On Wed, Mar 6, 2024 at 8:34 AM Michael S. Tsirkin <mst@redhat.com> wrote:
On Wed, Mar 06, 2024 at 08:07:31AM +0100, Eugenio Perez Martin wrote:
On Wed, Mar 6, 2024 at 6:34 AM Jason Wang <jasowang@redhat.com> wrote:
On Tue, Mar 5, 2024 at 3:46 AM Jonah Palmer <jonah.palmer@oracle.com> wrote:
The goal of these patches are to add support to a variety of virtio and
vhost devices for the VIRTIO_F_NOTIFICATION_DATA transport feature. This
feature indicates that a driver will pass extra data (instead of just a
virtqueue's index) when notifying the corresponding device.
The data passed in by the driver when this feature is enabled varies in
format depending on if the device is using a split or packed virtqueue
layout:
Split VQ
- Upper 16 bits: shadow_avail_idx
- Lower 16 bits: virtqueue index
Packed VQ
- Upper 16 bits: 1-bit wrap counter & 15-bit shadow_avail_idx
- Lower 16 bits: virtqueue index
Also, due to the limitations of ioeventfd not being able to carry the
extra provided by the driver, ioeventfd is left disabled for any devices
using this feature.
Is there any method to overcome this? This might help for vhost.
As a half-baked idea, read(2)ing an eventfd descriptor returns an
8-byte integer already. The returned value of read depends on eventfd
flags, but both have to do with the number of writes of the other end.
My proposal is to replace this value with the last value written by
the guest, so we can extract the virtio notification data from there.
The behavior of read is similar to not-EFD_SEMAPHORE, reading a value
and then blocking if read again without writes. The behavior of KVM
writes is different, as it is not a counter anymore.
Thanks!
I doubt you will be able to support this in ioeventfd...
I agree.
But vhost does not really need the value at all.
So why mask out ioeventfd with vhost?
The interface should not be able to start with vhost-kernel because
the feature is not offered by the vhost-kernel device. So ioeventfd is
always enabled with vhost-kernel.
Or do you mean we should allow it and let vhost-kernel fetch data from
the avail ring as usual? I'm ok with that but then the guest can place
any value to it, so the driver cannot be properly "validated by
software" that way.
vhost-vdpa is probably the only one that might need it...
Right, but vhost-vdpa already supports doorbell memory regions so I
guess it has little use, isn't it?
Thanks!
Thanks
A significant aspect of this effort has been to maintain compatibility
across different backends. As such, the feature is offered by backend
devices only when supported, with fallback mechanisms where backend
support is absent.
Jonah Palmer (8):
virtio/virtio-pci: Handle extra notification data
virtio-pci: Lock ioeventfd state with VIRTIO_F_NOTIFICATION_DATA
virtio-mmio: Handle extra notification data
virtio-mmio: Lock ioeventfd state with VIRTIO_F_NOTIFICATION_DATA
virtio-ccw: Handle extra notification data
virtio-ccw: Lock ioeventfd state with VIRTIO_F_NOTIFICATION_DATA
vhost/vhost-user: Add VIRTIO_F_NOTIFICATION_DATA to vhost feature bits
virtio: Add VIRTIO_F_NOTIFICATION_DATA property definition
hw/block/vhost-user-blk.c | 1 +
hw/net/vhost_net.c | 2 ++
hw/s390x/s390-virtio-ccw.c | 16 ++++++++++++----
hw/s390x/virtio-ccw.c | 6 ++++--
hw/scsi/vhost-scsi.c | 1 +
hw/scsi/vhost-user-scsi.c | 1 +
hw/virtio/vhost-user-fs.c | 2 +-
hw/virtio/vhost-user-vsock.c | 1 +
hw/virtio/virtio-mmio.c | 15 +++++++++++----
hw/virtio/virtio-pci.c | 16 +++++++++++-----
hw/virtio/virtio.c | 18 ++++++++++++++++++
include/hw/virtio/virtio.h | 5 ++++-
net/vhost-vdpa.c | 1 +
13 files changed, 68 insertions(+), 17 deletions(-)
--
2.39.3