On Tue, 2008-06-03 at 16:11 +1000, Nick Piggin wrote:

quoted

      - readl is synchronous (ie, makes the CPU think the
          data was actually used before executing subsequent
          instructions, thus waits for the data to come back,
          for example to ensure that a read used to push out
          post buffers followed by a delay will indeed happen
          with the right delay).

So your readl can pass an earlier cacheable store or earlier writel?

I forgot to mention that all MMIO are ordered vs. each other and I
do prevent readl from passing earlier cacheable stores too in my
current implementation but I'n not 100% we want to "guarantee" that,
unless we have stupid devices that trigger DMA's on reads with side
effects.. anyway, it is guaranteed in the current case.

Ben.