On Fri, Jan 16, 2015 at 03:40:28PM +0000, Arnd Bergmann wrote: > On Friday 16 January 2015 15:33:20 Will Deacon wrote: > > On Fri, Jan 16, 2015 at 03:14:13PM +0000, Arnd Bergmann wrote: > > > On Friday 16 January 2015 14:55:45 Will Deacon wrote: > > > > On Fri, Jan 16, 2015 at 02:45:30PM +0000, Tom Lendacky wrote: > > > > > I have tested ACPI-enablement patches for the amd-xgbe/amd-xgbe-phy > > > > > drivers that I'm about to submit upstream with the V7 patch series > > > > > on the AMD Seattle server platform. There does not appear to be support > > > > > for the _CCA attribute in this patch series. The amd-xgbe driver will > > > > > setup the device domain and cache attributes based on the presence of > > > > > this attribute, but it requires the arch support to assign the proper > > > > > DMA operations in order for it to all work correctly. > > > > > > > > > > Overriding the _CCA attribute in the driver, I was able to successfully > > > > > test the driver and this patch series. > > > > > > > > Hopefully this will all be addressed when the IORT parts of ACPI have > > > > settled down (the current proposal allows for these attributes to be > > > > described as well as their interaction with things like IOMMUs). > > > > > > > > In the meantime, are you falling back to non-coherent DMA? If so, what > > > > attributes have you settled on? We need to be really careful not to > > > > corrupt data during cache invalidatation when mapping a non-coherent > > > > buffer for the CPU. > > > > > > I think in case of ACPI we should use cache-coherent as the default, > > > as this is what all servers will use for DMA masters. > > > > I don't agree. The dma-coherent we have for device-tree isn't nearly > > expressive enough for the kind of things we want to describe and there's > > no reason to make the same mistake in ACPI, especially as it *is* being > > addressed by IORT. If we run with _CCA, then we're going to be stuck > > supporting something that isn't fit for purpose and which will likely be > > abused to describe both fixed features of the system and software > > configuration preferences. It also opens up a can of worms if we have to > > support a mixture of _CCA and IORT in the future. > > > > Or are you suggesting that we ignore _CCA and just assume cache-coherency? > > In that case, how do we support systems that aren't cache coherent, where > > not being cache coherent includes devices that require either device or > > IOMMU configuration to enable cacheable transactions? > > I was thinking we'd ignore _CCA because as you say a simple on/off flag > would not be enough to describe what we have to do for noncoherent > devices. I can't think of any reason why a server hardware would include > noncoherent devices, so if they are configurable they should be configured > into coherent mode by the firmware. The on-board ethernet on Seattle requires the driver to program its AXI attributes, so configuring it to be a coherent master actually means "program the same cacheable AXI settings as you have on the CPU". That sounds like Linux should be doing it to me, but even if the firmware takes a guess at "normal cacheable WBRWA", it's not clear to me whether that register persists across things like adapter reset. Tom? There's also the situation where the firmware hasn't initialised the register and Linux realises this during probe. What should it do then? Will -- To unsubscribe from this list: send the line "unsubscribe linux-acpi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
