On Mon, 15 Apr 2019 18:43:24 -0400 Tony Krowiak <akrowiak@xxxxxxxxxxxxx> wrote: > On 4/15/19 2:59 PM, Halil Pasic wrote: > > On Mon, 15 Apr 2019 12:51:23 -0400 > > Tony Krowiak <akrowiak@xxxxxxxxxxxxx> wrote: > > > >> Having said that, I understand your concern about a driver hogging > >> resources. I think I can provide a solution that serves both the > >> purpose of preventing problems associated with accidental removal > >> of AP resources as well as allowing root to remove them > >> forcefully. I'll work on that for v2. > > > > Can you tell us some more about this solution? Should we stop reviewing > > v1 because v2 is going to be different anyway? > > Patch 1 and 2 will be removed. There will not be a major design change > between these patches and v2. In order to avoid a long explanation of > my proposed changes, I'd prefer to state that the patch set will > establish and enforce the following rules: > > 1. An APQN can be assigned to an mdev device iff it is NOT > reserved for use by a zcrypt driver and is not assigned to > another mdev device. > > 2. Once an APQN is assigned to an mdev device, it will remain > assigned until it is explicitly unassigned. > > 3. A queue's APQN can be set in the guest's CRYCB iff the APQN is > assigned to the mdev device used by the guest; however, if the > queue is also in the host configuration (i.e., online), it MUST > also be bound to the vfio_ap device driver. > > 4. When a queue is bound to the vfio_ap driver and its APQN > is assigned to an mdev device in use by a guest, the guest will > be given access to the queue. > > 5. When a queue is unbound from the vfio_ap driver and its APQN > is assigned to an mdev device in use by the guest, access to > the card containing the queue will be removed from the guest. > Keep in mind that we can not deny access to a specific queue > due to the architecture (i.e., clearing a bit in the AQM > removes access to the queue for all adapters) > > 6. When an adapter is assigned to an mdev device that is in use > by a guest, the guest will be given access to the adapter. > > 7. When an adapter is unassigned from an mdev device that is in use > by a guest, access to the adapter will removed from the guest. > > 8. When a domain is assigned to an mdev device that is in use > by a guest, the guest will be given access to the domain. > > 9. When a domain is unassigned from an mdev device that is in use > by a guest, access to the domain will removed from the guest. > Based on our off-the-list chat and this list I think I know where are you heading :). I think it's actually the design that I currently prefer the most. But in that case, it may be wise to touch base with Reinhard -- AFAIR he was the strongest proponent of the 'do not let a[pq]mask changes take away queues from guests' design. Regards, Halil
