Hi Michael, VIRTIO_F_NOTIFICATION_DATA needs to be exposed to make Marvell's device works with Qemu. Any other better ways to expose VIRTIO_F_NOTIFICATION_DATA feature bit for vhost vdpa use case?
Thanks, Srujana. > Subject: RE: [EXT] Re: [PATCH] vhost_net: add NOTIFICATION_DATA and > IN_ORDER feature bits to vdpa_feature_bits > > Ping. > > > Subject: RE: [EXT] Re: [PATCH] vhost_net: add NOTIFICATION_DATA and > > IN_ORDER feature bits to vdpa_feature_bits > > > > > > > > > -----Original Message----- > > > From: Michael S. Tsirkin <[email protected]> > > > Sent: Monday, February 19, 2024 3:15 PM > > > To: Srujana Challa <[email protected]> > > > Cc: [email protected]; Vamsi Krishna Attunuru > > > <[email protected]>; Jerin Jacob <[email protected]>; Jason > > > Wang <[email protected]> > > > Subject: [EXT] Re: [PATCH] vhost_net: add NOTIFICATION_DATA and > > > IN_ORDER feature bits to vdpa_feature_bits > > > > > > External Email > > > > > > -------------------------------------------------------------------- > > > -- On Tue, Jan 02, 2024 at 04:44:32PM +0530, Srujana Challa wrote: > > > > Enables VIRTIO_F_NOTIFICATION_DATA and VIRTIO_F_IN_ORDER > feature > > > bits > > > > for vhost vdpa backend. Also adds code to consider all feature > > > > bits supported by vhost net client type for feature negotiation, > > > > so that vhost backend device supported features can be negotiated with > guest. > > > > > > > > Signed-off-by: Srujana Challa <[email protected]> > > > > --- > > > > hw/net/vhost_net.c | 10 ++++++++++ > > > > net/vhost-vdpa.c | 2 ++ > > > > 2 files changed, 12 insertions(+) > > > > > > > > diff --git a/hw/net/vhost_net.c b/hw/net/vhost_net.c index > > > > e8e1661646..65ae8bcece 100644 > > > > --- a/hw/net/vhost_net.c > > > > +++ b/hw/net/vhost_net.c > > > > @@ -117,6 +117,16 @@ static const int > > > > *vhost_net_get_feature_bits(struct vhost_net *net) > > > > > > > > uint64_t vhost_net_get_features(struct vhost_net *net, uint64_t > > > > features) { > > > > + const int *bit = vhost_net_get_feature_bits(net); > > > > + > > > > + /* > > > > + * Consider all feature bits for feature negotiation with vhost > > > > backend, > > > > + * so that all backend device supported features can be negotiated. > > > > + */ > > > > + while (*bit != VHOST_INVALID_FEATURE_BIT) { > > > > + features |= (1ULL << *bit); > > > > + bit++; > > > > + } > > > > return vhost_get_features(&net->dev, > vhost_net_get_feature_bits(net), > > > > features); > > > > } > > > > > > I don't think we should do this part. With vdpa QEMU is in control > > > of which features are exposed and that is intentional since features > > > are often tied to other behaviour. > > > > Vdpa Qemu can negotiate all the features which vdpa backend device > > supports with the guest right? > > Guest drivers (it could be userspace or kernel drivers) will negotiate > > their own features, so that frontend supported features will get the > > precedence. > > > > > > > > > diff --git a/net/vhost-vdpa.c b/net/vhost-vdpa.c index > > > > 3726ee5d67..51334fcfe2 100644 > > > > --- a/net/vhost-vdpa.c > > > > +++ b/net/vhost-vdpa.c > > > > @@ -57,7 +57,9 @@ typedef struct VhostVDPAState { > > > > */ > > > > const int vdpa_feature_bits[] = { > > > > VIRTIO_F_ANY_LAYOUT, > > > > + VIRTIO_F_IN_ORDER, > > > > VIRTIO_F_IOMMU_PLATFORM, > > > > + VIRTIO_F_NOTIFICATION_DATA, > > > > VIRTIO_F_NOTIFY_ON_EMPTY, > > > > VIRTIO_F_RING_PACKED, > > > > VIRTIO_F_RING_RESET, > > > > -- > > > > 2.25.1
