[dpdk-dev] [PATCH v9 0/6] Support TCP/IPv4, VxLAN, and GRE GSO in DPDK

Ferruh Yigit ferruh.yigit at intel.com
Sat Oct 7 01:34:14 CEST 2017


On 10/7/2017 12:32 AM, Ferruh Yigit wrote:
> On 10/5/2017 9:36 PM, Mark Kavanagh wrote:
>> Generic Segmentation Offload (GSO) is a SW technique to split large
>> packets into small ones. Akin to TSO, GSO enables applications to
>> operate on large packets, thus reducing per-packet processing overhead.
>>
>> To enable more flexibility to applications, DPDK GSO is implemented
>> as a standalone library. Applications explicitly use the GSO library
>> to segment packets. This patch adds GSO support to DPDK for specific
>> packet types: specifically, TCP/IPv4, VxLAN, and GRE.
>>
>> The first patch introduces the GSO API framework. The second patch
>> adds GSO support for TCP/IPv4 packets (containing an optional VLAN
>> tag). The third patch adds GSO support for VxLAN packets that contain
>> outer IPv4, and inner TCP/IPv4 headers (plus optional inner and/or 
>> outer VLAN tags). The fourth patch adds GSO support for GRE packets
>> that contain outer IPv4, and inner TCP/IPv4 headers (with optional 
>> outer VLAN tag). The fifth patch in the series enables TCP/IPv4, VxLAN,
>> and GRE GSO in testpmd's checksum forwarding engine. The final patch
>> in the series adds GSO documentation to the programmer's guide.
>>
>> Performance Testing
>> ===================
>> The performance of TCP/IPv4 GSO on a 10Gbps link is demonstrated using
>> iperf. Setup for the test is described as follows:
>>
>> a. Connect 2 x 10Gbps physical ports (P0, P1), which are in the same
>>    machine, together physically.
>> b. Launch testpmd with P0 and a vhost-user port, and use csum
>>    forwarding engine with "retry".
>> c. Select IP and TCP HW checksum calculation for P0; select TCP HW
>>    checksum calculation for vhost-user port.
>> d. Launch a VM with csum and tso offloading enabled.
>> e. Run iperf-client on virtio-net port in the VM to send TCP packets.
>>    With enabling csum and tso, the VM can send large TCP/IPv4 packets
>>    (mss is up to 64KB).
>> f. P1 is assigned to linux kernel and enabled kernel GRO. Run
>>    iperf-server on P1.
>>
>> We conduct three iperf tests:
>>
>> test-1: enable GSO for P0 in testpmd, and set max GSO segment length
>>     to 1518B. Run two iperf-client in the VM.
>> test-2: enable TSO for P0 in testpmd, and set TSO segsz to 1518B. Run
>>     two iperf-client in the VM.
>> test-3: disable GSO and TSO in testpmd. Run two iperf-client in the VM.
>>
>> Throughput of the above three tests:
>>
>> test-1: 9.4Gbps
>> test-2: 9.5Gbps
>> test-3: 3Mbps
>>
>> Functional Testing
>> ==================
>> Unlike TCP packets, VMs can't send large VxLAN or GRE packets. The max
>> length of tunneled packets from VMs is 1514B. So current experiment
>> method can't be used to measure VxLAN and GRE GSO performance, but simply
>> test the functionality via setting small GSO segment length (e.g. 500B).
>>
>> VxLAN
>> -----
>> To test VxLAN GSO functionality, we use the following setup:
>>
>> a. Connect 2 x 10Gbps physical ports (P0, P1), which are in the same
>>    machine, together physically.
>> b. Launch testpmd with P0 and a vhost-user port, and use csum forwarding
>>    engine with "retry".
>> c. Testpmd commands:
>>     - csum parse_tunnel on "P0"
>>     - csum parse_tunnel on "vhost-user port"
>>     - csum set outer-ip hw "P0"
>>     - csum set ip hw "P0"
>>     - csum set tcp hw "P0"
>>     - csum set tcp hw "vhost-user port"
>>     - set port "P0" gso on
>>     - set gso segsz 500
>> d. Launch a VM with csum and tso offloading enabled.
>> e. Create a vxlan port for the virtio-net port in the VM. Run iperf-client
>>    on the VxLAN port, so TCP packets are VxLAN encapsulated. However, the
>>    max packet length is 1514B.
>> f. P1 is assigned to linux kernel and kernel GRO is disabled. Similarly,
>>    create a VxLAN port for P1, and run iperf-server on the VxLAN port.
>>
>> In testpmd, we can see the length of all packets sent from P0 is smaller
>> than or equal to 500B. Additionally, the packets arriving in P1 is
>> encapsulated and is smaller than or equal to 500B.
>>
>> GRE
>> ---
>> The same process may be used to test GRE functionality, with the exception that
>> the tunnel type created for both the guest's virtio-net, and the host's kernel
>> interfaces is GRE:
>>    `ip tunnel add <gre tunnel> mode gre remote <remote IP> local <local_ip>`
>>
