On Tue, 8 Nov 2022 20:54:58 -0400 Jason Gunthorpe <jgg@xxxxxxxxxx> wrote: > On Tue, Nov 08, 2022 at 03:28:31PM -0700, Alex Williamson wrote: > > > Perhaps this should have been obvious, but I'm realizing that > > vfio-noiommu mode is completely missing without VFIO_CONTAINER, which > > seems a barrier to deprecating VFIO_CONTAINER and perhaps makes it a > > Yes, it is the same as the allow_unsafe_interrupts - it is something > that currently goes missing if you turn off VFIO_CONTAINER. > > This seems straightforward enough to resolve in a followup, we mostly > just need someone with an existing no-iommu application to test > compatability against. Keeping it working with the device cdev will > also be a bit interesting. If you have or know about some application > I can try to make a patch. DPDK supports no-iommu mode. > > question whether IOMMUFD should really be taking over /dev/vfio/vfio. > > No-iommu mode has users. > > I view VFIO_CONTAINER=n as a process. An aspiration we can work > toward. > > At this point there are few places that might want to use it. Android > perhaps, for example. It is also useful for testing. One of the main > values is you can switch the options and feed the kernel into an > existing test environment and see what happens. This is how we are > able to quickly get s390 mdev testing, for instance. > > We are not going to get to a widely useful VFIO_CONTAINER=n if we > don't have a target that people can test against and evaluate what > compatability gaps may exist. > > So, everytime we find something like this - let's think about how can > we make iommufd compatibility handle it and not jump straight to > giving up :) > > I'm kind of thinking v6.4 might be a reasonable kernel target when we > might have closed off enough things. I agree that it's very useful for testing, I'm certainly not suggesting to give up, but I'm not sure where no-iommu lives when iommufd owns /dev/vfio/vfio. Given the unsafe interrupts discussion, it doesn't seem like the type of thing that would be a priority for iommufd. We're on a path where vfio accepts an iommufd as a container, and ultimately iommufd becomes the container provider, supplanting the IOMMU driver registration aspect of vfio. I absolutely want type1 and spapr backends to get replaced by iommufd, but reluctance to support aspects of vfio "legacy" behavior doesn't give me warm fuzzies about a wholesale hand-off of the container to a different subsystem, for example vs an iommufd shim spoofing type1 support. Unfortunately we no longer have a CONFIG_EXPERIMENTAL option to hide behind for disabling VFIO_CONTAINER, so regardless of our intentions that a transition is some time off, it may become an issue sooner than we expect. Thanks, Alex
