 
> > favourite targets.
> 
> FWIW the etrax-fs machine works OK although I am seeing a noticeable slow-down
> after the patch.

I tracked this down and the MMIO patch was for some reason triggering a bug
in the ETRAX models causing the slowdown. ETRAX ethernet/dma works fine
again.

Cheers