[ovs-dev] Unwildcarding of fields modified by set_field action

[Jarno Rajahalme]

> On Jan 10, 2017, at 2:49 PM, Jan Scheurich <jan.scheurich at web.de> wrote:
> 
> Sorry for resending, but an earlier version of the message was blocked by list filters:
> 
> Hi,
> 
> 
> When testing the OVS datapath in the context of work on L3 tunneling and packet-type aware pipeline, we have observed that ofproto-dpif-xlate unwildcards match fields that are being modified by set_field actions, even when those fields are never matched on in the pipeline.
> 
> As a result the megaflow cache is bloated with unnecessarily many entries that only differ in such a non-matched field and the datapath classifier lookup cost increases due to larger miniflow keys/masks.
> 
> Example:
> 
> OpenFlow flow entry:
> 
> ovs-ofctl -Oopenflow13 add-flow br0 “in_port=2,actions=set_field:fe:4a:7a:ae:7c:fb->dl_dst,1”
> 
> Datapath flows:
> 
> recirc_id(0),in_port(3),eth(dst=fe:4a:7a:ae:7c:f1),eth_type(0x0800),ipv4(frag=no), packets:26, bytes:2548, used:0.571s, actions:set(eth(dst=fe:4a:7a:ae:7c:fb)),2
> 
> recirc_id(0),in_port(3),eth(dst=fe:4a:7a:ae:7c:f2),eth_type(0x0800),ipv4(frag=no), packets:26, bytes:2548, used:0.571s, actions:set(eth(dst=fe:4a:7a:ae:7c:fb)),2
> 
> recirc_id(0),in_port(3),eth(dst=fe:4a:7a:ae:7c:f3),eth_type(0x0800),ipv4(frag=no), packets:26, bytes:2548, used:0.571s, actions:set(eth(dst=fe:4a:7a:ae:7c:fb)),2
> 
> …
> 
> We do not understand why this is done and, if there are no good reasons to keep it, would suggest to remove the unwildcarding of fields modified by set_field actions.
> 

Unwildcarding of set fields is only done for practical reasons in the userspace code, and could be optimized away, especially for recent enough datapaths that implement the masked set action.

Currently, the same bit mask is used to mark what fields have been masked and what fields have been set. It would be possible to add a separate flow mask for setting and use the current one only for matching. It will cause some additional overhead on the upcall processing, but will generate better megaflows. I remember looking at this when implementing datapath masked set actions, and vaguely recall someone else also having a go at this later?

> What we do understand is that fields are unwildcarded when they are copied into packet registers or other packet fields. The reason is that these do/may affect the resulting packet. But that is not the case here.
> 
> In the context of L3 tunneling and PTAP the current behavior produces incorrect results because setting the dl_dst of a packet received from an L3 port adds a match on dl_dst==00:00:00:00:00:00 to the Megaflow which can never match because the packet received from the L3 tunnel does not have a dl_dst field.
> 

We might have some code sanitizing the mask that may take care of this, as there are also other cases where we are setting fields that do not exist in the incoming packet.

  Jarno

> Thanks, Jan
> 
> _______________________________________________
> dev mailing list
> dev at openvswitch.org
> https://mail.openvswitch.org/mailman/listinfo/ovs-dev