Re: [dpdk-dev] [PATCH 3/3] net/virtio: reject unsupported Rx multi queue modes

[Tiwei Bie <tiwei.bie@intel.com>]

On Thu, Oct 10, 2019 at 11:27:53AM +0300, Andrew Rybchenko wrote:
> On 10/10/19 11:23 AM, David Marchand wrote:
> > On Thu, Oct 10, 2019 at 10:14 AM Andrew Rybchenko
> > <arybchenko@solarflare.com> wrote:
> > > On 10/10/19 10:42 AM, Thomas Monjalon wrote:
> > > > 09/10/2019 13:24, Andrew Rybchenko:
> > > > > On 10/9/19 1:41 PM, Tiwei Bie wrote:
> > > > > > My understanding is that, setting mq_mode to ETH_MQ_RX_NONE means
> > > > > > no method is enforced on how to route packets to MQs.
> > > > > I'm not sure. It is definitely a place to be improved in
> > > > > ethdev documentation. Thomas, Ferruh, what do you think?
> > > > > Is it really a definition of ETH_MQ_RX_NONE?
> > > > I think it means everything go to queue 0.
> > > I understand it this way as well.
> > > 
> > > > The comment says no DCB, RSS or VMDQ.
> > > > It looks like the "NONE" value has been abused for some custom steering.
> > > > We have two options:
> > > >        - document NONE as a possible case of custom steering
> > > >        - add a new CUSTOM value
> > > I'd prefer to say that ETH_MQ_RX_RSS with rss_hf equal to 0 means
> > > unspecified/unknown steering. If application just want to spread
> > > traffic across many Rx queues, it is natural choice to say that
> > > it want RSS, but do not care about spreading algorithm etc.
> > > It allows driver use recommended defaults if rss_hf is controllable,
> > > or just spread in virtio case.
> > RSS is about maintaining affinity of a "flow" (as in packets sharing
> > the same l3/l4 tuples) to a specific queue.
> > Here, we can have packets from a same flow on any queue depending on
> > what happened on the vhost side.
> 
> Interesting. I'd like to know a bit more about it. I didn't know that
> it is so unstable. Could someone who knows the topic well add
> a bit more information about it.

The spec defined automatic steering [1], but it's not very strict
for the device. So it depends on the implementation.

[1] https://github.com/oasis-tcs/virtio-spec/commit/f7020384521e5938cdd1351270b32b143f070d00