Re: [dpdk-dev] [RFC PATCH v2] vhost: Add VHOST PMD

[Bruce Richardson <bruce.richardson@intel.com>]

On Mon, Oct 19, 2015 at 10:32:50AM +0100, Loftus, Ciara wrote:
> > On 2015/10/16 21:52, Bruce Richardson wrote:
> > > On Mon, Aug 31, 2015 at 12:55:26PM +0900, Tetsuya Mukawa wrote:
> > >> The patch introduces a new PMD. This PMD is implemented as thin
> > wrapper
> > >> of librte_vhost. It means librte_vhost is also needed to compile the PMD.
> > >> The PMD can have 'iface' parameter like below to specify a path to
> > connect
> > >> to a virtio-net device.
> > >>
> > >> $ ./testpmd -c f -n 4 --vdev 'eth_vhost0,iface=/tmp/sock0' -- -i
> > >>
> > >> To connect above testpmd, here is qemu command example.
> > >>
> > >> $ qemu-system-x86_64 \
> > >>         <snip>
> > >>         -chardev socket,id=chr0,path=/tmp/sock0 \
> > >>         -netdev vhost-user,id=net0,chardev=chr0,vhostforce \
> > >>         -device virtio-net-pci,netdev=net0
> > >>
> > >> Signed-off-by: Tetsuya Mukawa <mukawa@igel.co.jp>
> > > With this PMD in place, is there any need to keep the existing vhost library
> > > around as a separate entity? Can the existing library be
> > subsumed/converted into
> > > a standard PMD?
> > >
> > > /Bruce
> > 
> > Hi Bruce,
> > 
> > I concern about whether the PMD has all features of librte_vhost,
> > because librte_vhost provides more features and freedom than ethdev API
> > provides.
> > In some cases, user needs to choose limited implementation without
> > librte_vhost.
> > I am going to eliminate such cases while implementing the PMD.
> > But I don't have strong belief that we can remove librte_vhost now.
> > 
> > So how about keeping current separation in next DPDK?
> > I guess people will try to replace librte_vhost to vhost PMD, because
> > apparently using ethdev APIs will be useful in many cases.
> > And we will get feedbacks like "vhost PMD needs to support like this usage".
> > (Or we will not have feedbacks, but it's also OK.)
> > Then, we will be able to merge librte_vhost and vhost PMD.
> 
> I agree with the above. One the concerns I had when reviewing the patch was that the PMD removes some freedom that is available with the library. Eg. Ability to implement the new_device and destroy_device callbacks. If using the PMD you are constrained to the implementations of these in the PMD driver, but if using librte_vhost, you can implement your own with whatever functionality you like - a good example of this can be seen in the vhost sample app.
> On the other hand, the PMD is useful in that it removes a lot of complexity for the user and may work for some more general use cases. So I would be in favour of having both options available too.
> 
> Ciara
>

Thanks.
However, just because the libraries are merged does not mean that you need
be limited by PMD functionality. Many PMDs provide additional library-specific
functions over and above their PMD capabilities. The bonded PMD is a good example
here, as it has a whole set of extra functions to create and manipulate bonded
devices - things that are obviously not part of the general ethdev API. Other
vPMDs similarly include functions to allow them to be created on the fly too.

regards,
/Bruce