[ovs-dev] [PATCH] [RFC] flow: Do not clear L3+ fields of flow in flow_push_mpls()

[Ben Pfaff]

On Thu, May 01, 2014 at 08:21:36AM -0700, Jarno Rajahalme wrote:
> 
> > On May 1, 2014, at 7:53 AM, Ben Pfaff <blp at nicira.com> wrote:
> > 
> >> On Mon, Apr 28, 2014 at 03:57:58PM -0700, Jarno Rajahalme wrote:
> >> 
> >>> On Mar 20, 2014, at 10:05 AM, Ben Pfaff <blp at nicira.com> wrote:
> >>> 
> >>>> On Fri, Mar 14, 2014 at 04:19:52PM +0900, Simon Horman wrote:
> >>>> When creating a flow in the datapath as the result of an upcall
> >>>> the match itself is the match supplied in the upcall while
> >>>> the mask of the match, if supplied, is generated based on the
> >>>> flow and mask composed during action translation.
> >>>> 
> >>>> In the case of, for example a UDP packet, the match will include
> >>>> of L2, L3 and L4 fields. However, if the flow is cleared in
> >>>> flow_push_mpls() then the mask that is synthesised from it will
> >>>> not include L3 and L4 fields. This seems incorrect and the kernel
> >>>> datapath complains about this mismatch.
> >>>> 
> >>>> Signed-off-by: Simon Horman <horms at verge.net.au>
> >>> 
> >>> The goal of clearing the fields is to ensure that later flow tables
> >>> can't match on fields that aren't visible anymore.  That's important for
> >>> accurate OpenFlow implementation, so I'd rather not change it.  On the
> >>> other hand, I see the point you're making, but I don't immediately
> >>> understand why it happens that way.  After all, I can change fields with
> >>> OpenFlow actions and the datapath flows work out OK, why doesn't this
> >>> work out OK too?  Do you understand the reason?
> >> 
> >> As the flow?s dl_type is changed to an MPLS type, later non-MPLS rules
> >> will not match on the modified flow. AFAIK, you can match on L3/L4
> >> fields only by also matching on the corresponding dl_type as a
> >> prerequisite, no?
> > 
> > Yes, that's true.
> > 
> >> If this holds, I?d rather not clear the fields so we can properly do a
> >> set IPv4 action followed by an MPLS push action. Currently the the
> >> MPLS action clears the flow values at the translation time set in the
> >> preceding action, so that at the commit time the values intended for
> >> set IPv4 action are lost.
> > 
> > Are you sure?  compose_mpls_push_action() call commit_odp_actions() to
> > avoid this very problem.
> > 
> 
> I did not notice this, sorry about that.
> 
> > Assuming I'm right about that, at this point what I really want is an
> > example of a situation that's broken in the current code and not broken
> > with this patch applied, so that I can understand exactly what we're
> > getting at here.
> > 
> 
> It seems that a check on the nw_proto before committing set ports
> actions is enough to avoid introducing new problems as part of my
> masked set actions patch series.

OK.

Simon, I guess that you still see a problem that we should fix.  Can you
provide me an example that I can work through for myself?  I want
everything to work well.