Re: [dpdk-dev] [PATCH v6 1/9] librte_mbuf:the rte_mbuf structure changes

[Thomas Monjalon <thomas.monjalon@6wind.com>]

2014-11-13 11:24, Liu, Jijiang:
> From: Thomas Monjalon [mailto:thomas.monjalon@6wind.com]
> > 2014-11-13 03:17, Liu, Jijiang:
> > > From: Thomas Monjalon [mailto:thomas.monjalon@6wind.com]
> > > > 2014-10-23 02:23, Zhang, Helin:
> > > > > From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Thomas
> > > > > Monjalon
> > > > > > 2014-10-21 14:14, Liu, Jijiang:
> > > > > > > From: Thomas Monjalon [mailto:thomas.monjalon@6wind.com]
> > > > > > > > 2014-10-21 16:46, Jijiang Liu:
> > > > > > > > > +	uint16_t packet_type;
> > > > > > > >
> > > > > > > > Why not name it "l2_type"?
> > > > >
> > > > > 'packet_type' is for storing the hardware identified packet type
> > > > > upon different layers of protocols (l2, l3, l4, ...).
> > > > > It is quite useful for user application or middle layer software
> > > > > stacks, it can know what the packet type is without checking the
> > > > > packet too
> > > > much by software.
> > > > > Actually ixgbe already has packet types (less than 10), which is
> > > > > transcoded into
> > > > 'ol_flags'.
> > > > > For i40e, the packet type can represent about 256 types of packet,
> > > > > 'ol_flags' does not have enough bits for it anymore. So put the
> > > > > i40e packet types
> > > > into mbuf would be better.
> > > > > Also this field can be used for NON-Intel NICs, I think there must
> > > > > be the similar concepts of other NICs. And 16 bits 'packet_type'
> > > > > has severl
> > > > reserved bits for future and NON-Intel NICs.
> > > >
> > > > Thanks Helin, that's the best description of packet_type I've seen so far.
> > > > It's not so clear in the commit log:
> > > > 	http://dpdk.org/browse/dpdk/commit/?id=73b7d59cf4f6faf
> > > >
> > > > > > > In datasheet, this term is called packet type(s).
> > > > > >
> > > > > > That's exactly the point I want you really understand!
> > > > > > This is a field in generic mbuf structure, so your datasheet has no value here.
> > > > > >
> > > > > > > Personally , I think packet type is  more clear what meaning of this field is .
> > > > > >
> > > > > > You cannot add an API field without knowing what will be its generic meaning.
> > > > > > Please think about it and describe its scope.
> > > >
> > > > I integrated this patch with the VXLAN patchset in the hope that
> > > > you'll improve the situation afterwards.
> > > > This is the answer you recently gave to Olivier:
> > > > 	http://dpdk.org/ml/archives/dev/2014-November/007599.html
> > > > "
> > > > 	Regarding adding a packet_type in mbuf, we ever had a lot of
> > > > discussions as follows:
> > > > 	http://dpdk.org/ml/archives/dev/2014-October/007027.html
> > > > 	http://dpdk.org/ml/archives/dev/2014-September/005240.html
> > > > 	http://dpdk.org/ml/archives/dev/2014-September/005241.html
> > > > 	http://dpdk.org/ml/archives/dev/2014-September/005274.html
> > > > "
> > > >
> > > > To sum up the situation:
> > > > - We don't know what are the possible values of packet_type
> > > > - It's only filled by i40e, while other drivers use ol_flags
> > > > - There is no special value "unknown" which should be set by drivers
> > > >   not supporting this feature.
> > > > - Its only usage is to print a decimal value in
> > > > app/test-pmd/rxonly.c
> > > >
> > > > It's now clear that nobody cares about this part of the API.
> > > > So I'm going to remove packet_type from mbuf.
> > > > I don't want to keep something that we don't know how to use, that
> > > > is not consistent across drivers, and that overlap another API part (ol_flags).
> > >
> > > The packet type in 40e is very important for user, using packet type
> > > can help to speed up packet analysis/identification in their
> > > application, especially tunneling packet format.
> > > Now I'm working on implementing packet type definition in rte_ethdev.h
> > > file and  translation table in i40e, which is almost done.
> > > The packet type  definition in in rte_ethdev.h file like below.
> > > /*
> > >  * Ethernet packet type
> > >  */
> > > enum rte_eth_ptype {
> > >         /* undefined packet type, means HW can't recognise it */
> > >         RTE_PTYPE_UNDEF = 0,
> > > ...
> > >
> > >         /* IPv4 --> GRE/Teredo/VXLAN --> MAC --> IPv4 */
> > >         RTE_PTYPE_IPv4_GRENAT_MAC_IPv4FRAG_PAY3,
> > >         RTE_PTYPE_IPv4_GRENAT_MAC_IPv4_PAY3,
> > >         RTE_PTYPE_IPv4_GRENAT_MAC_IPv4_UDP_PAY4,
> > >         RTE_PTYPE_IPv4_GRENAT_MAC_IPv4_TCP_PAY4,
> > >         RTE_PTYPE_IPv4_GRENAT_MAC_IPv4_SCTP_PAY4,
> > >         RTE_PTYPE_IPv4_GRENAT_MAC_IPv4_ICMP_PAY4,
> > >
> > >         /* IPv4 --> GRE/Teredo/VXLAN --> MAC --> IPv6 */
> > >         RTE_PTYPE_IPv4_GRENAT_MAC_IPv6FRAG_PAY3
> > >         RTE_PTYPE_IPv4_GRENAT_MAC_IPv6_PAY3,
> > >         RTE_PTYPE_IPv4_GRENAT_MAC_IPv6_UDP_PAY4,
> > >         RTE_PTYPE_IPv4_GRENAT_MAC_IPv6_TCP_PAY4,
> > >         RTE_PTYPE_IPv4_GRENAT_MAC_IPv6_SCTP_PAY4,
> > >         RTE_PTYPE_IPv4_GRENAT_MAC_IPv6_ICMP_PAY4,
> > >
> > >         /* IPv4 --> GRE/Teredo/VXLAN --> MAC/VLAN */
> > >         RTE_PTYPE_IPv4_GRENAT_MACVLAN_PAY3,
> > > ...
> > > }
> > 
> > OK, it seems well abstracted.
> > I think the last part of these names (PAY3/PAY4) is useless.
> > 
> > When this patch for API and i40e will be ready?
> > I'd prefer fixing the API instead of removing it.
> 
> If needed, next week, I can send a patch for this.
> 
> > > Yes, we don't use packet type in many places now, which doesn't mean
> > > we don't use it  in the future (when supporting another tunneling packet).
> > >
> > > It is ok for me if you want to remove the packet_type filed in mbuf,
> > > but we will send a separate patch set for introducing packet type in
> > > the future, which includes 1g/10/40g PMD changes.
> > 
> > When the patches for igb/ixgbe will be ready?
> 
> We need some time to investigate this for igb/ixgbe, probably some
> example codes and test application codes need to changed. 
> You can assume that it cannot be done in DPDK1.8.
> 
> So here are my three suggestions:
> 
> 1. keep packet_type in mbuf and wait for all the igb/ixgb/i40e changes
> done in DPDK2.0. Now, I don't send a separate  patch set for it. 
> 2. keep  packet_type in mbuf, I just send i40e patch set for this in
> DPDK1.8. In DPDK2.0, we will send a patch set  for igb/ixgbe.
> 3. It can be removed now,  and we will send a separate patch set for
> introducing packet type in the future.

Option 2 please :)
My main concerns are:
- clearly document it
- have hardware abstraction

Thanks
-- 
Thomas