[dpdk-dev] [PATCH v5] app/testpmd: fix the default RSS key configuration

[Ferruh Yigit]

On 10/20/2020 10:00 AM, oulijun wrote:
> 
> 
> 在 2020/10/16 18:57, Ferruh Yigit 写道:
>> On 10/16/2020 11:04 AM, oulijun wrote:
>>>
>>>
>>> 在 2020/10/16 7:53, Ferruh Yigit 写道:
>>>> On 10/16/2020 12:21 AM, Ophir Munk wrote:
>>>>>> From: Ferruh Yigit <ferruh.yigit at intel.com>
>>>>>>>>> On 10/15/2020 1:41 PM, Lijun Ou wrote:
>>>>>>>>>> When start the testpmd, the pmd driver initializes the RSS
>>>>>>>>>> configuration. Generally, the recommended RSS hash key is used as
>>>>>>>>>> the default key in the driver. In addition, the default key is
>>>>>>>>>> different from the default RSS flow in testpmd without specifying
>>>>>>>>>> RSS hash key. So. if you do not specify the RSS key when creating
>>>>>>>>>> an RSS rule, the testpmd uses the default key as the default RSS
>>>>>>>>>> key of the RSS rule. As a result, you may mistakenly consider that
>>>>>>>>>> the RSS key in use is the valid default key of the NIC, actually,
>>>>>>>>>> the key and the valid default key of the NIC are two values.
>>>>>>>>>>
>>>>>>>>>> Consider the follow usage with testpmd:
>>>>>>>>>> 1. first, startup testpmd:
>>>>>>>>>> testpmd> show port 0 rss-hash key
>>>>>>>>>> RSS functions:
>>>>>>>>>>     all ipv4-frag ipv4-other ipv6-frag ipv6-other ip RSS key:
>>>>>>>>>> -
>>>>>> 6D5A56DA255B0EC24167253D43A38FB0D0CA2BCBAE7B30B477CB2DA3803
>>>>>> 0F
>>>>>>>>>> -20C6A42B73BBEAC01FA
>>>>>>>>>> 2. create a rss rule
>>>>>>>>>> testpmd> flow create 0 ingress pattern eth / ipv4 / udp / end
>>>>>>>>>> testpmd> actions rss \
>>>>>>>>>> types ipv4-udp end queues end / end
>>>>>>>>>>
>>>>>>>>>> 3. show rss-hash key
>>>>>>>>>> testpmd> show port 0 rss-hash key
>>>>>>>>>> RSS functions:
>>>>>>>>>>     all ipv4-udp udp
>>>>>>>>>> RSS key:
>>>>>>>>>> -
>>>>>> 74657374706D6427732064656661756C74205253532068617368206B65792C
>>>>>> 206F
>>>>>>>>>> -76657272696465
>>>>>>>>>>
>>>>>>>>>> In order to solve the above problems, it use the NIC valid default
>>>>>>>>>> RSS key instead of the testpmd dummy RSS key in the flow
>>>>>>>>>> configuration when the RSS key is not specified in the flow rule.
>>>>>>>>>> If the NIC RSS key is invalid, it will use testpmd dummy RSS key as the
>>>>>> default key.
>>>>>>>>>>
>>>>>>>>>> Fixes: ac8d22de2394 ("ethdev: flatten RSS configuration in flow
>>>>>>>>>> API")
>>>>>>>>>> Cc: stable at dpdk.org
>>>>>>>>>>
>>>>>>>>>> Signed-off-by: Lijun Ou <oulijun at huawei.com>
>>>>>>>>>> Reviewed-by: Phil Yang <phil.yang at arm.com>
>>>>>>>>>> ---
>>>>>>>>>> V4->V5:
>>>>>>>>>> -rewrite the commit log
>>>>>>>>>> -add reviewed-by
>>>>>>>>>
>>>>>>>>> Hi Lijun,
>>>>>>>>>
>>>>>>>>> There were multiple other comments, it seems they are not addressed
>>>>>>>>> but only updated the commit log, can you please check comments to
>>>>>> prev versions.
>>>>>>>>>
>>>>>>>>> Before going into the details, my question was what happens if
>>>>>>>>> default key not provided at all?
>>>>>>>>> It seems this has been already tried by Ophir [1], later reverted
>>>>>>>>> back [2] bringing the initial issue back.
>>>>>>>>>
>>>>>>>>> According commit, the reason of revert is to support following
>>>>>> command:
>>>>>>>>> "flow create 0 <pattern> actions rss queues 0 1 end key_len 40 / end"
>>>>>>>>>
>>>>>>>>> @Ophir, @Lijun,
>>>>>>>>> Can we ignore the 'key_len' if the 'key' is not supported and solve
>>>>>>>>> current issue as initially intended ([1])?
>>>>>>>>>
>>>>>>>> Hi, Ferruh
>>>>>>>>     I have discussed with Phil Yang about the problem in [1]. I think
>>>>>>>> there may be other problems with the idea and there is no better
>>>>>>>> solution. and we need to remove key_len definition from rte_rss_conf
>>>>>>>> structure. They don't have a plan. And [1] was eventually reverted.
>>>>>>>>
>>>>>>>
>>>>>>> Why ignoring 'key_len' (set it to zero) when there is no 'key'
>>>>>>> provided doesn't work?
>>>>>>>
>>>>>>
>>>>>> What do you think [1] + following update, will it work?
>>>>>>
>>>>>>    diff --git a/lib/librte_ethdev/rte_flow.c b/lib/librte_ethdev/rte_flow.c
>>>>>>    index ee4f3464fe..e7789c87b3 100644
>>>>>>    --- a/lib/librte_ethdev/rte_flow.c
>>>>>>    +++ b/lib/librte_ethdev/rte_flow.c
>>>>>>    @@ -535,7 +535,7 @@ rte_flow_conv_action_conf(void *buf, const size_t
>>>>>> size,
>>>>>>                               }),
>>>>>>                               size > sizeof(*dst.rss) ? sizeof(*dst.rss) : 
>>>>>> size);
>>>>>>                    off = sizeof(*dst.rss);
>>>>>>    -               if (src.rss->key_len) {
>>>>>>    +               if (src.rss->key_len && src.rss->key) {
>>>>>>                            off = RTE_ALIGN_CEIL(off, sizeof(*dst.rss->key));
>>>>>>                            tmp = sizeof(*src.rss->key) * src.rss->key_len;
>>>>>>                            if (size >= off + tmp)
>>>>>>
>>>>>
>>>>> Ferruh, your suggestion ([1] + update) looks correct. I also verified it on 
>>>>> mlx5 PMD.
