On Thu, Mar 29, 2018 at 10:19:41AM +0100, Will Deacon wrote: > On Wed, Mar 28, 2018 at 10:57:32AM -0600, Jason Gunthorpe wrote: > > On Wed, Mar 28, 2018 at 11:13:45AM +0100, Will Deacon wrote: > > > On Wed, Mar 28, 2018 at 09:01:27PM +1100, Benjamin Herrenschmidt wrote: > > > > On Wed, 2018-03-28 at 11:55 +0200, Arnd Bergmann wrote: > > > > > > powerpc and ARM can't quite make them synchronous I think, but at least > > > > > > they should have the same semantics as writel. > > > > > > > > > > One thing that ARM does IIRC is that it only guarantees to order writel() within > > > > > one device, and the memory mapped PCI I/O space window almost certainly > > > > > counts as a separate device to the CPU. > > > > > > > > That sounds bogus. > > > > > > To elaborate, if you do the following on arm: > > > > > > writel(DEVICE_FOO); > > > writel(DEVICE_BAR); > > > > > > we generally cannot guarantee in which order those accesses will hit the > > > devices even if we add every barrier under the sun. You'd need something > > > in between, specific to DEVICE_FOO (probably a read-back) to really push > > > the first write out. This doesn't sound like it would be that uncommon to > > > me. > > > > The PCI posted write does not require the above to execute 'in order' > > only that any bus segment shared by the two devices have the writes > > issued in CPU order. ie at a shared PCI root port for instance. > > > > If I recall this is very similar to the ordering that ARM's on-chip > > AXI interconnect is supposed to provide.. So I'd be very surprised if > > a modern ARM64 has an meaningful difference from x86 here. > > From the architectural perspective, writes to different "peripherals" are > not ordered with respect to each other. The first writel will complete once > it gets its write acknowledgement, but this may not necessarily come from > the endpoint -- it could come from an intermediate buffer past the point of > serialisation (i.e. the write will then be ordered with respect to other > accesses to that same endpoint). The PCI root port would look like one > peripheral here. That is basically the same as PCI - PCI has no write ACK, so all writes are buffered by the PCI interconnect and complete in some undefined temporal order when multiple end points are involved. This does not seem very different from what happens in x86.. > > When talking about ordering between the devices, the relevant question > > is what happens if the writel(DEVICE_BAR) triggers DEVICE_BAR to DMA > > from the DEVICE_FOO. 'ordered' means that in this case > > writel(DEVICE_FOO) must be presented to FOO before anything generated > > by BAR. > > Yes, and that isn't the case for arm because the writes can still be > buffered. The statement is not about buffering, or temporal completion order, or the order of acks returning to the CPU. It is about pure transaction ordering inside the interconnect. Can write BAR -> FOO pass write CPU -> FOO? Jason -- To unsubscribe from this list: send the line "unsubscribe linux-rdma" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
