On Tue, 2022-07-05 at 18:08 +0100, Robin Murphy wrote: > v2: https://lore.kernel.org/linux-iommu/cover.1650890638.git.robin.murphy@xxxxxxx/ > > Hi all, > > Here's v3, now with working x86! Having finally made sense of how I > broke Intel, I've given AMD the same fix by inspection. I'm still not > 100% sure about s390, but it looks like it should probably be OK since > it seems to register an IOMMU instance for each PCI device (?!) before > disappearing into PCI hotplug code, wherein I assume we should never see > a PCI device appear without its IOMMU already registered. Yes, this is a bit unusual as our PCI architecture doesn't really have a notion of an IOMMU device only of I/O translation tables. These are then registered per PCI function. PCI functions may share I/O translation tables and thus DMA address spaces but this is not done at the moment. As Matt already mentioned we do need a small change for this patch series. Since that was still mangled in his mail for me I just replied with that using "git send-email". With Matt's patch applied I can confirm that this works fine for us and does look like a useful simplification. So feel free to add my Tested-by: Niklas Schnelle <schnelle@xxxxxxxxxxxxx>
