On Wed, Aug 19, 2026 at 09:29:09AM -0700, Jacob Pan wrote: > Hi Jason, > > On Wed, 19 Aug 2026 09:46:52 -0300 > Jason Gunthorpe <[email protected]> wrote: > > > On Tue, Aug 18, 2026 at 05:13:29PM -0700, Mukesh R wrote: > > > On 8/18/26 16:48, Jason Gunthorpe wrote: > > > > On Tue, Aug 18, 2026 at 04:39:32PM -0700, Mukesh R wrote: > > > > > > > > > Thanks for the review. I know there is another set of patches > > > > > for pvIOMMU ongoing (both came from the same source) and you > > > > > are probably repeating things, and I appreciate your patience. > > > > > I'll also try to look for your comments in that patch series > > > > > going forward. > > > > > > > > Those patches look pretty good so you should try to copy their > > > > stuff :) > > > > > > > > They ended up never calling detach, I hope you can do that too. > > > > Just > > > > > > Unfortunately, my case is a bit different due to the "direct attach" > > > feature for guest VMs. When a device moves from say a paging domain > > > to be directly attached to a VM, today, it must first be detached. > > > But since that just moves to identity, I am discussing with hyp > > > team if detach is really required, and if we can get rid of it via > > > just attaching to identity/blocked first before doing the direct > > > attach. Hopefully, it can be done even if changes are needed in the > > > hypervisor. > > > > You definately cannot have it blip to identity, that is security > > unacceptable for VFIO. > > > > Once it is attached to VFIO it must only be blocked or under a > > translation controlled by VFIO. Never something else. > Based on this requirement, I don't see there is a difference between > this root driver and the guest driver in terms of VFIO DMA ownership > transition in the layer above the immediate hypercall site. i.e. > > old domain -> blocking domain -> new target domain > > Regardless of source or target domain type (e.g. direct attach as > the target domain), the blocking domain must be attached. > > If the hypervisor detach operation internally implies transition > through identity, then it needs to be hidden behind the blocking-domain > attach transaction.
The best thing is to do as the other driver did and just always and only do attach. Hypervisor just changes the attachment instead of demanding a pointless detach. Jason
