Re: [PATCH v3 1/6] doc: add RSS hash algorithm feature

[Ferruh Yigit <ferruh.yigit@amd.com>]

On 11/28/2023 1:34 AM, lihuisong (C) wrote:
> 
> 在 2023/11/28 0:35, Ferruh Yigit 写道:
>> On 11/27/2023 3:43 PM, Ferruh Yigit wrote:
>>> On 11/27/2023 1:12 PM, lihuisong (C) wrote:
>>>> 在 2023/11/27 20:19, Ferruh Yigit 写道:
>>>>> On 11/25/2023 1:47 AM, Huisong Li wrote:
>>>>>> Add hash algorithm feature introduced by 23.11 and fix some RSS
>>>>>> features
>>>>>> description.
>>>>>>
>>>>>> Fixes: 34ff088cc241 ("ethdev: set and query RSS hash algorithm")
>>>>>>
>>>>>> Signed-off-by: Huisong Li <lihuisong@huawei.com>
>>>>>> Acked-by: Chengwen Feng <fengchengwen@huawei.com>
>>>>>> ---
>>>>>>    doc/guides/nics/features.rst | 26 ++++++++++++++++++++++----
>>>>>>    1 file changed, 22 insertions(+), 4 deletions(-)
>>>>>>
>>>>>> diff --git a/doc/guides/nics/features.rst
>>>>>> b/doc/guides/nics/features.rst
>>>>>> index 1a1dc16c1e..0d38c5c525 100644
>>>>>> --- a/doc/guides/nics/features.rst
>>>>>> +++ b/doc/guides/nics/features.rst
>>>>>> @@ -277,10 +277,12 @@ RSS hash
>>>>>>    Supports RSS hashing on RX.
>>>>>>      * **[uses]     user config**: ``dev_conf.rxmode.mq_mode`` =
>>>>>> ``RTE_ETH_MQ_RX_RSS_FLAG``.
>>>>>> -* **[uses]     user config**: ``dev_conf.rx_adv_conf.rss_conf``.
>>>>>> +* **[uses]     user config**: ``rss_conf.rss_hf``.
>>>>>>
>>>>> Feature title is "RSS hash", it can be two things,
>>>>> 1. "Receive Side Scaling" support
>>>>> 2. Provide RSS hash to application
>>>>>
>>>>> When this document first prepared RSS hash value was always
>>>>> provided to
>>>>> the application when RSS enabled.
>>>>> So intention with this feature was "Receive Side Scaling" support,
>>>>> hence
>>>>> 'RTE_ETH_MQ_RX_RSS_FLAG' added.
>>>>>
>>>>> Later providing RSS has to the application separated as optimization,
>>>>> 'RTE_ETH_RX_OFFLOAD_RSS_HASH' & 'RTE_MBUF_F_RX_RSS_HASH' added for
>>>>> this
>>>>> support.
>>>> What should I do for above two comments？
>>>> To tell application how to use it?
>>>>
>>> Just tried to give some context.
>>>
>>>
>>>>> As the intention of this feature is "Receive Side Scaling" support, we
>>>>> shouldn't reduce configuration struct to 'rss_conf.rss_hf'.
>>>>>
>>>>> Instead perhaps can expand to:
>>>>> 'rte_eth_conf.rx_adv_conf.rss_conf', 'rte_eth_rss_conf'
>>>>   I just pick their common part.😁
>>>>
>>>> ok, will fix it.
>>>>
>>>>>
>>>>>>    * **[uses]     rte_eth_rxconf,rte_eth_rxmode**:
>>>>>> ``offloads:RTE_ETH_RX_OFFLOAD_RSS_HASH``.
>>>>>>    * **[provides] rte_eth_dev_info**: ``flow_type_rss_offloads``.
>>>>>>    * **[provides] mbuf**: ``mbuf.ol_flags:RTE_MBUF_F_RX_RSS_HASH``,
>>>>>> ``mbuf.rss``.
>>>>>> +* **[related]  API**: ``rte_eth_dev_configure``,
>>>>>> ``rte_eth_dev_rss_hash_update``
>>>>>> +  ``rte_eth_dev_rss_hash_conf_get()``.
>>>>>>    
>>>>> ack
>>>>>
>>>>>>      .. _nic_features_inner_rss:
>>>>>> @@ -288,7 +290,7 @@ Supports RSS hashing on RX.
>>>>>>    Inner RSS
>>>>>>    ---------
>>>>>>    -Supports RX RSS hashing on Inner headers.
>>>>>> +Supports RX RSS hashing on Inner headers by rte_flow API.
>>>>>>    
>>>>> This should be clarified with details below, not sure if it
>>>>> required to
>>>>> limit description to rte_flow.
>>>> But this block like rte_flow_action_rss is from rte_flow.
>>>> And ethdev ops doesn't support inner RSS.
>>>> So I think it is ok.
>>>>
>>> Yes it is supported by rte_flow, and '[uses]' information should already
>>> clarify it.
>>>
>>>
>>>>>
>>>>> And I guess similar confusion exist with the providing hash to user.
>>>>> Need to check if rte_flow implementation puts hash to mbuf along with
>>>>> doing the RSS, or if it checks 'RTE_ETH_RX_OFFLOAD_RSS_HASH' offload,
>>>>> and update below items accordingly.
>>>> Do we need to tell user how to use it here?
>>>> I feel this document is a little simple and main to list interface for
>>>> user.
>>>> In addition, it is better that the more detail about RSS should be
>>>> presented  in rte_flow features.
>>>>
>>> No, I am not suggesting to add more detail.
>>>
>>> My concern is 'RTE_ETH_RX_OFFLOAD_RSS_HASH' information may not be
>>> correct, ethdev APIs checks offload flags, but does rte_flow
>>> implementation check it?
>>>
>>> My suggestion is double check that piece of information and fix it if
>>> required.
>>>
>> Thinking twice, ethdev API or rte_flow or different ways to configure
>> RSS, but datapath that puts hash value to mbuf is same.
>> So same 'RTE_ETH_RX_OFFLOAD_RSS_HASH' check is used for both method, and
>> it is OK to have it documented.
> There was a check for  RTE_ETH_RX_OFFLOAD_RSS_HASH in
> rte_eth_dev_configure().
> This offload flag depends on the RTE_ETH_MQ_RX_RSS_FLAG mode.
> As far as I know, ethdev ops allows to enable RSS hash only when
> RTE_ETH_MQ_RX_RSS_FLAG mode is set.
> But rte flow API enable RSS hash doesn't depend on this MQ mode.
> So I guess that the mbuf.ol_flags:RTE_MBUF_F_RX_RSS_HASH and mbuf.rss
> also be set to report application when create RSS hash flow, even if
> application don't set RTE_ETH_MQ_RX_RSS_FLAG mode by ethdev ops.
>

