Re: [PATCH] net: support VLAN stacking packet type parsing

[Vladimir Medvedkin <medvedkinv@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 5691 bytes --]

Hi Morten, all,


пн, 7 июл. 2025 г. в 19:09, Morten Brørup <mb@smartsharesystems.com>:

> > From: Bruce Richardson [mailto:bruce.richardson@intel.com]
> > Sent: Friday, 4 July 2025 13.32
>
> > Hi all,
> >
> > this email discussion comes at a bit of a fortunate time for me, as I'm
> > currently looking at our vlan tag/qinq stripping behaviour in our Intel
> > NIC
> > drivers, and there is some discussion internally as to what our driver
> > behaviour should be compared to what it has historically been. :-)
> >
> > The documentation - both in the NIC guide [1] and the testpmd guide [2]
> > -
> > is rather short on detail as to what exactly the behaviour should be
> > when
> > vlan strip or qinq strip is implemented. Therefore, I'd hope that those
> > more familiar with networking than me would be able to help clarify
> > things
> > so we can document the correct behaviour precisely - and hopefully test
> > our
> > drivers against it in future!
> >
> > The simple cases are obvious (looking only at stripping behaviour here):
> > * no vlan stripping - nothing done to packet
> > * no vlan tag in pkg - nothing to do, irrespective of offload
> > * Vlan strip enabled and single vlan tag present - HW should strip the
> > tag and
> >   place it in descriptor for placing in mbuf.
> >
> > Now the questions I have:
> > * To handle questions with 2 vlan tags, the QinQ case - do we need to
> >   enable both vlan-strip and QinQ strip, or does QinQ strip imply
> > stripping
> >   both?
> >   - one suggested interpretation here, was that QinQ implies stripping
> > the
> >     tag with id EtherType 0x88a8, and vlan stripping implies taking off
> > the
> >     tag with 0x8100
> >   - another interpretation is vlan strip means just to take off one tag
> > (if
> >     present), and qinq strip means to take off both tags (if present).
> >
>
> First off, consider VLAN stripping...
> It strips the VLAN tag if present, but also allows (and parses) untagged
> packets.
> A link with a mix of tagged and untagged packets is called a "hybrid
> link", so this scenario is perfectly valid and common.
> Referring to this behavior, I would expect something similar for QinQ
> stripping, i.e. with QinQ stripping enabled, two, one or zero tags are
> allowed (and parsed).
> This makes the VLAN strip flag superfluous when the QinQ strip flag is set.
>
> You could have a QinQ trunk carrying only QinQ tagged packets and untagged
> Layer 2 Control Protocol packets (LACP etc.).
> In this case you might want the ability to drop VLAN tagged packets, which
> should not occur on the link.
>

That's not quite correct.
There are 2 valid usecases, that may bring some ambiguity:
    1. Some vendors may support mixing dual/single tagged packets on a
physical port, (for example refer to the JunOS flexible-vlan-tagging)
    2. Service provider(SP) provides L2 connectivity to a customer, and
customer is able to send non tagged frames via SP infrastructure.

Thus, upon receive single tagged packet at the SP exit node (the switch
customer is connected to) how does it distinguish (w/o reading local
configuration, i.e. VLAN A - QinQ outer tag, vlans B and C - regular VLANs)
whether the packet is non tagged encapsulated into SP's QinQ, or a regular
VLAN packet belonging to the internal SP infrastructure?
In each case, NIC has to place the VLAN tag in different places of the
descriptor/mbuf.


However, since we don't have such a feature for VLAN trunks, I wouldn't
> expect it for QinQ trunks either.
>
> Another important detail...
> Formally, QinQ is EtherType 0x88a8 with two VLAN tags.
> However, I think double-tagging with EtherType 0x8100 is still broadly in
> use (in old networks, where it is difficult to upgrade to the official QinQ
> EtherType), so I would also treat packets with two VLAN tags (of EtherType
> 0x8100) as QinQ.
> There was also an intermediate unofficial EtherType 0x9100 for QinQ
> tagging, before EtherType 0x88a8 was standardized... but I think we can
> ignore that.
>
> > The question above leads to other consequences:
> > * if we enable qinq strip, but get a single-vlan tagged frame, what is
> > the
> >   behaviour?
> > * if we get a qinq packet, but regular vlan strip is enabled, which tag,
> > if
> >   any, is stripped?
> > * should it be an error to enable both qinq strip and vlan strip at the
> >   same time? Should it be an error to enable qinq strip without vlan
> > strip?
> > * in the mbuf, we have a "vlan_tci" field, and an "vlan_tci_outer"
> > field.
> >   For single vlan strip, presumably only the vlan_tci field should be
> > used,
> >   and for qinq traffic stripped, it's obvious which field goes where.
> >   However, what if we have QinQ strip and we only receive a single vlan
> >   tag, where should that be put? Should it go in inner or outer?
>
> From a protocol parsing perspective, a single VLAN tagged packet has no
> "outer" tag.
>
> Also: Consider the link being configured as a "super-hybrid link"
> (probably not an official name for such a link, but expanding on the common
> term "hybrid link"), carrying a mix of untagged, VLAN tagged and QinQ
> tagged packets. In this case, a single VLAN tagged packet is just a normal
> VLAN tagged packet, with the VLAN ID obviously going to the ordinary
> vlan_tci field.
>
> >
> > Feedback welcome, and suggested doc updates welcome too.
> >
> > Thanks,
> > /Bruce
> >
> >
> > [1] https://doc.dpdk.org/guides/nics/features.html#vlan-offload
> > [2] https://doc.dpdk.org/guides/testpmd_app_ug/run_app.html
>


-- 
Regards,
Vladimir

[-- Attachment #2: Type: text/html, Size: 8738 bytes --]