Linux Virtualization - Jan 2018 - [virtio-dev] Re: [RFC PATCH net-next v2 2/2] virtio_net: Extend virtio to use VF datapath when available

On Sun, Jan 28, 2018 at 3:02 PM, Stephen Hemminger
<stephen at networkplumber.org> wrote:
> On Fri, 26 Jan 2018 18:30:03 -0800
> Jakub Kicinski <kubakici at wp.pl> wrote:
>
>> On Fri, 26 Jan 2018 15:30:35 -0800, Samudrala, Sridhar wrote:
>> > On 1/26/2018 2:47 PM, Jakub Kicinski wrote:
>> > > On Sat, 27 Jan 2018 00:14:20 +0200, Michael S. Tsirkin wrote:
>> > >> On Fri, Jan 26, 2018 at 01:46:42PM -0800, Siwei Liu
wrote:
>> > >>>> and the VM is not expected to do any
tuning/optimizations on the VF driver
>> > >>>> directly,
>> > >>>> i think the current patch that follows the netvsc
model of 2 netdevs(virtio
>> > >>>> and vf) should
>> > >>>> work fine.
>> > >>> OK. For your use case that's fine. But that's
too specific scenario
>> > >>> with lots of restrictions IMHO, perhaps very few
users will benefit
>> > >>> from it, I'm not sure. If you're unwilling to
move towards it, we'd
>> > >>> take this one and come back with a generic solution
that is able to
>> > >>> address general use cases for VF/PT live migration .
>> > >> I think that's a fine approach. Scratch your own
itch!  I imagine a very
>> > >> generic virtio-switchdev providing host routing info to
guests could
>> > >> address lots of usecases. A driver could bind to that one
and enslave
>> > >> arbitrary other devices.  Sounds reasonable.
>> > >>
>> > >> But given the fundamental idea of a failover was floated
at least as
>> > >> early as 2013, and made 0 progress since precisely
because it kept
>> > >> trying to address more and more features, and given
netvsc is already
>> > >> using the basic solution with some success, I'm not
inclined to block
>> > >> this specific effort waiting for the generic one.
>> > > I think there is an agreement that the extra netdev will be
useful for
>> > > more advanced use cases, and is generally preferable.  What
is the
>> > > argument for not doing that from the start?  If it was made I
must have
>> > > missed it.  Is it just unwillingness to write the extra 300
lines of
>> > > code?  Sounds like a pretty weak argument when adding kernel
ABI is at
>> > > stake...
>> >
>> > I am still not clear on the need for the extra netdev created by
>> > virtio_net. The only advantage i can see is that the stats can be
>> > broken between VF and virtio datapaths compared to the aggregrated
>> > stats on virtio netdev as seen with the 2 netdev approach.
>>
>> Maybe you're not convinced but multiple arguments were made.
>>
>> > With 2 netdev model, any VM image that has a working network
>> > configuration will transparently get VF based acceleration without
>> > any changes.
>>
>> Nothing happens transparently.  Things may happen automatically.  The
>> VF netdev doesn't disappear with netvsc.  The PV netdev transforms
into
>> something it did not use to be.  And configures and reports some
>> information from the PV (e.g. speed) but PV doesn't pass traffic
any
>> longer.
>>
>> > 3 netdev model breaks this configuration starting with the
creation
>> > and naming of the 2 devices to udev needing to be aware of master
and
>> > slave virtio-net devices.
>>
>> I don't understand this comment.  There is one virtio-net device
and
>> one "virtio-bond" netdev.  And user space has to be aware of
the special
>> automatic arrangement anyway, because it can't touch the VF.  It
>> doesn't make any difference whether it ignores the VF or PV and VF.
>> It simply can't touch the slaves, no matter how many there are.
>>
>> > Also, from a user experience point of view, loading a virtio-net
with
>> > BACKUP feature enabled will now show 2 virtio-net netdevs.
>>
>> One virtio-net and one virtio-bond, which represents what's
happening.
>>
>> > For live migration with advanced usecases that Siwei is
suggesting, i
>> > think we need a new driver with a new device type that can track
the
>> > VF specific feature settings even when the VF driver is unloaded.
>
> I see no added value of the 3 netdev model, there is no need for a bond
> device.
I agree a full-blown bond isn't what is needed. However, just forking
traffic out from virtio to a VF doesn't really solve things either.

