[ovs-dev] [PATCH 04/16] actions: Allow secondary decoding of a flow

Simon Horman horms at verge.net.au
Wed Feb 13 14:47:46 UTC 2013


On Tue, Feb 12, 2013 at 09:42:52AM -0800, Ben Pfaff wrote:
> On Wed, Feb 06, 2013 at 10:53:55PM +0900, Simon Horman wrote:
> > Actions may provide information to allow further decoding of
> > a flow.
> > 
> > For example:
> > 
> > In the case of MPLS, L3 and L4 information may not initially be decoded
> > from the frame as the ethernet type of the frame is an MPLS type and no
> > information is known about the type of the inner frame.
> > 
> > However, the type of the inner frame may be provided by an mpls_pop action
> > in which case L3 and L4 information may be decoded providing a finer
> > grained match than is otherwise possible.
> > 
> > Signed-off-by: Simon Horman <horms at verge.net.au>
> 
> Why does odp_flow_key_from_flow() take a separate encap_dl_type
> parameter instead of using flow->encap_dl_type?  There may be a good
> reason, but I don't understand yet.

Having looked over the code paths again I believe that is now an artifact
of previous development cycles and can be removed. I'll see about doing so.

> The actions_allow_l3_extraction() check is probably not sufficient in
> facet_check_consistency(), because it misses ofpacts found via
> "resubmit".  We will probably need to record this in xlate_actions().

Thanks. I had not considered resubmit.

To clarify your suggestion: as xlate_actions() traverses all rules we can
use it, perhaps using a return value, use it to determine if there l3
extraction is possible or not more accurately and possibly more efficiently
than using actions_allow_l3_extraction(). If so, yes, I agree and I'll make it
so.

In "ofproto: Allow actions richer matches based on actions"
is used again in handle_flow_miss_without_facet() via
handle_flow_miss_l3_extraction(). It seems that resubmit needs to be taken
into account there too.

> By the way, I think that we may need to adjust how MPLS information is
> included in ODP flow key data.  Reading datapath/README, I think that we
> may need to use a format like:
> 
>         eth(...), eth_type(0x8847), mpls(...), encap(ip(...), tcp(...))
> 
> whereas I believe we are currently omitting the encap and simply writing
> 
>         eth(...), eth_type(0x8847), mpls(...), ip(...), tcp(...)
> 
> I think the latter will make kernel ABI compatibility problematic.

Yes, the patch-set currently uses the latter.

I did once have a go at implementing the former, but without much success.
I can try again. But before I do, I would like to know if encap() can
be nested. I ask this because there can (in theory) be multiple MPLS LSEs.
And also I believe that VLAN and MPLS can co-exist in a single frame.



More information about the dev mailing list