On Fri, 17 Feb 2006, Jes Sorensen wrote: > > It's just horribly slow and an overkill if it's only ordering of consecutive > > writes rather than ordering of a read after a write that has to be > > guaranteed. But perhaps we could keep abusing wmb() for the former... > > Depends on where it's happening. If it's the CPU doing it before it hits > the bus then wmb() would probably be ok. These systems only allow the CPU to access devices other than RAM. But consistency between platforms is important as the same drivers may be used for systems that have different access rules. Examples are currently the defxx and the tgafb drivers. They should work correctly with the R3220 writeback buffer as well as with PCI. Maciej
