On Tue, Nov 24, 2015 at 06:00:01PM +0900, Tetsuya Mukawa wrote:
> The vhost PMD will be a wrapper of vhost library, but some of vhost
> library APIs cannot be mapped to ethdev library APIs.
> Becasue of this, in some cases, we still need to use vhost library APIs
> for a port created by the vhost PMD.
>
> Currently, when virtio device is created and destroyed, vhost library
> will call one of callback handlers. The vhost PMD need to use this
> pair of callback handlers to know which virtio devices are connected
> actually.
> Because we can register only one pair of callbacks to vhost library, if
> the PMD use it, DPDK applications cannot have a way to know the events.
>
> This may break legacy DPDK applications that uses vhost library. To prevent
> it, this patch adds one more pair of callbacks to vhost library especially
> for the vhost PMD.
> With the patch, legacy applications can use the vhost PMD even if they need
> additional specific handling for virtio device creation and destruction.
>
> For example, legacy application can call
> rte_vhost_enable_guest_notification() in callbacks to change setting.
TBH, I never liked it since the beginning. Introducing two callbacks
for one event is a bit messy, and therefore error prone.
I have been thinking this occasionally last few weeks, and have came
up something that we may introduce another layer callback based on
the vhost pmd itself, by a new API:
rte_eth_vhost_register_callback().
And we then call those new callback inside the vhost pmd new_device()
and vhost pmd destroy_device() implementations.
And we could have same callbacks like vhost have, but I'm thinking
that new_device() and destroy_device() doesn't sound like a good name
to a PMD driver. Maybe a name like "link_state_changed" is better?
What do you think of that?
On the other hand, I'm still thinking is that really necessary to let
the application be able to call vhost functions like
rte_vhost_enable_guest_notification()
with the vhost PMD driver?
--yliu
