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

["Liu, Jijiang" <jijiang.liu@intel.com>]

> -----Original Message-----
> From: Ananyev, Konstantin
> Sent: Friday, December 5, 2014 6:56 AM
> To: Olivier MATZ; Thomas Monjalon
> Cc: dev@dpdk.org; Liu, Jijiang
> Subject: RE: [dpdk-dev] [PATCH v5 2/3] mbuf:add three TX ol_flags and repalce
> PKT_TX_VXLAN_CKSUM
> 
> 
> 
> > -----Original Message-----
> > From: Olivier MATZ [mailto:olivier.matz@6wind.com]
> > Sent: Thursday, December 04, 2014 1:51 PM
> > To: Ananyev, Konstantin; Thomas Monjalon
> > Cc: dev@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,
> >
> > On 12/04/2014 12:03 PM, Ananyev, Konstantin wrote:
> > >>>>> 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 */
> >
> > For the user, it's clearer to have one information in a flag.
> > If you just look at the name of the flag, the natural meaning is 2/,
> > else we would need to rename them in:
> >    PKT_TX_IPV4_CKSUM
> >    PKT_TX_IPV4_NO_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.
> >
> > in solution 2/, PKT_TX_IP_CKSUM implies PKT_TX_IPV4 so checking
> > PKT_TX_IP_CKSUM is still enough in drivers.
> 
> Both 1 and 2 seems backward compatible.
> 
> >
> > In the driver, it is also simpler. With solution 1/:
> >
> > /* check if we need ipcsum */
> > if (flags & PKT_TX_IP_CKSUM)
> >
> > /* check if packet is ipv4, may be needed to set a hw field */ if
> > (flags & (PKT_TX_IP_CKSUM|PKT_TX_IPV4))
> 
> Do you really mean 1 here? When all 3 flags are mutually exclusive?
> If so, it doesn't look right. For 1 both (PKT_TX_IP_CKSUM|PKT_TX_IPV4) should
> never be up.
> 
> >
> >
> > With solution 2/
> >
> > /* check if we need ipcsum */
> > if (flags & PKT_TX_IP_CKSUM)
> >
> > /* check if packet is ipv4, may be needed to set a hw field */ if
> > (flags & PKT_TX_IPV4)
> 
> The thing is that it wouldn't be possible with FVL driver - it has to setup mutually
> exclusive fields for these 2 cases:
> PKT_TX_IP_CKSUM - ipv4 with HW checksum
> PKT_TX_IPV4 - ipv4 without HW checksum
> 
> So with #2, driver has either:
> if (flags & PKT_TX_IP_CKSUM) {...} else if (flags & PKT_TX_IPV4) {...} And always
> keep condition for PKT_TX_IP_CKSUM first.
> Or do:
> if (flags & PKT_TX_IPV4) {...} if (flags & PKT_TX_IP_CKSUM) {...} and in that case
> always keep condition for PKT_TX_IP_CKSUM last, so it always overwrite
> PKT_TX_IPV4 settings.
> 
> Basically with #2 PKT_TX_IPV4 is not enough to make a decision, even if it is set,
> we'll have to check for PKT_TX_IP_CKSUM anyway.
> 
> While with 1 we can put them in any order, both:
> If (flags & PKT_TX_IP_CKSUM) {...} else if (flags & PKT_TX_IPV4) {...} And If (flags
> & PKT_TX_IPV4) {...} else if (flags & PKT_TX_IP_CKSUM) {...} Will work.
> 
> Konstantin

Yes. I agree.

PKT_TX_IPV4      /* packet is IPv4 */ 
This flag don't have too much offload meaning for TX side,  because we can't use this information to set transmit descriptor ( or set offload register ) precisely , so it is not a real offload flag.

PKT_TX_IPV4      /* packet is IPv4, and we don't want hw cksum */
It  is a offload flag, we can use this flag to set  transmit descriptor precisely.  
Yes, the comments are different, the PKT_TX_IPV4 has different meanings, but we have to consider which comment will affect if offload work or not.

BTW, we pay too much time on this topic...


> >
> >
> > I agree it can looks like a detail, but I really think it's important
> > to have the most logical and straightforward api for mbuf, as it's the
> > core of DPDK.
> >
> > Regards,
> > Olivier