Re: [PATCH v2] virtio: refresh vring region cache after updating a virtqueue size

[Cornelia Huck <cohuck@redhat.com>]

On Fri, Mar 17 2023, Carlos López <clopez@suse.de> wrote:

> When a virtqueue size is changed by the guest via
> virtio_queue_set_num(), its region cache is not automatically updated.
> If the size was increased, this could lead to accessing the cache out
> of bounds. For example, in vring_get_used_event():
>
>     static inline uint16_t vring_get_used_event(VirtQueue *vq)
>     {
>         return vring_avail_ring(vq, vq->vring.num);
>     }
>
>     static inline uint16_t vring_avail_ring(VirtQueue *vq, int i)
>     {
>         VRingMemoryRegionCaches *caches = vring_get_region_caches(vq);
>         hwaddr pa = offsetof(VRingAvail, ring[i]);
>
>         if (!caches) {
>             return 0;
>         }
>
>         return virtio_lduw_phys_cached(vq->vdev, &caches->avail, pa);
>     }
>
> vq->vring.num will be greater than caches->avail.len, which will
> trigger a failed assertion down the call path of
> virtio_lduw_phys_cached().
>
> Fix this by calling virtio_init_region_cache() after
> virtio_queue_set_num() if we are not already calling
> virtio_queue_set_rings(). In the legacy path this is already done by
> virtio_queue_update_rings().
>
> Signed-off-by: Carlos López <clopez@suse.de>
> ---
> v2: use virtio_init_region_cache() instead of
> virtio_queue_update_rings() in the path for modern devices.
>
>  hw/s390x/virtio-ccw.c      | 1 +
>  hw/virtio/virtio-mmio.c    | 1 +
>  hw/virtio/virtio-pci.c     | 1 +
>  hw/virtio/virtio.c         | 2 +-
>  include/hw/virtio/virtio.h | 1 +
>  5 files changed, 5 insertions(+), 1 deletion(-)
>
> diff --git a/hw/s390x/virtio-ccw.c b/hw/s390x/virtio-ccw.c
> index e33e5207ab..f44de1a8c1 100644
> --- a/hw/s390x/virtio-ccw.c
> +++ b/hw/s390x/virtio-ccw.c
> @@ -237,6 +237,7 @@ static int virtio_ccw_set_vqs(SubchDev *sch, VqInfoBlock *info,
>                  return -EINVAL;
>              }
>              virtio_queue_set_num(vdev, index, num);
> +            virtio_init_region_cache(vdev, index);

Hmm... this is not wrong, but looking at it again, I see that the guest
has no way to change num after our last call to
virtio_init_region_cache() (while setting up the queue addresses.) IOW,
this introduces an extra round trip that is not really needed.

>          } else if (virtio_queue_get_num(vdev, index) > num) {
>              /* Fail if we don't have a big enough queue. */
>              return -EINVAL;
> diff --git a/hw/virtio/virtio-mmio.c b/hw/virtio/virtio-mmio.c
> index 23ba625eb6..c2c6d85475 100644
> --- a/hw/virtio/virtio-mmio.c
> +++ b/hw/virtio/virtio-mmio.c
> @@ -354,6 +354,7 @@ static void virtio_mmio_write(void *opaque, hwaddr offset, uint64_t value,
>          if (proxy->legacy) {
>              virtio_queue_update_rings(vdev, vdev->queue_sel);
>          } else {
> +            virtio_init_region_cache(vdev, vdev->queue_sel);
>              proxy->vqs[vdev->queue_sel].num = value;
>          }
>          break;
> diff --git a/hw/virtio/virtio-pci.c b/hw/virtio/virtio-pci.c
> index 247325c193..02fb84a8fa 100644
> --- a/hw/virtio/virtio-pci.c
> +++ b/hw/virtio/virtio-pci.c
> @@ -1554,6 +1554,7 @@ static void virtio_pci_common_write(void *opaque, hwaddr addr,
>          proxy->vqs[vdev->queue_sel].num = val;
>          virtio_queue_set_num(vdev, vdev->queue_sel,
>                               proxy->vqs[vdev->queue_sel].num);
> +        virtio_init_region_cache(vdev, vdev->queue_sel);
>          break;
>      case VIRTIO_PCI_COMMON_Q_MSIX:
>          vector = virtio_queue_vector(vdev, vdev->queue_sel);

OTOH, all other transports need this call, as setting the addresses and
setting num are two distinct operations. So I think we have two options:

- apply your patch, and accept that we have the extra round trip for ccw
  (which should not really be an issue, we're processing a channel
  command anyway)

- leave it out for ccw and add a comment why it isn't needed

(I think I'd prefer to just go ahead with your patch.)

Question (mostly for the other ccw folks): Do you think it is a problem
that ccw sets up queue addresses and size via one command, while pci and
mmio set addresses and size independently? I guess not; if anything, not
setting them in one go may lead to issues like the one this patch is
fixing.