>> As in the VxLAN testcase, the length of packets sent from P0, and received on
>> P1, is less than 500B.
>>
>> Change log
>> ==========
>> v9:
>> - fix testpmd build for i686 target
>> - change log level from WARNING to DEBUG in the case of unsupported packet
>>   (rte_gso_segment())
>>
>> v8:
>> - resolve coding style infractions (indentation).
>> - centralize invalid parameter checking for rte_gso_segment() into a single
>>   'if' statement.
>> - don't clear PKT_TX_TCP_SEG flag for packets that don't qualify for GSO
>>   on account of invalid params.
>> - allow GSO for tunneled packets only via gso_ctx (by correcting 'if'
>>   statement condition).
>>
>> v7:
>> - add RTE_GSO_SEG_SIZE_MIN macro; use this to validate gso_ctx.gso_segsz.
>> - rename 'ipid_flag' member of gso_ctx to 'flag'.
>> - remove mention of VLAN tags in supported packet types.
>> - don't clear PKT_TX_TCP_SEG flag if GSO fails.
>> - take all packet overhead into account when checking for empty packet.
>> - ensure that only enabled GSO types are enacted upon (i.e. no fall-through to
>>   TCP/IPv4 case from tunneled case).
>> - validate user-supplied gso segsz arg against RTE_GSO_SEG_SIZE_MIN in testpmd.
>> - simplify error-checking/handling for GSO failure case in testpmd csum engine.
>> - use 0 instead of !RTE_GSO_IPID_FIXED in testpmd.
>>
>> v6:
>> - rebase to HEAD of master (i5dce9fcA)
>> - remove 'l3_offset' parameter from 'update_ipv4_tcp_headers'
>>
>> v5:
>> - add GSO section to the programmer's guide.
>> - use MF or (previously 'and') offset to check if a packet is IP
>>   fragmented.
>> - move 'update_header' helper functions to gso_common.h.
>> - move txp/ipv4 'update_header' function to gso_tcp4.c.
>> - move tunnel 'update_header' function to gso_tunnel_tcp4.c.
>> - add offset parameter to 'update_header' functions.
>> - combine GRE and VxLAN tunnel header update functions into a single
>>   function.
>> - correct typos and errors in comments/commit messages.
>>
>> v4:
>> - use ol_flags instead of packet_type to decide which segmentation
>>   function to use.
>> - use MF and offset to check if a packet is IP fragmented, instead of
>>   using DF.
>> - remove ETHER_CRC_LEN from gso segment payload length calculation.
>> - refactor internal header update and other functions.
>> - remove RTE_GSO_IPID_INCREASE.
>> - add some of GSO documents.
>> - set the default GSO length to 1514 and fill PKT_TX_TCP_SEG for the
>>   packets sent from GSO-enabled ports in testpmd.
>> v3:
>> - support all IPv4 header flags, including RTE_PTYPE_(INNER_)L3_IPV4,
>>   RTE_PTYPE_(INNER_)L3_IPV4_EXT and RTE_PTYPE_(INNER_)L3_IPV4_EXT_
>>   UNKNOWN.
>> - fill mbuf->packet_type instead of using rte_net_get_ptype() in
>>   csumonly.c, since rte_net_get_ptype() doesn't support vxlan.
>> - store the input packet into pkts_out inside gso_tcp4_segment() and
>>   gso_tunnel_tcp4_segment() instead of rte_gso_segment(), when no GSO
>>   is performed.
>> - add missing incldues.
>> - optimize file names, function names and function description.
>> - fix one bug in testpmd.
>> v2:
>> - merge data segments whose data_len is less than mss into a large data
>>   segment in gso_do_segment().
>> - use mbuf->packet_type/l2_len/l3_len etc. instead of parsing the packet
>>   header in rte_gso_segment().
>> - provide IP id macros for applications to select fixed or incremental IP
>>   ids.
>>
>> Jiayu Hu (3):
>>   gso: add Generic Segmentation Offload API framework
>>   gso: add TCP/IPv4 GSO support
>>   app/testpmd: enable TCP/IPv4, VxLAN and GRE GSO
>>
>> Mark Kavanagh (3):
>>   gso: add VxLAN GSO support
>>   gso: add GRE GSO support
>>   doc: add GSO programmer's guide
> 
> Hi Mark, Jiayu,
> 
> I was about the get this to next-net, but recognized same problem with
> gro patch. There are uint8_t storage type usage for port_id. Port id is
> now 16bits. And for testpmd you can prefer "portid_t" storage type as well.
> 
> Another thing is this patch and GRO patch conflicts, because both
> touches same parts. Since Jiayu is the author of the gro patch, is it
> possible to define order between these two patches and it doesn't
> conflict while applying. Order doesn't matter as long as dependency
> defined in cover letter of the patch.
> 
> Can you please send a new version with above two items addressed?

btw, please keep Konstantin's and others Acks in next version of the
patches.

> 
> Thanks,
> ferruh
> 



More information about the dev mailing list