[dpdk-dev] [PATCH 1/3] mbuf:add two TX offload flags and change three fields

[Ananyev, Konstantin]

Hi Olver,

> -----Original Message-----
> From: Olivier MATZ [mailto:olivier.matz at 6wind.com]
> Sent: Friday, November 28, 2014 10:46 AM
> To: Ananyev, Konstantin; Liu, Jijiang
> Cc: dev at dpdk.org
> Subject: Re: [dpdk-dev] [PATCH 1/3] mbuf:add two TX offload flags and change three fields
> 
> Hi Konstantin,
> 
> On 11/27/2014 06:01 PM, Ananyev, Konstantin wrote:
> >> Yes, I think it could be done that way too.
> >> Though I still prefer to keep l4tun_len - it makes things a bit cleaner (at least to me).
> >> After all  we do have space for it in mbuf's tx_offload.
> >
> > As one more thing in favour of separate l4tun_len field:
> > l2_len is 7 bit long, so in theory it might be not enough, as for FVL:
> > 12:18 L4TUNLEN L4 Tunneling Length (Teredo / GRE header / VXLAN header) defined in Words.
> 
> The l2_len field is 7 bits long because it was mapped to ixgbe hardware.

Yes.

> If it's not enough (although I'm not sure it's possible to have a header
> larger than 128 bytes in this case), it's probably because we should
> not have been doing that.

I also can't imagine the L2 header being that long.
>From other side - I just don't like an idea of PMD stripping off HW capabilities.

> Maybe these generic fields should be generic :)
> If it's not enough, what about changing l2_len to 8 bits?

Yes, was thinking about the same thing.
Maybe instead of introducing l4tun_len, we should increase l2_len and outer_l2_len sizes to 8bits each? 
Though that would break the pair:
l2_len : 7
l3_len :9 
Which is quite useful, as it fits into 2B, and maps exactly to ixgbe TCD layout.
Wonder if we change it, would be there any performance penalty for ixgbe, and if yes how big.

> 
> Often in the recent discussions, I see as an argument "fortville needs
> this so we need to add it in the mbuf". I agree that currently
> fortville is the only hardware supported for the new features, so it
> can make sense to act like this, but we have to accept to come back
> to this API in the future if it makes less sense with other drivers.

Yes, it is sometime hard to keep a balance.
>From one side people would like to have a  clean and generic API, from other side
people would like to have ab ability to use all features that HW supports. 

> 
> Also, application developers can be annoyed to see that the mbuf flags
> and meta data are just duplicating information that is already present
> in the packet.
> 
> About the l4tun_len, it's another field the application has to fill,
> but it's maybe cleaner. I just wanted to clarify why I'm discussing
> these points.
> 
> Regards,
> Olivier