Re: [PATCH v2] virtio_net: CTRL_GUEST_OFFLOADS depends on CTRL_VQ

From: Jason Wang
Date: Mon Jan 06 2020 - 21:29:23 EST

On 2020/1/6 äå8:54, Michael S. Tsirkin wrote:
On Mon, Jan 06, 2020 at 10:47:35AM +0800, Jason Wang wrote:
On 2020/1/5 äå9:22, Michael S. Tsirkin wrote:
The only way for guest to control offloads (as enabled by
VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) is by sending commands
through CTRL_VQ. So it does not make sense to

The spec does not outlaw devices with such a configuration, so we have
to support it. Simply clear VIRTIO_NET_F_CTRL_GUEST_OFFLOADS.
Note that Linux is still crashing if it tries to
change the offloads when there's no control vq.
That needs to be fixed by another patch.

Reported-by: Alistair Delva <adelva@xxxxxxxxxx>
Reported-by: Willem de Bruijn <willemdebruijn.kernel@xxxxxxxxx>
Fixes: 3f93522ffab2 ("virtio-net: switch off offloads on demand if possible on XDP set")
Signed-off-by: Michael S. Tsirkin <mst@xxxxxxxxxx>

Same patch as v1 but update documentation so it's clear it's not
enough to fix the crash.

drivers/net/virtio_net.c | 9 +++++++++
1 file changed, 9 insertions(+)

diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 4d7d5434cc5d..7b8805b47f0d 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -2971,6 +2971,15 @@ static int virtnet_validate(struct virtio_device *vdev)
if (!virtnet_validate_features(vdev))
return -EINVAL;
+ /* VIRTIO_NET_F_CTRL_GUEST_OFFLOADS does not work without
+ * VIRTIO_NET_F_CTRL_VQ. Unfortunately spec forgot to
+ * specify that VIRTIO_NET_F_CTRL_GUEST_OFFLOADS depends
+ * on VIRTIO_NET_F_CTRL_VQ so devices can set the later but
+ * not the former.
+ */
+ if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
+ __virtio_clear_bit(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS);

If it's just because a bug of spec, should we simply fix the bug and fail
the negotiation in virtnet_validate_feature()?
One man's bug is another man's feature: arguably leaving the features
independent in the spec might allow reuse of the feature bit without
breaking guests.

And even if we say it's a bug we can't simply fix the bug in the
spec: changing the text for a future version does not change the fact
that devices behaving according to the spec exist.

Otherwise there would be inconsistency in handling feature dependencies for
ctrl vq.

That's a cosmetic problem ATM. It might be a good idea to generally
change our handling of dependencies, and clear feature bits instead of
failing probe on a mismatch.

Something like I proposed in the past ? [1]


It's worth thinking - at the spec level -
how we can best make the configuration extensible.
But that's not something spec should worry about.

if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
int mtu = virtio_cread16(vdev,
offsetof(struct virtio_net_config,