Re: [PATCH] virtio: Force DMA restricted devices through DMA API

["Michael S. Tsirkin" <mst@redhat.com>]

On Tue, Jul 19, 2022 at 02:05:58PM +0000, Keir Fraser wrote:
> On Tue, Jul 19, 2022 at 07:56:09AM -0400, Michael S. Tsirkin wrote:
> > On Tue, Jul 19, 2022 at 10:02:56AM +0000, Keir Fraser wrote:
> > > If virtio devices are tagged for "restricted-dma-pool", then that
> > > pool should be used for virtio ring setup, via the DMA API.
> > > 
> > > In particular, this fixes virtio_balloon for ARM PKVM, where the usual
> > > workaround of setting VIRTIO_F_ACCESS_PLATFORM in the virtio device
> > > doesn't work because the virtio_balloon driver clears the flag. This
> > > seems a more robust fix than fiddling the flag again.
> > > 
> > > Signed-off-by: Keir Fraser <keirf@google.com>
> > 
> > 
> > So the reason balloon disables ACCESS_PLATFORM is simply
> > because it passes physical addresses to device and
> > expects device to be able to poke at them.
> > 
> > I worry about modifying DMA semantics yet again - it has as much of a
> > chance to break some legacy configs as it has to fix some.
> > 
> > 
> > And I don't really know much about restricted-dma-pool but
> > I'd like to understand why does it make sense to set it for
> > the balloon since it pokes at all and any system memory.
> 
> So this is set in the device tree by the host, telling it to bounce all DMA
> through a restricted memory window (basically swiotlb). The original reason
> is simply to isolate DMA, to the extent possible, on IOMMU-less systems.
> 
> However it is also useful for PKVM because the host is not trusted to access
> ordinary protected VM memory.

I'll have to read up on pKVM. Will get back to you.

> To allow I/O via the host, restricted-dma-pool
> is used to cause a bounce aperture to be allocated during VM boot, which is
> then explicitly shared with the host. For correct PKVM virtio operation, all
> data *and metadata* (virtio rings and descriptors) must be allocated in or
> bounced through this aperture.
>
> Insofar as virtio device accesses to virtio rings in guest memory essentially
> *are* DMA (from the pov of the guest), I think it makes sense to respect the
> bounce buffer for those rings, if so configured by the device tree.
>
> > > ---
> > >  drivers/virtio/virtio_ring.c | 8 ++++++++
> > >  1 file changed, 8 insertions(+)
> > > 
> > > diff --git a/drivers/virtio/virtio_ring.c b/drivers/virtio/virtio_ring.c
> > > index a5ec724c01d8..12be2607c648 100644
> > > --- a/drivers/virtio/virtio_ring.c
> > > +++ b/drivers/virtio/virtio_ring.c
> > > @@ -12,6 +12,7 @@
> > >  #include <linux/hrtimer.h>
> > >  #include <linux/dma-mapping.h>
> > >  #include <linux/spinlock.h>
> > > +#include <linux/swiotlb.h>
> > >  #include <xen/xen.h>
> > >  
> > >  #ifdef DEBUG
> > > @@ -248,6 +249,13 @@ static bool vring_use_dma_api(struct virtio_device *vdev)
> > >  	if (!virtio_has_dma_quirk(vdev))
> > >  		return true;
> > >  
> > > +	/* If the device is configured to use a DMA restricted pool,
> > > +	 * we had better use it.
> > > +	 */
> > > +	if (IS_ENABLED(CONFIG_DMA_RESTRICTED_POOL) &&
> > > +	    is_swiotlb_for_alloc(vdev->dev.parent))
> > > +		return true;
> > > +
> > >  	/* Otherwise, we are left to guess. */
> > >  	/*
> > >  	 * In theory, it's possible to have a buggy QEMU-supposed
> > > -- 
> > > 2.37.0.170.g444d1eabd0-goog
> > 
> >