ioremap of pci region on pSeries LPAR vs SMP

[Anton Blanchard]

Hi,

> It was an enormous simplification and Linus was keen to do it.  He
> actually looks at our code from time to time now that his desktop
> machine is a G5. :)

Roland (the infiniband guy) and Linus were behind it:

http://marc.theaimsgroup.com/?l=linux-kernel&m=109579598620069&w=2

Looks like it was due to __raw_* not having any EEH checks.

As a side note its a worry that we dont have IO macros that order but
dont byte swap. __raw_* (which doesnt order) is going to catch out a
lot of driver writers I suspect.

> Some architectures do in fact use ioremap cookie poisoning for that
> reason.  We could do that as a debug option.

Ive seen HPC stuff that wants to be able to mmap a PCI cards resources into
userspace. Their hack on ppc64 was to look at the high nibble of the
address and convert it to a non EEH address if required :)

Im not sure how best to solve the userspace mmap issue but there are a
few groups wanting that.

Anton