Re: [PATCH v3 1/1] s390: virtio: let arch accept devices without IOMMU feature

From: Cornelia Huck <cohuck@redhat.com>
To: Halil Pasic <pasic@linux.ibm.com>
Cc: Pierre Morel <pmorel@linux.ibm.com>,
	linux-kernel@vger.kernel.org, borntraeger@de.ibm.com,
	frankja@linux.ibm.com, mst@redhat.com, jasowang@redhat.com,
	kvm@vger.kernel.org, linux-s390@vger.kernel.org,
	virtualization@lists.linux-foundation.org,
	thomas.lendacky@amd.com, david@gibson.dropbear.id.au,
	linuxram@us.ibm.com, heiko.carstens@de.ibm.com,
	gor@linux.ibm.com
Subject: Re: [PATCH v3 1/1] s390: virtio: let arch accept devices without IOMMU feature
Date: Fri, 19 Jun 2020 11:20:51 +0200	[thread overview]
Message-ID: <20200619112051.74babdb1.cohuck@redhat.com> (raw)
In-Reply-To: <20200618002956.5f179de4.pasic@linux.ibm.com>

On Thu, 18 Jun 2020 00:29:56 +0200
Halil Pasic <pasic@linux.ibm.com> wrote:

> On Wed, 17 Jun 2020 12:43:57 +0200
> Pierre Morel <pmorel@linux.ibm.com> wrote:
> 
> > An architecture protecting the guest memory against unauthorized host
> > access may want to enforce VIRTIO I/O device protection through the
> > use of VIRTIO_F_IOMMU_PLATFORM.
> > 
> > Let's give a chance to the architecture to accept or not devices
> > without VIRTIO_F_IOMMU_PLATFORM.
> >   
> [..]
> 
> 
> I'm still not really satisfied with your commit message, furthermore
> I did some thinking about the abstraction you introduce here. I will
> give a short analysis of that, but first things first. Your patch does
> the job of preventing calamity, and the details can be changed any time,
> thus: 
> 
> Acked-by: Halil Pasic <pasic@linux.ibm.com>
> 
> Regarding the interaction of architecture specific code with virtio core,
> I believe we could have made the interface more generic.
> 
> One option is to introduce virtio_arch_finalize_features(), a hook that
> could reject any feature that is inappropriate.

s/any feature/any combination of features/

This sounds like a good idea (for a later update).

> 
> Another option would be to find a common name for is_prot_virt_guest()
> (arch/s390) sev_active() (arch/x86) and is_secure_guest() (arch/powerpc)
> and use that instead of arch_needs_virtio_iommu_platform() and where-ever
> appropriate. Currently we seem to want this info in driver code only for
> virtio, but if the virtio driver has a legitimate need to know, other
> drivers may as well have a legitimate need to know. For example if we
> wanted to protect ourselves in ccw device drivers from somebody
> setting up a vfio-ccw device and attach it to the prot-virt guest (AFAICT
> we only lack guest enablement for this) such a function could be useful.

I'm not really sure if we can find enough commonality between
architectures, unless you propose to have a function for checking
things like device memory only.

> 
> But since this can be rewritten any time, let's go with the option
> people already agree with, instead of more discussion.

Yes, there's nothing wrong with the patch as-is.

Acked-by: Cornelia Huck <cohuck@redhat.com>

Which tree should this go through? Virtio? s390?

> 
> Just another question. Do we want this backported? Do we need cc stable?

It does change behaviour of virtio-ccw devices; but then, it only
fences off configurations that would not have worked anyway.
Distributions should probably pick this; but I do not consider it
strictly a "fix" (more a mitigation for broken configurations), so I'm
not sure whether stable applies.

> [..]
> 
> 
> >  int virtio_finalize_features(struct virtio_device *dev)
> >  {
> >  	int ret = dev->config->finalize_features(dev);
> > @@ -179,6 +194,13 @@ int virtio_finalize_features(struct virtio_device *dev)
> >  	if (!virtio_has_feature(dev, VIRTIO_F_VERSION_1))
> >  		return 0;
> >  
> > +	if (arch_needs_virtio_iommu_platform(dev) &&
> > +		!virtio_has_feature(dev, VIRTIO_F_IOMMU_PLATFORM)) {
> > +		dev_warn(&dev->dev,
> > +			 "virtio: device must provide VIRTIO_F_IOMMU_PLATFORM\n");  
> 
> I'm not sure, divulging the current Linux name of this feature bit is a
> good idea, but if everybody else is fine with this, I don't care that

Not sure if that feature name will ever change, as it is exported in
headers. At most, we might want to add the new ACCESS_PLATFORM define
and keep the old one, but that would still mean some churn.

> much. An alternative would be:
> "virtio: device falsely claims to have full access to the memory,
> aborting the device"

"virtio: device does not work with limited memory access" ?

But no issue with keeping the current message.