Re: [dpdk-dev] [EXT] Re: [RFC 1/3] ethdev: add ptype as an Rx offload

[Andrew Rybchenko <arybchenko@solarflare.com>]

On 8/6/19 6:45 PM, Stephen Hemminger wrote:
> On Tue, 6 Aug 2019 14:31:43 +0000
> Pavan Nikhilesh Bhagavatula <pbhagavatula@marvell.com> wrote:
>
>>> -----Original Message-----
>>> From: Andrew Rybchenko <arybchenko@solarflare.com>
>>> Sent: Tuesday, August 6, 2019 2:30 PM
>>> To: Pavan Nikhilesh Bhagavatula <pbhagavatula@marvell.com>; Jerin
>>> Jacob Kollanukkaran <jerinj@marvell.com>; John McNamara
>>> <john.mcnamara@intel.com>; Marko Kovacevic
>>> <marko.kovacevic@intel.com>; Thomas Monjalon
>>> <thomas@monjalon.net>; Ferruh Yigit <ferruh.yigit@intel.com>
>>> Cc: dev@dpdk.org
>>> Subject: [EXT] Re: [dpdk-dev] [RFC 1/3] ethdev: add ptype as an Rx
>>> offload
>>>
>>> External Email
>>>
>>> ----------------------------------------------------------------------
>>> On 8/6/19 11:02 AM, pbhagavatula@marvell.com wrote:
>>>> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
>>>>
>>>> Add ptype to DEV_RX_OFFLOAD_* flags which can be used to
>>> enable/disable
>>>> packet type parsing.
>>>>
>>>> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
>>> I like the idea. I think there are few more Rx features which
>>> lack Rx offload bit:
>>>   - delivery of RSS hash in mbuf (it is not always required when
>>>     RSS is used to distribute packets across Rx queues)
>> Especially when applications use custom hash functions to store flows.
>>
>>>   - maybe Rx mark, since it is an extra information which could
>>>     be passed by NIC to CPU and it is better to know in advance
>>>     at Rx queue setup if it should be requested and processed
>> Are you referring to RTE_FLOW_ACTION_TYPE_MARK?
>>
>>> API breakage should be considered here. I think it is OK to
>>> introduce it in the next release cycle in a dummy way which
>>> does not affect packet type delivery for existing PMDs
>>> (i.e. add offload capability and advertise in PMD, but do not
>>> take it into account when Rx mbuf is filled in) and
>>> submit deprecation notice that it may be taken into account
>>> by PMDs in 20.02 to avoid packet type delivery if the offload
>>> is not requested. It will allow applications to make transition
>>> smoother.
>> Couldn’t agree with you more. I could extend the current RFC to include
>> RSS and RX mark as we would be modifying the same offload fields across
>> all drivers. Easier for PMD maintainers too.
>>
>>> Acked-by: Andrew Rybchenko <arybchenko@solarflare.com>
> I would rather the ptype offload be always on and handled in software
> for drivers that don't do it.

It sounds like wasting of CPU cycle for nothing in some cases.
Also where should software stop? There are various tunnels etc.
If application is unhappy with supported classification provided
by the driver, it can always use software parser if really required.