Re: [dpdk-dev] [patch v4] doc: announce API change in ethdev offload flags

[Andrew Rybchenko <arybchenko@solarflare.com>]

On 8/9/19 12:13 PM, Tom Barbette wrote:
> I think the silent breaking is still not solved for 
> DEV_RX_OFFLOAD_RSS_HASH and DEV_RX_OFFLOAD_FLOW_MARK.
>
> An old application will still compile without any problem, but the RSS 
> hash will not be written and the app will break... They should be 
> negative. Eg DEV_RX_OFFLOAD_NO_RSS_HASH and DEV_RX_OFFLOAD_NO_FLOW_MARK?
>
> Then, regarding the idea, are we sure it's better to add a 
> configuration check/a branch than always copying a few bytes from a 
> warm cache line to a warm cache line? Or some HW could free some 
> internal resources? But drivers are free to ignore it so it is not a 
> problem as opposed to silent breaking.

It could be not delivered from NIC to host saving PCIe bandwidth. E.g. 
flow mark plus RSS hash is 8 byte and it is more than 10% for small packet.

> Tom
>
> On 2019-08-09 10:17, pbhagavatula@marvell.com wrote:
>> From: Pavan Nikhilesh <pbhagavatula@marvell.com>
>>
>> Add new offload flags ``DEV_RX_OFFLOAD_RSS`` and 
>> ``DEV_RX_OFFLOAD_FLOW_MARK``.
>> Add new function ``rte_eth_dev_set_supported_ptypes`` to allow 
>> application to
>> set specific ptypes to be updated in ``rte_mbuf::packet_type``
>>
>> Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
>> Acked-by: Andrew Rybchenko <arybchenko@solarflare.com>
>> Acked-by: Jerin Jacob <jerinj@marvell.com>
>> Acked-by: Hemant Agrawal <hemant.agrawal@nxp.com>
>> ---
>>   doc/guides/rel_notes/deprecation.rst | 23 +++++++++++++++++++++++
>>   1 file changed, 23 insertions(+)
>>
>> diff --git a/doc/guides/rel_notes/deprecation.rst 
>> b/doc/guides/rel_notes/deprecation.rst
>> index 37b8592b6..e4e2a85d7 100644
>> --- a/doc/guides/rel_notes/deprecation.rst
>> +++ b/doc/guides/rel_notes/deprecation.rst
>> @@ -78,3 +78,26 @@ Deprecation Notices
>>     to set new power environment if power environment was already 
>> initialized.
>>     In this case the function will return -1 unless the environment 
>> is unset first
>>     (using ``rte_power_unset_env``). Other function usage scenarios 
>> will not change.
>> +
>> +* ethdev: New offload flags ``DEV_RX_OFFLOAD_RSS_HASH`` and 
>> ``DEV_RX_OFFLOAD_FLOW_MARK``
>> +  will be added in 19.11.
>> +  This will allow application to enable or disable PMDs from updating
>> +  ``rte_mbuf::hash::rss`` and ``rte_mbuf::hash::fdir`` respectively.
>> +  This scheme will allow PMDs to avoid writes to ``rte_mbuf`` fields 
>> on Rx and
>> +  thereby improve Rx performance if application wishes do so.
>> +  In 19.11 PMDs will still update the fields even when the offloads 
>> are not
>> +  enabled.
>> +
>> +* ethdev: New function ``rte_eth_dev_set_supported_ptypes`` will be 
>> added in
>> +  19.11.
>> +  This will allow application to request PMD to set specific ptypes 
>> defined
>> +  through ``rte_eth_dev_set_supported_ptypes`` in 
>> ``rte_mbuf::packet_type``.
>> +  If application doesn't want any ptype information it can call
>> +  ``rte_eth_dev_set_supported_ptypes(ethdev_id, RTE_PTYPE_UNKNOWN)``
>> +  If application doesn't call ``rte_eth_dev_set_supported_ptypes`` 
>> PMD can
>> +  return ``rte_mbuf::packet_type`` with 
>> ``rte_eth_dev_get_supported_ptypes``.
>> +  If application is interested only in L2/L3 layer, it can inform 
>> the PMD to update
>> + ``rte_mbuf::packet_type`` with L2/L3 ptype by calling
>> + ``rte_eth_dev_set_supported_ptypes(ethdev_id, RTE_PTYPE_L2_MASK | 
>> RTE_PTYPE_L3_MASK)``.
>> +  This scheme will allow PMDs to avoid writes to ``rte_mbuf`` fields 
>> on Rx and
>> +  thereby improve Rx performance if application wishes do so.
>> -- 
>> 2.17.1
>>