Re: [PATCH v1 15/22] Add iommufd configure option

[Alex Williamson]

On Wed, 20 Sep 2023 14:49:19 -0300
Jason Gunthorpe <jgg@nvidia.com> wrote:

> On Wed, Sep 20, 2023 at 07:37:53PM +0200, Eric Auger wrote:
> 
> > >> qemu will typically not be able to
> > >> self-open /dev/iommufd as it is root-only.  
> > >
> > > I don't understand, we open multiple fds to KVM devices. This is the
> > > same.  
> > Actually qemu opens the /dev/iommu in case no fd is passed along with
> > the iommufd object. This is done in
> > [PATCH v1 16/22] backends/iommufd: Introduce the iommufd object, in
> > 
> > iommufd_backend_connect(). I don't understand either.  
> 
> The char dev node is root only so this automatic behvaior is fine
> but not useful if qmeu is running in a sandbox.
> 
> I'm not sure what "multiple fds to KVM devices" means, I don't know
> anything about kvm devices..

Looking at a local VM, the only kvm related open file is /dev/kvm,
which kvm_init() does directly open.  The other tun/tap/vhost files are
all passed by fd.  We have a bunch of anon_inodes representing eventfds
and vcpu source from /dev/kvm, but the only other direct files are disk
images and the created pid file.
 
> The iommufd design requires one open of the /dev/iommu to be shared
> across all the vfios.

"requires"?  It's certainly of limited value to have multiple iommufd
instances rather than create multiple address spaces within a single
iommufd, but what exactly precludes an iommufd per device if QEMU, or
any other userspace so desired?  Thanks,

Alex