UDP-GRO not working

Stephen Hemminger stephen at networkplumber.org
Fri Mar 11 18:15:40 CET 2022


On Thu, 10 Mar 2022 12:49:04 +0000
Kumara Parameshwaran <kparameshwar at vmware.com> wrote:

> Hi ,
> 
> I tried using the UDP GRO feature in DPDK recently and it did not see working. I understand the GRO for UDP is applicable only for fragmented packets, there is the following check in gro_udp4.c
> /*
> * Don't process non-fragment packet.
> */
> if (!is_ipv4_fragment(ipv4_hdr))
> return -1;
> 
> 
> There looks to be some conflict in the definition of RTE_PTYPE in rte_mbuf_ptype.h and the usage of this in GRO layer, rte_gro.c
> 
> The below are the definitions,
> 
> #define RTE_PTYPE_L4_TCP                    0x00000100
> #define RTE_PTYPE_L4_UDP                    0x00000200
> #define RTE_PTYPE_L4_FRAG                   0x00000300
> 
> Below is the check in GRO layer,
> 
> #define IS_IPV4_TCP_PKT(ptype) (RTE_ETH_IS_IPV4_HDR(ptype) && \
>         ((ptype & RTE_PTYPE_L4_TCP) == RTE_PTYPE_L4_TCP) && \
>         (RTE_ETH_IS_TUNNEL_PKT(ptype) == 0))
> 
> #define IS_IPV4_UDP_PKT(ptype) (RTE_ETH_IS_IPV4_HDR(ptype) && \
>         ((ptype & RTE_PTYPE_L4_UDP) == RTE_PTYPE_L4_UDP) && \
>         (RTE_ETH_IS_TUNNEL_PKT(ptype) == 0))
> 
> So, for a fragmented UDP packet both RTE_PTYPE_L4_TCP & RTE_PTYPE_L4_UDP would be set and the GRO layer would be not able to interpret the packet type right.
> 
> I am using rte_net_get_ptype API before the packet is being sent to the GRO subsystem as the DPDK PMD for the NIC I am using would not set the packet types as required by the GRO subsystem.
> 
> I would like to contribute a patch for this bug if this indeed is an issue, I was thinking if the GRO subsystem is L4 fragmented then in the GRO layer invoked the appropriate handler, either gro_tcp4_reassemble or gro_ud4_reassemble.
> 
> Please let me know if I am missing something here.
> 
> Thanks,
> Param.

Are you using RSS, perhaps the fragmented packet is arriving on a different queue.
Since fragments don't have UDP header, often the arrive on a default queue.


More information about the dev mailing list