04/03/2021 11:00, Ori Kam: > From: Thomas Monjalon > > 28/02/2021 20:48, Ori Kam: > > > Currently, DPDK application can offload the checksum check, > > > and report it in the mbuf. > > > > > > However, this approach doesn't work if the traffic > > > is offloaded and should not arrive to the application. > > > > > > This commit introduces rte flow item that enables > > > > s/rte flow/rte_flow/ > > > > Sure > > > > matching on the checksum of the L3 and L4 layers, > > > in addition to other checks that can determine if > > > the packet is valid. > > > some of those tests can be packet len, data len, > > > unsupported flags, and so on. > > > > > > The full check is HW dependent. > > > > What is the "full check"? > > How much it is HW dependent? > > > > This also relates to your other comments, > Each HW may run different set of checks on the packet, > for example one PMD can just check the tcp flags while > a different PMD will also check the option.
I'm not sure how an application can rely on such a vague definition. > > > + * RTE_FLOW_ITEM_TYPE_SANITY_CHECKS > > > + * > > > + * Enable matching on packet validity based on HW checks for the L3 and > > > L4 > > > + * layers. > > > + */ > > > +struct rte_flow_item_sanity_checks { > > > + uint32_t level; > > > + /**< Packet encapsulation level the item should apply to. > > > + * @see rte_flow_action_rss > > > + */ > > > +RTE_STD_C11 > > > + union { > > > + struct { > > > > Why there is no L2 check? > > > Our HW doesn't support it. > If other HW support it, it should be added. It would be an ABI breakage. Can we add it day one? > > > + uint32_t l3_ok:1; > > > + /**< L3 layer is valid after passing all HW checking. */ > > > + uint32_t l4_ok:1; > > > + /**< L4 layer is valid after passing all HW checking. */ > > > > l3_ok and l4_ok looks vague. > > What does it cover exactly? > > > It depends on the HW in question. > In our case it checks in case of L3 > the header len, and the version. > For L4 checking the len. If we don't know exactly what is checked, how an application can rely on it? Is it a best effort check? What is the use case? > > > + uint32_t l3_ok_csum:1; > > > + /**< L3 layer checksum is valid. */ > > > + uint32_t l4_ok_csum:1; > > > + /**< L4 layer checksum is valid. */ What worries me is that the checksum is separate but other checks are in a common bucket. I think we should have one field per precise check with a way to report what is checked.