Hi Dumitru

On Mon, Jan 6, 2025 at 3:55 PM Dumitru Ceara <[email protected]> wrote:

> On 1/6/25 3:25 PM, Xavier Simonart wrote:
> > Hi Numan, Dumitru
> >
>
> Hi Xavier,
>
> > Thanks for your feedback and sorry for the confusion.
> > Comment embedded below.
> >
> > Thanks
> > Xavier
> >
> > On Mon, Dec 23, 2024 at 12:58 PM Dumitru Ceara <[email protected]>
> wrote:
> >
> >> On 12/20/24 10:31 PM, Numan Siddique wrote:
> >>> On Thu, Nov 14, 2024 at 8:58 AM Xavier Simonart <[email protected]>
> >> wrote:
> >>>> Signed-off-by: Xavier Simonart <[email protected]>
> >>>> ---
> >>>>  controller/physical.c | 15 ++++++++++++---
> >>>>  1 file changed, 12 insertions(+), 3 deletions(-)
> >>>>
> >>>> diff --git a/controller/physical.c b/controller/physical.c
> >>>> index 8d5065098..6b93696fa 100644
> >>>> --- a/controller/physical.c
> >>>> +++ b/controller/physical.c
> >>>> @@ -2194,15 +2194,15 @@ consider_mc_group(struct ovsdb_idl_index
> >> *sbrec_port_binding_by_name,
> >>>>          }
> >>>>
> >>>>          int zone_id = ct_zone_find_zone(ct_zones,
> port->logical_port);
> >>>> -        if (zone_id) {
> >>>> -            put_load(zone_id, MFF_LOG_CT_ZONE, 0, 16, &ofpacts);
> >>>> -        }
> >>> Hi Xavier,
> >>>
> >>> I'm a little confused with this patch.  Can you please explain what is
> >>> the improvement here ?
> >>
> > The goal of the patch is minor: it avoids some differences between flows
> > created by I+P and recompute.
> > Fixing this will let us (in a later patch) automatically compare I+P and
> > recompute flows and catch issues.
> >
> > Without the patch I+P might produce flows as:
> > ... table=43 ... priority=100,reg15=0x8000,metadata=0x1
> > actions=load:0->NXM_NX_REG6[],
> >
> load:0x3->NXM_NX_REG13[0..15],load:0x1->NXM_NX_REG15[],resubmit(,44),load:0x8000->NXM_NX_REG15[]
> > and recompute:
> > ... table=43 ... priority=100,reg15=0x8000,metadata=0x1
> > actions=load:0->NXM_NX_REG6[]*,load:0x4->NXM_NX_REG13[0..15]*
> >
> ,load:0x3->NXM_NX_REG13[0..15],load:0x1->NXM_NX_REG15[],resubmit(,44),load:0x8000->NXM_NX_REG15[]
> > So, in one case we set register REG13 twice: first it is set to 4, and
> then
> > overwritten to 3; in the other case we set it directly to 3.
> > Goal of the patch is to avoid the extra load:0x4->NXM_NX_REG13[0..15].
> >
> >>
> >>> Also can you please update the commit message with more details ?
> >>
> > I'll do it if we agree that the patch makes sense.
> >
> >>>
> >>> From what I understand prior to this patch,  for all the port bindings
> >>> in a multicast group,
> >>> zone id of the pb (if present) is loaded to the ct_zone register.
> >>>
> >>> How is this patch avoiding setting the same register again ?
> >>
> > The issue happens for some port types, such as virtual or localport.
> > Those ports usually have a ct-zone assigned, so we were issuing
> > "put_load(zone_id, MFF_LOG_CT_ZONE, 0, 16, &ofpacts);" for those ports.
> > However, we were not calling "local_output_pb" for those ports (so, no
> > resubmit), and hence the ct_zone register was either overwritten when
> > handling another port in the loop, or was useless (no resubmit after
> > setting the ct_zone register).
> >
>
> Thanks for the explanation but I'm still trying to understand why
> recompute and incremental processing would produce different results.  I
> can see how the order can change but I don't understand why recompute
> generates the "useless" register set operation and the incremental
> processing call wouldn't.
>
The issue occurs when a ct_zone is allocated, for virtual and localport
types.

For virtual ports, this is a side effect of having ct_zone allocated when
recomputing, but not in I+P.
Hence, when handling the virtual port while recomputing, we set the ct_zone
register, do not add any resubmit and start handling the next port.

They both call the same consider_mc_group().  I'm probably missing
> something but it seems to me that at least for virtual/localport types
> that should generate the same openflow actions.
>
Localport types have a slightly different root cause: when the localport is
created in sb (w/o iface-id in ovs),
consider_mc_group is called (sb_mc_group updated) and ct_zone register is
not set (as no zone yet).
Then, when iface-id is set, a zone_id is created... but consider_mc_group
is not executed anymore in I+P.
When in recompute for localport types, the ct_zone has already been
allocated when we execute consider_mc_group (sb_mc_group was not updated),
and the flows are different,
(ct_zone reg13 set for the localport type, no resubmit, and then
potentially other ct_zone register and a resubmit).
I am not sure whether the fact that we do not consider_mc_group when
ct_zone is updated for the localport is an issue as such,
as the only flow difference it creates is meaningless (ct zone set twice
for instance).

>
> > With the patch, we do not set the ct_zone register for those ports.
> >
>
> I'm not completely opposing the patch though.  However, maybe we should
> wrap setting the zone into its own function, to avoid code duplication.
>
The issue occurs when we do not run local_output_pb for a port for which we
have a ct_zone allocated: we set the ct_zone for the port,
then move to the next port, w/o resubmit. So, it might make sense to have a
function like
local_set_ct_zone_and_output_pb. WDYT?

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

Reply via email to