On Wed, 18 Nov 2026 at 09:28, Isaac Boukris <[email protected]> wrote:
>  drivers/net/virtio/virtio_ethdev.c | 18 ++++++++++++++----
>  1 file changed, 14 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/net/virtio/virtio_ethdev.c 
> b/drivers/net/virtio/virtio_ethdev.c
> index f2302089ee..8600f3899a 100644
> --- a/drivers/net/virtio/virtio_ethdev.c
> +++ b/drivers/net/virtio/virtio_ethdev.c
> @@ -1455,10 +1455,20 @@ virtio_configure_intr(struct rte_eth_dev *dev)
>                 return -ENOTSUP;
>         }
>
> -       ret = rte_intr_efd_enable(dev->intr_handle, dev->data->nb_rx_queues);
> -       if (ret < 0) {
> -               PMD_INIT_LOG(ERR, "Fail to create eventfd");
> -               return ret;
> +       /* On reinitialization, release the eventfds if the number of Rx 
> queues changed. */
> +       if (rte_intr_type_get(dev->intr_handle) == RTE_INTR_HANDLE_VFIO_MSIX 
> &&

This still does not feel right to me.


> +           rte_intr_dp_is_en(dev->intr_handle) &&
> +           rte_intr_nb_efd_get(dev->intr_handle) != dev->data->nb_rx_queues) 
> {
> +               virtio_intr_disable(dev);
> +               rte_intr_efd_disable(dev->intr_handle);

So what happens here iiuc.

For vdev, event FDs are filled by the driver (contrary to MSIX /
rte_intr_efd_enable) in the virtio-user probe function for example.
But we have an asymmetry in EAL, where rte_intr_efd_enable is not
changing the handle for VDEV, but calling rte_intr_efd_disable breaks
it resetting nb_efd to 0.

Not resetting to 0 in rte_intr_efd_disable for vdev would break
rte_intr_dp_is_en.
And it probably has subtle implications on other part of this API...

We should not be building a workaround in drivers, but correctly
fixing this now in net/virtio is risky:
- the virtio driver should probably not be calling rte_intr_efd_enable
here in code common to PCI and VDEV ethdevs.
  VDEV sets up the interrupt handle once and for all during probing,
and PCI should be doing the same.
- or.. calls to rte_intr_efd_disable are missing in the
virtio_dev_init common code.


Taking a step back now.
The initial eventfd leak fix in EAL changed assumptions for drivers,
with the expectation now being that the drivers properly
enable/disable.
I saw one driver (net/iavf) change on that topic, which is really
little when compared to the number of users in-tree.
We likely have bugs in other drivers that were not caught by validation.

Given that we are in rc3 (rc4 by eow), I am for deferring the virtio
changes to the next release and instead reverting the EAL fix in this
release until we get a proper pass on each maintainer before
reapplying.
I'll send a patch in that direction.


-- 
David Marchand

Reply via email to