RE: [PATCH v3] net/ice: add MAC anti-spoof option

["Morten Brørup" <mb@smartsharesystems.com>]

> From: Bruce Richardson [mailto:bruce.richardson@intel.com]
> Sent: Wednesday, 17 December 2025 14.46
> 
> On Wed, Dec 17, 2025 at 01:37:16PM +0100, Morten Brørup wrote:
> > > From: Bruce Richardson [mailto:bruce.richardson@intel.com]
> > > Sent: Wednesday, 17 December 2025 12.53
> > >
> > > On Thu, Dec 11, 2025 at 03:22:32PM +0000, Bruce Richardson wrote:
> > > > On Wed, Dec 03, 2025 at 03:47:08PM +0100, Morten Brørup wrote:
> > > > > > From: Mandal, Anurag [mailto:anurag.mandal@intel.com]
> > > > > > Sent: Wednesday, 3 December 2025 15.36
> > > > > >
> > > > > > Hi Morten Brørup,
> > > > > >
> > > > > > From: Morten Brørup <mb@smartsharesystems.com>
> > > > > > Sent: 03 December 2025 17:11
> > > > > > > @@ -1761,13 +1763,39 @@ ice_setup_vsi(struct ice_pf *pf,
> enum
> > > > > > > ice_vsi_type type)
> > > > > > >  		/* Source Prune */
> > > > > > >  		if (ad->devargs.source_prune != 1) {
> > > > > > >  			/* Disable source prune to support VRRP
> > > > > > > -			 * when source-prune devarg is not set
> > > > > > > +			 * when source-prune devargs is not set
> > > > > > >  			 */
> > > > > > >  			vsi_ctx.info.sw_flags =
> > > > > > >  				ICE_AQ_VSI_SW_FLAG_LOCAL_LB;
> > > > > > > -			vsi_ctx.info.sw_flags |=
> > > > > > > +		} else { /* Enable Source Prune in Rx */
> > > > > > > +			vsi_ctx.info.sw_flags =
> > > > > > >  				ICE_AQ_VSI_SW_FLAG_SRC_PRUNE;
> > > > > > >  		}
> > > > > >
> > > > > > This looks like a bug fix related to Source Prune?
> > > > > >
> > > > > > Ans: Not exactly.
> > > > > > Initially, Source Prune was disabled, and MAC Anti-spoof
> check
> > > was
> > > > > > enabled by default. This was done by following:-
> > > > > > Source Prune is disabled by setting local loopback with
> > > > > > ICE_AQ_VSI_SW_FLAG_LOCAL_LB flag in the Rx direction.
> > > > > > ICE_AQ_VSI_SW_FLAG_SRC_PRUNE is added to prevent transmitted
> > > packets
> > > > > > from being looped back in some circumstances.
> > > > > > Now, MAC Anti-spoof check can be disabled by clearing both
> > > > > > ICE_AQ_VSI_SW_FLAG_SRC_PRUNE and
> > > > > > ICE_AQ_VSI_SEC_FLAG_ENA_MAC_ANTI_SPOOF flags and setting Tx
> > > loopback
> > > > > > with
> > > > > > ICE_AQ_VSI_SW_FLAG_ALLOW_LB flag in the Tx direction.
> > > > > >
> > > > > > As we moved to making both source prune and mac anti-spoof
> check
> > > > > > disabled by default, I thought no point to set
> > > > > > ICE_AQ_VSI_SW_FLAG_SRC_PRUNE during source prune disable and
> then
> > > > > > clearing it to disable mac anti-spoof.
> > > > >
> > > > > OK. Thank you for elaborating.
> > > > >
> > > > > >
> > > > > > Thank you.
> > > > > >
> > > > > > Regards,
> > > > > > Anurag M
> > > > >
> > > > > Note to maintainers:
> > > > > This devarg is like the Source Prune devarg.
> > > > > If we want to elevate these exotic features into proper Ethdev
> > > APIs, it should be done for both devargs in a separate patch.
> > > > >
> > > > > Acked-by: Morten Brørup <mb@smartsharesystems.com>
> > > > >
> > > > Applied to dpdk-next-net-intel.
> > > >
> > > Unfortunately, this patch causes changes in the driver behaviour
> > > leading to
> > > CI failures. These issues can be seen with testpmd where packets
> are
> > > looping back inside a nic port unexpectedly.
> >
> > Can you please elaborate "packets are looping back"?
> >
> 
> When testpmd is configured for mac forwarding, sending in a single
> packet
> leads to a constant stream of packets being handled by testpmd.
> 
> > If the packets egress on one physical port, they certainly shouldn't
> ingress back on the same physical port.
> >
> > However, if they egress on one virtual port, and are internally
> switched to ingress on another virtual port on the same physical port,
> I would consider that expected behavior - the same would happen if
> those ports were physical and connected to the same physical switch.
> >
> > If they are ingressing on the same virtual port they were sent on,
> that would seem like a bug in the NICs virtual switch. A physical
> switch normally wouldn't transmit packets back out on the port they
> ingressed on.
> >
> 
> Not exactly sure what is happening internally, it needs some
> investigation.

Sure sounds like it.
Let's hope it's only testpmd, and not a broader issue with the NIC/driver.

> 
> > > Therefore, this patch
> > > needs to
> > > be dropped from next-net-intel.
> > >
> > > Can you please do a new version adding the feature you require
> while
> > > still
> > > keeping the existing default behaviour.  I'm going to move the
> patch
> > > status
> > > from accepted to "changes requested" in patchwork, in anticipation
> of a
> > > new
> > > version.
> > >
> > > Regards,
> > > /Bruce
> >
> > This sounds like the CI needs to be fixed.
> > Why does the CI expect this kind of filtering to be enabled by
> default?
> > I wouldn't expect other NICs to perform similar filtering.
> >
> It could well be a testing issue, or a combination of incorrect default
> behaviour and a sub-optimal test case. However, until that is fully
> root-caused, I'm backing out the patch for safety.

A new patch with reverted default behavior seems like the proper short-term solution.
While investigating, we can live with the wrong default behavior, as a "well known bug".
And the configurability will be helpful for root-causing.

> 
> /Bruce