Re: Virtual RDMA device

[Marcel Apfelbaum <marcel@xxxxxxxxxx>]

On 01/08/2017 0:52, Michael S. Tsirkin wrote:
> On Mon, Jul 31, 2017 at 10:10:15PM +0300, Marcel Apfelbaum wrote:

[...]

> > This seems to be the mail thread with most CCs.

Hi Michael,

> Please copy the upstream lists. virtio-dev, virtualization and
> qemu-devel for starters. I'd say linux-rdma as well.

Sure, adding all the mentioned lists.
To recap:

Yuval and I succeeded to implement a QEMU pvrdma device that works
(with some limitations) with VMware's Linux driver.
The git can be found at:
    https://github.com/yuvalshaia/qemu/tree/pvrdma.master.ibverbs
The QEMU RFC at:
    https://patchwork.kernel.org/patch/9653893/

While we do plan to merge it to QEMU, as a next step it seems
to be a consensus that virtio should be used to implement guest<->host
communication; since the project is not small, we need a plan.

We will do our best to attend the Micro RDMA conference:
  - 
https://www.linuxplumbersconf.org/2017/rdma-microconference-accepted-into-the-linux-plumbers-conference/

The attached presentation draft is intended to be a starting point
for a discussion on how to continue the project.

> > I attached a presentation
> > draft with *current* implementation of the pvrdma
> > device.(please excuse my presentation skills).
> >
> > It is supposed to be a starting point for discussing
> > a virtio-rdma device and for a presentation at
> > the RDMA Micro-conference.
> >
> > Please feel free to review/add and point to
> > what important things have I missed and how
> > we can continue.
>
> I'd suggest ignoring migration initially, just get
> rdma with memory overcommit working. For that:
> - come up with a timeline
> - decide who works on each of 3 things
>    1. extend libibverbs
>    2. layer virtio in qemu on top
>    3. write guest driver
>
>
> > I want to mention there is another discussion on this matter:
> >    https://lists.gnu.org/archive/html/qemu-devel/2017-07/msg06012.html
>
> That simply lists what is missing in libibverbs to be able to implement
> a PV RDMA device with memory overcommit and migration on top.
>
> > We should converge somehow.
> > Once Yuval returns from his vacation we will send a v2.
> >
> > Thanks,
> > Marcel
> >
> > >

[...]
Thanks,
Marcel

Attachment: pvrdma.odp
Description: application/vnd.oasis.opendocument.presentation