On Thu, Jan 01, 2004 at 11:31:17PM +0100, Andi Kleen wrote:
> Arjan van de Ven <arjanv@redhat.com> writes:
> >
> > This is the first time I've seen this bug (thanks again for reporting it)
> > and it's clearly an API violation of the ONE API in linux which is very well
> > documented so no other platform will have copied this bad bug... Adaptect
> 
> It may be well documented, but ports do not follow the documentation
> very well in that area. In particular i386 handles the consistent dma mask
> completely different than all the 64bit ports. Small overview:
> 
> Architecture                  pci_alloc_consistent uses
> i386                          normal dma mask
				.... but clipped to 32 bit

> most 64bit archs/old x86-64   hardcoded 0xffffffff
> newer IA64/some x86-64        consistent dma mask (2) 
> newest x86-64                 normal dma mask & consistent dma mask (1) (2)
> old redhat IA64               hardcoded 0xffffffffffffffff (?)

not just old RH; At least one old SuSE too :0 (in fact I found a previous
report claiming this only affected SuSE but it seems we're in the same boat
now)

> 
> (1) to be bug-to-bug compatible with i386, some drivers rely on that.
> (2) consistent dma mask is normally equivalent to hardcoded 0xffffffff
> because that is the default value
>                                     
> It's quite a mess unfortunately. Due to (2) the 64bit architectures
> are consistent, except for that old IA64. But i386 doing something
> different is quite a problem.

but i386 will *never* return > 32 bit at least.