On Tue, Mar 14, 2023 at 06:49:25PM +0100, Ilya Maximets wrote: > On 3/14/23 12:15, Simon Horman wrote: > > On Mon, Mar 13, 2023 at 09:34:30PM +0100, Ilya Maximets wrote: > >> On 3/13/23 11:27, Roi Dayan wrote: > >>> From: Oz Shlomo <o...@nvidia.com> > >>> > >>> Currently jumping over a output-to-port action is translated to tc > >>> mirror action and stolen control action. > >>> However, the tc control action is not propagated to the hw offload action, > >>> thus the hardware action will mirror the packet and continue to the next > >>> action. > >> > >> This sounds like a bug in the hardware offloading in the kernel/driver. > >> Shouldn't this be fixed in the kernel? > > > > Hi Ilya, > > > > I think what we are seeing is a mismatch between how > > OVS translates flows to TC, and how TC actually works (since forever). > > > > It seems to me that fixing the translation, as this patch does, > > is the right approach. But I could be missing something (often the case). > > I'm not a TC expert, but the description sounded like the driver > or some other code in the kernel is ignoring the TC_ACT_STOLEN. > If that's correct, that must be a bug in the kernel. > > We could still change the way OVS translates flows for TC in order > to avoid the kernel bug as long as we're not introducing anything > nasty in the flow translation logic. But it still remains a kernel > bug that needs to be fixed regardless. So, I was just making sure > that the issue is worked on from the kernel side. And if it can > be fixed in the kernel in a reasonable time frame, then maybe we > don't need to change OVS after all. This jumping code is fairly > complex already.
Thanks Ilya, I understand your point. And I agree that the kernel treatment of TC_ACT_STOLEN needs to be understood/investigated. _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev
