ioremap with cache on 44x

Hi,

I want to ioremap some memory on a PCI board with caching from a 440EPx
CPU. Typically ioremap() ends in __ioremap() with the flags set to _PAGE_NO_CACHE.
But even without this flag, the remapped range does not use the cache.
Using a BDI2000 debugger I always see the corresponding TLBs with 
caching inhibited.

How can I access PCI memory with caching?

Matthias