Re: [dpdk-dev] [PATCH v3 08/13] testpmd: rework csum forward engine

["Ananyev, Konstantin" <konstantin.ananyev@intel.com>]

> -----Original Message-----
> From: Olivier MATZ [mailto:olivier.matz@6wind.com]
> Sent: Wednesday, November 26, 2014 11:15 AM
> To: Ananyev, Konstantin; dev@dpdk.org
> Cc: Walukiewicz, Miroslaw; Liu, Jijiang; Liu, Yong; jigsaw@gmail.com; Richardson, Bruce
> Subject: Re: [PATCH v3 08/13] testpmd: rework csum forward engine
> 
> Hi Konstantin,
> 
> On 11/26/2014 11:10 AM, Ananyev, Konstantin wrote:
> > As I can see you removed code that sets up TX_PKT_IPV4 and TX_PKT_IPV6  of ol_flags.
> > I think that we need to keep it.
> > The reason for that is:
> > With FVL, to make HW TX checksum offload work, SW is responsible to provide to the HW information about L3 header.
> > Possible values are:
> > - IPv4 hdr with HW checksum calculation
> > - IPV4 hdr (checksum done by SW)
> > - IPV6 hdr
> > - unknown
> > So let say to for the packet: ETHER_HDR/IPV6_HDR/TCP_HDR/DATA
> > To request HW TCP checksum offload,  SW have to provide to HW information that it is a packet with IPV6 header
> > (plus as for ixgbe: l2_hdr_len, l3_hdr_len, l4_type, l4_hdr_len).
> > That's why TX_PKT_IPV4 and TX_PKT_IPV6   were introduced.
> >
> > Yes, it is  a change in public API for HW TX offload, but I don't see any other way we can overcome it
> > (apart from make TX function itself to parse a packet, which is obviously not a good choice).
> > Note that existing apps working on existing HW (ixgbe/igb/em) are not affected.
> > Though apps that supposed to be run on FVL HW too have to follow new convention.
> >
> > So I suggest we keep setting these flags in csumonly.c
> 
> Right, I missed these flags.
> It's indeed an API change, but maybe it makes sense, and setting it
> is not a big cost for the application.
> 
> So I would also need to slightly modify the API help in the following
> patches:
>   - [04/13] mbuf: add help about TX checksum flags
>   - [10/13] mbuf: generic support for TCP segmentation offload
> 
> I'll send a v4 this afternoon that integrates this change.

Ok, thanks.

> 
> Do you know precisely when the flags PKT_TX_IPV4 and PKT_TX_IPV6 must
> be set by the application? Is it only the hw checksum and tso use case?

Yes, I believe it should be set only for hw checksum and tso.

> If yes, I'll add it in the API help too.
> 
> By the way (this is probably off-topic), but I'm wondering if the TX
> flags should have the same values than the RX flags:
> 
>    #define PKT_TX_IPV4          PKT_RX_IPV4_HDR
>    #define PKT_TX_IPV6          PKT_RX_IPV6_HDR

Thought about that too.
>From one side,  it is a bit out of our concept: separate RX and TX falgs.
>From other side, it allows us to save 2 bits in the ol_flags.
Don't have any strong opinion here.
What do you think?  

> 
> > Apart from that , the patch looks good to me.
> > And yes, we would need to change the  the way we handle TX offload for tunnelled packets.
> 
> Thank you very much Konstantin for your review.
> 
> Regards,
> Olivier