On Thu, Jul 19, 2012 at 08:53:23PM -0700, Ben Pfaff wrote: > On Fri, Jul 20, 2012 at 11:23:02AM +0900, Isaku Yamahata wrote: > > This patch adds instruction OF11 apply-actions/goto-table supports. > > > > I rewrite the patch series based on Ben's suggestion. > > Now introduce ofpact_instruction to represent instruction delimiter, > > and avoid nesting. > > Ben, do you like this approach? > > Oh, gosh, I was going to reply to your question from yesterday but > didn't get a chance yet. I didn't expect you to work on it before I > did. I hope I didn't waste a bunch of your time. > > The high-level principle I have in mind is to reconcile the various > ideas of actions across OpenFlow versions by implementing them as > extensions in the protocols that don't support them natively. > > Joe Stringer has already submitted a patch that implements > Write-Metadata almost the way I want it, in the patch here: > http://openvswitch.org/pipermail/dev/2012-July/018897.html > There's a lot of discussion of approach afterward, ending up here: > http://openvswitch.org/pipermail/dev/2012-July/018997.html > > So, let me describe what I have in mind overall for when we're finished. > OpenFlow has the following instructions, which may appear at most once > each per set of instructions and which the spec says are executed in the > order below: > > Apply-Actions > Clear-Actions > Write-Actions > Write-Metadata > Goto-Table > > My thought is that each of these instructions should correspond to > one or more ofpact types, which will be put into the list of ofpacts in > the order above: > > - Apply-Actions: the existing ofpacts. > > - Clear-Actions: add a new ofpact_clear_actions. > > - Write-Actions: see below. > > - Write-Metadata: add a new ofpact_write_metadata. > > - Goto-Table: add a new ofpact_goto_table (or possibly reuse > ofpact_resubmit if that seems suitable after some thought). > > I don't think that any of these needs nesting, with the possible > exception of Write-Actions. I'm not sure about that one; I could see an > argument each way. What do you think?
Yes, either way is possible, and I implemented the both way. Given that you're thinking that Apply-Actions shouldn't use nesting, I'd like not to use nesting for Write-Actions. If Write-Actions uses nesting, it would be natural for Apply-Actions to use nesting. At least for me. What concerns me is how complex xlate_actions() would be with Write-Actions. Anyway we can re-consider it when I (or someone else) attack Write-Actions. It is internal representation, not exposed API/ABI. > For the string representations, as I told Joe Stringer, I'm leaning > toward requiring them to be in the correct order for instructions listed > above. Otherwise it seems risky that one could confuse the user. > Similarly for the OF1.0 extension representation of OF1.1 instructions. > > Comments? > > Thanks, > > Ben. > -- yamahata _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
