Re: [Qemu-devel] [PATCH v2 4/5] vhost: move virtio 1.0 check to cross-endian helper

[Greg Kurz]

On Tue, 12 Jan 2016 01:23:56 +0100
Laurent Vivier <address@hidden> wrote:

> On 11/01/2016 17:13, Greg Kurz wrote:
> > Indeed vhost doesn't need to ask for vring endian fixing if the device is
> > virtio 1.0, since it is already handled by the in-kernel vhost driver. This
> > patch simply consolidates the logic into the existing helper.
> > 
> > Signed-off-by: Greg Kurz <address@hidden>
> > Reviewed-by: Cornelia Huck <address@hidden>
> > ---
> >  hw/virtio/vhost.c |    9 +++++----
> >  1 file changed, 5 insertions(+), 4 deletions(-)
> > 
> > diff --git a/hw/virtio/vhost.c b/hw/virtio/vhost.c
> > index 2e1e792d599e..aef750df22ad 100644
> > --- a/hw/virtio/vhost.c
> > +++ b/hw/virtio/vhost.c
> > @@ -750,6 +750,9 @@ static void vhost_log_stop(MemoryListener *listener,
> >  
> >  static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
> >  {
> > +    if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
> > +        return false;
> > +    }
> >  #ifdef TARGET_IS_BIENDIAN
> >  #ifdef HOST_WORDS_BIGENDIAN
> >      return !virtio_is_big_endian(vdev);  
> 
> I'm wondering if in term of performance you can remove the call of
> virtio_vdev_has_feature() as it is done in virtio_is_big_endian().
> 

We cannot do this: in the case of a big endian host and a virtio 1.0 device,
virtio_is_big_endian() would return false and vhost_needs_vring_endian()
would return true. This is wrong: the byteswapping for virtio 1.0 is already
handled in the linux vhost driver since it has the feature bits. Only legacy
devices may need to call vhost_virtqueue_set_vring_endian_legacy() in a
cross-endian situation.

This isn't really a performance issue since this function is never called on
a hot path.

But indeed, your comment raises a concern about clarity. After the the call
to virtio_vdev_has_feature(), we are only caring for legacy devices: we
don't need the extra check in virtio_is_big_endian().

Also the device_endian field is always set (VIRTIO_DEVICE_ENDIAN_UNKNOWN may
only happen on a virtio_load() path that cannot lead here). So we don't need
the assert() either.

We simply need to open code the device_endian checking instead of calling
virtio_is_big_endian(). Like this:

/* Only legacy devices offered to an endian ambivalent guest may
 * face a cross-endian situation. All other cases return false.
 */
static inline bool vhost_needs_vring_endian(VirtIODevice *vdev)
{
    if (virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1)) {
        return false;
    }
#ifdef TARGET_IS_BIENDIAN
#ifdef HOST_WORDS_BIGENDIAN
    return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_LITTLE;
#else
    return vdev->device_endian == VIRTIO_DEVICE_ENDIAN_BIG;
#endif
#else
    return false;
#endif
}

I'll make it a separate patch.

Thanks !

--
Greg

> > @@ -811,8 +814,7 @@ static int vhost_virtqueue_start(struct vhost_dev *dev,
> >          return -errno;
> >      }
> >  
> > -    if (!virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1) &&
> > -        vhost_needs_vring_endian(vdev)) {
> > +    if (vhost_needs_vring_endian(vdev)) {
> >          r = vhost_virtqueue_set_vring_endian_legacy(dev,
> >                                                      
> > virtio_is_big_endian(vdev),
> >                                                      vhost_vq_index);
> > @@ -908,8 +910,7 @@ static void vhost_virtqueue_stop(struct vhost_dev *dev,
> >      /* In the cross-endian case, we need to reset the vring endianness to
> >       * native as legacy devices expect so by default.
> >       */
> > -    if (!virtio_vdev_has_feature(vdev, VIRTIO_F_VERSION_1) &&
> > -        vhost_needs_vring_endian(vdev)) {
> > +    if (vhost_needs_vring_endian(vdev)) {
> >          r = vhost_virtqueue_set_vring_endian_legacy(dev,
> >                                                      
> > !virtio_is_big_endian(vdev),
> >                                                      vhost_vq_index);
> > 
> >   
>