compose_output_action__(), in ofproto/ofproto-dpif-xlate.c, clears
metadata for patch port traversals in the block that begins "if
(xport->peer)", with this code:

        flow->metadata = htonll(0);
        memset(&flow->tunnel, 0, sizeof flow->tunnel);
        memset(flow->regs, 0, sizeof flow->regs);


On Thu, Apr 05, 2018 at 08:52:31PM +0200, Alan Kayahan wrote:
> That....is my question and theory :) How can I see that? Can you point me
> to the code where such decision is taken?
> 
> 2018-04-05 20:39 GMT+02:00 Ben Pfaff <b...@ovn.org>:
> 
> > What metadata is propagating?
> >
> > On Thu, Apr 05, 2018 at 08:14:51PM +0200, Alan Kayahan wrote:
> > > I introduced a new 32bit field in flow.h to match on. A push_header
> > action
> > > appends a header, which contains TLVs, to the packet and sets this field
> > > for the first time. An increment action sets this field to a value that
> > > resides in the first TLV, if increment action is called again, the field
> > is
> > > set to the value in the second TLV and so on.
> > >
> > > I have a chain of br0-br1-br2 connected via patch ports. In br0 flow rule
> > > calls push_header and out to br1. In br1, the flow rule matches on the
> > port
> > > connected to br0, plus match on the 32bit field.The action for this rule
> > is
> > > increment, then out to br2. Just like br1, br2 matches on the port from
> > br1
> > > plus the match on the 32bit field for the value in the TLV.
> > >
> > > Everything works well until br2. But the match on the field for the value
> > > in the TLV doesn't work. If I remove the match on the field at br2 and
> > just
> > > use in_port and redirect the traffic to packet hex printer, I do see that
> > > the value in the TLV has been set correctly at the offset of this 32bit
> > > field.
> > >
> > > The increment action depends on the iteration of TLVs in the header at
> > each
> > > packet, therefore is at datapath and can not update the flow context. So
> > if
> > > the flow metadata is propagating (which seems to be the only
> > explanation),
> > > the new field is set to the old value, which is the culprit. Perhaps a
> > > better approach is to create an own context just like nat, iterate the
> > TLVs
> > > and populate the context with all available values when push_header is
> > > first called, then redesign the increment action so that it just pops the
> > > next value from the context and emits the data-plane action. Eliminates
> > > TLVs iteration per packet at the dataplane.
> > >
> > > But just to save the day, do you have a trick to stop this propagation?
> > Or
> > > do you think the problem might be something else?
> > >
> > > Thanks!
> > >
> > > 2018-04-05 19:41 GMT+02:00 Ben Pfaff <b...@ovn.org>:
> > >
> > > > On Thu, Apr 05, 2018 at 07:31:35PM +0200, Alan Kayahan wrote:
> > > > > OVS patch ports allow the propagation of the metadata (e.g. flow
> > context)
> > > > > across the connected switches.
> > > > >
> > > > > Is there an option to disable the metadata propagation feature? I
> > need
> > > > this
> > > > > for my research to benchmark certain behavior. Or patch ports become
> > > > > nothing but veth pairs when this feature is disabled?
> > > >
> > > > Patch ports zero out most metadata.  What additional metadata do you
> > > > want them to clear?
> > > >
> >
_______________________________________________
discuss mailing list
disc...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-discuss

Reply via email to