On Fri, Nov 6, 2026 at 2:04 AM Eelco Chaudron <[email protected]> wrote:
>
>
>
> On 4 Nov 2026, at 17:49, Han Zhou via dev wrote:
>
> > On Fri, Oct 16, 2026 at 9:41 AM Aaron Conole <[email protected]> wrote:
> >>
> >> Han Zhou via dev <[email protected]> writes:
> >>
> >>> One OpenFlow flow's actions are translated into one datapath flow.
> >>> "resubmit" is followed during that translation, so every action it
> >>> reaches is appended to the same action list.  A pipeline that fans a
> >>> packet out to many destinations can therefore build an action list that
> >>> the datapath refuses, because a Netlink attribute carries a 16-bit
> >>> length.
> >>>
> >>> Recirculation is the only thing that ends one action list and begins
> >>> another, but a controller has had no way to ask for it.  Open vSwitch
> >>> recirculates on its own behalf for MPLS, for "ct(table=N)", and for a
> >>> paused "controller" action, and "debug_recirc" exposes it directly, but
> >>> that action is intentionally undocumented, subject to change, and
> >>> rejected unless ovs-vswitchd runs with --enable-dummy.
> >>>
> >>> Add "recirc", a documented extension action that ends the datapath flow
> >>> being built and resumes the actions that follow it in a new one.
> >>> Registers, metadata, the stack, and the remaining actions cross the
> >>> boundary, so a pipeline may be divided across several datapath flows
> >>> without changing what it computes.  Translation reuses the existing
> >>> freeze path, which already advances past the action before freezing so
> >>> that the continuation makes progress.
> >>>
> >>> Assisted-by: Claude Opus 5.0, Cursor
> >>> Signed-off-by: Han Zhou <[email protected]>
> >>> ---
> >>
> >> Recheck-request: github-robot
> >
> > Hi Ilya, this feature is required by the OVN fix of broadcast scaling 
> > problem:
> > https://patchwork.ozlabs.org/project/ovn/list/?series=531675
> >
> > I understand the review queue is quite long, but it would be
> > appreciated if you could give a quick look to see if the approach is
> > reasonable, before taking a detailed review, so that we could make
> > some decisions accordingly downstream.
>
> Hi Han,
>
> Find my thoughts on this below, but I'm also interested in other
> maintainers' feedback.
>
> To me this feels like a hack.  From an OVS perspective we try to
> minimise recirculations as much as possible, as they have performance
> impacts (an additional datapath lookup per recirculation) as well as
> preventing offloads from happening.  Adding the ability for users to
> insert actions that only make sense at the datapath level sounds
> wrong, especially to work around architectural issues at a higher
> level.
>

Hi Eelco,

Thanks for your feedback. I'll drop the action.  An OpenFlow action
that just means
"end the datapath flow here" really is a datapath detail leaking up.

But I don't think we're trading away a fast path here.  These flows
don't work at all today, and they fail quietly.

xlate_resubmit_resource_check() refuses every resubmit once the action
list passes 64 kB:

    } else if (ctx->odp_actions->size > UINT16_MAX) {
        xlate_report_error(ctx, "resubmits yielded over 64 kB of actions");
        /* NOT an error, as we'll be slow-pathing the flow in this case? */
        ctx->exit = true; /* XXX: translation still terminated! */

Each refusal throws away whatever that resubmit would have added, but
translation keeps going and the flow still gets installed.  And since
the list is clamped it never actually exceeds the limit, so
nl_attr_oversized() at the end of xlate_actions() never fires and the
flow never gets slow-pathed.  It just goes in missing some of its
actions.  The comments above look like they already suspected this.

Numbers from an OVN multicast flood to 800 remote ports: 712 reached, 88 not.

So I'd rather split this in two.

(1) Correctness.  Delete those four lines and let translation finish.
The list grows to its real size (~247 kB here), nl_attr_oversized()
catches it, SLOW_ACTION gets set, all 800 make it.  Nothing new - that
path has been there since 2013, just unreachable.  I posted a patch
separately:

https://patchwork.ozlabs.org/project/openvswitch/patch/[email protected]/

(2) Performance.  (1) leaves these on the slow path.  If that's not
good enough, I think the answer is OVS inserting the recirculation
itself when the list would overflow, rather than an action users can
place. I can follow-up with this if it is reasonable.

What do you think?

Best,
Han

> Cheers,
>
> Eelco
>
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to