One of the issues as I see it is the fact that the qdisc model with
the merged device gets to be pretty ugly. There is the fact that every
packet that goes to the VF has to pass through the qdisc code twice.
There is the dual nature of the 2 netdev solution that also introduces
the potential for head-of-line blocking since the virtio could put
back pressure on the upper qdisc layer which could stall the VF
traffic when switching over. I hope we could avoid issues like that by
maintaining qdiscs per device queue instead of on an upper device that
is half software interface and half not. Ideally the virtio-bond
device could operate without a qdisc and without needing any
additional locks so there shouldn't be head of line blocking occurring
between the two interfaces and overhead could be kept minimal.

Also in the case of virtio there is support for in-driver XDP. As
Sridhar stated, when using the 2 netdev model "we cannot support XDP
in this model and it needs to be disabled". That sounds like a step
backwards instead of forwards. I would much rather leave the XDP
enabled at the lower dev level, and then if we want we can use the
generic XDP at the virtio-bond level to capture traffic on both
interfaces at the same time.

In the case of netvsc you have control of both sides of a given link
so you can match up the RSS tables, queue configuration and everything
is somewhat symmetric since you are running the PF and all the HyperV
subchannels. Most of the complexity is pushed down into the host and
your subchannel management is synchronized there if I am not mistaken.
We don't have this in the case of this virtio-bond setup. Instead a
single bit is set indicating "backup" without indicating what that
means to topology other than the fact that this virtio interface is
the backup for some other interface. We are essentially blind other
than having the link status for the VF and virtio and knowing that the
virtio is intended to be the backup.

We might be able to get to a 2 or maybe even a 1 netdev solution at
some point in the future, but  I don't think that time is now. For now
a virtio-bond type solution would allow us to address most of the use
cases with minimal modification to the existing virtio and can deal
with feature and/or resource asymmetry.

- Alex

On Sun, Jan 28, 2018 at 08:26:53PM -0800, Alexander Duyck
wrote:
> >> > For live migration with advanced usecases that Siwei is
suggesting, i
> >> > think we need a new driver with a new device type that can
track the
> >> > VF specific feature settings even when the VF driver is
unloaded.
> >
> > I see no added value of the 3 netdev model, there is no need for a
bond
> > device.
> 
> I agree a full-blown bond isn't what is needed. However, just forking
> traffic out from virtio to a VF doesn't really solve things either.
> 
> One of the issues as I see it is the fact that the qdisc model with
> the merged device gets to be pretty ugly. There is the fact that every
> packet that goes to the VF has to pass through the qdisc code twice.
> There is the dual nature of the 2 netdev solution that also introduces
> the potential for head-of-line blocking since the virtio could put
> back pressure on the upper qdisc layer which could stall the VF
> traffic when switching over. I hope we could avoid issues like that by
> maintaining qdiscs per device queue instead of on an upper device that
> is half software interface and half not. Ideally the virtio-bond
> device could operate without a qdisc and without needing any
> additional locks so there shouldn't be head of line blocking occurring
> between the two interfaces and overhead could be kept minimal.
> 
> Also in the case of virtio there is support for in-driver XDP. As
> Sridhar stated, when using the 2 netdev model "we cannot support XDP
> in this model and it needs to be disabled". That sounds like a step
> backwards instead of forwards. I would much rather leave the XDP
> enabled at the lower dev level, and then if we want we can use the
> generic XDP at the virtio-bond level to capture traffic on both
> interfaces at the same time.
I agree dropping XDP makes everything very iffy.
> In the case of netvsc you have control of both sides of a given link
> so you can match up the RSS tables, queue configuration and everything
> is somewhat symmetric since you are running the PF and all the HyperV
> subchannels. Most of the complexity is pushed down into the host and
> your subchannel management is synchronized there if I am not mistaken.
> We don't have this in the case of this virtio-bond setup. Instead a
> single bit is set indicating "backup" without indicating what
that
> means to topology other than the fact that this virtio interface is
> the backup for some other interface. We are essentially blind other
> than having the link status for the VF and virtio and knowing that the
> virtio is intended to be the backup.
Would you be interested in posting at least a proof of concept
patch for this approach? It's OK if there are some TODO stubs.
It would be much easier to compare code to code rather than
a high level description to code.
> We might be able to get to a 2 or maybe even a 1 netdev solution at
> some point in the future, but  I don't think that time is now. For now
> a virtio-bond type solution would allow us to address most of the use
> cases with minimal modification to the existing virtio and can deal
> with feature and/or resource asymmetry.
> 
> - Alex

