[dpdk-dev] [PATCH v8 10/10] app/testpmd:test VxLAN Tx checksum offload

[Liu, Jijiang]

Hi Olivier,

> -----Original Message-----
> From: Olivier MATZ [mailto:olivier.matz at 6wind.com]
> Sent: Thursday, November 6, 2014 9:09 PM
> To: Liu, Jijiang
> Cc: dev at dpdk.org
> Subject: Re: [dpdk-dev] [PATCH v8 10/10] app/testpmd:test VxLAN Tx checksum
> offload
> 
> Hello Jijiang,
> 
> On 11/06/2014 12:24 PM, Liu, Jijiang wrote:
> >> Is it possible to have a more formal definition? For instance, is the
> >> following definition below correct?
> >>
> >>   "the PKT_RX_TUNNEL_IPV4_HDR flag CAN be set by a driver if the packet
> >>    contains a tunneling protocol inside an IPv4 header".
> >
> > Yes, correct.
> >
> >> If the definition above is correct, I don't see how this flag can
> >> help an application to run faster. There is already a flag telling if
> >> there is a valid IPv4 header (PKT_RX_IPV4_HDR). As the
> >> PKT_RX_TUNNEL_IPV4_HDR flag does not tell what is ip->proto, the work
> >> done by an application to dissect a packet would be exactly the same with or
> without this flag.
> >
> > If the PKT_RX_TUNNEL_IPV4_HDR flag is set, which means driver tell
> application that incoming packet is encapsulated packet, and application will
> process / analyse the packet according to tunneling format indicated by
> packet_type.
> 
> Where is it written that when the PKT_RX_TUNNEL_IPV4_HDR flag is set, the
> packet_type is also set?
> 
> To which header packet_type refers to? Inner or Outer? Depends?
> 
> What are the possible values for packet_type?
> 
> Is the PKT_RX_TUNNEL_IPV4_HDR flag set in mbuf related to the commands
> rx_vxlan_port add|del? If yes, it should be written in the API!
> (assuming this is the right API design)
> 
> When the PKT_RX_TUNNEL_IPV4_HDR flag is set, does PKT_RX_IPV4_HDR or
> PKT_RX_VLAN_PKT concerns the inner or outer headers? I hope it still concerns
> the first one, else it would break many applications relying on the these flags.
> 
> As you can see, today, an application cannot use PKT_RX_TUNNEL_IPV4_HDR or
> m->packet_type because it is not documented.
> 
> 
> > In terms of VXLAN packet format (MAC,IPv4,UDP,VXLAN,MAC,IP,TCP,PAY4), if
> only the PKT_RX_IPV4_HDR flag is set, and application regard its payload as "from
> VXLAN to PAY4", but actually, the real payload is PAY4.
> >
> >> Please, can you give an example showing in which conditions this flag
> >> can help an application?
> >
> > http://dpdk.org/ml/archives/dev/2014-October/007151.html
> > http://dpdk.org/ml/archives/dev/2014-October/007156.html
> >
> > We used the PKT_RX_TUNNEL_IPV4_HDR in the two patches to help
> application identify incoming packet is tunneling packet.
> 
> As you agreed on "the PKT_RX_TUNNEL_IPV4_HDR flag CAN be set by a driver",
> it means that if the flag is not present, the application should do the check in
> software. And there are several reasons why the flag may not be present:
>  - the packet is not a VxLAN packet
As long as it is tunneling packet with IPv4/6 header, the flag should be set by driver.

>  - the hw or driver was not able to recognize it (I don't know, maybe
>    if there are IP options the hw will not recognize it?) 
>  - the hw or driver does not support it (all drivers except i40e)
E1000/ixgbe don't support VXLAN packet and another tunneling packet, so driver don't need to set this flag.
As to other NICs that support tunneling packet , I don't why HW or driver can't recognize it.

> So the application has to provide the software equivalent code to process PAY4.
> 
> The "csum" testpmd forwarding engine is now a bad example because it is not
> able to do the same processing in software or hardware. It now only works with
> an i40e driver, which was not the case before. Also, the semantic of the command
> line arguments changed. Before, the meaning was "if the flag is set, process the
> checksum in the NIC, else in SW".
> Now, it's "huh... it depends on the flag."


Currently, If the packet is non-tunneling packet, I believe the  "csum" testpmd forwarding engine also works well as before.
we changed the engine as follows, which is compatible with previous implementation.
-		if (pkt_ol_flags & PKT_RX_IPV4_HDR) {
+		if (pkt_ol_flags & (PKT_RX_IPV4_HDR | PKT_RX_TUNNEL_IPV4_HDR)) {
...

-		else if (pkt_ol_flags & PKT_RX_IPV6_HDR) {
+		} else if (pkt_ol_flags & (PKT_RX_IPV6_HDR | PKT_RX_TUNNEL_IPV6_HDR)) {


> I will submit a rework of the csum fowarding engine to clarify its behavior.
OK. good.

> Regards,
> Olivier