[ovs-dev] [PATCH 2/2] netdev-dpdk: add support for vhost IOMMU feature

Maxime Coquelin maxime.coquelin at redhat.com
Tue Nov 28 13:03:30 UTC 2017


Hi Sean,

On 11/27/2017 03:25 PM, Mooney, Sean K wrote:
>> Sean can confirm, but my reading is that Sean's primary concern is with
>> the proposal to add any config option to enable iommu - not with the
>> proposal to change from a per port to per OVS basis.
> [Mooney, Sean K] so I had a quick chat with mark about the background to this.
> My current understanding is that form qemu 2.6-2.9 there is a bug that causes the
> Guest to hang due to a lack of support for a required ack to correctly negotiate the
> Use of the iommu feature. As a result a config option either global or per interfaces is
> Required. As an aside OpenStack does not currently support requesting iommu virtualization so
> I don’t see why we could not put a requirement on our side that you use qemu 2.9.1 or newer
> To avail of that feature. Going back to ovs then assuming we used a single global config vaule
> In for example the other_config dictionary of the Open_vSwitch table in the ovsdb.
> e.g. ovs-vsctl set Open_vswitch . other_config:vhost-iommu-support=true
> I would be fine with that approach if we can confirm 1 thing first.

I confirm you understanding:
1. Before Qemu v2.9.1, we must not enable iommu-support option because
because it enables a vhost-user protocol feature (reply-ack) that is
suffers from a bug in these Qemu version. And this is not a problem
because iommu isn't anyway supported in these Qemu versions.
2. Since Qemu v2.9.1, the reply-ack bug has been fixed, so we can enable
iommu-support option safely, using iommu or not.

Regards,
Maxime


More information about the dev mailing list