[dpdk-dev] [PATCH v5 2/3] mbuf:add three TX ol_flags and repalce PKT_TX_VXLAN_CKSUM

[Ananyev, Konstantin]

Hi Thomas,

> -----Original Message-----
> From: Thomas Monjalon [mailto:thomas.monjalon at 6wind.com]
> Sent: Thursday, December 04, 2014 10:45 AM
> To: Ananyev, Konstantin; Olivier MATZ
> Cc: dev at dpdk.org; Liu, Jijiang
> Subject: Re: [dpdk-dev] [PATCH v5 2/3] mbuf:add three TX ol_flags and repalce PKT_TX_VXLAN_CKSUM
> 
> Hi,
> 
> 2014-12-04 10:23, Ananyev, Konstantin:
> > From: Liu, Jijiang
> > > From: Olivier MATZ [mailto:olivier.matz at 6wind.com]
> > > > On 12/03/2014 01:59 PM, Ananyev, Konstantin wrote:
> > > > >> I still think having a flag IPV4 + another flag IP_CHECKSUM is not
> > > > >> appropriate.
> > > > >
> > > > > Sorry, didn't get you here.
> > > > > Are you talking about our discussion should PKT_TX_IP_CKSUM and
> > > > > PKT_TX_IPV4 be mutually exclusive or not?
> > > >
> > > > Yes
> > > >
> > > > >> I though Konstantin agreed on other flags, but I may have
> > > > >> misunderstood:
> > > > >>
> > > > >> http://dpdk.org/ml/archives/dev/2014-November/009070.html
> > > > >
> > > > > In that mail, I was talking about my suggestion to make  PKT_TX_IP_CKSUM,
> > > > PKT_TX_IPV4 and PKT_TX_IPV6 to occupy 2 bits.
> > > > > Something like:
> > > > > #define	PKT_TX_IP_CKSUM	(1 << X)
> > > > > #define	PKT_TX_IPV6		(2 << X)
> > > > > #define 	PKT_TX_IPV4		(3 << X)
> > > > >
> > > > > "Even better, if we can squeeze these 3 flags into 2 bits.
> > > > > Would save us 2 bits, plus might be handy, as in the PMD you can do:
> > > > >
> > > > > switch (ol_flags & TX_L3_MASK) {
> > > > >      case TX_IPV4:
> > > > >         ...
> > > > >         break;
> > > > >      case TX_IPV6:
> > > > >         ...
> > > > >         break;
> > > > >      case TX_IP_CKSUM:
> > > > >         ...
> > > > >         break;
> > > > > }"
> > > > >
> > > > > As you pointed out, it will break backward compatibility.
> > > > > I agreed with that and self-NACKed it.
> > > >
> > > > ok, so we are back between:
> > > >
> > > > 1/ (Jijiang's patch)
> > > > PKT_TX_IP_CKSUM  /* packet is IPv4, and we want hw cksum */
> > > > PKT_TX_IPV6      /* packet is IPv6 */
> > > > PKT_TX_IPV4      /* packet is IPv4, and we don't want hw cksum */
> > > >
> > > > with PKT_TX_IP_CKSUM and PKT_TX_IPV4 exclusive
> > > >
> > > > and
> > > >
> > > > 2/
> > > > PKT_TX_IP_CKSUM  /* we want hw IP cksum */
> > > > PKT_TX_IPV6      /* packet is IPv6 */
> > > > PKT_TX_IPV4      /* packet is IPv4 */
> > > >
> > > > with PKT_TX_IP_CKSUM implies PKT_TX_IPV4
> > > >
> > > >
> > > > Solution 2/ looks better from a user point of view. Anyone else has an opinion?
> > >
> > > Let's think about these IPv4/6 flags in terms of checksum and IP version/type,
> > >
> > > 1. For IPv6
> > > IP checksum is meaningful only for IPv4,  so we define 'PKT_TX_IPV6      /* packet is IPv6 */' to tell driver/HW that this is IPV6
> packet,
> > > here we don't talk about the checksum for IPv6 as it is meaningless. Right?
> > >
> > > PKT_TX_IPV6      /* packet is IPv6 */         ------ IP type: v6;  HW checksum: meaningless
> > >
> > > 2. For IPv4,
> > > My patch:
> > >
> > > PKT_TX_IP_CKSUM  /* packet is IPv4, and we want hw cksum */--------------------------IP type: v4;  HW checksum: Yes
> > > PKT_TX_IPV4      /* packet is IPv4, and we don't want hw cksum */ ----------------------- IP type: v4;  HW checksum: No
> > >
> > > You want:
> > > PKT_TX_IP_CKSUM  /* we want hw IP cksum */-------------------------- IP type: v4;  HW checksum: Yes
> > > PKT_TX_IPV4      /* packet is IPv4*/ ------------------------  IP type: v4; HW checksum: yes or no?
> > >                                                                                                        driver/HW don't know, just know this is packet with IPv4 header.
> > >                                                                                                        HW checksum: meaningless??
> >
> > Yep, that's why I also don't like that suggestion: PKT_TX_IPV4 itself doesn't contain all information.
> > PMD will have to check PKT_TX_IP_CKSUM  anyway.
> 
> I prefer solution 2 because a flag should bring only 1 information.

Why is that? For example in mbuf we already have a flag that brings 2 things:
PKT_TX_IP_CKSUM  /* packet is IPv4, and we want hw cksum */

If it would be possible to compress 10 meanings into 1 bit, I would happily do that.
Unfortunately, it is rarely possible :)

> It's simply saner and could fit to more situations in the future.

Could you give an example of such situation?
I personally couldn't come up with the case where #2 would have any real advantage. 
Konstantin

> 
> --
> Thomas