> From: Jason Gunthorpe <jgg@xxxxxxxx> > Sent: Tuesday, April 1, 2025 10:12 PM > > On Wed, Mar 12, 2025 at 07:53:17AM +0000, Tian, Kevin wrote: > > > Probably we should not allow device-specific mode unless the user is > > capable of CAP_SYS_RAWIO? It allows an user to pollute caches on > > CPUs which its processes are not affined to, hence could easily break > > SLAs which CSPs try to achieve... > > I'm not sure this is within the threat model for VFIO though.. > > qemu or the operator needs to deal with this by not permiting such > HW to go into a VM. it could be used by native app e.g. dpdk. > > Really we can't block device specific mode anyhow because we can't > even discover it on the kernel side.. > hmm the TPH capability reports which steering modes (no st, irq vector, or device specific) are supported by a device. and the mode must be selected explicitly when sw enables the capability. so policy-wise vfio could advocate/enforce that only the interrupt vector mode is supported for non-privileged users.
