Re: [RFC PATCH 2/5] conf: Add an iommufd member struct to virDomainIOMMUDef

> Add support for an "iommufd" virDomainIOMMUDef member that will be > translated to a qemu "-object iommufd" command line argument. This > iommufd struct has an "id" member to specify the iommufd object id > that can be associated with a passthrough device, and an "fd" > member to specify the fd for externally opening the /dev/iommu and > VFIO cdev by a management layer. > > @@ -28310,6 +28406,16 @@ virDomainIOMMUDefFormat(virBuffer *buf, > > virXMLFormatElement(&childBuf, "driver", &driverAttrBuf, NULL); > > + if (iommu->iommufd) { > + virBufferAddLit(&childBuf, "<iommufd>\n"); > + virBufferAdjustIndent(&childBuf, 2); > + virBufferAsprintf(&childBuf, "<id>%s</id>\n", iommu->iommufd->id); > + if (iommu->iommufd->fd) > + virBufferAsprintf(&childBuf, "<fd>%s</fd>\n", iommu->iommufd->fd); I'm not convinced we should be exposing the QEMU "id" in the XML. It is an identifier libvirt uses internally for configuring QEMU, but I don't see a reason why the end user needs to control its name. I'm similarly not sure we should expose an FD to users eitherm, just use that internally when talking to QEMU. IOW, why aren't we just having "iommu='yes'" as an attribute on the <iommu> device, avoiding all the logic below that tries to force all devices to use the same configuration.