On Mon, Jan 29, 2018 at 10:24 AM, Michael S. Tsirkin <mst at redhat.com>
wrote:
> On Sun, Jan 28, 2018 at 08:26:53PM -0800, Alexander Duyck wrote:
>> >> > For live migration with advanced usecases that Siwei is
suggesting, i
>> >> > think we need a new driver with a new device type that
can track the
>> >> > VF specific feature settings even when the VF driver is
unloaded.
>> >
>> > I see no added value of the 3 netdev model, there is no need for a
bond
>> > device.
>>
>> I agree a full-blown bond isn't what is needed. However, just
forking
>> traffic out from virtio to a VF doesn't really solve things either.
>>
>> One of the issues as I see it is the fact that the qdisc model with
>> the merged device gets to be pretty ugly. There is the fact that every
>> packet that goes to the VF has to pass through the qdisc code twice.
>> There is the dual nature of the 2 netdev solution that also introduces
>> the potential for head-of-line blocking since the virtio could put
>> back pressure on the upper qdisc layer which could stall the VF
>> traffic when switching over. I hope we could avoid issues like that by
>> maintaining qdiscs per device queue instead of on an upper device that
>> is half software interface and half not. Ideally the virtio-bond
>> device could operate without a qdisc and without needing any
>> additional locks so there shouldn't be head of line blocking
occurring
>> between the two interfaces and overhead could be kept minimal.
>>
>> Also in the case of virtio there is support for in-driver XDP. As
>> Sridhar stated, when using the 2 netdev model "we cannot support
XDP
>> in this model and it needs to be disabled". That sounds like a
step
>> backwards instead of forwards. I would much rather leave the XDP
>> enabled at the lower dev level, and then if we want we can use the
>> generic XDP at the virtio-bond level to capture traffic on both
>> interfaces at the same time.
>
> I agree dropping XDP makes everything very iffy.
>
>> In the case of netvsc you have control of both sides of a given link
>> so you can match up the RSS tables, queue configuration and everything
>> is somewhat symmetric since you are running the PF and all the HyperV
>> subchannels. Most of the complexity is pushed down into the host and
>> your subchannel management is synchronized there if I am not mistaken.
>> We don't have this in the case of this virtio-bond setup. Instead a
>> single bit is set indicating "backup" without indicating what
that
>> means to topology other than the fact that this virtio interface is
>> the backup for some other interface. We are essentially blind other
>> than having the link status for the VF and virtio and knowing that the
>> virtio is intended to be the backup.
>
> Would you be interested in posting at least a proof of concept
> patch for this approach? It's OK if there are some TODO stubs.
> It would be much easier to compare code to code rather than
> a high level description to code.
That is the general idea. I was hoping to go the bonding route last
week but I got too snarled up trying to untangle the features we
didn't need. I have some code I am working on but don't have an ETA
for when it will be done.

At this point I am hoping we can build something based on Sridhar's
original patch that can addresses the items I brought up and shifts to
more of a 3 netdev model. If I am not mistaken we might be able to do
it as a separate driver that has a pair of calls that allow for
adding/removing a virt-bond that is provided with a MAC address and a
device pointer so that we can do the bits necessary to get ourselves
swapped with the original virtio device and identify the virtio as the
backup channel.

Thanks.

- Alex