Even rte_flow is used for RSS, application may want to control exposing
hash value via 'RTE_ETH_RX_OFFLOAD_RSS_HASH' flag, so lets keep the
documentation as it is.

But in implementation it is possible that some drivers may not be using
the offload flag, as you explained above, this confused me but we still
should document what is expected from driver.


>>
>>
>>
>>>>>
>>>>>>    * **[uses]    rte_flow_action_rss**: ``level``.
>>>>>>    * **[uses]    rte_eth_rxconf,rte_eth_rxmode**:
>>>>>> ``offloads:RTE_ETH_RX_OFFLOAD_RSS_HASH``.
>>>>>> @@ -303,9 +305,25 @@ RSS key update
>>>>>>    Supports configuration of Receive Side Scaling (RSS) hash
>>>>>> computation. Updating
>>>>>>    Receive Side Scaling (RSS) hash key.
>>>>>>    -* **[implements] eth_dev_ops**: ``rss_hash_update``,
>>>>>> ``rss_hash_conf_get``.
>>>>>> +* **[implements] eth_dev_ops**: ``dev_configure``,
>>>>>> ``rss_hash_update``, ``rss_hash_conf_get``.
>>>>>> +* **[uses]     user config**: ``rss_conf.rss_key``,
>>>>>> ``rss_conf.rss_key_len``
>>>>>>    * **[provides]   rte_eth_dev_info**: ``hash_key_size``.
>>>>>> -* **[related]    API**: ``rte_eth_dev_rss_hash_update()``,
>>>>>> +* **[related]    API**: ``rte_eth_dev_configure``,
>>>>>> ``rte_eth_dev_rss_hash_update()``,
>>>>>> +  ``rte_eth_dev_rss_hash_conf_get()``.
>>>>>> +
>>>>> ack
>>>>>
>>>>> There is an inconsistency in the documentation but I think it is
>>>>> good to
>>>>> use '()' when documenting API, like: 'rte_eth_dev_configure()'
>>>> +1 will fix it.
>>>>>
>>>>>> +
>>>>>> +.. _nic_features_rss_hash_algo_update:
>>>>>> +
>>>>>> +RSS hash algorithm update
>>>>>> +-------------------------
>>>>>> +
>>>>>> +Supports configuration of Receive Side Scaling (RSS) hash algorithm.
>>>>>> Updating
>>>>>> +RSS hash algorithm.
>>>>>> +
>>>>>> +* **[implements] eth_dev_ops**: ``dev_configure``,
>>>>>> ``rss_hash_update``, ``rss_hash_conf_get``.
>>>>>> +* **[uses]     user config**: ``rss_conf.algorithm``
>>>>>> +* **[provides]   rte_eth_dev_info**: ``rss_algo_capa``.
>>>>>> +* **[related]    API**: ``rte_eth_dev_configure``,
>>>>>> ``rte_eth_dev_rss_hash_update()``,
>>>>>>      ``rte_eth_dev_rss_hash_conf_get()``.
>>>>>>      
>>>>> This document describes features listed in the 'default.ini', so we
>>>>> shouldn't have above.
>>>>>
>>>>> And I don't think RSS hash algorithm update is a big enough feature to
>>>>> list in the feature list, perhaps it can be embedded in the RSS
>>>>> support
>>>>> block, what do you think?
>>>> Yes it is not a bit feature.
>>>> so put it to RSS hash, right?
>>>>
>>> Yes please.
>>>
>> .