On 9/23/22 08:44, Ashok Raj wrote: >> But, the point that Kirill and I were getting at is still that devices >> *have* a role to play here. The idea that this can be hidden at the >> IOMMU layer is pure fantasy. Right? > If you *can't* send tagged memory to the device, what is the > role the device need to play? > > For now you can only send proper VA's that are canonical. Today, yes, you have to keep tagged addresses away from devices. They must be sequestered in a place that only the CPU can find them. The observation that Kirill and I had is that there are thing that are done solely on the device today -- like accessing a translated address twice -- without IOMMU involvement. We were trying to figure out how that would work in the future once tagged addresses are exposed to devices and they implement all the new PCI magic. After our private chat, I think the answer is that devices *have* a role to play. Device-side logic must know how to untag memory before asking for translation or even *deciding* to ask for address translation. But, hopefully, the communicating that untagging information to the device will be done in a device-agnostic, standardized way, just how PASIDs or ATS are handled today.
