Hi Chenmin, Regarding your second question, some of the device drivers could mark buffers for tx csum offload which is honored only if data is about to be delivered to the network (or if some other action performed to the buffer requires csums to be computed). Otherwise, if packets end up consumed locally, e.g., tunnel or host stack termination, the data is assumed to be correct and no checksum is computed.
So, for instance, if udp/tcp packets received from a tap interface are terminated locally, we don’t try to compute/validate checksum, since there’s nothing to validate as the checksum has not yet been filled in. Regards, Florin > On Apr 13, 2020, at 12:52 AM, Sun, Chenmin <chenmin....@intel.com> wrote: > > [Edited Message Follows] > > Hi experts, > > I am confusing about the L4 checksum behavior in VPP. I noticed the code was > there for a very long time(more than 3 years) so I open this topic to see if > anybody knows the history and why. > > The first problem is in DPDK plugin and was introduced in commit d81566ff92: > Disable for-us udp/tcp checksum validation by default > in Current DPDK plugin, the code clears flags > VNET_BUFFER_F_L4_CHECKSUM_COMPUTED and VNET_BUFFER_F_L4_CHECKSUM_CORRECT when > "enable-tcp-udp-checksum" option is enabled. While these two flags are used > to identify if the L4 checksum is checked and correct, respectively. So I > think this is a wrong behavior - are we doing the opposite? > The current DPDK plugin assumes all the packets from DPDK are correct but in > fact we should check per packet and set the corresponding bits. I've pushed a > patch to fix this issue and Dave invited experts to help to review - Thanks > Dave :-) > > Another problem in the IP4 graph node. The following mechanism was first > introduced in commit 96be8e88109b3e1 and got refactored in commit 1b25552eb. > > #define ip4_local_csum_is_offloaded(_b) \ > _b->flags & VNET_BUFFER_F_OFFLOAD_TCP_CKSUM \ > || _b->flags & VNET_BUFFER_F_OFFLOAD_UDP_CKSUM > > #define ip4_local_need_csum_check(is_tcp_udp, _b) \ > (is_tcp_udp && !(_b->flags & VNET_BUFFER_F_L4_CHECKSUM_COMPUTED \ > || ip4_local_csum_is_offloaded (_b))) > > #define ip4_local_csum_is_valid(_b) \ > (_b->flags & VNET_BUFFER_F_L4_CHECKSUM_CORRECT \ > || (ip4_local_csum_is_offloaded (_b))) != 0 > > In my understanding, the Marcos ip4_local_need_csum_check and > ip4_local_csum_is_valid are used to check if the L4 checksum is computed and > checked before ip4-local node. While the Macro ip4_local_csum_is_offloaded is > to check if the L4 checksum can be offloaded to NIC on the TX side. I am not > very clear why do we need to check the tx offload capability in the RX path - > maybe we are expecting the NIC can help to re-calculate the checksum before > sending out? Think about if an APP(local APP running upon TCP/UDP stack or a > tunnel like VxLAN/GTPU/etc...) is running on L4, it would get invalid > checksum packets(Considering the DPDK plugin marked all the packets are > COMPUTED and CORRECT as I described in the first problem). > > see in: > commit 96be8e88109b3e166b76f58e552dbe438d73bb73 > Author: Jakub Grajciar <jakub.grajc...@pantheon.tech> > Date: Mon Oct 30 14:56:17 2017 +0100 > > vnet: ip4/6_local->don't drop packet if marked for TCP/UDP offload cksum > calculation > >
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#16057): https://lists.fd.io/g/vpp-dev/message/16057 Mute This Topic: https://lists.fd.io/mt/72982275/21656 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-
