> From: Alex Williamson > Sent: Friday, January 26, 2024 5:20 AM > > On Thu, 25 Jan 2024 09:18:02 +0100 > Eric Auger <eau...@redhat.com> wrote: > > > Hi Vivek, > > > > On 1/18/24 20:20, Vivek Kasireddy wrote: > > > > > > + if (iommu_phys_bits && phys_bits > iommu_phys_bits) { > > > + phys_bits = iommu_phys_bits; > > are you allowed to change the host cpu characteristics without taking > > care of compats for migration? > > Not only is migration an issue, but so is hotplug. Anything that > aligns the VM configuration to the host is going to have migration > issues and anything that does so conditionally based on the current set > of attached devices will have hotplug issues.
Does it make more sense to print out a warning so it's admin to figure out the proper phys_bits workable with the iommu restriction, instead of forcing/hard-coding any value implicitly? > > It'd be more prudent to find the least common denominator under > /sys/class/iommu/dmar*/intel-iommu/cap regardless of attached devices, > but it still affects migration compatibility between hosts. > > Also note that vfio-pci can specify the device with either host= or > sysfsdev= and with vfio cdev support and iommufd the file descriptor > for the vfio device might be passed via SCM rights, so we may not have > a reference to it in sysfs. The above appears to only work with the > host= device specification. Thanks, > iommufd supports a IOMMUFD_CMD_GET_HW_INFO cmd which can also report cap/ecap on intel iommu.