>>>>> Advantage: it's a generic fix for all dpdk applications using rte_flows 
>>>>> (not just testpmd).
>>>>> It reduces code.
>>>>> With this fix the responsibility of handling key==NULL and/or len==0 is 
>>>>> moved to the PMDs (which is good).
>>>>> With regard to Lijun patch - I liked the approach of overriding the default 
>>>>> testpmd key with the default PMD key.
>>>>> But it only addresses testpmd. More code was added.
>>>>> It seems OK to call rte_eth_dev_rss_hash_conf_get() as part of parsing RSS, 
>>>>> but it would feel more confident if we could confirm it for all the PMDs 
>>>>> (by testing) or at least review the PMDs rss_hash_conf_get() implementations.
>>>>>
>>>>
>>>> Lijun's idea can work. There was a problem in implementation related to the 
>>>> key size assumption, which can be fixed.
>>>>
>>>> Even it is fixed, when user gives a rss rule without a key, we are getting 
>>>> key from device and feeding same key back to device, this is unnecessary I 
>>>> think.
>>>> When user didn't provide a key, rss rule shouldn't touch the key at all.
>>>>
>>> Do you mean that the driver should not configure the key to the hardware when 
>>> the RSS key is not specified?
>>
>> We are talking about 'key' in RSS rte flow rule, not generally configuring the 
>> device for RSS.
>>
> Yes, I know. However, I am talking about some implementations. The configuration 
> interface may be the same as that for configuring RSS
> parameters of the device in general mode.
>> If a specific rss rte flow rule, lets say to filter some defined packets to 
>> some specific queues, don't have 'key' as a part of rule, I am suggesting not 
>> touch 'key' configuration of the device.
>>
> Yes, I agree. I think your point of view is very reasonable and a more
> appropriate implementation. However, for the sake of simplicity and
> other considerations in the architecture design of some devices, the
> hardware may support reasonable hybrid configuration for paramter configuration 
> in the RSS config, rather than independent configuration
> for hw, that is, maybe touch A of rss configuration of the device and
> must touch B of rss configuration. As a result, If the preceding scenario 
> occurs, it is reasonable to configure an RSS config that does not change and 
> specified.What do you think?
>>> If so, we cannot identify when the user does not specify the RSS key to 
>>> determine whether to configure the key.
>>
>> I think we can. If the rule has 'key' attribute, driver can use that key to 
>> program the device, if 'key' attribute is missing, don't do anything related 
>> to the rss key.
>>
> Yes, if 'key' attribute is missing, don't do anything related to the rss key. 
> It's more reasonable.
> But, If he can't do that, I think it is reasonable to configure the default key 
> that already exists on the device. The only disadvantage is wasteful.
>>> In hardware design, most vendors do not configure keys for hardware 
>>> independently, which may be associated with other RSS config parameters.
>>> I think it would be reasonable for us to reconfigure the RSS key with the 
>>> default value configured and valid in the hardware as the default value if 
>>> the user does not specify the RSS key.
>>
>> There are already APIs to update the RSS configuration, like RSS key, hash 
>> functions etc.. They are independent from the rte flow RSS rule.
>> Application should configure RSS according needs using those APIs.
>>
> Yes, should do this. Are there any unreasonable users who simply do not know 
> these APIs or do not want to use them and want to configure some parameters 
> through the rte flow RSS rule without changing other parameters?
>> The question here is about each RSS rte flow rule that can update the key. Am 
>> I missing something?
> I don't think you've misunderstood the general situation.
> Even if the RSS key is not specified in a rule, the driver uses the default key 
> value to re-touch the device.
>>
>>>> Complication was when user provides key_len without a key, I think both 
>>>> ignoring or returning error in this case is OK.
>>> I agree. However, I think it is meaningless to expose the RSS key length to 
>>> users. Do you consider deleting the RSS key length directly?
>>
>> Isn't both 'key' and 'key_len' needed to program the RSS key? Can the driver 
>> know the size of the 'key' without 'key_len'?
>> And driver should verify these provided values.
>>
> I know and agree.It is only during the analysis of the [1] scheme, from the time 
> it was proposed to the time it was withdrawn, that I found some guys in the 
> community questioned the significance of key_len.
> 
> Back to the subject, what is your plan for the solution in the patch?

I think, Ophir's old reverted patch [1] + 'rte_flow_conv_action_conf()' update I 
proposed above can work. Is this working for you? If so can you please send a patch?


[1] https://git.dpdk.org/dpdk/commit/?id=a4391f8bae8