[ovs-discuss] packet reorderoing

Ben Pfaff blp at nicira.com
Thu Mar 5 19:37:43 UTC 2015


On Wed, Mar 04, 2015 at 07:08:03PM +0000, Gray, Mark D wrote:
> > From: Ben Pfaff [mailto:blp at nicira.com]
> > Sent: Wednesday, March 4, 2015 6:12 PM
> > On Wed, Mar 04, 2015 at 05:05:33PM +0000, Gray, Mark D wrote:
> > > > From: Ben Pfaff [mailto:blp at nicira.com]
> > > > Sent: Wednesday, March 4, 2015 4:59 PM
> > > >
> > > > On Wed, Mar 04, 2015 at 04:53:45PM +0000, Gray, Mark D wrote:
> > > > > > From: Ben Pfaff [mailto:blp at nicira.com]
> > > > > > Sent: Wednesday, March 4, 2015 4:51 PM On Wed, Mar 04, 2015 at
> > > > > > 04:42:36PM +0000, Gray, Mark D wrote:
> > > > > > > >
> > > > > > > > On Wed, Feb 25, 2015 at 11:28:56AM -0600, swair shah wrote:
> > > > > > > > > When packets which does not match any rule in ovs, are
> > > > > > > > > buffered at the ingress switch, say first n packets are
> > > > > > > > > buffered. Switch sends a flow_mod and message and
> > > > > > > > > packet_out for the buffered packets, meanwhile if some
> > > > > > > > > more packets of the same flow arrive (say n+1 to m),
> > > > > > > > > they'll also get buffered (as
> > > > flow_mod) hasn't arrived yet.
> > > > > > > > >
> > > > > > > > > Once flow_mod message arrives at the switch then
> > > > > > > > > subsequent packets of the flow will be forwarded but
> > > > > > > > > packets (n+1 to m) will still be buffered waiting for
> > > > > > > > > their packet_outs. This results in reordering of
> > > > > > > > packets.
> > > > > > > > >
> > > > > > > > > Is there a way around this, besides from proactively
> > > > > > > > > setting up a
> > > > flow?
> > > > > > > > >
> > > > > > > > > Is there a way to match flow_mod to existing buffered packets?
> > > > > > > >
> > > > > > > > Reordering can happen.  There isn't a way to avoid it
> > > > > > > > entirely (besides proactive flow setup).  It isn't usually a
> > > > > > > > problem, though, because most protocols only send one packet
> > > > > > > > before they receive a reply from the opposite end.
> > > > > > >
> > > > > > > The only way that I can think to nail up a flow is via dpctl.
> > > > > > > Is this how you suggest to do this?
> > > > > >
> > > > > > Can you be more specific about what you want to do?  I can think
> > > > > > of a few possibilities.
> > > > >
> > > > > I don't know what Swair wants but I would like to be certain to
> > > > > avoid reordering of packets traversing the vswitch. I suspect the
> > > > > way to do this is to add a flow into the datapath and make it
> > > > > persistent so that upcalls are never triggered.
> > > >
> > > > I don't see how that can solve the problem that he is reporting.
> > > > Did you read his whole scenario?  If we knew how to handle the
> > > > packet in the datapath at the beginning of the scenario then we
> > > > would not be sending it to the controller.
> > >
> > > I see the confusion, he is referring to using a controller and it not
> > > matching the switch, I am thinking of missing the datapath. My bad ..
> > > >
> > > > I doubt that this is a real problem that needs to be solved, because
> > > > it's been in the back of my mind since about 2008 as a possible
> > > > concern but this is the
> > >
> > > It has been a concern of mine for a while and I also expected it to
> > > come up at some stage but it never did. I don't think it is an issue
> > > in most cloud deployments but could see it be an issue in NFV
> > > deployments as usually there would be an expectation that a router or
> > > another network element wouldn't reorder.
> > >
> > > However, hypothetically, how would you resolve this?
> > 
> > One way would be to send only the first packet in a given microflow to the
> > controller and buffer the rest.  OpenFlow provides three ways for controllers
> > to handle packet-ins:
> 
> I was talking about the similar problem across the dpif interface. 
> 
> * Packet group A arrives at NIC and get sent to ofproto as an upcall. 
> 
> * Upcall processes some packets from A and calls operate() from dpif.
> operate() will execute ops, a FLOW_PUT followed by EXECUTEs.
> However, some packets may not have been executed are still
> buffered in the vswitchd.
> * At this point, packet group B arrives at the NIC but matches on the fast path
> as the flow has already been "put" and are sent directly without having an 
> upcall. 
> 
> The remaining packets from group A that are buffered in the
> vswitchd will arrive out of order. How could we resolve this?

I don't have a solution I really like.  The solutions I've thought of
have significant downsides.  I'd love to hear of a good solution.  In
the absence of both a good solution and a situation where there's a
real problem to solve, I don't intend to worry about it.



More information about the discuss mailing list