Recovering from IOMMU and/or KVM MMU faults

Recovering from IOMMU and/or KVM MMU faults

[Demi Marie Obenour]

[-- Attachment #1.1: Type: text/plain, Size: 833 bytes --]

I'm designing a new type of virtio device (vhost-guest, was virtio vhost-user), and I'm running into an interesting design question.

Specifically, the frontend (a vhost-user client) can ask the device (a vhost-user server) to ummap memory.  This is handled by replacing the mapping with a PROT_NONE one.  However, the guest might be accessing that memory.  Furthermore, that memory might also be mapped into the IOVA of a VFIO device for zero-copy I/O.

If the guest is accessing the memory, KVM will get a fault.  If the PCI device is accessing it, there will be an IOMMU fault.  Is it possible to recover from this?

If not, then the only option is to wait for the guest to release the memory itself.  If it does not do so, that's a memory leak (host denial of service).
-- 
Sincerely,
Demi Marie Obenour (she/her/hers)

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]