On 10 August 2016 at 14:55, Eric Garver <e...@erig.me> wrote:
> Hi,
>
> See two replies below.
>
> Thanks again!
>
> On Fri, Aug 05, 2016 at 10:03:24AM -0400, Eric Garver wrote:
>> Joe,
>>
>> Thanks for further review. I'll add the changes you have below to the
>> series.
>>
>> I'll take a look at the "check-system-userspace" failure. The "802.1ad
>> - push/pop outer tag" test fails on at least one of my setups.
>
> I fixed the remaining test failures.
> There is some work to feature flag 802.1ad support to preserve backwards
> compatibility. I'll hold off on sending out another revision until I can
> make the necessary test changes to support the feature flag. The feature
> flag probably also means additional tests to cover both operating modes.
OK, fair enough. Whenever you're ready, CC me and I can provide
feedback or apply these as appropriate.
>> On Thu, Aug 04, 2016 at 05:53:59PM -0700, Joe Stringer wrote:
>> > Thanks for updating the series.
>> >
>> > With the incremental patch below this is looking pretty reliable for
>> > check-kmod/check-kernel on the platforms I can test on, although
>> > there's still some issue with "make check-system-userspace". It seems
>> > like the userspace datapath cannot receive double-tagged packets from
>> > AF_PACKET properly; it's unclear where the issue is yet.
>
> The frames were being sent with the wrong TPID, 0x8100 instead of
> 0x88a8. So the kernel's VLAN device with proto 802.1ad was dropping
> them.
> With the below kernel commit, it works as expected.
>
> a0cdfcf39362 ("packet: deliver VLAN TPID to userspace")
>
> I believe this commit went into the 3.14 kernel release.
OK, this would affect Ubuntu14.04 (kernel 3.13) for all of the cvlan
tests and I see that those tests fail on that kernel.
I also see the failure below in the IPv6 case, with >MTU size packets
on more recent kernels, do you see this also?
2016-08-04T23:39:12.846Z|00029|netdev_linux|WARN|error sending
Ethernet packet on ovs-p1: Message too long
Daniele and I took a bit of a look at this, and the simplest path
forward is just to drop the MTU on the CVLAN interfaces to 1492. For
some reason, some part of the netdev transmit path for AF_PACKET
socket doesn't understand the cvlan and is appears to be treating both
VLAN tags + the payload as the ethernet payload for MTU calculation,
then because 1508 > 1500, returning -EMSGSIZE. For the OVS testsuite,
we should probably adjust the MTU like this. If you'd like to follow
up on why this is and whether a better fix can be made, that would be
great too and a better long-term solution.
> This can probably be fixed in OVS userspace dataplane by detecting the
> double tag by peaking at the ethertype. If the ethertype is 0x8100 and
> the packet auxdata says there is another tag, then we know it's double
> tagged and can push the 802.1ad tag. Currently this code just assumes
> 0x8100 if the kernel didn't pass the TPID. This is all within
> netdev_linux_rxq_recv_sock().
Thanks for the detailed investigation!
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
