Re: [dpdk-dev] [patch v3] doc: announce API change in ethdev offload flags

[Jerin Jacob Kollanukkaran <jerinj@marvell.com>]

> -----Original Message-----
> From: Ananyev, Konstantin <konstantin.ananyev@intel.com>
> Sent: Thursday, August 8, 2019 10:24 PM
> To: Jerin Jacob Kollanukkaran <jerinj@marvell.com>; Pavan Nikhilesh
> Bhagavatula <pbhagavatula@marvell.com>; stephen@networkplumber.org;
> arybchenko@solarflare.com; hemant.agrawal@nxp.com;
> thomas@monjalon.net; Yigit, Ferruh <ferruh.yigit@intel.com>; Richardson,
> Bruce <bruce.richardson@intel.com>; Neil Horman
> <nhorman@tuxdriver.com>; Mcnamara, John <john.mcnamara@intel.com>;
> Kovacevic, Marko <marko.kovacevic@intel.com>
> Cc: dev@dpdk.org
> Subject: [EXT] RE: [dpdk-dev] [patch v3] doc: announce API change in ethdev
> offload flags
> > > > Since application has two knobs rte_eth_dev_get_supported_ptypes()
> > > > and DEV_RX_OFFLOAD_PTYPE. We may not need to new ol_flags for
> this
> > > change. Right?
> > > > i.e if application sets the DEV_RX_OFFLOAD_PTYPE, The application
> > > > will get the parsed ptypes by the driver(=
> > > rte_eth_dev_get_supported_ptypes()).
> > > > So there is no scope ambiguity. Right?
> > >
> > > I still think there is:
> > > Imagine user has 2 eth devices, one does support
> > > DEV_RX_OFFLOAD_PTYPE, second doesn't.  Now he has a mix of packets
> > > from both devices, that you want t process.
> > > How would he figure out for which of them ptype values are valid,
> > > and for each are not?
> > > Trace back from what port he has received them?
> > > Not very convenient, and not always possible.
> >
> > I thought so. But in that case, application can always set
> > DEV_RX_OFFLOAD_PTYPE Flags for all the ethdev ports. Right? Rather
> > having any complicated ol_flags or port based parsing. If limit the
> _contract_ to following, we are good. Right?
> > # when DEV_RX_OFFLOAD_PTYPE is set, mbuf.packet_type will be valid
> and
> > mbuf.packet_type will have parsed packet type
> 
> Yes sure in principle user can calculate smallest common subset of RX
> offloads supported by all devs in the system and use only  them.
> Then he can use some global value for ol_flags that will be setup at
> initialization time, instead of checking ol_flags for every mbuf.
> Though inside DPDK we don't use that method for other offloads (cksum,
> vlan, rss).
> Why we should do different here?

I agree. We don't need to.

> Again how to deal with hot-plugged devices with such approach?
> 
> >
> > or the negative offload(This contract is pretty clear, I don't think
> > any ambiguity at all) # when DEV_RX_OFFLOAD_NO_PTYPE(something
> > similar) is set, mbuf.packet_type will be invalid.
> >
> > > I think we need either to introduce new ol_flag value (as we usually
> > > do for other RX offloads), or force PMD to always set ptype value.
> >
> > Setting new  ol_flag value may effect performance for existing drivers
> > which don't planning to use this offload
> 
> If the driver doesn't support DEV_RX_OFFLOAD_PTYPE, it wouldn't need to
> set anything (neither ol_flags, neither packet_type).

Yes

> 
> > and it complicates the
> > application to have additional check based on ol_flag. If you see any
> > corner case with above section,
> >
> > How about just setting as ptype as 0 incase it is not parsed by driver.
> 
> As I said above - ok by me.
> AFAIK, this is current behavior, so no changes in PMD will be required.
> 
> > Actual lookup is the costly one, writing 0 to pytpe is not costly as
> > there are plenty of writes in Rx and it will be write merged(No CPU
> > stall)
> 
> Yes packet_type is at first 64B, so shouldn't cause any extra overhead.
> 
> >
> > I did not get the complete picture of
> > "rte_eth_dev_set_supported_ptypes(uint16_t port_id, uint32_t
> ptype_mask); instead of DEV_RX_OFFLOAD_PTYPE? scheme", Does it help?
> 
> I thought about it as just a different way to disable(/limit) requested by user
> PTYPE support.
> If let say user is not interested in ptype information at all, he can ask PMD to
> just always set ptype value to 0:
> rte_eth_dev_set_supported_ptypes(port, RTE_PTYPE_UNKNOWN);
> 
> if he is interested just in L2/L3 layer info, he can ask PMD to provide ptype
> information only for L2/L3:
> rte_eth_dev_set_supported_ptypes(port, RTE_PTYPE_L2_MASK |
> RTE_PTYPE_L3_MASK);
> 
> Or to enable all supported by PMD ptypes:
> rte_eth_dev_set_supported_ptypes(port, UINT32_MAX)


The API looks good to me. We need to document the  rte_eth_dev_set_supported_ptypes()
must  be called when device is in stop state to allow PMD do slow path configuration.

To summarize:
Two options to control PTYPE lookup:
Option 1:
- If DEV_RX_OFFLOAD_PTYPE set, PMD returns mbuf->packet_type with rte_eth_dev_get_supported_ptypes()
- If DEV_RX_OFFLOAD_PTYPE is not set, PMD still can return  mbuf->packet_type with rte_eth_dev_get_supported_ptypes()
But if PMD can do some optimization, it can avoid ptype lookup and return mbuf->packet_type as zero.

Option 2:
- Introduce rte_eth_dev_set_supported_ptypes(port, needed_ptypes).

I think, rte_eth_dev_set_supported_ptypes() is better option As Konstantain suggested to
have selective control of ptype parsing by PMD at the cost of adding new API.

I think, we can avoid breaking exiting application by, If rte_eth_dev_set_supported_ptypes() is not called,
PMD must return mbuf->packet_type with rte_eth_dev_get_supported_ptypes().
If rte_eth_dev_set_supported_ptypes() and successful, PMD must return
mbuf->packet_type with rte_eth_dev_set_supported_ptypes()

If there no objection to this API, We can send updated deprecation notice.

>