Re: [PATCH v2 1/1] s390: virtio: let arch accept devices without IOMMU feature

[Pierre Morel]

On 2020-06-16 11:52, Halil Pasic wrote:
> On Mon, 15 Jun 2020 14:39:24 +0200
> Pierre Morel <pmorel@xxxxxxxxxxxxx> wrote:
>
> I find the subject (commit short) sub optimal. The 'arch' is already
> accepting devices 'without IOMMU feature'. What you are introducing is
> the ability to reject.
>
> > 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 don't particularly like the commit message. In general, I believe
> using access_platform instead of iommu_platform would really benefit us.

IOMMU_PLATFORM is used overall in Linux, and I did not find any 
occurrence for ACCESS_PLATFORM.


> > Signed-off-by: Pierre Morel <pmorel@xxxxxxxxxxxxx>
> > ---
> >   arch/s390/mm/init.c     | 6 ++++++
> >   drivers/virtio/virtio.c | 9 +++++++++
> >   include/linux/virtio.h  | 2 ++
> >   3 files changed, 17 insertions(+)
> >
> > diff --git a/arch/s390/mm/init.c b/arch/s390/mm/init.c
> > index 87b2d024e75a..3f04ad09650f 100644
> > --- a/arch/s390/mm/init.c
> > +++ b/arch/s390/mm/init.c
> > @@ -46,6 +46,7 @@
> >   #include <asm/kasan.h>
> >   #include <asm/dma-mapping.h>
> >   #include <asm/uv.h>
> > +#include <linux/virtio.h>
>
> arch/s390/mm/init.c including virtio.h looks a bit strange to me, but
> if Heiko and Vasily don't mind, neither do I.

Do we have a better place to install the hook?
I though that since it is related to memory management and that, since 
force_dma_unencrypted already is there, it would be a good place.

However, kvm-s390 is another candidate.

> >   
> >   pgd_t swapper_pg_dir[PTRS_PER_PGD] __section(.bss..swapper_pg_dir);
> >   
> > @@ -162,6 +163,11 @@ bool force_dma_unencrypted(struct device *dev)
> >   	return is_prot_virt_guest();
> >   }
> >   
> > +int arch_needs_iommu_platform(struct virtio_device *dev)
>
> Maybe prefixing the name with virtio_ would help provide the
> proper context.

The virtio_dev makes it obvious and from the virtio side it should be 
obvious that the arch is responsible for this.

However if nobody has something against I change it.

> > +{
> > +	return is_prot_virt_guest();
> > +}
> > +
> >   /* protected virtualization */
> >   static void pv_init(void)
> >   {
> > diff --git a/drivers/virtio/virtio.c b/drivers/virtio/virtio.c
> > index a977e32a88f2..30091089bee8 100644
> > --- a/drivers/virtio/virtio.c
> > +++ b/drivers/virtio/virtio.c
> > @@ -167,6 +167,11 @@ void virtio_add_status(struct virtio_device *dev, unsigned int status)
> >   }
> >   EXPORT_SYMBOL_GPL(virtio_add_status);
> >   
> > +int __weak arch_needs_iommu_platform(struct virtio_device *dev)
> > +{
> > +	return 0;
> > +}
> > +
>
> Adding some people that could be interested in overriding this as well
> to the cc list.

Thanks,

> >   int virtio_finalize_features(struct virtio_device *dev)
> >   {
> >   	int ret = dev->config->finalize_features(dev);
> > @@ -179,6 +184,10 @@ int virtio_finalize_features(struct virtio_device *dev)
> >   	if (!virtio_has_feature(dev, VIRTIO_F_VERSION_1))
> >   		return 0;
> >   
> > +	if (arch_needs_iommu_platform(dev) &&
> > +		!virtio_has_feature(dev, VIRTIO_F_IOMMU_PLATFORM))
> > +		return -EIO;
> > +
>
> Why EIO?

Because I/O can not occur correctly?
I am open to suggestions.

> Overall, I think it is a good idea to have something that is going to
> protect us from this scenario.

It would clearly be a good thing that trusted hypervizors like QEMU 
forbid this scenario however should we let the door open?

Thanks,
Pierre

--
Pierre Morel
IBM Lab Boeblingen